Niko McCarty

Writing on biology.

Twitter ↗
LinkedIn ↗
E: nsmccarty3@gmail.com

  • This OpenAI Wet-Lab Blog is Pretty Good

    There’s a recent blog from OpenAI where they used GPT-5 to optimize a common biology experiment, called Gibson Assembly. I’ve seen criticisms online from people who say things like, “Who cares? A human totally could have done that” or whatever. And that’s true. But I still think this blog is nice for a couple reasons.

    First, faster iterations is one of the best ways to accelerate biotechnology progress more broadly. Experiments take much too long, and are often much too unreliable, for scientists to move quickly. Therefore, we should invest more resources toward optimizing and improving common methods that seem “mundane”.

    Second, this is a simple experimental system in which to test AI; indeed, that’s the whole point! Gibson Assembly has been around for nearly two decades, is widely-used, and only requires three enzymes. It is therefore a natural fit for AI companies to benchmark their models on biological questions. (The parameter space is not too large!)

    To understand what OpenAI actually did, I first need to tell you about Gibson Assembly, a common method biologists use to stitch DNA molecules together. Originally developed in 2009, most scientists use Gibson because it’s dead simple: Everything works at one temperature (50°C) and it requires only three enzymes. The DNA molecules to be joined together are designed such that they have 15-40 nucleotides, at either end, which overlaps with the other DNA molecule. All the DNA is then added to a tube and an enzyme, exonuclease, “chews back” several dozen nucleotides from the 5’ ends of each molecule, leaving behind long single-stranded “arms.” These arms float around in the liquid, collide with a matching arm in another DNA sequence, and hug each other tightly. A second enzyme, DNA polymerase, runs along these touching DNA strands and fills in parts of the arms that don’t overlap or are still single-stranded. Finally, DNA ligase seals the “nick” and heals the strands, thus forming a newly assembled, double-stranded piece of DNA.

    OpenAI collaborated with a new biosecurity startup, Red Queen Bio (co-founded by Hannu Rajaniemi, an excellent science fiction writer), to build the evaluation framework. The metric they settled on is called cloning efficiency, which just means this: For a fixed amount of input DNA (like one picogram) transformed into cells, how many colonies successfully grow and contain the correctly assembled DNA molecule? By the end of their blog post, the OpenAI team claims that they were able to boost this number 79x relative to a “baseline protocol” from New England Biolabs, or NEB, a common purveyor of the Gibson enzymes.

    An important note is that OpenAI says no humans were involved in optimizing the reaction; all the humans did was carry out protocols generated by GPT-5, and also upload experimental results back into the model. They repeated this several times, coaxing the model to iterate each time. Their Gibson Assembly was remarkably simple, involving just two DNA molecules: a gene encoding a fluorescent protein and a plasmid to hold the gene.

    (The OpenAI team, intriguingly, also set up a set up a robot to automate the Gibson Assembly and transformation, but couldn’t get it to work as well as a human. “We compared the robot’s work to human-performed experiments at each step. The robot successfully handled the transformation process…When compared directly with human-performed transformations, the robot generated similar quality data with equivalent improvements over baseline, showing early potential for automating and accelerating biological experiment optimization.” However. “while the fold-changes between the robot and human experiments were similar, absolute colony counts from the robot were approximately ten-fold lower than manual execution.”)

    After several rounds of iteration, the model made two notable proposals:

    First, it added two additional enzymes to the normal Gibson Assembly reaction. Specifically, it added “the recombinase RecA from E. coli, and phage T4 gene 32 single-stranded DNA–binding protein (gp32).” The blog continues: “Working in tandem, gp32 smooths and detangles the loose DNA ends, and RecA then guides each strand to its correct match.” This tweak improved the “cloning efficiency” metric by 14x over the standard NEB protocol.

    Second, it made a subtle change to how the assembled DNA molecules were inserted into living cells. Specifically, GPT-5 told the humans to spin down cells in a centrifuge, thus forming a pellet, prior to transforming them. This is typically not recommended because competent cells are “fragile,” but the OpenAI team writes that “the cells tolerated concentration well and the increased molecular collisions boosted transformation efficiency substantially (>30-fold on final validation).”

