While working in Paul Berg’s laboratory, Mertz found that a restriction enzyme called EcoRI could cut DNA and, in the process, leave behind “sticky ends.” Mertz used the enzyme to cut two pieces of DNA—each isolated from a distinct organism—and then joined their ends together to create a new strand of genetic material. Thus was born recombinant DNA, the backbone of modern biotechnology.

The prior summer, Mertz had mentioned her intention to do this precise experiment while attending a summer course at Cold Spring Harbor Laboratory, a small research institute on the northern lip of Long Island. Mertz’s experimental plans frightened other scientists, who quickly enacted a voluntary moratorium on “the cloning of any DNA that might contain potentially biohazardous materials.”

In 1973, two professors—Herb Boyer at UCSF and Stanley Cohen at Stanford University—finished what Mertz started. The duo cut up DNA from E. coli and Staphylococcus and recombined them into a single loop, which they then inserted into E. coli cells. As the cells divided, they propagated the recombinant DNA and passed it down to their progeny.

Berg shared the 1980 Nobel Prize in Chemistry, while Boyer and Cohen became co-authors on one of the most lucrative biotechnology patents of all time (namely, that of recombinant DNA), which earned $250M in lifetime licensing and royalty fees.

However, back when Cohen and Boyer first told people about their chimeric DNA molecules, scientists became fearful that the technology could be used to make dangerous biological agents, saying “Oh my God, you guys can really make some dangerous things,” as Berg later recalled. The scientific community recognized that this was a key moment: Prior to EcoRI, potential hazards from intermingling genetic material from different organisms seemed like a distant possibility, because technological barriers made it extremely difficult to recombine arbitrary strands of DNA. Mertz’s discovery removed these barriers.

In response, Berg and colleagues organized a large meeting in 1975 to decide what to do. More than 150 scientists, philosophers, lawyers, and journalists gathered at the Asilomar resort near Monterey Bay to devise rules governing recombinant DNA research. The conference ended with a set of recommendations, including a moratorium on some experiments.

Asilomar’s echoes are felt in every biology lab today. The biosecurity guidelines that emanated from that meeting are a big part of why, even now, molecular biologists use weakened forms of E. coli for experiments and laboratories have six or more different waste bins.

But Asilomar’s most impactful legacy, arguably, is that a coalition recognized a pivotal moment of innovation and gathered to think through its possible risks. This legacy is worth revisiting because it seems like new technologies are announced almost daily: gene-editing tools, organisms with synthetic genomes, computationally-designed proteins, large language models, and more. Like EcoRI, each of these brings great benefit, but may also inadvertently enable harm.

Starting Point

Not all innovations create new risks. When scientists and policy makers today need to evaluate a new innovation, we’ve found that a good place to start is to answer a single question: “If successful, would this technology overcome protections that currently keep things safe?”

Arriving at an answer requires at least two steps. First, determine what currently keeps us safe, and then evaluate whether an innovation would substantively remove those protections.

Consider viruses: One of the many phenomena that limit the danger of viruses is the rate at which they become zoonotic, which is the time it takes for them to jump from animals to humans. 75 percent of new infectious diseases arise this way, according to the CDC. As a result, technologies that can circumvent or accelerate this process are considered high risk by most researchers, and the CDC has issued moratoriums on “gain of function” research at times even though the technology can enhance our understanding of disease.

Another example is germline genome editing, or changing DNA such that it’s passed onto an organism’s offspring. This carries risks, too. Prior to the discovery of gene drives, CRISPR and other gene-editing tools, germline edits were not a big concern because technological barriers made these experiments impossible. Now all the enabling technologies are within reach (and have been unethically demonstrated in human children). Germline editing has been outlawed in over 70 different countries and by international treaties.

In some situations, safety relies on restricting access to information. A controversial paper published in 2012 showed that just four mutations could convert the H5N1 bird influenza virus into a deadly pathogen capable of infecting ferrets. Because the technical barriers to perform mutagenesis are relatively low, restricting access to knowledge of these zoonotic mutations is a key way to mitigate risks. This is why researchers are often encouraged to consider whether their work could create so-called “information hazards.”

Perhaps the most dramatic example of “restricting access” concerns DNA manufacturing itself. Prior to the availability of à la carte DNA synthesis, getting hold of a DNA sequence encoding a dangerous toxin or organism required physical access to existing genetic material or the original organism —a bad “actor” had to find a tube and steal it, basically. Physical protections on such samples are typically strong, so risks were low.

All of this changed with on-demand DNA synthesis. Now, all that is needed to get DNA is a data file with the sequence. DNA can be ordered from at least a dozen different companies or printed using a benchtop synthesis machine. Large swaths of the DNA synthesis industry have responded by screening the sequences that researchers order, and will refuse to fill orders—and even notify law enforcement—when appropriate.

Notably, DNA screening started as a voluntary initiative, self-imposed by the industry, much like Asilomar’s first set of self-imposed restrictions on recombinant DNA research. And recently, the White House issued an executive order to require it of all providers.

Despite these examples, it’s worth emphasizing that research can touch on sensitive areas without creating new risks. One could argue, for example, that publishing a recipe to make methamphetamine on the internet is not an information hazard because the same information already exists elsewhere. Making information easier to access, in other words, doesn’t necessarily increase risks.

But then, of course, there are all those technologies that exist in a nebulous zone of gray. These are the technologies that might overcome existing protections, but have unclear risks.

Groups from RAND, MIT, and OpenAI have, for example, all sought to evaluate whether large-language models like ChatGPT create new bioengineering and biodesign risks. Despite the similarities of their studies, they each arrive at very different conclusions regarding the risks posed by these models. Their differences seem to stem from disagreements around what currently keeps things safe in the areas they tested, particularly concerning the importance of information hazards.

And what about AI tools to design proteins? It is often argued that these tools can be used to make proteins that don’t exist anywhere in nature—and thus potentially sidestep technological difficulties in making toxins and infectious agents. Moreover, biological surveillance systems could be blind to them in the worst case scenario, because the products could look very different from any known proteins.

As a result, many scientific leaders think this technology requires special attention. David Baker, George Church, and other protein designers have published plans to regulate the technology—including mandatory DNA screening and the storage of all synthesis DNA in secure databases—but also acknowledged that “an international group…should take the lead” on implementing the policies. The details there still need to be ironed out.

As a final set of examples, let’s think through efforts to engineer microbes to perform chemistry.

A few years ago, our Asimov Labs team (while working at MIT) and others engineered microbes to biosynthesize benzylamine, a molecule used to make therapeutics, textiles, paints, and a rocket propellant that is among the highest-energy materials known to man, called CL-20.

Reflecting on the key starting question, we first asked ourselves and a panel of external biosecurity consultants, “What protections currently limit the illicit manufacture of explosives?” And the answer, it turned out, was not the availability of benzylamine. The same molecule made by our engineered microbes is already cheap and widely-available online.

Rather, the main obstacle is converting benzylamine into its explosive form, which requires special facilities and formulation know-how that is not available online. On the flipside, our work made it possible to make benzylamine in an environmentally responsible way, without creating the toxic byproducts that plague chemical synthesis.

A final case study is yeast engineered to make opioids.

In 2015, Christina Smolke’s group at Stanford University engineered yeast to make two types of opioids, thebaine and hydrocodone, from sugar. The scientists did this by adding 21 enzymes—taken from plants, mammals, bacteria, and other organisms—to the yeast cells. These strains have the potential to unlock cheaper and more effective medicines, but could also be used to manufacture illicit drugs. The engineered cells made thebaine at a titer of 6.4 μg/L, and hydrocodone at 0.3 μg/L. There are obviously laws that limit the possession and manufacture of drugs. But when this paper was published, the scientists considered whether the engineered yeast cells would, nonetheless, remove obstacles blocking people from getting their hands on illicit narcotics.

At the low titers reported in the paper, the authors estimated that it would take thousands of liters of yeast cells to make enough hydrocodone for a single dose of Vicodin. In other words, scaling up production—and not information or access—was the main bottleneck limiting risk. For that reason, they didn’t deem the engineered cells to be a major risk at the time.

Antheia, a startup company co-founded by Smolke, is now making thebaine at commercial scales. Even so, their large manufacturing capacity doesn’t necessarily remove barriers to access; the company’s approach is proprietary and their factory has the same physical protections as any chemical company would.

Throughout all of these examples, we’ve left out an important detail. Namely, whose responsibility is it to answer these questions? Should we all gather again at a resort near Monterey Bay?

Ironically, another Asilomar is unlikely to be as effective, according to Paul Berg.1 Asilomar worked because a coalition formed “before an entrenched, intransigent, and chronic opposition developed,” Berg said. Technologies today are “qualitatively different: they are often entwined with economic self-interest and increasingly beset by nearly irreconcilable ethical, religious, and legal conflicts, as well as by challenges to deeply held social values.” This makes it difficult to organically build a cross-societal coalition.

There is also now a diverse, distributed community working to safeguard technologies. A growing cadre of organizations including the American Society for Microbiology, the Danish Centre for Biosecurity and Biopreparedness, iGEM, the Joint Genome Institute, Science Magazine, and Asimov Labs (previously housed at MIT) have all published detailed frameworks and approaches to help other laboratories and scientists evaluate and navigate complex biorisks. Moreover, dozens of non-profit organizations (e.g. IBBIS, NTI, SecureDNA), commercial entities (e.g. Aclid, Battelle, BBN, Gryphon), IGOs, philanthropies, and academics have established efforts that are dedicated to evaluating, detecting, and addressing biosecurity challenges.

We believe that biotechnology researchers of all walks should engage with this community and stay apprised of its work. In fact, when Asimov Labs was at MIT, we established a program to extend biosecurity training and awareness down to the level of the individual researcher: In this program, every member of the lab is required to perform a biosecurity self-assessment of their work, and to report their assessments to the wider group in lab meetings every few months. These assessments were also reviewed by an external biosecurity advisory committee. We established a framework to structure this self-assessment, and help ensure nothing fell through the cracks.

On the government side, emerging directly from the events around Asilomar, the National Institutes of Health (NIH) established the Recombinant DNA Advisory Committee, which has been operating for almost 50 years. It was renamed the Novel and Exceptional Technology and Research Advisory Committee (NExTRAC) in 2019. The NIH also convenes the National Science Advisory Board for Biosecurity (NSABB), an expert panel that recommends policies to prevent biotechnology from aiding terrorism. The Office of Science and Technology Policy (OSTP), which reports to the President, has recently released specific guidance for DNA screening. These groups and others at NIH, NIST, DOE, and DOD study the gamut of technologies, from those that could harm human health (engineered viruses) to environmental risks (gene drives).

A thorough treatment of biosecurity is challenging. As in the time of Asilomar, it intermingles science, society and policy, and requires a mix of different perspectives and competencies. What’s more, progress in biotechnology is still swift and the status quo changes over time. Grappling with biosecurity risks isn’t a one-time exercise, but an ongoing process. The single question posed in this essay isn’t a panacea for biosecurity, but we still think it’s a good place to start.

Contributors: Ben Gordon, Niko McCarty, Alec Nielsen & Arturo Casini.

It’s Time to Scale

The F.D.A. has approved dozens of cell therapies, clinical interventions that involve placing cells into a person’s body to treat a disease. Nearly half of these approved cell therapies are engineered, which means that the cells are altered to carry added transgenes; a chimeric antigen receptor added to T-cells to treat cancer (Kymriah®), perhaps, or a hemoglobin gene added to blood-making stem cells to treat sickle-cell disease (Lyfgenia®).

Engineered cell therapies are typically made by packaging the transgene into a lentivirus — a type of retrovirus — and then using that lentivirus to deliver the transgene into a patient’s cells. The engineered cells, thus altered, are reinfused back into the body. But there’s a problem: It’s costly and complicated to make lentivirus at scale, and we will need to get better at making these viruses if we stand any chance of making cell therapies more widely available, for more diseases, at a lower cost.

A typical CAR-T therapy costs anywhere from $375,000 and $1,000,000 for a single infusion. This fluctuates from one hospital to the next, of course, but it’s an eye-watering price tag. Biotechnology can bring this cost down, in part by making it easier to make the lentiviruses that are used to make cell therapies in the first place. 

Earlier this month, we announced our LV Edge system. It’s an entire suite of tools, including cell lines and software, to help streamline and lower the costs of lentivirus manufacturing. We think it will help the cell therapy space scale up, and also make it simpler for startups and smaller companies to compete. 

Bottlenecks

First, a quick recap of how lentiviruses are made, and the bottlenecks we’re trying to solve. (If you read part I, you can skip over this and move to the next section.)

Four plasmids, or loops of DNA, are typically used to make a lentivirus. Each plasmid carries genes that encode different parts of the virus. The genes are split across plasmids to reduce the likelihood that they will recombine to form a replicating, infectious virus. The env plasmid encodes glycoproteins, which sit in the virus’ outer envelope and help it target a patient’s cells. The gag-pol and rev plasmids encode retrovirus structural proteins and enzymes like reverse transcriptase, which converts the lentivirus’ RNA payload into DNA before integration into the patient cells’ genome. The fourth plasmid encodes the transgene of interest; a CAR for cancer therapies, a hemoglobin gene for blood therapies, and so on.

When scientists want to make lentivirus, they often start by taking the well-studied human cell line HEK293, immortalized cells derived from a female fetus in 1973, and transfecting them with the four plasmids. Transfection is just a fancy way of saying that the plasmids are inserted into the HEK293 cells, using either chemicals or electricity. Cells that take up the four plasmids begin to make lentiviruses, which emerge from the cells by cloaking themselves in lipids and budding away into the extracellular environment.

Those are the basics. Now here are the problems.

If a lentivirus is to be used in an F.D.A.-approved therapy, then the plasmids are considered a ‘raw material’ to make that therapy and must be certified as GMP-grade. This certification ensures that the plasmids are safe, pure, effective, and traceable.

But GMP-grade plasmids are inordinately expensive. They can cost hundreds of thousands of dollars per batch, and it can take anywhere from six months to a year to receive them after ordering. Only a few companies make GMP-grade plasmids, so this is a major bottleneck in lentivirus manufacturing.

A second problem is the limitations of transient transfection. When a scientist makes lentiviruses at clinical or commercial scale, they grow normal HEK293 or related cells in a large bioreactor; anywhere from 50 to 1,000 liters in volume. The scientist then floods the cells with the four plasmids, mixing them into the bioreactor. These plasmids enter the cells and coax them to make lentiviruses. The transfected cells die, and the lentiviruses are isolated and purified.

This process is crude. The plasmids don’t always mix evenly in the bioreactor (especially at larger volumes), not every cell receives an equal number of plasmids, and the large bioreactor means that many grams of DNA are required. This manufacturing process has poor scalability and high costs.

Fortunately, both problems can be resolved with advanced genetic design. Our LV Edge system includes a HEK293-derived cell line that already has all the lentivirus genes stably integrated into its genome, and a sequence optimizer tool to boost the expression of lentivirus payloads. But getting these tools to work just right was a tricky engineering challenge.

Engineered Cells

Asimov’s mission is to advance humanity’s ability to design living systems, making it easier to build biotechnologies with outsized societal benefits. We do this by developing four core technology “stacks”: Engineered host cells, genetic systems, CAD software for genetic design, and data-driven models.

The LV Edge system takes advantage of technology we’ve already developed for other applications, such as for therapeutic antibodies. It includes a HEK293-derived host cell line that makes the lentiviruses, and a sequence optimizer that modifies the transgene — or lentivirus “payload,” be it a CAR or something else — to boost expression.

Our engineered cells are called the “packaging line,” because they physically package and make lentiviruses. The packaging cells are HEK293 cells that have all three of the viral plasmids —  env, gag-pol, and rev — stably integrated into their genome under inducible control. This solves a key bottleneck in the current way that lentiviruses are made; instead of growing cells in a large bioreactor and then transfecting them with all four plasmids at once, a company using our LV Edge system would only need to transfect one plasmid encoding their transgene of interest. GMP-grade plasmids are still required, but the overall cost of manufacturing is lower.

