2,500 Compounds and a 51 Percent Hit Rate: How a Small Graph Net Found New Antibiotics

Nathan Labenz opens by noting that even the hyper-informed crowd at The Curve had barely heard this story. Jim Collins, Termeer Professor of Medical Engineering at MIT, has spent five-plus years refining an AI pipeline that does not look like AGI: small convolutional graph neural networks, a few thousand labeled structures, modest compute, and a multi-step filter that scores novelty, stability, synthesizability, and human toxicity. The first public win was a 2020 *Cell* paper. From a 110-million-compound in silico library the model helped pull out a new antibiotic, named halicin after HAL 9000. In the film, HAL kills humans; halicin kills bacteria.

The training set was 2,500 compounds. Colleagues in AI, Collins says, dismissed the plan and told the team not to start: far too little data to do anything meaningful. Labenz admits that, coming from trillion-token language models, he would have guessed the library was at least two orders of magnitude too small. When candidates were actually synthesized and tested, the true-positive rate was about 51 or 52 percent. That would be modest for telling a cat from a dog. For a new antibiotic, where a large random screen usually hits well under 1 percent, it was a shock — a pleasant one.

The same workflow has since been pointed at several bacterial targets, with papers in *Nature* and *Nature Chemical Biology*. Some compounds work against resistant strains, some by previously unknown mechanisms, and some without wrecking gut commensals. They are moving toward clinical trials. Barring an Operation Warp Speed for antibiotics, Labenz notes, patients will wait years. What struck him was the bill: a pipeline of 15 to 20 new antibiotics, including trials, on the order of $20 billion — cheap next to the datacenter deals that dominate AI headlines, and a win that does not require superintelligence.

Collins starts with a primer. Antibiotics are typically oral small molecules, aspirin-class drugs that disrupt a bacterial protein tied to cell division, protein production, or DNA replication. That disruption sets off downstream stress, energetic demand, and toxic metabolic byproducts that damage DNA, RNA, protein, and membranes — a vicious cycle. Effectiveness is not a cartoon missile. Screens almost always measure growth inhibition, not killing. Bactericidal drugs aim to kill at a human-safe dose; bacteriostatic drugs stop growth and lean on the immune system. Concentrations at an infection site are uneven, so many bugs see sub-lethal levels. The ideal molecule hits the pathogen and spares the host — and, if possible, spares healthy gut and skin bacteria too.

Resistance is not mysterious. Bacteria exist to survive and replicate. Mutations appear with some frequency at every division, and antibiotic stress selects changes in the target or in downstream machinery that make the cell less susceptible. Mutants live; the rest die. Overuse in human medicine and in agriculture — prophylaxis and growth promotion — has moved so-called superbugs out of hospitals onto playing fields, childcare centers, schools, and shopping centers. Collins tells students the worst place to be when you are sick is a hospital; now the bugs are in the community. About one million to 2.5 million people die of bacterial infections globally each year. A 2016 UK commission estimated 10 million deaths a year by 2050 if the crisis is not addressed, outpacing cancer.

Pharma largely walked away for economic reasons. Alexander Fleming found penicillin in September 1928; Oxford turned it into a manufactured drug in the early 1940s. The discovery heyday was the 1940s through the 1960s, before the microbiology, biotech, genomics, and AI revolutions. Since then, a discovery winter. It costs as much to develop an antibiotic as a cancer or blood-pressure drug, but the antibiotic sells for a few dollars and is taken for a day or a few days; the others sell for thousands and may be taken for life. Worse, companies that reached approval watched doctors shelf the product as a last line of defense, so revenue never came and several firms went bankrupt after the milestone. Collins wants Operation Warp Speed-style public-private partnerships and philanthropy. There are walks and ribbons for cancer, he notes, and none for resistant infections, even though every listener has lost someone. His pitch to the ultra-wealthy: about $20 billion could address antibacterial resistance for decades — history without a return.