    Now, recall that at the start of this little blog I said I really liked this experiment! (Do not crucify me, ye AI optimists.) But no internet commentary is truly complete without some nitpicking, so here goes.

    One criticism is that the largest improvement made by the model was not related to Gibson Assembly at all! It was related to how the DNA gets delivered into cells. And, indeed, prior studies have shown something similar. (This research paper, for example, says that one of the best ways to improve transformation is to concentrate cells beforehand. Fair play to the OpenAI team for linking to this in their blog post.) And if you are a human reading this blog, and you are planning to spin down your competent cells before transformation, just be sure to aliquot everything into small tubes first; repeated spins will, over time, kill everything.

    Another issue is that adding RecA and gp32 to a Gibson Assembly reaction complicates things quite a bit. For a normal Gibson reaction, everything comes in a single kit from NEB with the enzymes, and the whole experiment is done at one temperature: 50°C! But doing a Gibson Assembly this way would require one to buy purified RecA and gp32, and also change incubation temperatures to get everything working (RecA and gp32 work best at 37°C.) This is more expensive and more complicated, but maybe worthwhile in some cases.

    And lastly, the selected metric — namely, how many colonies one gets from a given amount of DNA — doesn’t actually seem all that useful in most scenarios. A scientist stitching together two strands of DNA doesn’t actually care if they only get five colonies because, often, they only need to get ONE colony that works, and then they can grow up those cells in large beakers and extract a huge amount of the plasmid. A more useful metric might be to increase the total number of unique DNA strands that can be joined together in a single Gibson Assembly reaction, without reducing overall quality, instead.

    Still, I liked this blog post as a whole. I’m glad people are optimizing the “small” things, and I don’t blame OpenAI for not trying to solve cancer, in its overwhelming magnitude of manifestations, on their first attempt! Gibson Assembly is a much better starting point.

  • Enzymes from Random Molecules

    A new paper in Nature shows that enzymes can be made by mixing just four molecules together, none of which are amino acids. The four molecules randomly link together to form long polymer chains, some of which catalyze chemical reactions.

    Though this sounds impressive, the paper itself is quite strange. For one, it is extremely short (only about 2,700 words) and has no discussion section. The text is also absurdly dense; likely designed to be read by materials or physics people, rather than biologists. And lastly, I think the paper is most interesting for the things it leaves unwritten — the ideas left out rather than put in. Understanding why this paper matters, then, is mostly an exercise in speculation.

    For context, scientists have been trying to design new enzymes for decades. But this “design” has traditionally been done by searching for amino acid sequences which then fold into a 3D shape with some desired function. Computational biologists tend to fixate on the sequence; they tend to consider proteins as individuals rather than as populations of molecules.

    Enzyme design is also a really hard problem. An enzyme’s interior holds amino acids in a precise way, such that the amino acid(s) in the active site latch onto substrates and convert them into new molecules. This “active site” is surrounded by other amino acids that create a microenvironment suited to the reaction. If the substrate is negatively-charged, for example, the microenvironment works to exclude positively-charged molecules.

    Despite their complexity, biologists have designed viable enzymes computationally. Last year, David Baker’s group at the University of Washington designed a serine hydrolase that breaks down ester groups, or chemicals made by joining together an acid and an alcohol. This AI-designed enzyme has an active site made from three amino acids (a “catalytic triad”) that work together to catalyze the reactions. But it was quite slow, completing just one reaction per second, compared to the thousands of reactions per second that is typical of natural serine hydrolases. Enzyme design thus remains a mostly unsolved problem.

    This new Nature paper, though, took a completely different approach. The key breakthrough, in my eyes, is its focus on populations of polymers rather than in trying to create one perfect polymer. The authors created enzymes using a statistical or probabilistic approach, rather than a deterministic one.

    The researchers focused on metalloenzymes, which are arguably simpler than serine hydrolases because they only have a single amino acid in their active site, rather than a ‘triad’. Metalloenzymes hold metal ions (often zinc, iron, or copper) in that active site; hence the name. The researchers made two types of metalloenzymes: terpene cyclases, which take a string of carbons as substrate and “loop” them into a circle, and peroxidases, which use the iron in heme to oxidize substrates, like hydrogen peroxide. I’ll just focus on the terpene cyclase, as the approach taken was largely identical in both cases.