Soon, we’ll also roll out a “producer line” cell line. These are HEK293 cells that have all of the viral plasmids and the transgene stably integrated into the genome. These cells will fully eliminate the need for companies to transfect cells with GMP-grade plasmids to make lentivirus. It takes a few months for us to integrate a transgene, so the advantage of the packaging line is that it’s more flexible and enables faster timelines.

We’re excited about these cell lines because they remove one of the barriers that make it difficult for the cell therapy sector to scale.

In 1986, the F.D.A. approved the first “recombinant biotherapeutic protein” made by CHO cells. Now, CHO cells make about 70 percent of all biologics drugs sold on the market, including bestsellers like Humira®, whose sales exceeded $21 billion in 2022, and Keytruda®, a cancer therapy with $20.9 billion sold in the same year. Biologics drugs are so popular that, to treat all the patients who need them, the drug industry manufactures many metric tons of these molecules every year. Genentech alone makes 10 metric tons of biologic drugs each year, about 20 percent of the global total.

If cell therapies ever become even a fraction as popular as biologics, then we’ll need to scale manufacturing by a lot. And we think that the only way to reach expected demands in the next decade will be to use stable cell lines and get rid of plasmid transfections entirely.

We had to overcome three challenges to make our cell lines: Stability, titers, and leakiness. 

First, stability. Both of our cell lines — the “packaging” and “producer” — were made with an engineered transposase. A transposase is an enzyme that “cuts” and “pastes” chunks of DNA into semi-random sites in a cell’s genome. They’re often used to genomically integrate plasmid DNA. Our engineered transposase has a high efficiency, such that we can stably insert multiple plasmids into HEK293 cells in a single experiment. ‘Stably’ means that the engineered system remains functional over the course of dozens of cell divisions, which matches the number of times a cell will divide while growing in a bioreactor. When a system is unstable, it becomes less and less functional over time, often by epigenetic silencing of the integrated plasmids. After transfection and integration, we screen hundreds of clones and check the top ones to ensure they remain stable over dozens of generations. 

Second, titers. One of the main concerns with doing a stable integration — rather than transient transfections — is that the latter can introduce many more copies of the plasmids into cells, and thus enable HEK293 cells to make a larger amount of active lentivirus. A good titer in the industry is around 1E8 TU/mL before purification, meaning that every milliliter of HEK293 cells makes about 100 million “transducing” units, or viruses.

But there is no point in using a stable cell line, and saving money on GMP-grade plasmids, if the manufacturing titer goes down a hundred-fold to, say, 1E6 TU/mL. That kind of drop in production would force companies to use one hundred-times the bioreactor volume to produce the same amount of lentivirus. That’s not sustainable.

Fortunately, our “packaging” and “producer” cell lines both have titers of 1E8 TU/mL or higher. There is no drop-off in production compared to transient transfections. We tested titers across several transgenes, including the CAR used in Kymriah, and achieved titers of 1.5E8 TU/mL with our “packaging” cell line after sequence optimization (more on that below). 

And third, leakiness. Remember how we said that as soon as you put the lentivirus genes into a cell, the cells begin to make viruses and die from the effort? Well, this is a major problem if you’ve integrated all the lentivirus genes into the HEK293 genome, as we have. It’s important to make sure that lentiviruses only get made if cells are first induced with a small molecule, because any “leakiness” of the lentivirus genes could kill the cells. Striking an appropriate balance between zero leakiness and high production was really tricky.

We ultimately figured it out by using a few genetic design and process development tricks. First we made the system controllable by replacing all the viral gene promoters (which initiate RNA transcription) with characterized promoters that can be triggered by an external molecule, without much leak in its absence. Next, we had to shuffle genes around onto different plasmids and make the entire system "integrate-able". Then we cleaned up the system by removing viral elements that may cause uncontrolled plasmid replication. And lastly, we reduced aberrant RNA splicing in our plasmids using transcriptomics data to guide our debugging.

Computational Tools

The LV Edge system also includes computational tools to optimize the lentivirus payload sequence. Kernel is our drag-and-drop, browser-based genetic design software. With Kernel, you can search through hundreds of thousands of genetic parts and insert them onto a canvas to design genetic systems. Kernel also has lab notebooks, an AI assistant, and built-in sequence optimization tools that we’ve developed in-house.

At first, when we tested our cell lines’ ability make lentiviruses packed with various CAR transgenes, including Kymriah, anti-BCMA (used to treat multiple myeloma) and Yescarta (used to treat B-cell lymphoma), we noticed that the titer for Yescarta was substantially lower than the other CARs.

We had previously developed a sequence optimization tool to increase the expression of transgenes, and we tested it out on Yescarta, keeping its amino acid sequence unchanged. Surprisingly, not only did this increase Yescarta’s expression in recipient cells as a payload, it also boosted the lentiviral titer by several fold. This was unexpected; optimizing the sequence of the payload appeared to cause cells to make more packaged virus. 

We see this effect for other CARs as well, but we don’t yet fully understand why this happens. Why would optimizing the payload, but not the viral genes involved in packaging, cause a cell to make more virus? We have some hypotheses and we’re running more experiments to better our understanding. In the meantime, we offer our sequence optimization tool to our partners as an additional tool in the toolkit to improve lentivirus manufacturing.

If you’d like to learn more, we’ll be at Advanced Therapies Congress in London in March, at ASGCT in Baltimore in May, and at ESACT in Edinburgh in June. You can also watch our upcoming webinar or visit our website to learn more about our lentivirus engineering efforts.

***

We’re hiring for computational roles. Come work with us! Or subscribe for updates.

Contributors: Jeremy Gam, Alec Nielsen, Mark Stockdale, Michelle Chang, Brianna Jayanthi, Ben Gordon, Jamie Freeman, Raja Srinivas, Chris Thorne, Martín Cárcamo. Words by Niko McCarty.

I. Cell Therapy

In 2010, two patients with chronic lymphocytic leukemia, a form of blood cancer, received an experimental therapy.

T-cells, a type of white blood cell that helps the body to mount an immune response, were carefully extracted from each person’s blood. The T-cells were then engineered to express a protein known as a CAR, which stands for chimeric antigen receptor. To accomplish this, a gene expression cassette for the CAR protein was packaged as RNA into a lentiviral vector (a type of engineered HIV that can’t replicate, and where the HIV genome is replaced with a synthetic genetic payload, in this case the CAR), and then mixed with the T-cells. The lentiviruses entered the patients’ T-cells and delivered the payload-containing RNA molecules, which were then reverse transcribed into DNA by lentiviral enzymes and integrated into the recipient cells’ genomes.

After several days, the T-cells began to make the CAR protein. These particular CAR proteins were engineered to recognize a specific antigen associated with leukemia cells, converting that into a signal that the T-cell should attack. The T-cells, thus re-wired, were placed back into the bodies from which they came. In this way, the patient's immune system had effectively been taught how to eradicate cancer cells as though they were any other intruder.

Carl June, the oncologist in charge of the therapy, was not initially optimistic that the therapy would work long-term, however. He told reporters that he expected that the CAR-T cells “would be gone in a month or two,” and the cancers would likely persist or come back. But that’s not what happened.

Both patients had complete remission. Their cancer melted away. And after more than a decade, things remained that way. The engineered T-cells lived on in the body and continued to fight back against rogue cancer cells. (One of the two patients, unfortunately, recently died from COVID-19.) 

June’s study was an early demonstration of the power of engineered cell therapy. It suggested, for the first time, that engineered immune cells could hunt down and destroy cancer. And it set off a blitz to search for other forms of cancer — or other diseases entirely — to go after.

The Food and Drug Administration (F.D.A.) has now approved dozens of cell therapy products, nearly half of which are engineered, meaning that the cells are altered to carry added transgenes — a CAR, a toll-like receptor, or something else. Engineered cell therapies are any clinical intervention in which modified cells are placed into the body to correct a disease or replace existing tissues. CAR-T therapy is probably the most notable example, but there are other applications, too.

At least 3,900 cell, gene, and RNA therapies are in clinical or preclinical development, according to a 2023 report from the American Society of Gene + Cell Therapy.

Lyfgenia, for example, is a cell therapy that is used to treat sickle-cell disease. One of the terrible things about this particular disease is the pain it inflicts upon a person. The condition impairs blood cells from carrying enough oxygen through the body, due to a genetic mutation that causes abnormal hemoglobin protein. It also makes those cells become misshapen and ‘sticky,’ such that they begin to clump together. The sickle-shaped cells can block vessels and cause shooting pains or, in some cases, death.

The way that Lyfgenia works is relatively straightforward: Scientists collect the stem cells that make blood from a patient’s body, and then genetically alter them so that they make a functional form of the hemoglobin protein. The engineered stem cells are then reinfused back into the body, and begin to make healthy blood cells. In a clinical trial of 47 people who received Lyfgenia, the episodes of searing pain “were eliminated or significantly reduced in all patients.”

Apologies for the editorializing, but that’s incredible.

Despite the many cell therapies that are now coming onto the market, CAR-T therapies remain the most popular. There is, for example, an FDA-approved CAR-T therapy that can induce complete cancer remission in up to 90 percent of patients who have relapsed or refractory B-cell acute lymphoblastic leukemia, or B-ALL.

Physicians who would once hesitate to use the word “cure” for cancer treatments are now lobbing it about in press releases. Over the next two decades, millions of people will likely receive cell therapy. That’s a great thing. But it will also come with its fair share of challenges.

Among them: How do we safely, reliably, and efficiently get DNA into a patient's cells? You can't just inject DNA into a patient, or even into a cell, so we need another way. Viruses are the natural solution to this problem, but they come with their own challenges, too.

Many cell therapies today are made by using lentiviruses, which enter the patient’s cells and act as a sort of vessel that endows them with new genetic material. Currently, this step is performed outside the body and the engineered cells are infused back into the patient. In the future, this delivery vessel may be able to be injected straight into the bloodstream.

Lentiviruses are an entire genus of viruses that infect humans, cows, pigs, and other mammals. The lentivirus used to make cell therapies came from HIV, and was tested in a clinical trial for the first time in 2003, with a total of 65 patients. In that trial, the lentivirus was used to introduce a ‘genetic payload’ into T-cells to treat — of all things — an HIV infection. None of the patients had an adverse event as a result of the lentivirus, even after 8 years of monitoring.

And then, in 2017, Kymriah became the first F.D.A.-approved cell therapy made using a lentivirus. Kymriah is a treatment for acute lymphoblastic leukemia and non-Hodgkin lymphoma. In the next few years, many more lentivirus-made cell therapies will get F.D.A. approval. Current cell therapies are typically autologous, which means they use a patient’s own cells to fix a disease. This means that the entire supply chain is effectively patient-specific.

Demand for lentiviral vectors is expected to grow in the next few years, according to McKinsey & Co., especially as dozens of lentivirus-made cell therapies that are currently in clinical trials could soon reach the market. When that happens, “manufacturers could struggle to meet demand,” according to a recent report by Endpoints News. 

It’s one thing to manufacture a small number of cell therapies for participants in a clinical trial, but a different beast entirely to scale up manufacturing for tens of thousands of patients. We need to make lentivirus manufacturing both more affordable and easier to scale if we ever want engineered cell therapies to help millions of people.

II. Bring the Payload

The term “lentivirus” derives from the Greek word for lenti-, meaning “slow.” These viruses were first discovered circulating as “a slowly progressive disorder” amongst sheep flocks in Iceland. A long period of time, ranging from several months to years, often elapses between the initial infection of lentivirus and the onset of a disease. But although lentiviruses are ancient — at least 7 million years old — most of the progress in adapting them to cure diseases has happened in the last two decades.

Each lentivirus carries strands of RNA that are reverse transcribed into DNA and then integrated into a host’s genome after infection. This feature is precisely why lentiviruses were adopted for cell therapies in the first place: Their genetic payloads weasel their way into the host genome and thus pass on “stable” copies of genetic information. If a scientist uses a lentivirus to engineer a stem cell, for example, then that stem cell will continue to divide into otherwise normal cells carrying the transgene.

Another advantage of using lentiviruses, as opposed to adeno-associated viruses, or AAVs, is that they are much larger. A typical lentivirus can carry about 10,000 nucleotides, more than twice as much as AAV, which is useful when the payload you want to deliver is a larger gene.

Still, lentiviruses have their own problems. For one, they tend to integrate genetic material in a semi-random manner, which means there is always a chance — however slight — that a lentivirus will insert its ‘payload’ into an important part of the cell’s genome, and cause cancer. The risk is not zero. AAVs, for context, are typically non-integrating, which means their genetic material payloads don’t get stably passed on to a cell’s progeny.

By December 2023, the F.D.A. had received reports of at least “22 cases of T-cell cancers that occurred after treatment with CAR-T products,” according to a recent article in The New England Journal of Medicine. The agency has introduced additional monitoring requirements as a result, but it’s not clear exactly what causes these cancers.

Lentiviruses are also expensive to make. A typical CAR-T therapy costs anywhere from $375,000 to $1,000,000, a price that is being paid by some insurance companies in the U.S., but that is causing European health systems to balk. Part of this high cost stems from the ways in which a lentivirus is made in the laboratory.

Four plasmids, or loops of DNA, are required to make a lentivirus. Each of these plasmids carries genes that encode a different part of the virus. The env plasmid encodes glycoproteins, which sit in the virus’ outer envelope and help it target certain cells. The gag-pol and rev plasmids encode retrovirus structural proteins and things like reverse transcriptase, the enzyme required for the virus to convert its RNA payload into DNA before integration into the genome. The fourth plasmid encodes the gene of interest; a CAR or something else. Lentiviruses are made with four individual plasmids to reduce the risk that a lentivirus, once created, can continue to reproduce on its own and cause disease. Splitting up the plasmids makes the viruses safer; it deactivates them.

All of these plasmids are wrapped up in a polymer bubble and then transfected, or delivered across the cell membrane and into the nucleus of production cells. Most companies use HEK293 or its derivatives to make lentiviruses. HEK293 cells were made in 1973 by culturing kidney cells from a human embryo. When lentivirus plasmids are added to HEK293 cells, they begin to make the viruses. And that’s the basic gist of lentivirus manufacturing.

But here’s the problem: The plasmids themselves are a ‘reagent’ in the manufacturing. In order for the lentivirus to be used in an F.D.A.-approved therapy, the plasmids must be certified as GMP-grade, meaning they are safe, pure, effective, and traceable. GMP-grade plasmids are inordinately expensive, costing in the neighborhood of hundreds of thousands of dollars per batch. It also takes between six months to a year, after ordering, to receive GMP-grade plasmids. This is a major bottleneck in lentivirus manufacturing.

Lentivirus manufacturing is typically done in 200 liter or 1,000 liter bioreactors. Each bioreactor run costs somewhere in the mid-seven figure range, and has to be carefully managed to ensure that there is relative consistency between one run and the next.

All of these challenges make it difficult for companies, especially startups, to compete in this space. There are not enough companies making GMP-grade plasmids, and the scalability of the production process isn't good enough to meet future demand for cell therapies.. We need to reduce costs, improve consistency, and boost scales.

III. LV Edge

Between 2017 and 2023, at least 20,000 people received CAR-T therapy. But now, CAR-Ts are being engineered to target an even wider range of cancers. To keep costs in check and reach a wider number of people, lentivirus production will need to scale accordingly. 

All of the current F.D.A.-approved cancer cell therapies go after blood-based cancers. Carvykti is used to treat multiple myeloma, Kymriah is used to treat B-cell lymphomas, and so on. Mechanistically, Kymriah works by reprogramming a person’s T-cells to express a CD19-specific CAR. This CAR protein is embedded in the T-cell membrane and sticks out like a thumb. When the CAR protein binds to a cancer cell or a B-cell with a specific antigen, the T-cell becomes “activated” and mounts an immune response. The cancer cells are killed, cytokines are released, macrophages are activated.