The Collins lab had used machine learning on antibiotics for some twenty years, first to reverse-engineer biomolecular networks and boost existing drugs. In March 2018 MIT launched a campus-wide AI initiative, admitting it had slept through the third wave after the Minsky–Papert and Winston eras. Collins's office sat next to Regina Barzilay, who works on biomedical AI; they shared an interest in drug discovery and brought in Tommi Jaakkola. There was no dedicated money. Collins's phrase is that they bootlegged the project. The first training library was 2,500 compounds: 1,700 FDA-approved drugs, including the known universe of antibiotics, plus 800 natural products, all tested on *E. coli* for growth inhibition. *E. coli* is both a model organism and a pathogen behind urinary-tract infections and food poisoning.

They could have kept continuous inhibition — 90 percent versus 30 percent. They discretized at 80 percent: one if you hit the cutoff, zero if not. Labenz asks whether they tried to predict the scalar. Collins suspects they did, and the results were not good. A correlative model did not have enough data to place a brand-new structure on a 0–1 line; that, he thinks, would take hundreds of thousands or millions of compounds. Binarizing asked a sharper question: which structural features make a *good* antibiotic. The architecture was a graph neural net with convolutions, Barzilay and Jaakkola's Chemprop platform, built earlier by people including Yang, Kyle Swanson, and Wengong Jin. Atoms are nodes, bonds are edges; the net learns substructures associated with the binary label. Late 2018, transformers were just appearing and not yet that popular. The graph representation matched high-school chemistry diagrams.

They train an ensemble of 20 identical models with different random initializations and average the crowd. That model was then pointed at the Broad Institute's 6,100-compound drug-repurposing library, with three simultaneous demands: predicted antibacterial, predicted not toxic to human cells, and structurally unlike existing antibiotics. Only one molecule hit all three. That was halicin. Small-molecule language models such as ChemBERT, plus NVIDIA and IBM variants, have since been tried; they learn slightly different schemes and have not outperformed the graph net. A multimodal hybrid gave a small bump. More recently the lab implemented MiniMol, another graph net pretrained with quantum-mechanical calculations; upcoming work, Collins says, shows Chemprop significantly outperforms it. They are beginning to think about 3D representations rather than 2D.

After halicin, scale went up. The first computational screen was about 110 million compounds in three days on the platform they had — a library no wet lab would ever assay empirically. They now work with Enamine, a synthesis company in Kyiv, still operating with a war outside. Enamine's make-on-demand "real space" is on the order of 65 to 70 billion compounds, built from building blocks and recipes they are confident they can run; a larger "unreal" space is around 220 billion. Academic labs keep tens of thousands of vials; big centers hundreds of thousands to a million; pharma low millions. The Broad's Center for the Discovery of Therapeutics has a robotic store of roughly 800,000 to a million on-site. Enamine can ship perhaps 4 to 4.5 million immediately. Theoretical chemical space is often quoted at 10^60. Collins has never seen the calculation; a trillion feels pragmatic. BioSolveIT, a colleague mentioned, claims trillions from its own blocks.

The model does not emit a single score. After antibacterial prediction, filters run mostly in sequence: difference from known antibiotics, toxicity to human cells, chemical stability, synthesizability. Multitask models have shown some advantage where features overlap across pathogens, but the practical job is still to shrink computer hits into things you will pay to make and test. A 2024 *Nature* paper led by postdoc Felix Wong used Monte Carlo tree search to find substructure "rationales" over-represented among top-scoring molecules — explainable clusters that may mark a new structural class and a shared mechanism. Collins calls it powerful: it showed what the models had actually learned in antibiotic space. Hits include abaucin against *Acinetobacter baumannii* and a compound against *Neisseria gonorrhoeae*, with others for additional pathogens, some in preclinical development.

Connor Coley's synthesizability models matter because generative design keeps proposing molecules you cannot make in a reasonable number of steps for reasonable money. Given that pain, Collins now prefers screening Enamine-style libraries where synthesis is guaranteed, even if some recipes cost extra, over an exotic generated structure of unknown makeability and unknown odds. The remaining gap is downstream of the dish. The models are strong at killing bacteria in vitro. They are not yet good at predicting a mouse. In antibiotics, unlike oncology's "cured 900 times in a mouse" joke, a molecule that works by systemic delivery in a mouse is estimated to have about a 90 percent chance of working in humans. That dish-to-mouse step is the next data problem. With Phare Bio they are gathering ADMET and PK/PD labels for what they will probably call DrugProp AI.