    In nature, terpene cyclases take a straight chain of ten carbon atoms — a molecule called citronellal — and fold them into a ring. If all goes well, the enzyme makes isopulegol, which is a carbon ring with one alcohol group. But if water gets into the active site, this reaction is disrupted and the enzyme instead makes menthoglycol, which is the same carbon loop but with two alcohol groups.

    Natural terpene cyclases have aspartate in their active site. The aspartate donates a proton to citronellal, thus making one of its carbon atoms positively charged. This triggers cyclization into a ring, as the “activated” carbon joins the carbon at the other side of the chain. The aspartate is surrounded by a hydrophobic shell, which keeps water out so that isopulegol gets made selectively instead of menthoglycol.

    Seeking to create random polymers which could mimic a terpene cyclase, the researchers first analyzed 1,300 metalloproteins, looking for commonalities between them. They found two things: First, metalloproteins tend to have one “key” amino acid in their interior — often histidine or aspartate — which latches onto the metal ion, locking it in place, so that it can perform the chemical reaction. Second, metalloenzymes tend to surround their active sites with hydrophobic amino acids, which exclude water molecules. To make a metalloenzyme, then, one basically just needs to situate a single amino acid, or electron donor, inside a hydrophobic shell.

    Next, the authors scoured chemical databases for molecules with these same properties, meaning they are hydrophobic or similar in shape and charge to histidine or aspartate. They ultimately settled on four molecules:

    • Methyl methacrylate (MMA), a hydrophobic molecule.
    • 2-ethylhexyl methacrylate (EHMA), an even more hydrophobic molecule.
    • Oligo(ethylene glycol) methyl ether methacrylate (OEGMA), a hydrophilic molecule.
    • 3-sulfopropyl methacrylate potassium salt (SPMA), which mimics aspartate as an electron donor; the active site surrogate.

    (Note: You need some hydrophilic molecules, even when trying to build a hydrophobic active site, because the polymers won’t dissolve in water without them. Instead, they will aggregate or precipitate out of the solution. Hence the inclusion of OEGMA.)

    Then, the researchers mixed these four molecules together, and each molecule randomly linked with others to create long and unique polymers. The hope was that some of these “pseudo-random” polymers would position a SPMA amid hydrophobic molecules, thus creating a terpene cyclase mimic. Initially, things did not go to plan.

    In their first trial, the researchers mixed 50% MMA, 20% EHMA, 25% OEGMA, and 5% SPMA and added the resulting polymers to citronellal. After 24 hours, the polymers cyclized citronellal, but poorly. About half of the citronellal molecules were converted, and only 55 percent of products were isopulegol. In other words, the polymers could slowly catalyze reactions, but not selectively.

    So the authors iterated. To optimize their reaction, they used a Monte Carlo algorithm to generate 100,000 polymer sequences based on each molecule’s ratio and reactivity. By tinkering with the molecular ratios and re-running these simulations, they figured out they could improve the odds that SPMA would be surrounded by hydrophobic residues — and thus act like a terpene cyclase — if they increased SPMA’s concentration (to 15%) while decreasing OEGMA (to 5%).

    This yielded much better results. In a second round, the polymers converted 91 percent of citronellal after 24 hours, with a selectivity for isopulegol of 76 percent.

    So why does any of this matter? Well, the paper doesn’t really say, outside of some vague or indirect commentary. So what follows is mostly speculation…

    I think one reason this paper is important is because it does away with the outdated notion that enzymes must be tuned at the sequence-level. The study shows, rather, that enzymes can be made spontaneously using pseudo-random populations of molecules, much like the earliest cells on Earth probably did. Early lifeforms didn’t need to evolve the perfect enzyme; they just needed to find concoctions of molecules that were “good enough” for a particular function.

    The study also suggests that the 20 amino acids used by cells are not particularly special, and their functions can be replaced with other molecules carrying the same properties — like “charged” or “hydrophobic” or “flexible” and so on.