It is more difficult to make CAR-T therapies that target solid tumors, though, for a few reasons. Among them: Many of the proteins that a CAR recognizes, on the surface of a cancer cell, also show up at a low level in healthy cells. That makes it challenging to make a CAR-T therapy against solid tumors that kills cancer cells but leaves everything else unscathed. Solid tumors also shroud themselves in an acidic environment that helps to keep out non-cancer cells, and diminishes the ability of T-cells to differentiate and mount an immune response.

But this is changing, and many more engineered CAR-T therapies are expected to hit the market soon. An early-stage clinical trial in 16 patients who had a solid tumor expressing a specific antigen, called CLDN6, found that CAR-T treatment led to a partial remission of testicular or ovarian cancer in six of them. There are dozens of other ongoing clinical trials for CAR-T therapies in solid tumors (including some in which macrophages, another type of white blood cell, are engineered to express CARs), for everything from lung to brain to liver cancers.

It would be great if these trials worked out. We really hope they do. CAR-T therapy is akin to engineering the immune system to recognize and destroy the cancer cells that it had failed to detect. This will help millions of people, but only if we’re ready to meet the demands. LV Edge can help with some of these problems. 

As a start, we’re rolling out a line of HEK293 cells in which the important genes needed to make lentivirus (env, gag-pol, and rev) are stably integrated into the genome. Companies who use our platform will receive access to this cell line, which could save them hundreds of thousands of dollars in GMP-grade plasmids per batch. One only needs to transiently transfect the HEK293 cells with a plasmid encoding the gene of interest (a particular CAR, say), and then induce, or “switch on,” expression of the lentivirus-making machinery. We call this our “Packaging” cell line.

We’ll soon roll out a “Producer” cell line, too, that has all of the lentivirus-making genes — including the gene of interest — stably integrated in the genome, without impacting titer. LV Edge could drastically reduce costs for companies seeking to move into the cell therapy space. Our cell lines routinely achieve titers between 1E8 and 1E9 TU/ml (transducing units per milliliter, where a single TU corresponds to one integration event in a target cell). Our cell lines also maintain >90% stability for dozens of generations.

Our LV Edge system includes a suite of computational tools to optimize the expression of your CAR (or any other gene of interest, for that matter) and ready-to-transfer processes and protocols for companies to scale up lentivirus production. We’ll share data on all of this, and more, in part II.

Join us or subscribe for updates.

Contributors: Jamie Freeman, Raja Srinivas, Martín Cárcamo, Jeff McMahan, Brendan Eder, Chris Thorne, Jeremy Gam, Ben Gordon, Alec Nielsen. Words by Niko McCarty.

Image Credit: The header image is a scanning electron micrograph of a human T lymphocyte, or T-cell, from a healthy person. Credit: NIAID

The history of molecular biology is rife with eccentric scientists who drummed up creative experiments to study unseen molecules, and then used deductive reasoning to piece a larger puzzle together. Mapping the Central Dogma is their crowning achievement.

The Central Dogma was first described by Francis Crick, the Cambridge scientist who solved DNA’s structure with James Watson, based on x-ray images obtained by Rosalind Franklin. In 1958, Crick wrote that once genetic information has passed into protein, “it cannot get out again.”

Although students typically learn the Central Dogma as something like DNA → RNA → protein, or “DNA is transcribed to RNA which is translated to protein,” this is not what Crick originally said. There are also exceptions to the oft-mentioned DNA→RNA→protein depiction; RNA is reverse transcribed into DNA, for example, and prions are protein aggregates that replicate themselves. Crick regretted naming his idea the ‘Central Dogma,’ he wrote in Nature, because the idea itself was speculative. Crick had misunderstood the definition of the word dogma.

Still, the ways in which cells read instructions encoded in DNA to create all the proteins necessary for life is the cornerstone of modern molecular biology. The scientists who cracked this code were often brilliant thinkers, and their experiments ought to be an inspiration for future genetic designers hoping to make discoveries in areas where we are currently most blind.

In this essay, we highlight 7 experiments that elucidated the Central Dogma and information processing in cells. These experiments include those that first isolated the intermediate molecule between DNA and proteins, called messenger RNA, cracked the genetic code, and solved the basic mechanism for DNA replication in living cells.

Experiments described in this essay are important, but not exhaustive. Biological knowledge is built up, slowly, by the collective efforts of hundreds of scientists. Only a book like The Eighth Day of Creation, by Horace Judson, could even begin to do justice to the rich and beautiful history of molecular biology. This essay focuses on a few important years, and is inspired by The Generalist’s article on the history of AI.

The Structure of DNA (1953)

Friedrich Miescher, a Swiss chemist, was the first person to isolate DNA. In 1869, he collected pus-covered bandages from patients at a university hospital and extracted a sticky substance from them. Miescher called this substance nuclein.

For decades after, most biologists believed that Miescher’s discovery was little more than a quaint curiosity. Early molecular biologists (the coin was termed in 1938) thought that proteins, rather than DNA, was the genetic material of living cells. Proteins are built from many different amino acids, and appear in all kinds of different shapes and sizes. This made them seem like the likelier option for genetic material.

By 1944, though, this view began to crumble when three scientists at the Rockefeller Institute in New York City, named Oswald Avery, Colin MacLeod, and Maclyn McCarty did an experiment to identify the molecule responsible for carrying genetic information. Their results pointed to DNA. 

The trio isolated protein and DNA from different strains of Streptococcus pneumoniae, a virulent bacterium, and then used enzymes to break down the molecules. The digested proteins and DNA were inserted into a harmless strain of bacteria, and the scientists waited to see if either molecule would turn the harmless cells virulent.

When digested DNA was added to the cells, the harmless bacteria adopted the traits of the virulent strain, but this did not happen with digested proteins. These results suggested that DNA, and not proteins, was a carrier of hereditary information.

A few miles north, at Columbia University, a biochemist named Erwin Chargraff read the Avery-MacLeod-McCarty paper and was "deeply moved by the sudden appearance of a giant bridge between chemistry and genetics," as he later wrote. Chargaff had an academic background in molecular chemistry. He realized that, if DNA was indeed the genetic material, then perhaps a chemist could dissect how it differs across organisms and thus explain the rich diversity of the natural world.

Chargaff’s team spent several years chewing up DNA sequences, separating out the individual nucleotides on pieces of paper, and exposing the nucleotides to a UV spectrophotometer. They repeated this for DNA molecules harvested from yeast, bacteria, beef spleens, and calf thymus. By 1949, Chargaff had cracked a basic principle of the DNA code:

"The desoxypentose nucleic acids from animal and microbial cells contain varying proportions of the same four nitrogenous constituents, namely adenine, guanine, cytosine, thymine…Their composition appears to be characteristic of the species, but not of the tissue, from which they are derived."

In other words, Chargaff correctly determined that every organism on Earth uses DNA molecules that are made from the same four letters. Genetic material only differs, from one species to the next, by the order in which the four nucleotides appear. Chargaff also noted that "the molar ratios of total purines to total pyrimidines, and also of adenine to thymine and of guanine to cytosine, were not far from 1." Said another way, the amount of ‘A’ in DNA is always equal to the total amount of ‘T’. Ditto for ‘G’ and ‘C’ nucleotides.

Chargaff shared his results in a lecture at Cambridge University in 1952. Watson and Crick were in attendance. The following year, using x-ray diffraction images first obtained by Rosalind Franklin at King’s College London, and perhaps also Chargaff’s observations, Watson and Crick assembled a biophysically accurate model of DNA. Their model was made from crude, metal sheets, but clearly depicted a right-handed double-helix in which ‘A’ connects to ‘T’, and ‘G’ connects to ‘C’. The model was published in Nature on 25 April 1953.

Recent revelations have revised Rosalind Franklin’s role in solving DNA’s structure. In the classic telling of this tale, Franklin is “portrayed as a brilliant scientist, but one who was ultimately unable to decipher what her own data were telling her about DNA,” according to an article by Matthew Cobb & Nathaniel Comfort in Nature. “She supposedly sat on the image for months without realizing its significance, only for Watson to understand it at a glance.”

But this tale is not accurate. Newly unearthed documents, including a shelved article that Franklin wrote with Crick and Watson for Time magazine in 1953, now suggest that “Franklin did not fail to grasp the structure of DNA. She was an equal contributor to solving it.”

DNA Replication (1958)

Watson and Crick’s 1953 Nature paper concludes with one of the most famous passages in biology’s history: 

“It has not escaped our notice that the specific pairing we have postulated immediately suggests a possible copying mechanism for the genetic material.” 

The Cambridge duo’s model correctly depicted DNA as a molecule composed from two interlocking strands, wherein ‘A’ always connects to ‘T’ and ‘G’ always connects to ‘C’. If the two strands were to unwind and detach from each other, Watson and Crick noted, it should be possible to recreate the original strand merely by pairing up each base in the separated strand with its appropriate nucleotide. This idea was called the semi-conservative model of replication.

Other eminent scientists attacked this idea. Max Delbrück was a renowned physicist at the California Institute of Technology who, together with Salvador Luria, had discovered that bacteria resist phage attacks via random mutations. He penned an article arguing that the semi-conservative model could not be correct because too much energy would be required to unwind the two DNA strands. 

Delbrück favored a different model, called dispersive replication, in which small chunks of a DNA molecule are broken up, and then matching DNA sequences are synthesized directly in the broken regions to create an intact, double-stranded helix. A third group of scientists favored a conservative replication model, which theorized that the entire DNA molecule is somehow copied without unwinding whatsoever.

Thanks to a particularly innovative experiment devised by two young scientists at Caltech, named Matthew Meselson and Franklin Stahl, Watson and Crick’s semi-conservative model was ultimately vindicated.

It would be relatively simple to figure out how DNA replicates if one could directly observe these molecules. But that was not possible in 1958. Instead, Meselson and Stahl devised a clever experiment, based on spinning molecules quickly in a centrifuge, to test the three models.

Meselson and Stahl’s key insight was to tag DNA strands undergoing replication with heavy atoms, such as nitrogen (N¹⁵) that carries an extra neutron. The scientists grew bacterial cells in a growth medium containing this heavy nitrogen, waited for the N¹⁵ to incorporate into all of the cells’ molecules, and then quickly transferred the ‘heavy’ microbes into growth media with normal nitrogen.

As the DNA molecules replicated, Meselson and Stahl killed the cells and used a centrifuge to spin down the molecules. As the tubes spin, heavier DNA moves toward the bottom and lighter DNA stays closer to the  top. Before the cells replicated their DNA, all of the DNA molecules contained heavy nitrogen. After one round of DNA replication, the DNA strands contained half-heavy and half-light nitrogen atoms (Meselson and Stahl saw two ‘bands’ begin to appear in their centrifuged tubes.) And after two rounds of DNA replication, only one-in-four DNA molecules contained heavy nitrogen, suggesting that the semi-conservative model was correct.

This experiment is renowned for its simplicity and clever approach – it is now called “the most beautiful experiment.” Delbrück was wrong; DNA replication occurs when the two interlocking strands unwind, and each strand is then used as a ‘template’ to remake a double helix.

The Central Dogma (1958)

After publishing his 1953 Nature paper about DNA’s structure, Francis Crick toured the world to lecture on an idea that "permanently altered the logic of biology," according to Horace Judson, author of The Eighth Day of Creation.

During his lectures, Crick would often draw a diagram on the auditorium’s blackboard. His diagram depicted how information flows through living cells; DNA is somehow converted into an intermediate molecule, which Crick called ‘template RNA’, that somehow encoded the amino acids in a protein molecule. Crick correctly predicted the basic details of protein synthesis years before direct experimental evidence had confirmed the existence of mRNA or tRNA.

In 1958, Crick adapted his lecture into a published article, called On Protein Synthesis. His target audience was “a general reader rather than the specialist." The article gave two hypotheses to explain the relationship between DNA and proteins, called the Sequence Hypothesis and the Central Dogma.

“The direct evidence for both of them is negligible,” Crick wrote, “but I have found them to be of great help in getting to grips with these very complex problems.”

The sequence hypothesis, in its simplest form, “assumes that the specificity of a piece of nucleic acid is expressed solely by the sequence of its bases, and that this sequence is a (simple) code for the amino acid sequence of a particular protein.” In other words, the bases in a strand of DNA or RNA corresponds to the amino acids in a protein.

The Central Dogma, Crick wrote, “states that once 'information' has passed into protein it cannot get out again.” Stated another way, “the transfer of information from nucleic acid to nucleic acid, or from nucleic acid to protein may be possible, but transfer from protein to protein, or from protein to nucleic acid is im­possible.”

This passage marked the first time that the Central Dogma, the defining idea of molecular biology, had been published. But this is not why Crick’s article was so prescient.

In the article, Crick used scattered experimental evidence and anecdotal observations, including the fact that "spermatozoa contain no RNA," to correctly predict that there must be a messenger RNA molecule in the cytoplasm that is produced by "the DNA of the nucleus."

Crick’s astounding ability to theorize was most prominently displayed, though, when he correctly inferred the existence of tRNAs, predicted what they were made of, and explained how they likely became 'charged' with amino acids for protein synthesis.

Molecular biologists knew that proteins were made from 20 amino acids, but most other details of protein synthesis were a mystery. Today, we know that tRNA molecules get 'loaded' with the correct amino acid via the action of specific enzymes, and that this is how a message encoded in a strand of RNA is used by the ribosome to build a protein. But Crick had little evidence for any of this. And yet, in his 1958 paper, he wrote:

"Granted that...[mRNA]...is the template, how does it direct the amino acids into the correct order? One's first naive idea is that the RNA will take up a configuration capable of forming twenty different 'cavities', one for the side-chain of each of the twenty amino acids. If this were so, one might expect to be able to play the problem backwards -- that is, to find the configuration of RNA by trying to form such cavities. All attempts to do this have failed, and on physical­ chemical grounds the idea does not seem in the least plausible...

Apart from the phosphate-sugar backbone, which we have assumed to be regular and perhaps linked to the structural protein of the particles, RNA presents mainly a sequence of sites where hydrogen bonding could occur. One would expect, therefore, that whatever went on to the tem­plate in a specific way did so by forming hydrogen bonds. It is therefore a natural hypothesis that the amino acid is carried to the template by an 'adaptor' molecule, and that the adaptor is the part which actually fits on to the RNA. In its simplest form one would require twenty adaptors, one for each amino acid.

What sort of molecules such adaptors might be is anybody's guess. They might, for example, be proteins...though personally I think that proteins, being rather large molecules, would take up too much space. They might be quite unsuspected molecules, such as amino sugars. But there is one possibility which seems inherently more likely than any other-that they contain nucleotides. This would enable them to join on to the RNA template by the same 'pairing' of bases as is found in DNA, or in polynucleotides.

If the adaptors were small molecules one would imagine that a separate enzyme would be required to join each adaptor to its own amino acid and that the specificity required to distinguish between, say, leucine, iso­leucine and valine would be provided by these enzyme molecules instead of by cavities in the RNA. Enzymes, being made of protein, can probably make such distinctions more easily than can nucleic acid."

This paper is a tour-de-force of logical reasoning. It became the focal point, a rallying cry, for molecular biologists seeking to crack the genetic code and resolve the cell’s mysteries. Crick, fortunately, would not have to wait long for his ideas to be vindicated. A ‘template RNA,’ or messenger RNA as it’s now called, was discovered just three years later.

Isolation of messenger RNA (1961)

Messenger RNA was first isolated by two separate research groups in 1961. Their results appeared back-to-back in the 13 May issue of Nature.

At the Institut Pasteur in Paris, the French scientists François Jacob and Jacques Monod had discovered that the enzymes required to break down a sugar in bacterial cells were only made after cells were exposed to that sugar. In other words, cells somehow “process” an external cue and make proteins in response. This marked the discovery of genetic regulation, but also raised a slew of questions. 