The first surprise was spectrum. Narrow-spectrum antibiotics kill the target pathogen and spare gut commensals. The models were trained only on a single pathogen; they were never counter-trained to avoid other bacteria. Yet abaucin and the gonorrhea compound both came back narrow-spectrum, leaving most "good guys" in the gut alone. Collins calls it more or less luck in the design sense. His mechanistic guess is that phenotypic screens plus novelty are landing on lipoproteins and lipoprotein transport, targets that differ sharply across pathogens. The team is now asking whether AI can start from those proteins as targets rather than from whole-cell phenotypes.

The second surprise was resistance. They compared halicin to ciprofloxacin, a widely used quinolone, on *E. coli* for 30 days. Cipro showed significant resistance within a few days and many-hundred-fold levels by day 30. Halicin showed none at a few days and none at 30. Collins leaves the usual caveat: run long enough and resistance will appear. The line he wants remembered: if an antibiotic researcher tells you they found a drug with no resistance, they are lying to themselves or lying to you. He suspects halicin is slow to evade because it hits multiple membrane-level protein targets. The bug would need mutations at three or four independent sites. That is an explicit AI design goal going forward: multi-target molecules that lengthen the resistance runway.

The third limit is honesty about what the models cannot do. Killing in a dish is solved at an acceptable rate; killing in a mouse is not. Labenz's societal read is that the workflow already works well enough to clone: pick a pathogen, run the panel, train, screen chemical space, synthesize. Collins agrees the success rate is decent and that early discovery is cheaper and higher-hit because of AI. Multi-objective optimization for truly great starting leads — not just any hit — still needs those missing drug-like properties. From society's standpoint, he says, other groups have shown the tool is valuable and should be used more. Improving the models would cut the cost from mouse to trial. It would not be a reason to wait before scaling the labs.

Labenz asks whether human judgment still has a seat. Absolutely, Collins says. Medicinal chemists remain skeptical of AI; decades of tweaking molecules can still outpace a model in some cases. No chemist he knows can review 110 million structures. They ran a head-to-head: a gifted medicinal-chemistry postdoc versus Coley's synthesizability model, both ranking 30,000 compounds against what Enamine would actually say. The human won. Andreas Luthin, now a young professor at Karolinska Institutet in Stockholm, left-swiped and right-swiped the entire set. Collins, married 35 years, jokes that he is unclear which swipe is a like. Asked what guided him, Luthin said he looked for liabilities. Bad, bad, bad — and if he could not find a liability, good. That is not how Collins would have set up a model.

The obvious next step is panels of chemists in the loop, reinforcement learning from human feedback, to capture intuition they may not quantify. Collins reaches for AlphaGo: DeepMind had trained on synthetic and human games, then jumped when they brought in a human expert — perhaps the thousandth-ranked player — for strategy. Drug discovery needs the analogue. Hostility and dismissiveness in medicinal chemistry are real; he still wants those people in the code. Pre-pandemic he met someone tied to Baidu who said the company had hired 15,000 people to label data, including 2,000 medical students on pathology images. He does not need that scale. He needs chemists willing to let him peer into how they act on what they have learned.

On compute versus synthesis, compute is almost a background fixed cost. Larger training libraries are expensive; making molecules is the commitment — money, time, and probability. Animal models are not cheap. Inside a small academic lab the trade-off is what you will advance to analog generation. Because generative AI still struggles with synthesis, Collins would rather pay Enamine than bet on a molecule he cannot make. After filters they still hold many hundreds before trying to synthesize, then a few that work at least in a mouse. AI has done well at early discovery across drug spaces, he says, and has barely been used further downstream, mostly for lack of data. The companies that pay for those labels will have the advantage.