    When I first discussed this paper with a friend, a protein biochemist, they urged me not to write about it. They said that metalloenzymes are not particularly difficult to make, and so this paper’s outcomes aren’t all that surprising. They pointed to another studydemonstrating that it’s possible to make functional metalloenzymes simply by mixing purified phenylalanine with zinc ions.

    My retort to their criticism, though, is that the authors have already used this same “random polymer” approach to make other types of proteins. In 2020, for example, they made protein channels that were exquisitely sensitive to protons and, during our conversation, hinted that they have also made other classes of enzymes, including hydrolases.

    But still, this paper leaves so much left unsaid. I suspect many protein biochemists reading this blog still won’t find the work impressive or useful or surprising or whatever. It takes a long time to overturn dogma, after all, and it’ll be an uphill battle to change peoples’ perceptions of enzymes and how one can make them.

    The paper itself also took seven years of work, according to a corresponding author, and involved many back-and-forth debates with a “hostile” reviewer. The manuscript was cut nearly in half (from nearly 5,000 words to 2,700), losing much of its philosophical framing. “This was the hardest paper I’ve ever published,” the authors told me. And after spending a week wrestling with whether to write about it, I understand why.

  • Brick Technology Addictions

    There’s a YouTube channel I like called Brick Technology. The videos are simple: a machine made from LEGOs must conquer an obstacle, like a wall, moat, or soapy ramp. When the machine inevitably fails, an engineer (always offscreen) modifies it to overcome the obstacle. Then, the obstacle gets bigger, and the engineering must continue. This happens again and again. The walls get bigger, the moats get wider, and the machines keep evolving in iterative loops. It is addicting to watch.

    In one video, a LEGO car is engineered to scale a wall. At first, the wall is only two bricks high; the car climbs it easily. Then the wall goes up to 25 bricks, and the car is affixed to a ladder. Using an electronic motor, it throws this ladder against the wall and pulls itself up the rungs. Then the wall goes to 50 bricks (or 55, or 60; too many to count on-screen) and even the ladder now falls short. The engineer mounts a propeller on top of the car, but the battery can’t generate enough lift. So they add a grappling hook instead. This LEGO car, now mounted with cannon, shoots the hook up and over the wall. The hook sticks, and the car pulls itself up at a 90 degree angle. All of this happens in the span of nine minutes.

    Why LEGOs? My guess is because they are easy to move around, to assemble and disassemble. Bricks can be snapped, changed, replaced. It’s much faster to iterate with plastic bricks than metal pieces, and faster still to iterate with metal than, say, a living organism. (Everything I write is ultimately about biology, isn’t it? Surely you didn’t think this was actually about LEGOs…)

    This YouTube channel got me thinking about fast iterations and the sense of accomplishment we feel upon finding solutions to a difficult problem. The feeling is addicting, not only for scientists but for little kids, too. At the start of any given Brick Technology video, it’s hard to imagine how a little machine, so humble at first, could possibly be modified to scale a towering wall. And yet, by video’s end, it has. Each iteration, every change, every solution is recorded and made visible. The fast cycles of trial and error are mesmerizing. The car scales the wall.

    Kids seem to like LEGOs and computers because these things move quickly. The kids can iterate upon ideas, try stuff that doesn’t work, and learn from failures in a few minutes. But in biology, everything takes too long. Cloning a single gene takes about a week, and we rarely disclose our failures and iterations; papers merely report the final solutions. Everything in biology is portrayed as a clear “story,” with all five figures neatly planned out months in advance (and no failed experiments, of course.)

    Just because biology is slow and opaque, though, doesn’t mean we can’t emulate faster fields; if not in practice, then at least in educational terms. I’d like to see a YouTube channel that recreates Brick Technology for biology. After all, this YouTube channel is the encapsulation of what biology research actually is! All a bioengineer is really doing is asking a question and seeking a solution. We start with a goal, like “engineer a gene circuit to make cells flash in oscillating patterns of green and red,” and begin building towards it. Most experiments end in failure: the cells don’t flash at all, or they only flash green, or everything is dead. But I think it’d be wonderful to watch a video where a scientist sets such a challenge, tries and fails to solve it, records everything, and explains how each iteration is made. Imagine how good it would feel, as a viewer, to see those engineered cells at the end, beautifully flashing. It would be addicting as hell to watch, too; a vision of the future made tangible, approachable, common.