Among them: How does a cell know which genes to turn on at any given time? Why doesn’t the whole genome “turn on” at the same time? Several answers were proposed. Maybe there is a custom ribosome corresponding to each gene, some said. Or maybe, as Crick had proposed in 1958, there is an intermediate molecule – a “template RNA” – that transmits messages between DNA and proteins.

In 1960, two groups set out to isolate this mystery molecule. The first group rallied around Matthew Meselson’s laboratory at the California Institute of Technology, and included Sydney Brenner and François Jacob. A second group rallied around Wally Gilbert’s group at Harvard, and included James Watson and François Gros, a French biologist who had worked with Jacob.

Both groups turned to a compelling, experimental model: bacteriophages. When E. coli bacteria are infected with a phage, many scientists had noted that the cells stop making their own proteins, and quickly switch over to making the phage proteins. “This system thus provides an ideal model for observing the synthesis of new proteins following the introduction of specific DNA,” Gilbert’s team noted in their 1961 paper.

To isolate messenger RNA, the Caltech scientists grew bacteria in a growth medium with heavy isotopes, much like Meselson had done with Stahl several years earlier to validate the semi-conservative model of replication. These ‘heavy’ bacteria were then infected with phage and immediately transferred into a growth medium with light isotopes. Infected cells were finally lysed open at regular time points and spun down in Meselson’s ultracentrifuges.

The bands that emerged from Meselson’s centrifuges confirmed a few things. First, bacterial cells did not make new ribosomes after they were infected. This observation was evidence against the fact there is a unique ribosome for each gene. Second, the results confirmed that a new type of RNA molecule was swiftly made after phage infection, and that this new RNA quickly attached to existing ribosomes in the cell. This suggested that the DNA in phages was being quickly transcribed into messenger RNA. And third, the bacterial cells began to make phage proteins using their existing ribosomes. 

They had discovered messenger RNA. There is an excellent, and much richer, account of this history by the scientific historian, Matthew Cobb.

Mapping a Codon (1961)

Crick’s 1958 paper made a series of predictions about messenger RNA, transfer RNAs, and how a code embedded in a DNA molecule could possibly encode a protein. But one longstanding question in molecular biology had to do with the nature of the genetic code itself. Namely, how do the nucleotides in a strand of RNA encode the amino acids in a protein? What does UAG mean, or GAA, or UUU, or any other codon, for that matter?

The first triplet codon to be mapped to an amino acid was ‘UUU’ to phenylalanine. This connection was made by two young researchers at the National Institutes of Health (NIH) in Bethesda, Maryland.

Heinrich Matthaei was a post-doctoral fellow working in the laboratory of Marshall Nirenberg, a new researcher at the Institutes. The two scientists were interested in the Central Dogma – they had read Crick’s paper – and aimed to understand the connection between RNA and proteins, often by running experiments on cell-free extracts, a liquid made by grinding up living cells in a mortar and pestle. This enabled the two scientists to study cell biochemistry without having to deal with living organisms.

At 3 o’clock in the morning of May 27th, the two scientists took some of these ‘cell guts’ and added a few drops of synthetic RNA with the sequence: 

UUUUUUUUUUUUUUUUU

Their concoction was next added to 20 different tubes, each of which held a different amino acid; valine, alanine, glutamine, and so on. One of the tubes contained phenylalanine amino acids that had been labeled with a radioactive isotope.

“The results were spectacular and simple at the same time,” according to a brief history from the NIH. “After an hour, the control tubes showed a background level of 70 counts, whereas the hot tube” – with the radioactive phenylalanine – “showed 38,000 counts per milligram of protein.”

In other words, when the synthetic RNA molecule was added to a tube of phenylalanine amino acids, the cell-free extract began to churn out radioactive peptides. This singular experiment suggested that the nucleotides UUU somehow encode phenylalanine during protein synthesis.

Over the next several years, Nirenberg and other researchers would go on to map all 64 codons, including the codon that signals the start of translation, AUG. Nirenberg shared the 1968 Nobel Prize in Physiology or Medicine.

Cracking the Genetic Code (1961)

The year 1961 was molecular biology’s annus mirabilis. Messenger RNA was isolated for the first time and Nirenberg and Matthaei decoded the ‘meaning’ of the first codon – UUU. Even after those papers were published, though, mysteries remained. Among them: Is the genetic code overlapping or non-overlapping? And is it actually made from doublet, triplet, or quadruplet codons?

A messenger RNA sequence that reads ‘AUGACC’ could be read by the ribosome as 'AUG' and then 'ACC,' or it could be read by the ribosome as ‘AUG’, ‘UGA,’ ‘GAC’, ‘ACC’. The former is a non-overlapping code, and the latter is an overlapping code. Similarly, the code could be read as ‘AU’ and then ‘GA’ and then ‘CC’ if codons were doublets, or ‘AUGA’ and then ‘UGAC’ if they were quadruplets, and so on. Nirenberg and Matthaei’s experiment did not help to answer either of these questions, because their synthetic RNA had a repetitive sequence: UUUUUUUU. 

In the waning weeks of 1961, Sydney Brenner, Lesie Barnett, Francis Crick, and R.J. Watts-Tobin used fragmentary experimental evidence and thought experiments to conclude that each amino acid in a protein is encoded by a triplet code, and that the letters in this code do not overlap. Their ideas were published in a paper entitled, “General nature of the genetic code for proteins.” 

Their experiments hinged on two things: A bacteriophage, called T4, that infects bacteria, and a particular type of dye, an acridine called proflavin, that precisely mutates DNA by adding or removing a single nucleotide. 

Crick, the ever-careful thinker, had a beautiful idea. He decided to take some T4 bacteriophage and then expose it to proflavin, such that the phage lost its ability to make a particular protein. If Crick added one base and then removed one base, using the acridine, he noted that the phage were able to make the protein. But if he used acridine to add two bases, the phage did not make the protein. When three bases were added, the phage made the protein again. From these observations, the scientists argued that the genetic code must use triplets to encode each amino acid. Their takeaway was based on this fragmented, experimental evidence.

Even though the "combination of mutations strongly suggested that the code was based on units of three bases, the experiments could not prove that to be the case – a code using groups of six bases was consistent with the results," wrote Matthew Cobb in a 2021 history of this paper. 

Today, we know that there are 64 codons in total, and that codons appear as ‘triplets’ to encode amino acids in a final protein chain. Codons made of six bases “would raise all sorts of problems,” as Cobb notes, “by massively increasing the number of either meaningless or degenerate sequences (there would be 4096 possible combinations of bases, rather than a mere 64).”

As Crick later said: This was “hardly likely to be taken seriously.”

Translation via a Single Ribosome (2008)

aBy 1961, the basic contours of the Central Dogma had been resolved. But that doesn’t mean all work has since abated, nor that the years from 1953 to 1961 are all-encompassing. Linus Pauling at Caltech predicted the main structural motifs of proteins as early as 1951. A ‘stop’ codon that halts protein synthesis was identified in 1965. The ribosome’s structure was solved in 2000, after decades of work, and culminated in the 2009 Nobel Prize in Chemistry.

Today, synthetic biologists continue to expand the Central Dogma using technologies that Francis Crick, in 1958, could only have dreamed of. And yet, the molecular choreography that underlies the Central Dogma continues to surprise. There are far more enzymes and components involved than early molecular biologists ever could have realized. Transfer RNAs carry amino acids to the ribosome, proteins interact with the ribosome to push it off the RNA strand, and dozens of proteins are involved in transcription initiation, elongation, and termination in human cells. 

Molecular biologists continue to resolve this complexity today. In a 2008 study, called “Following translation by single ribosomes one codon at a time,” chemists at the University of California, Berkeley studied individual ribosomes as they moved along a single messenger RNA molecule. Their experiment revealed the stochastic starts and stops of a ribosome during translation.

For this experiment, each end of an mRNA molecule was attached to a polystyrene bead. One of the beads was then placed in a laser trap, holding it in place. The middle of the mRNA molecule contained a long loop, which slowly unwound as the ribosome traversed along its length. As the mRNA molecule stretched out, this elongation could be directly measured by measuring the distance between the two beads.

The chemists repeated this experiment several times, and measured the rate at which the mRNA molecule stretched out each time. Their key result was this: Ribosomes do not glide along the mRNA at a steady pace (which would stretch out the molecule in a linear fashion), but rather jump from one codon to the next in time steps of around 0.1 seconds. The ribosome occasionally pauses between jumps. Each ribosome, then, translates a strand of mRNA in a slightly different amount of time.

This experiment is one of thousands that have been applied to study the Central Dogma in the last two decades. Crick’s 1958 article continues to inspire generations of molecular biologists, who have found his ideas to be rich fodder for a lifetime of scientific work. We now know a shocking amount about transcription, translation, and the genetic code; bacteria add about eight amino acids to a protein each second, human cells add about five amino acids in the same length of time, and DNA is transcribed to RNA at a rate of about 40 nucleotides per second.

Conclusion

The ways in which cells process information is a biophysical marvel that has slowly unraveled over the last 70 years. The Central Dogma, and the 7 seminal experiments described in this essay, are the basis for most everything we do in genetic engineering. But there are still many instances in which we, as genetic designers, place a gene into a cell expecting one thing to happen, but observe something entirely unexpected instead. In other words, biology does not always behave as we expect. 

Though useful, the Central Dogma is an incomplete way to think of a living cell. DNA is not always transcribed to RNA, and RNA is not always translated into protein. Sometimes RNA goes back to DNA. The only rule in biology is that there are exceptions to every rule. In future posts, we’ll continue our exploration of the Central Dogma and explain many of these exceptions.

Come join us or subscribe for updates.

Contributors: Ben Gordon and Alec Nielsen. Words by Niko McCarty.

Lead image: Marshall Nirenberg poses next to his personalized license plate in 1987. Credit: National Institutes of Health

Read this post on our website for a more aesthetic experience.

I. Computational Sport

A Formula 1 car is built from about 14,500 individual parts. Each one is designed on a computer, simulated, and then built in a factory. By modeling cars on silicon, before building them in the real-world, F1 teams save money and (ideally) stay below their $135 million cost-cap.

The simulation and modeling tools used by F1 teams are incredible, in part, because they span every scale. CAD (computer-aided design) tools are used to design a wing, for example, whereas CFD (computational fluid dynamics) tools predict the aerodynamic properties of that wing. 

Racing teams simulate every screw, every system, and then the entire car as it races around a track at 220+ miles per hour. Simulators can anticipate the specific lap that a driver will need to come in for a pit stop, due to tire wear. Teams even simulate how other teams will fare in a race, based on their prior performances. F1 racing is a computational sport.

Biology, of course, is not F1. Living cells are often described as “wet,” “messy,” or “unpredictable.” Most mathematical models that attempt to predict biology operate at a single level, and don’t extend much beyond. Scientists may model a lone genetic circuit, or perhaps the growth rate of a particular type of cell. But there is no unifying model that integrates everything, from a cell’s genome sequence all the way up to its behavior in a bioreactor, for mammalian cells. (Markus Covert’s group at Stanford University is building multi-scale cell physiology models for bacteria.)

In a prior blog post, we described our software tool, called Kernel, that can help design DNA constructs in silico. It’s analogous to how F1 teams use CAD tools to design a wing or engine. In this blog post, we reveal our suite of simulators that can predict how engineered cells will perform when cultured in a bioreactor. This is more analogous to the CFD tools that F1 teams use to simulate a wing’s performance on-track.

Our long-term goal is to bring all these tools together — Kernel and the simulators — into a holistic platform to model biological systems. And there’s a good reason for that.

Consider what happens when a scientist engineers a cell to make a medicine. They first design a plasmid, a loop of DNA, that contains all the genes needed to make that medicine. They then put that plasmid into the cells, and the cells make the medicine. Some cells do this better than others. The scientist removes the “losers” and selects the “winners,” the cells that make the most medicine. The winners are then placed into a bioreactor, and the scientist tunes and tweaks the bioreactor conditions until hitting upon the “optimal” conditions for that specific cell.

In other words, the conditions for each step in a bioprocess are dependent upon the conditions that came before. The conditions of a bioreactor are based on the characteristics of the “winning” cell. And the “winning” cell is based on the characteristics of the plasmid that went into it. This means that each “optimization” is biased by the step that preceded it.

A Formula 1 team would never design an engine, simulate it, fabricate it, and then build the rest of the car around it. They instead design the wing, the engine, and the suspension all together, holistically, to get the best performance out of the entire car. We should do the same for biology.

II. CHO Simulator

Chinese Hamster Ovary, or CHO cells, make roughly 70 percent of all F.D.A. approved biologics sold on the market. This includes bestselling antibodies like Humira®, whose sales exceeded $21 billion in 2022, and Keytruda®, a cancer therapy with $20.9 billion sold in the same year. 

There is a long history of mathematical models designed to make CHO cells better at making medicines. Indeed, our own simulator stands on the shoulders of giants, adapting insights from prior work developed by researchers in Ireland and the U.K. Scientists at Pfizer and elsewhere have similarly built simulators to predict CHO cell metabolism, and entire companies have spun up to enhance biologics manufacturing using data-driven CHO metabolism models.

Our CHO simulator is built from smaller, interlocking models that, together, predict how a cell’s metabolism will shift during the entire course of a bioreactor experiment. The model can identify the best way to culture cells, including what to feed them (and when), temperature, and agitation rates to boost their medicine-making potential.

We use a dynamic flux-balance analysis model, or dFBA, to predict how metabolites will change over time. A dFBA model essentially “looks” at a cell’s genome, uses algorithms to determine which enzymes the cell expresses, and then anticipates how those enzymes will interconvert molecules in the cell. It considers the concentrations of metabolites outside the cell, and supposes the cell has a particular objective it’s trying to fulfill, such as: “Grow as quickly as possible!” or “Make as much medicine as possible!”

A dFBA model can predict, for example, how quickly a sugar molecule will break down, and how the carbons from that sugar molecule will wend their way through a cell’s metabolic pathways. These models can also predict how the extracellular concentrations of molecules will change over the time of a process, capturing how metabolism shifts in response to a dynamic environment.

There are a couple limitations to traditional dFBA models, however; they suppose the cell’s objective doesn’t change over time, and their awareness of the conditions in the bioreactor is limited. To address these limitations, we’ve wrapped our dFBA model in a phenotype, or cell-state, model. This “outside” layer is made from a combination of artificial neural networks and ordinary differential equations that consider not only the environment of the bioreactor – such as the concentrations of metabolites within it – but also the internal states of the cells, including which enzymes are likely to be expressed.

By layering the two models, we are able to connect the external variables in a bioreactor to the internal dynamics of a living cell. And the two layers communicate both ways: Outputs calculated by the dFBA are continuously fed back into the phenotype model, and vice versa, in a closed loop of computational calculations.

Our model also has some unique features that make it suitable for holistic optimizations, ranging from genetic design all the way up through a cell’s performance in a bioreactor. The simulator is built upon a “right-sized” metabolic model, for example, that includes all the most important pathways without being too complicated. The model is also “aware” of more details of the bioreactor environment than your typical model. The “outside layer,” or cell-state model, can also represent how big the cells are, whether they’re growing, and what they want to eat. We’re now working to combine the CHO simulator with a model for gene expression, so that we can predict how a specific genetic design — say, a plasmid carrying a DNA sequence, added to a cell  — will affect how that cell grows and performs in a bioreactor. More on this in a future post.

We’re only able to build these models because we work with an excellent team of synthetic biologists who have devised methods to measure many variables, in real bioprocesses and for a large variety of cell lines, at our Boston laboratories. Our bioreactors are equipped with dozens of sensors that enable us to monitor cells in real-time, collect data, and transmit it to our biostatistics pipelines.