The cost sheet is small by AI-industry standards. A later 37,000-compound library cost about $150,000 to assemble, bordering on $5 each; vendors charge $10 to $20, or around $100 for special structures. Liquid-handling robots run about $20,000 to screen 40,000 compounds. That 40,000-ish library has been applied to seven bacterial pathogens and three human cell lines — ten experiments. From early hit through preclinical to IND-ready, Collins estimates under $2 million per compound. Phare Bio, a nonprofit the Antibiotics-AI project helped launch, has a $27 million ARPA-H grant to take 15 antibiotics through preclinical development, a generative-AI-driven pipeline. Clinical phases dominate the rest of the bill. Fifteen to 20 drugs, trials included, is the $20 billion pitch — under a tenth of a percent of U.S. GDP, and the same number Labenz had just seen attached to an xAI round. On most existential-risk lists, antibacterial resistance is the cheapest item. Poverty, hunger, and climate are multi-trillion problems.

Most of the work is single-use: help humans against pathogens. One dual-use exception woke Collins up after the fact. Toxicity models are trained to find molecules that are *not* toxic to human cells. The same weights, with the threshold flipped, find highly toxic ones, especially mechanisms without countermeasures. After the 2024 *Nature* paper, friends in the federal government asked whether he had considered it. He had not, and has since. Labenz asks about a version that only emits "non-toxic" and hides the score. Collins says no: anyone can pull the model apart and shift the threshold. Interpretability would not save you either; the internal representation is still there if someone is determined.

The same phenotypic-plus-graph approach is already leaving bacteria. Felix Wong, a postdoc also leading integrative biosciences in the Bay Area, used the platform to find senolytics that clear so-called zombie cells tied to neurologic conditions, scarring, and skin aging — it "worked beautifully." Collaborators are trying antifungals, antivirals, and anti-parasitics. Collins sees potential in cancer, neurology, and metabolic disease, with the caveat that a human cancer cell is much closer to a healthy human cell than a bacterium is. Next-generation wishes include finer AlphaFold structures for target screens, better small-molecule binding affinities, and tools that treat targets as context-dependent networks rather than single proteins. Automated "AI scientist" labs, he thinks, are premature: they run recipes we already know. *E. coli* has 4,000 genes; 1,500 still have unknown function. Ask an AI to annotate them and validate, and we cannot come close. He still wants AI as a thought partner, human in the loop.

Labenz's practical lesson for the AI crowd is that nobody waited for a bigger dataset, a bigger model, or AGI. They defined a coarse, clean problem, ran it on a narrow task, then expanded targets, libraries, and filters. When could this first have worked? Collins thinks the techniques were sitting there for a few years. Graph neural nets were the differentiating trick: the ability to break chemical structure into learned substructures and associate them with antibacterial or non-toxic labels. Deep nets had been popular for about a decade on images, from Bengio, Hinton, and LeCun. What the field needs now is more data, more talent, bigger chemical space, more biological and chemical context for mechanism, and generative design that aims at multiple targets from the start. He expects that in the coming few years.

If you work in biology or drugs, the reusable object is the pipeline: activity as the first gate, then novelty, toxicity, stability, synthesizability, each layer cutting the space. Collins's current preference is guaranteed-synthesis virtual libraries rather than unconstrained generation. "I'm more comfortable there than coming up with an exotic new molecule that I'm not sure I'm gonna synthesize." If you have money, his line is that an uber-wealthy individual could single-handedly address this challenge for the lifetime of everybody on the planet, at about $20 billion, without needing to make a dollar back. Labenz would also like a US–China race — or a team-up — with gold medals for new antibiotics instead of weapons.

The closing call is to young people as much as to philanthropists. Take the problem from AI, microbiology, or drug discovery; it is an exciting practical problem that can make a difference. It is a great time to be young in science even if the world seems on fire, because the technologies are arriving and the field is short of people. As for *E. coli*, 1,500 of 4,000 genes still have no known function, and AI has not solved that either. Labenz hopes a few listeners move a little in that direction. Collins thanks him for the conversation.