  • Moving to the Bay

    I moved to the Bay Area to join Astera Institute as an independent fellow.

    This job has no requirements. I will spend all my time writing, reading, and researching my own interests. I’ll host dinners, start a podcast, and finish my first book. I’ll talk to people, travel back to China, and build out my biotechnology manifesto. 

    I’ll obsess over big problems, like how to cure all infectious diseases; how to make life multiplanetary; how to safeguard life here on Earth; how to lower the cost and increase the speed of experiments; how to understand a single organism such that we can accurately simulate it; how to tell better stories. 

    I’ll write up my findings and share them freely. And when I spot problems that feel particularly important yet tractable, I’ll try to raise money and give it to scientists trying to solve them.

    Astera seems like a good fit for this work because they are broadly funding many different things in bio- and neurotech. I’ve enjoyed getting to know people like Michael Nielsen, Eli Dourado, and Prachee Avasthi over the years. And, being situated in the Bay Area, it feels like a good “hub” from which to instigate and build teams around big, important problems.

    Asimov Press will continue. We have articles finished and scheduled through May. We are also 80% done with our third book, Making the Modern Laboratory, which tells the story of why research laboratories look the way they do and how they could be better.

    Running Asimov Press is the most important thing I’ve done, and it wasn’t an easy decision to let it go, even if only for awhile.

    I still remember that first COVID summer, when I moved from Los Angeles to New York to become a writer. I enrolled in NYU’s science journalism program and rented a small apartment around the corner from Katz’s Deli, at 188 Orchard St, for $2650 per month. 

    As a young journalist, I had idealistic notions about writing and reporting; about making a career as a “pure” writer telling unapologetic stories. Holding these lofty dreams in my mind wasn’t easy, though.

    First-year journalism students were told to lower their expectations, both in terms of impact and career. Professors told me not to expect my published articles to have tangible impacts. Many pieces have been written about pollution in the Hudson, they said, but few have led politicians to actually act.

    The job market was also bad and scary. Staff positions were elusive even then, and are moreso today. Healthcare and benefits? Absolute luxury! Dreams of becoming an editor? Maybe in five years! Starting your own media company? Fuggedaboutit.

    That Asimov funded me to start a small media company just two years out of journalism school, then, feels insane. Here was a for-profit company that gave me the reins to their name — and reputation — without ever dictating the stories I publish. I’ll always feel grateful to Alec Nielsen, the company’s CEO, for believing in me more than anybody had before. 

    As I quickly discovered, though, running even a small media group — with just two full-time employees — takes a lot of time. We constantly have about 20 articles in the publishing pipeline. Every piece needed to be written, edited (usually 4-5 rounds of feedback), fact-checked, and copyedited. We also make artwork for each piece, record voiceovers, mix the audio, scour the web for images (or make our own) and publish. Writers must be paid; freelancers must be commissioned; social media posts must be written. 

    It is not easy to manage all these little pieces while trying to write my own things; especially a big book project. It is also difficult to hew to a grand ambition or “path” when you’re balancing opinions from lots of readers and teammates. I’m proud of Asimov Press, but do think we steered too much toward history and not enough toward pieces that shape discourse. Moving to Astera is a chance for me to rectify this.

    The Bay area will also be a nice change. I’ve spent the last year-and-a-half living in Kansas, on the border line where suburbs meet farmland. I’ve tried all the restaurants in a five-mile radius at least three times. I’ve shipped thousands of books out of my garage. I’ve enjoyed hundreds of quiet afternoons listening to the birds or flapping wings of bats, which flock through the air at night. And yet, I’ve been socially starved, aching for face-to-face meetings and candlelit dinners and monthly Poker tournaments. If you’re reading this, let’s meet up:  My email is nsmccarty3@gmail.com.

    I’m not entirely sure what the future holds, but I have recently found inspiration in Stewart Brand. At the bottom of Brand’s website he writes: “What do I usually do? I find things and I found things.”

    “Things I find include tools, ideas, books, and people, which I blend and purvey.  Things I’ve founded and co-founded include the Trips Festival (1966), Whole Earth Catalog (1968), Hackers Conference (1984), The WELL (1984), Global Business Network (1988), and The Long Now Foundation (1996).”