Now that we’ve got our engine modeled, and we’ve started integrating it with the other components of our system, let’s see how we use this model to actually predict how much medicine a cell will be able to make.

III. On-Track Performance

The complete CHO model predicts the rate of production and rate of consumption of various metabolites, including amino acids, sugar, and lactate. It also predicts titers for the actual medicine (such as an antibody) and the growth rate of the cell population.

Our predicted results are quite good. We achieved an error of roughly the same as the day-to-day variation in the data across the 31 variables over a 14-day bioreactor experiment. (For those with a statistical bent, that is to say we achieved a mean absolute scaled error of 1.3. A mean absolute scaled error value is expressed relative to the sum of the changes between each measurement and the next.) Between the low error and the wide range of metabolites we predict, this puts our model in good company amongst the best-performing models published in the literature.

The simulator can reveal metabolic tradeoffs in CHO cells, and we’re already using it to select additives (literally, nutrients added to a bioreactor) that can boost antibody production by 25%. Our first simulations of optimal feeding strategies also identified osmolality, or the concentration of dissolved particles, as a factor that needs to be controlled early on in a bioreactor experiment. The model is already being used to untangle the complex tradeoff between oxidizing and reducing reactions, nitrogen metabolism, and carbon metabolism that limit growth and productivity in a cell culture.

The CHO simulator model will soon be integrated into Kernel, our computer-aided design software for biology, where it will join other tools for plasmid engineering, gene therapy, and more. You’ll be able to use it there, and refine the model’s parameters by inputting your own datasets.

If this work sounds exciting, come join us. Or, subscribe for updates.

***

Contributors: Bronson Weston, Will Johnson, Alec Nielsen & Ben Gordon. Words by Niko McCarty.

This is Part 1 in a series on the tools Asimov is building to engineer and model mammalian cells. It’s meant to be a broad introduction to our ideas. Consider reading this article on our (new) website.

In 1948, a rodent breeder in New York named Victor Schwentker sent a message to a colleague in China: Could you get some hamsters for me? 

Schwentker had heard tales of a special hamster, native to northern China and Mongolia, with short gestation periods and natural resistance to human viruses — traits that would make them ideal for scientific research. Chinese scientists had used these hamsters to study pathogens since the 1910s, but Schwentker feared that the ongoing Chinese civil war would soon make them impossible to retrieve.

In the waning months of 1948, Schwentker sent a letter to Robert Briggs Watson, a Rockefeller Foundation field staff member, and asked him to bring (or, erm, smuggle) some hamsters back to New York.

Watson collected twenty hamsters – ten males and ten females, packed in a wooden crate – with help from a Chinese physician. He smuggled them out of the country via a Pan-Am flight from Shanghai, just before Communists took control of the country. (An excellent account of this story can be read online.)

Back in New York, Schwentker bred the hamsters and distributed them to other researchers, one of whom was a geneticist named Theodore Puck. In 1957, Puck took a small piece from a hamster’s ovary, plated the cells onto a dish, and passaged them. He isolated a clone that could divide again and again; an “immortalized” Chinese Hamster Ovary (CHO) cell with a genetic mutation that rendered them immune to normal senescence. A decade later, Puck also isolated a second clone that was unable to synthesize proline, an amino acid. Those cells could only be grown in media with added proline, thus paving the way for selection methods.

By 1986, the F.D.A. had approved the first “recombinant biotherapeutic protein” made by CHO cells. Today, CHO cells make roughly 70 percent of all F.D.A. approved biologics — molecules made by living cells — sold on the market. This includes bestsellers like Humira^®, whose sales exceeded $21 billion in 2022, and Keytruda^®, a cancer therapy with $20.9 billion sold in the same year.

It is remarkable, in a way, that a cell derived from a hamster, smuggled out of China halfway through the 20th century, has dominated biologics manufacturing for the last forty-odd years. But there’s a simple reason for its success: CHO cells are really good at making medicines.

They make lots of protein, routinely reaching production titers of multiple grams per liter. They also divide slightly faster than most human cells (18 to 24 hours) and grow at high densities, which means more cell ‘factories’ can be packed into the same volume. And CHO cells modify proteins in ways that microbial cells cannot; they tag proteins with sugars, or create disulfide bonds in proteins, much like human cells do.

CHO cells are also somewhat — but not always! — resistant to human viruses. There have been at least two cases where a rogue virus shut down a biologics factory. Genzyme facilities in Belgium and Allston, Massachusetts were infected with Vesivirus 2117, an RNA-packed, icosahedral virus, that cost the company “$100–300 million in lost sales,” according to reporting in Nature Biotechnology in 2009. (Other cell lines, such as NS0 and SP2/0, have also been used to make medicines, but they can add alpha-gal residues to proteins that cause severe immune reactions in people.)

Although CHO cells make many types of molecules, they are especially good at making monoclonal antibodies; ‘Y’ shaped proteins made from interwoven subunits, called heavy chains and light chains. Antibodies bind to molecules and proteins in the body with exceedingly high specificity. They are part of the body’s natural immune system, but are also used to treat everything from inflammation (Humira^®) to COVID-19 (Actemra^®). About 40% of all drugs are biologics, and antibodies alone will account for an estimated $300 billion in sales by 2025.

Coaxing CHO cells to make antibodies has followed basically the same recipe for the last 20 years. Each protein chain is first encoded as a DNA sequence and then placed downstream from a promoter that ‘drives’ its expression (historically, a promoter from human cytomegalovirus, called CMV). The DNA ’payload,’ encoding the antibody subunits, integrates into the genome. The cells begin to make antibodies.

Now here’s the problem. In the last two decades, individual pharma companies have relied on static methods — same genetic parts, same plasmid architecture, same host cells — to make antibodies with CHO cells. We now know, though, that the ratio of heavy and light chains, the number of DNA copies integrated into the genome, and many other factors in a dizzying constellation of possible “design space” can influence how much antibody a cell makes. The same factors can also determine how “good” the antibodies are: What fraction fold properly? Do the antibodies ‘clump’ up? Do they carry the correct sugar tags? 

This is really important to understand, because the normal way to make antibodies today is with a “one-size-fits-all” approach. And that doesn’t work for every antibody.

If you were to take two different antibody sequences, place them in otherwise identical plasmids – with the same promoters and terminators – and drop them into CHO cells with identical genomes, those two antibodies might be produced at vastly different titers. In other words, the same plasmid doesn’t work for every antibody.

Even the codons within a DNA sequence (CUC, UUG and CUG all encode leucine, for example) can influence how many antibodies a cell makes. And these effects can be extreme; we’ve personally seen 3x boosts in antibody titers from codon optimization alone.

We suspect that it’s possible to boost both the quality and amount of antibodies that a cell makes, from a few grams per liter to more than 10 grams per liter, for even difficult-to-make molecules. And we aim to do that by building synthetic biology and computational models that can more efficiently explore the “genetic design space” of living cells. We call this technology stack for genetic design — which works across multiple cell types and applications, including CHO — our “CHO Edge” system.

CHO Edge includes different host cell lines and a library of thousands of characterized genetic elements — promoters, terminators, and many other “parts.” It also includes computational tools and a hyperactive transposase that tunably integrates between 20 and 60 copies of the plasmid into the CHO genome. (A transposase is an enzyme that cuts out a DNA sequence and then “pastes” it at sites across the genome.)

We’ve also made a custom codon optimizer and signal peptide predictor. Signal peptides are short sequences that cause a protein to be exported, or secreted, from the cell. Our tool predicts how well a given signal peptide will work with a given antibody, and we’ll explain how it works in an upcoming blog.

All of these tools are used, together, to search through “genetic design space” and optimize therapeutics manufacturing. We routinely achieve titers between 6 and 10 grams of antibody per liter. In one case, we achieved 11 grams per liter for a 3-chain bispecific antibody — a more complex protein architecture than standard 2-chain monoclonal antibodies — with 85 percent heterodimerization, a measure of a molecule’s quality. We guarantee a minimal titer of 4 grams per liter, or our work is free.

Boosted titers are useful not just because they cut down on manufacturing costs, but because they can literally determine whether or not a life-saving medicine makes its way to patients. It’s impossible to know for sure, but many antibodies that could save lives seem not to make it to the clinic, simply because they are deemed “un-manufacturable” with existing tools.

We’re excited about our progress so far, but this is just the beginning. 

Asimov Labs is already collecting multi-modal measurements on transcription and translation rates, protein folding, protein-protein interactions, and energy consumption for thousands of different DNA sequences, including promoters and terminators, in CHO cells. They are also designing new types of assays to study DNA context and plasmid architectures, including how the position in which DNA is inserted into the genome influences its ultimate expression. These datasets are then used to improve our engineering tools. Our modeling teams have also made a CHO simulator tool to predict how metabolite levels shift as cells make antibodies. (We’ll write about this in our next blog.)

In the next few years, we’d like to collect data and build models that can predict the optimal genetic design and bio-manufacturing conditions of a therapeutic molecule merely by looking at its sequence. Imagine if you could take an antibody sequence (say, for Humira^®), type it into a computer, and then get an output that says, “Build this exact plasmid sequence, put it into a cell with this genome, and grow the cells under these conditions.”

Our CHO Edge tools are being used for antibodies first, but we plan to expand to materials, enzymes, and other molecules in the future. 

If this work sounds exciting, come join us. Or, subscribe for updates.

Contributors: HaeWon Chung, Kevin Smith, Lila Wroblewska, Martín Cárcamo, Jamie Freeman, Scott Estes, Rachel Kelemen, Alec Nielsen & Ben Gordon. Words by Niko McCarty.

Read this post on the Asimov website.

Is this the era of biology? It certainly seems that way.

For at least a century, scientists have tinkered with nature to solve problems. Two young researchers at the University of Toronto extracted insulin from dogs in 1921 and used the molecule to treat people with diabetes. The U.S.D.A. bred screwworms, sterilized them with x-rays, and airdropped them over Texas in the 1950s to decimate screwworm populations, which killed hundreds of thousands of cattle each year. The human insulin gene was cloned into bacteria in 1978. Dolly the Sheep was cloned in 1996. 

For a long time, such stories were relatively rare. Now they seem to happen every month.

Dozens of cell and gene therapies have been FDA-approved. Engineered microbes convert steel factory waste into ethanol. Some vaccines are designed on computers. Transgenic goats make lysozyme-laced milk that could help prevent childhood diarrheal diseases. And 95 percent of livestock in America are fed with genetically-modified crops. Our food and medical systems are already reliant upon biotechnology. In a few decades, so too will just about everything else.

When progress is swift, though, stories lag. Too many people are unaware of biology’s promise, or hold only vague thoughts about its perils. We need deeper, data-driven stories that make sense of the past, present, and future of biotechnology.

That’s why we’re building Asimov Press.

Asimov Press will be editorially-independent from Asimov, the company. It will publish essays, themed magazine issues, and books about engineered biology, especially at the intersection of AI and living systems. It is inspired by Stripe Press, a publishing venture that is funded by the payments company and aims to share “ideas for progress by publishing books about economic, technological, and societal advancements.” Asimov Press will cover diverse stories, people, and places by publishing work that spans history, biology, software, and ethics. 

Asimov Press will be led by founder and editor, Niko McCarty. We’re planning to hire an additional Editor/Writer to join the team, and also to expand our network of writers and data scientists to contribute articles and graphics. If this sounds like the job for you, please get in touch.

Our ambition is to grow the bioeconomy by sharing good ideas that attract talent, inspire research, and crowdsource solutions to global challenges. We’ll try to make sense of biosecurity, protein design, genome editing, and whatever comes next. Instead of being passive participants in this strange era, we’ll use stories to make sense of, and shape, our collective future.

Asimov Press launches in late 2023. Subscribe here to be notified.

Read this blog post on the Asimov website.

We’ve repeatedly touted the idea of genetic design, a research field that aims to apply biophysical insight to compose and layer genetically-encoded functions to achieve cellular behaviors.

If you scratch out the jargon, that means we want to work backwards from an objective — synthetic photosynthesis or the eradication of a cancerous cell in the body — to a DNA sequence that encodes all the biochemistry needed to make that happen.

Genetic design overlaps with synthetic biology in many ways, but the two aren’t the same. Genetic design is deeply rooted in biophysics and mathematics. It’s about understanding how different cellular components can work together to accomplish a desired behavior, and then building it.

Progress in biology is arguably improving faster than any other time in human history, but even advanced models can’t capture all the complexity of a single cell. We cannot fully model biology, and thus our designs often fail. That’s because there is much we don’t understand about life; thousands of genes in a bacterium as simple as E. coli still have unknown functions. 

Further complicating matters, cells don’t behave like little machines, despite the metaphors we ascribe to them. Biology is stochastic; probabilistic. Genes are not “switched on” at a steady rate, but fire instead in frenzied bursts of activity. Some proteins behave more like squiggly liquids than solids, or entirely change their function when moved from one environment to another (this is known as moonlighting).

And still, even with this foggy and incomplete view, our ability to do genetic design is rapidly improving. This is the story of the field, as told through seven seminal experiments. It is inspired by The Generalist’s essay on the history of AI. 

1. The lambda OR control system (1985)

Stéphane Leduc, a French chemist, was first to coin the phrase “synthetic biology.” In the early 1900s, he described biology as a series of chemical and physical forces that could be modeled with mathematics, and attempted to make life anew by mixing various chemicals. He was unsuccessful, but did make some pretty sculptures.

Over the last century, scientists have devised new tools to unravel the inner workings of living cells (including DNA sequencing and -omics technologies), and the data they’ve collected suggest that Leduc was, perhaps, ahead of his time. Biology can be modeled using mathematics, and cells — even if not built from scratch — can at least be endowed with new functions.

Early work in genetic design focused on bacteriophage lambda, or lambda phage, a tiny virus that infects E. coli bacteria. This phage can exist in two different states: lysogeny, where it incorporates genetic material into a bacterium, or lytic, where it reproduces and then bursts from the cells in a violent wave. 

The switch from one state to the other is controlled by just three genes, and the outcome can be watched with microscopes. Together, these traits make lambda phage an ideal test case to compose mathematical equations and test them in the laboratory.

In a 1985 paper, two biophysicists at Johns Hopkins devised a mathematical model to study lambda phage’s “switching behavior.” Madeline Shea (now a professor at the University of Iowa, and a former mentor of mine) and Gary K. Ackers’ predicted whether a phage would enter the lytic or lysogeny state based on the positions of little DNA sequences within one of the gene’s promoters. 

They used statistical mechanics (probability theory to understand how large groups of molecules behave) to calculate how often proteins and strands of DNA came together, and then differential equations to predict whether that would be sufficient to switch from one state to the other. 

Calculations were done on a Hewlett-Packard 1000, a $23,900 computer with 768 kilobytes of memory. At the time of its release, these computers “were so fast, they were bound by special US government export restrictions,” according to the HP Computer Museum.

And the Shea-Ackers model worked. It predicted phage behaviors over extended time periods, and later became the mathematical precipice upon which synthetic biology was founded.

2. Circuit simulation of genetic networks (1995)

Open a biochemistry textbook and flip to a section on metabolism. Peer upon the spaghetti monster diagrams that depict biochemical pathways, in which lines connect rectangles to lines and more rectangles. These diagrams are a chaotic effort to turn a living cell’s complexity into clean lines and digestible iconography.

The lambda phage genome has merely 48,500 bases of DNA, and its life cycle is controlled by just three genes. But much of biology is not like this; it is far messier. Most biochemical networks are built from hundreds of interacting pieces. And, just thirty years ago, there were no tools or frameworks to help us understand the complex dynamics of such large networks.

But then, in the mid-1990s, Harley McAdams and Lucy Shapiro devised a “hybrid modeling approach” to depict genetic networks. They combined “conventional biochemical kinetic modeling within the framework of a circuit simulation.” For perhaps the first time, biophysicists began to sketch out the complex interactions between biochemical “stuff” — DNA, promoters, genes, proteins — as circles and lines. Each biochemical object was then attached to another using the logic gate symbols from electrical engineering, where a triangle means “NOT,” a half-circle means “AND,” and so on. 

By modeling biology in a way akin to electrical circuits, it became far easier to build a predictive software for living cells. Each biochemical “thing” was stored as an object, and could be manipulated with a suite of functions to piece together a full biochemical network. 

“We are optimistic that libraries of generic object-oriented software models of common genetic mechanisms,” wrote McAdams and Shapiro, “can be developed to provide geneticists the type of user-friendly simulation tools that electrical circuit analysts now take for granted.” 

This was a prediction decades ahead of its time.

3. Early gene circuits (2000)

Much of the history of molecular biology was written by physicists. Erwin Schrödinger popularized a physical view of genetics. Francis Crick co-discovered the structure of DNA and was first to theorize the Central Dogma (he got his start in physics, and later designed acoustic mines for the British army during World War II.) Walter Gilbert studied exclusively physics through his PhD, co-founded Biogen, and developed new DNA sequencing technologies.

The same can also be said for synthetic biology. The field was founded mainly by physicists who modeled biological networks on a computer, took them apart, and rebuilt them anew to endow cells with new behaviors.

In 2000, back-to-back papers appeared in Nature describing early examples of synthetic gene circuits. A Princeton group, made up of a physics Ph.D. student and theoretical physicist, endowed bacteria with a gene circuit, called the repressilator, that made cells flash a green color – on and off – at regular intervals. A Boston University group, led by another physicist, introduced the toggle switch on the same day.

The toggle switch is made from just two genes; let’s call them A and B. Each gene encodes a protein that represses the other, and can itself be controlled by an external factor. Protein A is switched “off” by a small molecule, for example, and protein B is switched “off” by heat. This simple co-repressive design makes it possible to control “toggle” cells between two states: Add some heat and protein A dominates. Add the small molecule, and the opposite happens.

The repressilator was made from three genes, each repressing another. Gene A encodes a protein that blocks Gene B, which encodes a protein that blocks Gene C, and so on, in a perpetual inhibitory loop (in electrical engineering, this is called a ring oscillator). Each protein is also attached to a small peptide that speeds up its degradation, so that the cycle can continue to progress. This causes each protein, in the cells, to come up-and-down in its levels at a predictable frequency; 150 minutes.

These papers laid the foundation for synthetic biology. They showed that it’s possible to use simple mathematical models to build DNA sequences that endow living cells with new behaviors. A third paper, published just a few weeks later, reported a third synthetic gene circuit made from just one, self-repressing gene.

And from there, explosive growth ensued.

4. Programmed pattern formation (2005)

Technologies tend to follow a predictable arc. The first demonstration is usually primitive. But vigilant people take note, tinker around with the new technology, and gradually push it to new extremes. This happened with CRISPR gene-editing: After the early papers, hundreds of scientists quickly pivoted into the field, discovered other gene-editing proteins, made dozens of computational tools, and built an entire ecosystem of genetic engineers. 

The early genetic circuit papers may have initiated synthetic biology, but it was other scientists who pushed designs to new extremes — more equations, more biochemical components, more computer simulations.

The first synthetic biology conference was held at MIT in 2004. In the following year, an electrical engineering team at Princeton, led by Ron Weiss, described a genetic circuit that programmed groups of cells to form patterns in a petri dish. This paper was historically notable for a few reasons. 

Whereas early genetic circuits were confined to individual cells, this version added a “communication” component; all the cells could talk to each other. The design was also incredibly fine-tuned. It required 100 experiments and 2,000 computer simulations to arrive at a solution, and only 30% of the designs produced the pattern. 

First, the researchers placed ‘sender’ cells, programmed to make a small molecule called AHL, in the middle of a petri dish. The molecule diffuses out to nearby cells, called ‘receivers,’ that detect the AHL. Specifically, the molecule latches onto a protein called LuxR. When this happens, LuxR “activates” two branches of a circuit, each of which regulates a green fluorescent protein output.

Now here’s the clever bit. When receiver cells sense either very little or a lot of AHL (which happens at short and long distances from the senders), then the signals from the two circuit branches "clash": one branch is on and one is off. This shuts off expression of the fluorescent output.

But when the receiver cells are located at an intermediate distance, the signals traveling through the two branches of the circuit match, triggering the cells to glow green. This interplay of proteins, and a clever circuit architecture, creates a band detector that causes cells to form a glowing ring pattern in a petri dish.

This paper is one of the earliest demonstrations of true genetic design. Mathematical equations gave way to a computational model, which was then used to simulate thousands of designs until hitting upon those that could coax cells to perform an intricate, finely-balanced behavior.

5. Synthetic Biology Open Language (2009)

In 1886, an American mathematician, named Charles Sanders Peirce, described how logical operations — AND, NOR, OR, and so on — could be performed by “electrical switching circuits.” Prior to this letter, addressed to one of his former students, engineers had built logic machines with purely mechanical mechanisms.

For the next hundred years, logicians and engineers described electrical circuits using all kinds of notations and symbols. Many of these symbols — such as the two parallel lines that represent a capacitor — were adopted because the symbol was representative of the device itself; a capacitor accumulates charge on two surfaces placed near one another. But there were no standards; only norms.

In 1984, though, the Institute of Electrical and Electronics Engineers implemented a standard set of “graphic symbols for logic functions.” The manual includes dozens of little symbols and rules to ensure that an electrical circuit described in a Chinese paper could be immediately recreated by an American academic. Standard notation transcends language barriers. They make it possible to do industrial design and rapidly recreate, reconfigure, and prototype new designs.

But all of this was missing in biology, until about 2009, when a group of 76 people across 37 organizations came together to develop the Synthetic Biology Open Language, or SBOL. The group introduced dozens of symbols to depict various DNA sequences, including promoters (depicted as arrows), translation start sites, and even protein degradation tags.

“Such standards could enable synthetic biology companies to offer catalogs of devices and components by means of computer-readable data sheets, just as modern semiconductor companies do for electronics,” the authors wrote. And they further envisioned a future in which these tools would “enable a synthetic biologist to develop portions of a design using one software tool, refine the design using another tool, and finally transmit it electronically to a colleague or commercial fabrication company.”

6. Design of individual genetic parts (2009)

In 1957, a General Electric employee named Patrick Hanratty developed PRONTO, the first numerical control programming software. It was used to control machining tools; telling them where to cut, how deeply, and so on. 

Three years later, a PhD student at MIT, named Ivan Sutherland (who was working with Claude Shannon, the “father of information theory”) created the first true graphical, computer-aided design program. Called Sketchpad, it made it possible to draw shapes on a screen, adjust them, and export them to be reused by others. It laid the entire foundation of graphical-user interfaces.

Today, CAD tools are vastly more complicated and powerful. Many electrical engineers design circuit boards using AutoCAD Electrical, which has built-in tools to pick, say, wires that have continuous load correction and can operate at high temperatures. In other words, modern CAD tools will help you pick the right parts for a job.

This is missing in biology. About 15 years ago, there were no software tools to design biological sequences with predictable outputs, which could then be compiled into more complex genetic circuits. The Ribosome Binding Site (RBS) Calculator did that.

In cells, ribosomes are the big RNA:protein complexes that make other proteins. In bacteria, they latch onto a small snippet of messenger RNA, called the RBS, and begin translation. This short sequence plays a large role in how often a piece of RNA is translated into proteins. A ‘weak’ RBS coaxes ribosomes to make a little bit of protein, whereas a ‘strong’ RBS coaxes them to make a lot. The RBS Calculator took the biophysical and thermodynamic equations describing the interactions between ribosomes and RNA molecules, and packed them into a web tool that could design new RBS sequences

If a specific amount of protein is needed, the tool will suggest an RBS design that is accurate within a variance factor of 2.3. In other words, if you were to target a protein production rate of 100, the RBS Calculator outputs sequences that achieve a range between 43 and 230.

This tool matters because it directly outputs a DNA sequence based on a need. And, in engineering, actionability matters. It’s one thing to publish a mathematical equation, and quite another to use it to build a software tool that others can use. Simplicity and actionable results are what will make genetic design blossom into a field where anyone, anywhere, can design cells with custom behaviors on a computer.

7. Automated design of complex gene circuits (2016)

Not that long ago, genetic design was made up of miscellaneous and one-off tools, like the RBS calculator, that were useful but separated. A biologist could design an RBS sequence, or sketch out circuit designs using standard notations, but there still wasn’t really anything that could put all these features together, into one software package.

“Genetic circuit design automation,” a 2016 Science paper, marked a seminal shift. Developed by teams at MIT and Boston University, it described a programming language to design entire logic circuits made from DNA. In the spirit of Ivan Sutherland or AutoCAD, it suddenly made it possible to go from ad hoc designs to all the DNA instructions needed to create a genetic circuit that “functioned as specified.”

The programming language, called Cello, inspired the technological foundation of Asimov. At the time of its publication, Cello was used to design the most complex genetic circuits ever built — ten proteins that work together to encode complex logic functions — with a 75% first-try success rate. In other words, Cello directly outputs a DNA sequence, based on a user specification, that works as expected three-quarters of the time. 

In a series of follow-up papers, Cello was used to build sequential logic, a form of circuit that in real computers underpins memory, timing, and feedback. And, in a separate paper, it was used to program E. coli bacteria “to function as a digital display,” by encoding an entire electronic chip, across multiple strains, in 76,000 bases of DNA. 

Genetic design is getting better, fast. A couple decades ago, the earliest synthetic gene circuits were made from two or three genes. Now, genetic circuits are designed on computers and contain dozens of individual parts. If history is any indication, then we may be able to design new organisms, with entirely distinct biochemistries, in forty years. But we’ll first have to study biology as an intricate whole, and understand how our own designs fit into the context of lifeforms that have evolved, without us, over billions of years. Time will tell.

***

Come join us or subscribe for updates.

Contributors: Ben Gordon and Alec Nielsen. Text by Niko McCarty.

I. Launch

Today, we announced that the MIT-Broad Foundry team has joined Asimov and set up shop in our main office. This move happened a while ago, but we held off on announcing it until now. We are also retiring the name “Foundry” and switching over to Asimov Labs. The rebrand is inspired by the pioneering work of historic teams at Bell Labs, HP Labs, and Xerox PARC, for reasons that we hope will become clear in this blog.

The MIT-Broad Foundry formed in 2012. Over the next decade, its team designed and engineered pathways and living cells to make over 1,200 different molecules and materials, as well as hundreds of genetic devices, such as circuits and sensors. The Foundry was designed to take on wildly different challenges under tight time constraints, like making ten different molecules in 90 days without knowing them ahead of time.

By retiring the Foundry name and shifting over to Asimov Labs, we are reimagining how high-throughput experiments can drive biological progress. Our goal with Asimov Labs is not to use robots to screen thousands of strains, but rather to solve genetic design. We want to reduce the implementation risk in engineering a biological system — a bacterium, mammalian cell, or anything else — to zero.

We think of genetic design as applying biophysical insight to compose and layer genetically-encoded functions to achieve a cellular behavior. It means working backwards from, say, synthetic photosynthesis or the eradication of a cancerous cell in the body to a DNA sequence that encodes all the biochemistry to make it happen. We’re currently able to design systems with 10-100 discrete genetic functions in mammalian cells; but not with zero implementation risk. Design tools and models help, but there is still a lot of trial and error that we are trying to minimize. Our goal is to engineer increasingly complex behaviors until, ultimately, we can design entire genomes. 

This is our moonshot. It may sound outlandish, but our ability to design large and complex circuits using computer-based models has improved massively in the last decade. And now, Asimov Labs will help us move even faster.

II. Moonshot

In the mid-1990s, it often cost more than a thousand dollars to sequence one gene. To cut costs, universities set up their own DNA facilities and offered Sanger sequencing services to academics. They focused on the mundane sequencing needed to run a lab — a student’s plasmid or enzyme mutant, for example.

But as the Human Genome Project took off, these academic facilities saw an opportunity to come together and advance biology in ways that hadn’t been done before. Today, we take it for granted that we can look up any gene in thousands of different species. But twenty-five years ago, this was a pipe dream, realized only because big thinkers asked the right "what-if" questions at the right time.

Twenty universities and research centers drafted the initial sequence of the human genome. The Whitehead Institute/MIT Center for Genome Research, which later became the Broad Institute, sequenced much of the mouse genome, together with scientists at 26 other institutions. That work was published in 2002. The Broad Institute later led an effort to sequence 1,000 human genomes and, after four years of work, surpassed that goal and identified 38 million single-nucleotide differences across 14 human populations.

In other words, academic sequencing centers went after big moonshot projects. This is the context in which biofoundries first emerged. When synthetic biology ramped up in the mid-2000s, many people thought that some projects would be too big for individuals to tackle. Robots would be needed, they thought, to engineer biology in sophisticated ways.

An early idea for a synthetic biology “Foundry” came in 2009, when Chris Voigt (MIT; also an Asimov co-founder) and Michael Fischbach (Stanford) proposed “a high-throughput natural product discovery pipeline,” called FabPharm, at the University of California - San Francisco. They wanted to use robots to assemble genes and discover valuable molecules in nature. Their work built on an earlier effort by Drew Endy (Stanford) and Adam Arkin (UC Berkeley) to create a DNA factory that could print and screen genetic parts.

There are now dozens of biofoundries around the world, including several at companies. Many of them use robots and automated tools to parallelize microbial genome editing, to build genomes, or to coax microbes to make molecules. The Edinburgh Genome Foundry can do more than 2,000 DNA assembly reactions each week, which is 20-times the throughput of a lone scientist. The iBioFab Foundry at the University of Illinois can run 192 CRISPR-Cas gene-editing experiments at the same time. 

Still, it’s rare for biofoundries to go after moonshot projects. They could engineer cells to make every antibiotic found in the human microbiome, or chemically synthesize the entire human genome (see GP-Write) to learn more about how it’s regulated and controlled. Biofoundries have already banded together to do this for the yeast genome. But most others mostly provide small-scale services, such as DNA synthesis or automated experiment pipelines.

Asimov Labs is taking all the capabilities of a typical biofoundry — the robots, people, and high-throughput tools — and applying them to the genetic design moonshot. Our team collects multi-modal measurements on transcription and translation rates, protein folding, protein-protein interactions, and energy consumption across different cell types for thousands of different DNA sequences.

These data are then used to build biophysical cell simulations, which incorporate into generative design algorithms: think of it as a "compiler" that can automatically design multi-gene plasmids. We’re also training deep-learning models to predict DNA functions from sequences in vivo. As these models mature, we’ll give them to scientists via Kernel, our software to design, simulate, and optimize biological systems. We’re inspired by Ivan Sutherland’s work in creating Sketchpad, the first CAD program, in 1963. Sutherland went on to pioneer human-computer interactions at Xerox PARC.

Our ultimate goal is not only to use computers to design DNA sequences that work as expected in individual cells, but to do the same in complex environments at scale, such as large-volume bioreactors. 

Again, this is a moonshot; a work in progress. But our ability to design living systems is expanding quickly, and that gives us hope.

III. Plausible

It is rare for a genetic design — a logic circuit, metabolic pathway, or anything else — to work the first time. Cells are complex and not fully understood, and that’s why biological engineers often test dozens to thousands of designs to find one that works. 

But as we try to engineer more complex systems, from large metabolic pathways to groups of cells that work in symphony, we don’t think high-throughput experiments will find solutions. Instead, we should try to understand biology so deeply that one could predict its behaviors before any engineering begins.

We have a ways to go, but it's undeniable that our ability to design biology is improving fast.

Just look at prime editing, a ‘search-and-replace’ genome editing tool that can insert, delete, or swap bases in DNA. These proteins were made through genetic design, even if they’re not branded that way. The prime-editing protein was made by fusing a Cas9 nickase to a reverse transcriptase taken from a virus. This fusion protein was tested in cells and then rationally engineered to make better variants. Prime editing was not developed by brute force screening; it came from deep, molecular insights. 