    It was Brand who brought The Grateful Dead to San Francisco and mentored Kevin Kelly, the former executive editor of WIRED magazine. Brand has also written many books; about The MIT Media Lab, about Maintenance, and even about architecture and buildings (something he apparently knew little about when he began, but mastered over time). 

    In other words, Brand seems to be comfortable being uncomfortable; in moving around and trying new things. His ethos, of finding and founding, seems to describe a life well lived.

  • Why Vibrio Never Caught On

    There is a microbe, discovered in 1958 on Sapelo Island in Georgia, that divides every 10 minutes. This is the fastest division time ever observed in any bacterial species, and it belongs to Vibrio natriegens. (It divides so quickly, in part, because its genome is split across two chromosomes, each of which gets copied simultaneously.)

    Some idealistic researchers in George Church’s lab at Harvard spent several years making genetic tools for Vibrio. These scientists thought that if they made it simple to work with this microbe — engineer it, grow it, that sort of thing — then everyone would surely use it instead of E. coli. After all, Vibrio divides twice as fast as E. coli (10 instead of 20 minutes), meaning that you can get visible colonies on a plate in four hours instead of eight. This, in turn, means that common experiments become much quicker. Church’s group released these genetic tools in 2019, waited for others to use them, and quickly became disappointed (I’d imagine).

    Few researchers read the article and switched over to Vibrio. And if you go to most academic laboratories today, researchers will likely admit they know about Vibrio but then gleefully point you to their massive stockpile of E. coli strains instead. If you ask why they don’t use Vibrio, which would speed up their experiments, they will usually answer with some variation of, “Well, there’s little difference between 10 and 20 minutes. And besides, I just grow my E. coli cells overnight and continue my work the next day. Faster division times don’t really make a difference for me.”

    Humans can reasonably make this argument! But it will, I think, become an increasingly moot point in a world where AI tools propose hypotheses. As intelligence becomes cheaper, the bottleneck for discoveries will increasingly shift toward wet-lab experiments. And when wet-lab experiments become a bottleneck, more money will flow toward automating experiments. Human scientists can certainly design experiments around their own schedules, but robots don’t sleep or eat lunch. When experiments run fully automated, then, a time savings of ten minutes per cell division becomes a big deal indeed.

    DNA cloning is a method scientists use in nearly every biotechnology experiment. (And, unlike most other methods, researchers can actually automate it fairly easily.) Cloning GFP into a plasmid, for example, has a handful of basic steps. First, scientists use PCR to amplify the GFP gene (~3 hours). Then, they digest and ligate GFP into the plasmid (2.5 hours). Third, they transform this DNA into bacterial cells (1 hour). Next, they grow the transformed cells on agar plates and wait to see whether colonies grow. (It takes at least 24 cell divisions to see a colony and each division takes 20 minutes with E. coli, so this step alone takes 8 hours.) Finally, researchers pick these colonies and send them for sequencing, which can take a day or more.

    Assuming all these steps happen sequentially, without any breaks, then DNA cloning to sequencing takes a minimum of about 15 hours. Note that more than half of this time goes to waiting for cells to grow! Cell division times literally limit the rate of DNA cloning, an experiment that scientists collectively run millions of times each year.

    If scientists swapped E. coli for Vibrio, this DNA cloning example would drop by four hours, to about 11 hours in total. (Visible colonies of Vibrio appear on an agar plate in about four hours.) This represents a time savings of 27 percent, which may seem small. But again, imagine repeating this for thousands of plasmids across years of work. If a theoretical scientist had one robot cloning one gene at a time full time, autonomously, then switching to Vibrio would basically give three months of “free” productivity in the first year.

    Even if you don’t think that AI will generate meaningful hypotheses, progress in biotechnology remains broadly limited by wet-lab capabilities — specifically, the speed and cost of experiments. Whereas many PhD students aim to change the world by inventing a useful new method or medicine, they ought to consider instead how even a marginal improvement to a ubiquitous method can have far more consequential effects, mainly by speeding up discoveries at scale.

    (Thanks to Henry Lee for inspiring this article.)

    Enter your email

    Get new essays sent to your inbox.