Asimov Labs has also used AI models and simulations to make tissue-specific promoters for gene therapies. We did not do any screening for that work. The same team also developed a codon optimizer that outperforms every alternative on the market, at least for antibody production. A codon optimizer swaps codons in a gene sequence to boost gene expression, without changing amino acids in the final protein.

And finally, Asimov was founded, in part, based on work at MIT and Boston University that produced Cello, a programming language that interfaces with E. coli models to design genetic circuits. At the time of its publication, Cello had designed the most complex genetic circuits ever built — ten proteins that work together to encode complex logic functions — with a 75% first-try success rate. In other words, Cello directly outputs a DNA sequence based on a user specification that works as expected three-quarters of the time. 

For a follow-up paper, thousands of high-throughput experiments enabled us to build reliable models to extend Cello to sequential logic, a form of circuit that in real computers underpins memory, timing, and feedback.

So this is the future Asimov Labs is pushing towards. We’re fusing experimental data with biophysical models to design biology in silico. All of this is about building the foundation on which we hope to solve genetic design. We’ll be sure to update you on our progress in future blogs.

Come join us or subscribe for updates.

Contributors: Ben Gordon, Will Johnson, Chris Voigt, Anaïs Moisy, Arturo Casini, Alec Nielsen. Text by Niko McCarty.

Read this post on the Asimov website.

I.

GitHub launched in 2008. Their mission was to build a global platform for developer collaboration. What that means, basically, is they wanted computer programmers to work on projects together by giving them tools to store code (called repositories), track changes to that code, and edit other peoples’ code.

Mission accomplished. The company (now owned by Microsoft) is proof that, if you build useful tools and empower a community, people will come. GitHub is used by more than 100 million developers who have created nearly 400 million repositories. 

These repositories contain templates, guides, and code snippets to help any programmer do just about anything. Want to spin up a website? You can find thousands of demos on GitHub. How about a web scraper? There are hundreds of ‘em.

We are inspired by GitHub and believe that biologists should have access to similar tools, but for DNA sequences and experiments instead of computer code. That’s why we’re building a tool, called Kernel, for scientists to store DNA sequences and associated data, share them with others, and track how they change over time. 

At a micro-scale, we want to help biologists access the world's growing corpus of biochemical functions and genetic systems, and seamlessly collaborate on projects. At a macroscale, we are trying to accelerate biological progress.

II.

Most genetic engineering projects are ad hoc. If you ask ten genetic engineers to design a biological system to detect lead, you will get ten very different solutions, only some of which are likely to work. Biological progress would move faster if scientists knew what has been tried before, and could track how solutions to problems had evolved over time. This will be especially important as projects grow in sophistication and we try to engineer living cells to do more complex things. 

But biologists, and those who seek to engineer cells, are missing basic functionalities that software engineers take for granted. GitHub has three main features that have become a cornerstone of modern programming, and we think similar tools would help scientists, too.

The first feature is repositories, a kind of storage locker for code. Repositories can be copied or shared across the Internet, and marked as public or private. The public ones are searchable, and open-source communities on GitHub are built upon them. Recent research suggests there are more than 28 million public repositories. Developers rarely have to start a project from scratch; they can start with code that someone else has already written.

GitHub repositories also have an Issues tab (but the concept pre-dates the company). This is where people ask questions about code, or request new features. In 2008, though, GitHub introduced pull requests, enabling anyone with access to a repository to write their own changes and then “request” to add them to the project. All of a sudden, a random developer in Kazakhstan could spot a bug in someone else’s repository and write a fix for it, seamlessly. This feature spurred the growth of open-source communities. It shifted collaboration from passive issues into active solutions.

And the final feature is version control. Every time code is altered on GitHub, a digital record is added to the file. The genesis and evolution of every line is meticulously documented, which makes it much easier to find the source of an error or the origin of an idea. Developers can also revert to earlier versions when something goes wrong.

Now, what about biology? Are there any existing parallels?

Not really. There are open-source biology projects, of course, and many benevolent scientists share their notebooks and ideas. But most papers are locked behind paywalls (including nearly half of biomedical papers) and there aren’t really any tools to share, fork, collaborate, or actively make edits on a DNA sequence. 

There are many tools to digitally edit DNA, but they lack flexibility in controlling broader access. These tools don’t allow a scientist to, say, open-source a project with all of its DNA sequences and experimental data, giving the entire world the ability to natively and seamlessly fork and modify its contents.

Ideally, it would be as seamless to view DNA sequences, edit them, and share them with others as it is to do all those things with code on GitHub. That’s what better software tools for genetic engineers would enable.

Imagine if any scientist could make a DNA sequence, or fork someone else’s, and store it in a repository. Sequences could be marked as public or private, and shared with any number of individual collaborators who can view, make comments, or edit. This could help grow the open-source community. Every change to a DNA sequence could also be tracked and cataloged over time. The intellectual origins of every sequence would be documented. This would be transformative.

Often, when we modify a DNA sequence in our laboratory, we encounter mundane problems like, “Where did this mutation come from?” or “Why did we add this bit of DNA to the end of this gene?” The answer is rarely obvious. We have spent hours in Slack channels, pinging teammates about where this or that sequence came from. A version control system for genetic engineering would solve this. Every DNA sequence would have a detailed family tree of derivations, highlighting how parts have been combined, modified, or mutated over time.

Finally, our software tool would store all experimental data associated with a DNA sequence — how many copies of an RNA it makes, or how its final protein interacts with others in the cell — in the repository itself. By linking a DNA sequence to outcomes in living cells, it will be easier to design and debug DNA that works in predictable ways in the future.

III.

We’re working on a software tool that will encapsulate all of these features. It’s called Kernel, and we recently launched a beta version to select students. 

Kernel has a curated database of 650,000 searchable DNA sequences. Each sequence can be dragged-and-dropped onto a DNA assembly canvas to construct plasmids with custom functions. Kernel also includes a simulator tool that predicts how these DNA constructs will behave inside cells; their transcription rates, protein expression levels, and so on.

But there’s still a lot of work ahead. We are actively adding version control tools to track how DNA sequences are used or altered over time, for instance, and we are preparing to launch new permissions tools that will make it easier for scientists to collaborate on private projects, or to open-source entire collections of DNA sequences.

Our goal is to build a community around this tool. But we know that’s dependent upon us releasing useful features that make it easier for biologists to do all the things they already do. As the community grows, we hope others will build a suite of custom extensions or plugins to make it easier to design, or debug, biological sequences. 

It has not escaped our notice that a GitHub-inspired platform for genetic engineering could enable something analogous to Copilot: an AI-powered tool that generates biological sequences with desired functions. More on that in a future post.

If you’d like to build software tools for genetic engineers, come join us. Or, subscribe for updates.

Contributors: Ben Gordon, Kevin LeShane, Anaïs Moisy, Alec Nielsen & Layton Wedgeworth. Text by Niko McCarty.

Read this post on the Asimov website. Read Part I.

Strings of DNA

Every cell in the human body has the same genome. There are slight variations between each one — so-called somatic mutations — but the basic gist holds true.

This raises a difficult question: If every genome in every cell is (basically) the same, what accounts for the remarkable diversity of form and function in a single person? A neuron fires action potentials up to 100 times per second; the signals ripple through the brain at speeds exceeding 500 miles per hour. Osteoclast cells break down bones and can contain up to 100 nuclei, instead of just one. Skin cells in the feet make toenails, but skin on the elbow doesn’t. 

How does a cell know what it is, or what it does?

The answer is genome regulation. Individual cell types form during development and then become rooted — fixed — in their identities. Each cell type coils up its genome in a distinct way. Large segments of genes are ‘shut down’ in neurons, but those same genes remain open and active in toenail-making cells. Every genome also has epigenetic differences; chemical tags (like acetyl or methyl groups) tune gene expression. 

In short, two cells with identical DNA will package and read the genetic code in different ways.

It’s not yet possible to look at an arbitrary genome region and say, “this will be active in the retina,” or “this will be active in the heart.” But we are genetic designers, and we want to reliably engineer biology. By designing tools that can analyze a DNA sequence and predict its expression across the body, it becomes possible to also do the reverse: design custom DNA that is active only in specific tissues. Such a tool would be transformative for gene therapies. 

Our previous blog post explained how viral vectors are packaged with DNA and injected into the human body to treat disease. Gene therapies are a molecular marvel, but they also have problems. Viral vectors often transduce cells that we don’t want them to. A gene therapy for the brain often ends up in the liver. These off-target effects can cause side effects and are sometimes lethal. By packaging gene therapies with DNA sequences that only ‘switch on’ in specific tissues, though, we can improve their safety and efficacy.

In this blog, we explain how we are using transformer-based AI models to study gene regulation and then design synthetic promoters that accomplish this aim.

AI-Guided Design

There are about 20,000 protein-coding genes in the human genome. Deciphering how all of them switch ‘on’ and ‘off’ has challenged scientists for the last six decades. So let’s focus, instead, on just one gene.

In human cells, regulatory sequences surround the coding sequence that encodes a protein. Upstream of the gene, a promoter acts as an ignition switch for transcription. Downstream, a terminator signals the end of transcription. Enhancer sequences boost a gene’s expression and can appear anywhere, upstream or downstream, even millions of bases away.

This model of a gene is simplistic. It makes it sound as if we’ve got everything figured out! But even in cases where we know the identities of all the proteins that interact with a gene, and have mapped the sequence and position of every element — promoters, terminators, enhancers, and more — we still don’t fully understand how all the molecules come together and actually dance to control a gene’s expression.

AI tools can help. They’re already used to study complex phenomena, such as computer vision or natural language. But perhaps they can also predict how genes are regulated, and thus make it possible to design new promoters that behave in desired ways.

Consider transformers, the same type of neural network underlying ChatGPT. Originally developed to translate languages in 2017, transformers quickly surpassed recurrent neural networks in their ability to both translate language and infer the semantic connections between words.

Whereas a recurrent neural network ‘reads’ and ‘processes’ one word at a time — and quickly forgets words as it moves along a paragraph — a transformer can be parallelized across many words while learning and tracking the semantic connections between each one.

Consider two sentences: “Server, can I have the check?” and “Looks like I just crashed the server.” The word server appears in both sentences, but it has a different meaning in each case. The intended meaning of the word can be deciphered by looking at adjacent words, such as check in the first sentence and crashed in the second. Transformers pick up on these nuances.

We are fortunate, in a way, that biological sequences are reminiscent of written languages. DNA and proteins are represented as strings of letters, and these letters have regulatory and functional meanings. The letters TATAAA in a DNA sequence has a physical meaning: It’s a binding site for transcription factors that activate or repress a gene. The similarities between words and DNA mean that transformers can be brought to bear on biology.

AlphaFold used a transformer to predict 200 million protein structures from strings of amino acids. Other groups have used transformers to predict gene expression from DNA sequences or to predict which proteins will get ‘tagged’ with sugars in a cell.

At Asimov, we recently developed a computational tool to design synthetic promoters that are only active in specific tissues. The computational tool merges a transformer-based model, tissue RNA-sequencing datasets (a measure of gene expression in different parts of the body), and evolutionary data on sequence conservation (i.e. gene sequences that appear in both mice and humans). 

The generated sequences successfully confined gene expression to our target tissues. And although the synthetic promoters do not naturally exist in nature, the evolutionary data used in their design means that they are effectively "humanized." We expect that these promoters will translate from mice into humans without a loss of function.

The Experiment

Using our transformer-based design tool, our engineering team modeled and synthesized 17 synthetic promoters to target the heart, brain, liver, or muscles. Each promoter was fused to a luciferase coding sequence. The luciferase protein emits light, and it’s the same molecule that fireflies use to flash and communicate.

Luciferase will emit light in the presence of luciferin, ATP, magnesium, and oxygen. When we add luciferin to an animal’s drinking water, their organs literally glow if the luciferase payload is expressed within. We chose to use luciferase in these experiments because it has a higher sensitivity than fluorescent reporters. We are able to measure luciferase even if it is only weakly expressed. 

Each DNA sequence — a promoter, luciferase coding gene, terminator, and other bits and pieces — was packaged into a recombinant AAV9 capsid to simulate their delivery as a gene therapy. An AAV is made from two parts: a 20-sided icosahedral protein shell, called a capsid, and up to 4,700 bases of single-stranded DNA packed inside.

Three mice, all with an identical genetic background, were intravenously injected with the engineered AAVs. We used two well-studied promoters as controls: CMV, which is understood to express in all tissues, and MHCK7, which expresses in both heart and muscle. 

Every animal was given the same dosage of viral vector: 6 x 10¹³ viral genomes per kilogram of body mass. (A typical mouse weighs less than one ounce, so the animals were each injected with a few billion viral vectors.)

And then, we waited.

After two weeks, we imaged each animal and directly measured the bioluminescence of each tissue using a camera that counts photons. In the fourth week, we repeated the imaging and collected tissues for DNA quantification and RT-qPCR, a technique that counts the number of mRNA transcripts for a gene. We divided the total number of RNAs for luciferase by the number for GAPDH, a protein that has consistent gene expression throughout the day. We repeated this for every mouse and every promoter.

The synthetic promoter that we designed for the heart worked as expected. The luciferase imaging data and RNA quantification results both indicated that the promoter expressed luciferase at levels 120-times higher in the heart compared with the liver (see the Heart panel in the figure.)

The packaging limit of a recombinant AAV is quite small — just 4,700 bases of DNA — so we also wanted to see whether it’d be possible to truncate, or shorten, our synthetic promoter while maintaining its specificity for the heart. And it was. A shortened version of the heart-specific promoter had 29-times higher expression in the heart versus liver (see the Heart / Truncated panel in the figure), even when we cut the promoter to 40 percent of its original length. 

Doubling the gene therapy’s dose also did not substantially increase the promoter’s expression in off-target tissues. It resulted, instead, in 208-times higher expression in the heart versus liver (see the High dose panel). We’ve already taken these observations and fed them back into our model to improve predictions for future design cycles.

Results for other promoters were equally intriguing. 

A promoter designed to express in the muscles, for example, did not work at all. There was no bioluminescence and very little RNA expression (see the Muscle panel). But when we chopped off a small part of the promoter’s sequence, it became 37-times more active in the intestines compared to the liver (see the Muscle / Truncated panel.) We don’t fully understand why this happened, but the intestine contains another type of muscle: smooth muscle. Could this result be due to our synthetic promoter getting its wires crossed between these two muscle types? We’re not sure.

The transformer model did remarkably well overall. Computational tools don’t need to be perfect every time. They just need to be good enough to increase the probability of getting a winner within the confines of an experimental budget (a few dozen mice, a 96-well plate, and so on).

A promoter sequence that is 100 bases in length has 4¹⁰⁰ possible permutations. This number far exceeds the number of atoms in the universe. Our transformer-based design tool shrinks the space of combinatorial possibilities. It finds winners faster than random screens or high-throughput experiments. And, we think, it will help us create better gene therapies.

If you’re excited about the intersection of AI and biology, come join us. Or, subscribe for updates.

Contributors: Alina Ferdman, Ben Gordon, Will Johnson, Alec Nielsen, Raja Srinivas, Stephen von Stetina & and Dinghai Zheng. Text by Niko McCarty.

Read this post on the Asimov website. Read Part II.

We are living through a gene therapy renaissance. The F.D.A. has approved treatments for spinal muscular atrophy, multiple myeloma, and to restore sight in those with Leber congenital amaurosis, an inherited form of blindness. One-time treatments to lower blood cholesterol are likely on the way.

At least ten thousand people have already received a gene therapy. It is a modern miracle that we can package custom DNA inside of viral vectors, inject them into the body, and treat genetic disorders by supplying cells with “repaired” genes.

Just consider the F.D.A.-approved treatments for hemophilia A and B, genetic conditions that keep blood from clotting. In the early 1900s, most people with the condition died by 13 years of age, partly because there was no way to store blood. The only treatment, at that time, was direct transfusion from a family member. Today, people with hemophilia can be cured after a single injection.

But scattered amongst these success stories are reports of patient deaths and adverse events. An 18-year-old named Jesse Gelsinger was the first to die from a gene therapy, back in 1999. The F.D.A. swiftly shut down the clinical trial and launched investigations into 69 others, kicking off a ‘gene therapy winter.’ Biotech companies shuttered their doors and funding dwindled.

More than twenty years on, the biotech industry has adapted and boomed. Gene therapies are being approved at a rapid clip. Dozens of companies are searching for ways to make them safer and more effective.

In the next blog, we’ll dive into our recent work on using AI-guided design to create more precise gene therapies. But we first need to survey the current landscape: How are gene therapies made, how do they work, and what problems need solving? Only then can we appreciate the problems and perils of delivering genes to cure disease.

Reality

Jesse Gelsinger had a mutation in a single gene, called ornithine transcarbamoylase. This mutation rendered the protein — part of the urea cycle, which turns ammonia into urea — defective, causing ammonia to build up in his bloodstream. 

To (hopefully) fix the problem, researchers at the University of Pennsylvania packaged a ‘correct’ version of the OTC gene into an attenuated cold virus — an adenovirus — and injected it into Gelsinger’s artery. He died 4 days later after experiencing “a severe immune reaction to the [viral] vector.” 

In the aftermath of Gelsinger’s death, many biotech companies shifted toward adeno-associated viruses, or AAVs, which don’t replicate in human cells on their own, and are not associated with any human disease. An AAV is made from just two parts: a 20-sided icosahedron, called a capsid, and up to 4,700 bases of single-stranded DNA.

AAVs are also naturally diverse. At least 13 serotypes have been discovered so far. Each one has a unique capsid, dotted with sugars and proteins, that together determine which tissues the AAV targets. AAV9 delivers its genetic payload to the liver, muscles, and lungs, whereas AAV2 targets skeletal muscle, retina, liver, and neurons.

If a gene therapy ends up in the wrong places, side effects can ensue. Over 30% of gene therapy trials — including those that use AAVs — report a serious adverse event. Four patients died in a recent trial for X-linked myotubular myopathy (in which muscles slowly waste away) and another died in Pfizer’s trial for Duchenne muscular dystrophy. A 27-year-old named Terry Horgan recently died from a separate gene therapy for Duchenne. An intense immune response against the virus was the likely cause, according to a recent preprint.

An FDA advisory panel recently recommended SRP-9001, an AAV-based gene therapy for Duchenne muscular dystrophy, for approval. This committee doesn’t make final decisions, but the vote was close — 8-to-6 — and the same panel had previously planned to reject the therapy, according to STAT reporting, after a participant’s blood magnesium levels unexpectedly dropped. The F.D.A. paused the trial last June, and it resumed three months later. (Update: The F.D.A. approved the therapy, now called Elevidys, on 22 June 2023.)

To reduce safety issues, AAV gene therapies often target cells in an immune-privileged part of the body (like the eye, which limits inflammatory immune responses “so that vision isn’t harmed by swelling”), or they are given at a low dose.

Luxturna, an AAV gene therapy for blindness, is injected straight into the eye. Other gene therapies that target tissues around the body use a dose as low as 5.8 billion viral genomes per kilogram of body weight (which is not as high as it sounds!) Zolgensma, a one-time therapy for spinal muscular atrophy, is injected into the bloodstream at doses thousands of times higher. Last August, Novartis reported that two patients given Zolgensma died from acute liver failure.

These constraints are problematic. Often, we want to cure more than just genetic forms of blindness, and we often need high doses to do it. Fortunately, a cadre of companies are now engineering AAV capsids, or the DNA inside them, to make safer and more precise gene therapies.

Safety First

To engineer a gene therapy, one first needs to understand how they get made.

Most companies make AAVs by transfecting plasmids, or loops of DNA, into cells. These plasmids encode genes that make functional AAVs. One plasmid encodes the ‘payload’ — usually a human gene, designed to treat disease — which gets coiled up and packaged into the capsid. By tweaking the sequences encoded on these plasmids, one can also alter an AAV’s shell or the payload inside.

Two types of cells are often used to make AAVs: HEK293 (an immortalized human kidney cell) or SF9, from armyworm moths. But many factors determine a gene therapy's safety and efficacy. We focus here on three: Product quality, immune reactions, and precise expression. All three are also intertwined: lower AAV product quality often means a higher dose, which exacerbates immune reactions.

Let’s look at each problem, one at a time.

The first, product quality, is all about manufacturing. When an immortalized insect or kidney cell starts churning out AAVs, not every capsid gets filled, and not every capsid is created equally. Sometimes the cells make empty protein shells, with no genetic payload. Ideally, the ratio of full-to-empty capsids would be as high as possible, so that a gene therapy’s dose can be kept as low as possible. In nature, AAVs package capsids nearly perfectly, which means almost every particle is infectious. But when a scientist uses recombinant plasmids to produce AAV, the fraction of empties goes up, and nobody seems to fully understand why. 

Immortalized insect cells make greater quantities of virus, compared to kidney cells, but the AAVs seem to be less effective at delivering their payloads. Recent work from Nicole Paulk’s group (now the CEO of Siren Biotechnology) suggests two reasons for this: either the insect cells add post-translational modifications — chemical ‘tags’ — to the outer shell of the AAV, or they add epigenetic marks to the single-stranded DNA within. Some of these chemical tags may also cause an immune response. Either way, AAVs made by armyworm moth cells seem to be less potent.

The second problem is immune reactions. Most people are exposed to AAVs during their lives, and more than 60% of adults have antibodies against them. If an AAV capsid is recognized by the body as “foreign,” cytotoxic T lymphocytes fight them off by attacking human cells infected with the virus. Inflammation kicks in and can put a patient at risk. This is why many patients with preexisting antibodies to AAV are excluded from gene therapy trials, according to a recent study. Those antibodies can ‘shut down’ the therapy, and also cause an immune overreaction, which makes a trial less likely to succeed. It also means that lots of people who might need gene therapy can’t get it.

The final problem is precise expression. Remember that AAVs are ‘dumb’ protein shells. They don’t have a brain. They can’t think for themselves. So an AAV floats through the blood and infects many different cells along the way. An AAV designed to target the heart often ends up delivering its DNA to the liver and intestines, too, which can cause side effects later on.

These are the basic problems, then, and each one is connected to all the others. At Asimov, we believe that safer, more precise gene therapies are achievable if we adopt an engineering mindset. We think it’s possible to tweak proteins or alter DNA in juuuuuuust the right way to minimize immune reactions and off-target expression. 

And we are not alone. Lots of clever people are coming up with nifty tools to do precisely that.

Engineering Mindset

Each problem — full/empty ratio, immune reactions, and off-target expression — is solvable. 

The first problem is partially solvable during the purification process. After cells make AAVs, the capsids can be harvested by centrifugation, for example, which means they’re spun at very high speeds. Heavy things (like filled AAVs) settle to the bottom of the tube, while lighter things (like proteins and water) stay at the top. In this way, it’s also possible to enrich filled capsids, compared to empty ones, by about 50 percent. Another option is to use tangential flow filtration, in which molecules flow past a thin membrane and the small bits get removed. Both methods are imperfect at removing empty capsids.

The immune response and off-target expression problems are more difficult to solve, in part, because they’re so tightly connected. Solving one problem would mitigate the other, and two solutions have risen to the top: Either engineer the AAV capsid to ‘evade’ the immune system or target more precise tissues, or engineer the DNA inside to achieve a similar outcome. 

Emeryville-based 4D Molecular Therapeutics engineers AAV capsids. More specifically, they take the plasmids encoding the protein shell and randomly mutate them to make billions of AAV variants. These variants are then screened in monkeys to find capsids that infect cells only in select tissues, like the heart or brain. It seems to be working! Their mutated capsids are being tested in at least five phase I/II clinical trials for everything from eye disorders to Fabry disease, a rare genetic disorder that causes fat build-up in cells.

At the University of California, Berkeley, David Schaffer’s group (a founder of 4D) has helped to unravel how conserved epitopes — a molecule that antibodies bind to — on various AAV serotypes can be mutated to cloak viruses from the immune system. Other groups are also searching for AAV capsids from non-human animals, in nature, that can help gene therapies avoid antibodies in the body.

Other companies, opting for a less “random” approach, are building machine learning tools to design AAV capsids in a more rational way. Dyno Therapeutics recently reported that their AI-designed AAV capsids can pass through the blood-brain barrier in monkeys and deliver genetic payloads 100x better than recombinant AAV9 vectors to the central nervous system.

Besides engineering the AAV shell, one can also tweak DNA directly to help a gene therapy evade the immune system or target precise parts of the body. George Church and Ying Kai Chan at Harvard have crafted AAVs that “evade” the body’s immune and inflammatory responses, simply by adding a short snippet of DNA into the genetic payload. The DNA interacts with toll-like receptor 9, a protein that senses foreign DNA, and dampens its effect.

Other teams are altering DNA to ensure that gene therapies only become active in the ‘correct’ place within the body. A gene therapy for heart disease, in principle, shouldn’t be active in the liver, and a gene therapy for the brain shouldn’t be expressed in the muscle. But this is easier said than done.

A typical recombinant AAV genome has four components: A promoter, the transgene (flanked by short DNA snippets called inverted terminal repeats), and a termination signal. The transgene is the bit of DNA designed to ‘fix’ the broken gene, but the promoter determines how much transgene gets expressed and in which cells.

It is not, contrary to popular belief, always a good idea to crank up transgene levels. Sometimes, the cure for a genetic disease is toxic, which was indeed the case for two gene therapies tested for Rett syndrome in mice and GM2 gangliosidosis, a rare disorder that destroys nerve cells in the brain, in non-human primates. Transgene levels can be tuned by using a ‘weak’ promoter that doesn’t get transcribed into mRNA quite so much, or by packaging AAVs with short, regulatory RNAs that silence DNA.

Figuring out ways to confine transgene expression to specific parts of the body is a more challenging problem. Here, again, there are two common solutions: Either package AAVs with microRNA binding sites that shut down transgene expression in cells with a matching microRNA (a decades-long area of study), or design tissue-specific promoters that only ‘switch on’ in certain cells.

Let’s focus on that second bit: Tissue-specific promoters. There are dozens of them. One, called CD68, confines transgene expression to macrophages, a type of immune cell. Another, called MHCK7, is most active in the heart and muscles. But these naturally-occurring promoters often ‘leak’ into other tissues; when we packaged MHCK7 into AAV9 and injected mice, we found transgene expression not only in heart and muscles, but also in liver, lungs and brain. The problem of making a gene therapy just for the heart or just for the brain, then, remains unsolved.

The process to invent new tissue-specific promoters is also, unfortunately, ridiculously artisanal.

Most tissue-specific promoters are made by smashing a promoter and enhancer (a distant sequence that boosts a promoter’s activity) together. The new sequence is then packed into an AAV and tested in animals. Often, it doesn’t work, because that’s the nature of trial-and-error. 

A better way to get tissue-specific expression, we think, is to design promoters that don’t naturally exist in nature. And this has been done before! An MIT group previously made synthetic promoters that target only cancer cells, while avoiding expression in healthy calls. It was impressive work. But cancer cells express lots of genes that healthy cells don’t, and so there are lots of promoters that differentiate the two. It’s challenging to make tissue-specific promoters that distinguish between two healthy cells in the body, like those in the stomach and intestines (which are both made of muscle, after all!)

In the next blog post, we’ll describe how we used an AI transformer model to create 17 synthetic promoters with tissue-specific expression, including one that is 208x more active in the heart than the liver. We think this tool could go a long way toward making gene therapies safer and more precise.

But these are the early days of a new gene therapy era. Jesse Gelsinger was the first person to die from a gene therapy, nearly 25 years ago. And, over the last two decades, biotechnology has made incredible strides. Several gene therapies now have F.D.A. approval and dozens more are likely on the way. 

Meanwhile, datasets have ballooned in size. There are millions of genome sequences, transcriptomics datasets, and protein structures. We have atomic-resolution images of AAV capsids. Scientists are tweaking and testing dozens of tissue-specific promoters. All of this progress coincides with a veritable explosion in AI tools. The key will be to bring these advances together to cure once-intractable diseases.

We think that fewer deaths and more success stories will follow, and thus complete this modern miracle.

Join us or subscribe for updates.

Blog Contributors: Scott Estes, Ben Gordon, Mike Leonard, Alec Nielsen, Kevin Smith, Chris Stach, Chris Thorne & Lila Wroblewska. Text by Niko McCarty.

Read this post on our website.

Biology is the world’s most advanced technology. It is precise, diverse, and operates on wide scales.

A cell can rearrange molecules with atomic precision, using enzymes that perform millions of chemical reactions each second. A redwood tree compiles 250 tons of carbon into roots, leaves, and other structures during its lifetime; the entire tree grows using the DNA instructions encoded in a seed. At a smaller scale, there are billions of microbes in a small test tube culture, and some microbes can divide in less than 10 minutes. A single microbe can become a billion in under five hours.

All this to say: Biology is capable of incredible feats. Organisms can be harnessed to do great things in the world, if we could reliably engineer them. Genetic engineering, though, is still in the dark ages compared to most other technical fields.

Today, cells are engineered mostly through trial-and-error. Scientists use robots, and rely on the natural multiplicity of organisms, to perform millions of experiments in a single vial. But as we try to design increasingly complex biology, this approach will likely fail. We cannot test even a small fraction of the combinatorial space within biochemistry. (A protein with four amino acids has 160,000 possible permutations. The average human protein has 430 amino acids.) If we truly want to harness living cells to “make just about anything,” then we need to make genetic engineering a quantitative, predictable field. 

To do this, Asimov is building a suite of tools, cell lines, genetic parts, and CAD software. We use them to engineer cells to perform functions that would otherwise be impossible. Our CAD software is used to design DNA constructs made of individual genetic parts, and also to model how the DNA will work in cells. When we invent a new tool, improve a predictive modeling algorithm, or collect data on new DNA sequences, everything is added to the platform. We license everything to scientists.

A key goal here is democratization: We want anyone to be able to use engineering-grade tools to reliably design biology in sophisticated ways. If we’re successful, much of what we use in our everyday world — food, medicines, and clothes — could be a product of genetic engineering. Things that are already produced by genetic engineering — insulin, CAR-T cancer therapies, pest-resistant crops, and mosquitoes that curb disease — will become cheaper and more accessible.

We know that this vision will be difficult to achieve. And, like other audacious goals, it will require a veritable village to make it happen. That’s why we are launching a technical blog to document our ideas, progress, and impressions.

We will share our excitement, reveal our challenges, and explain what we are doing to move things forward. We also want to push beyond the hype and hyperbole that is so common in synthetic biology, and openly discuss “how the sausage gets made.”

This blog is not a place for press releases. It is inspired by the technical blogs published by Netflix, Square, and Stripe. We will do our best to build credibility over time, by sharing as many details as we possibly can.

Not all of the blogs will be about Asimov, either. Some will be deep explainers about emerging topics within genetic design, like deep learning for protein function or how to make more precise gene therapies. Other pieces will discuss our work in software design, automated bioreactor data processing, biophysical modeling, and therapeutics biomanufacturing. Each post will be co-written by scientists working on those projects.

We know that the key to credibility is specificity, so we’ll do our best to explain everything in an open and transparent way. When we don’t know the answer to a question, we’ll tell you. When we can’t talk about something because it’s proprietary, we’ll tell you. But we will always strive to be deeply technical in our effort to demystify biology.

Join us or subscribe to get updates twice per month.

Subscribe now