The Antibiotic Winter - How Artificial Intelligence Is Ending Seventy Years of Silence
In earlier blog posts, we explored how AI models are being used for protein folding and DNA language modeling. Machine learning is driving rapid progress in another important area of biology: antibiotic discovery. Rushil Mohan, a smart high-school student, worked with me and wrote a fantastic review post to introduce you to this fascinating area of research.
The Antibiotic Winter: How Artificial Intelligence Is Ending Seventy Years of Silence
- Rushil Mohan
Every seventy seconds, a person dies from an infection that antibiotics can no longer treat. In 2019, drug-resistant bacteria were responsible for approximately 1.27 million deaths, more than HIV/AIDS, more than malaria. By 2050, if current trends hold, that number is projected to surpass ten million annually, making antimicrobial resistance the leading cause of death on Earth. We are not approaching a crisis. We are already witnessing one.
Penicillin’s discovery in 1928 and its mass production in the 1940s effectively made previously fatal infections survivable. Surgical procedures became routine. Childbirth became dramatically safer. The pharmaceutical industry, recognizing the opportunity, poured resources into antibiotic development for several decades. Then, around the 1980s, things froze up and shifted to a crisis. The “easy” antibiotics, the molecules that were structurally similar to what we already had, and therefore straightforward to develop and test, had mostly been found. New drug classes became rarer. Investment shifted toward more profitable chronic-disease treatments. And the bacteria kept evolving.
This period, as some have called it, became the Antibiotic Winter. A season of stagnation, during which resistance accumulated while discovery stalled. What ended it, or what is in the process of ending it, is not a new class of chemist or a new understanding of bacterial biology, but a technology that didn’t exist when the crisis began: artificial intelligence. This is the story of how that happened, in roughly the order it happened, and of what it means.
Early Discoveries
The field of AI-driven antibiotic discovery did not begin with a single success. It began with a series of incremental, largely unglamorous advances that laid the conceptual and technical foundation for everything that followed. Understanding what those advances were and why they mattered is necessary for grasping the significance of the breakthrough that came after them.
The earliest computational approaches operated on a simple premise: if we can characterize what makes a peptide lethal to bacteria, we can predict which sequences are worth synthesizing and testing. Antimicrobial peptides, or chains of amino acids that can disrupt or penetrate bacterial membranes, were a natural starting point. They are produced naturally by almost every living organism as a frontline immune defense, they are structurally diverse, and their small size makes them computationally tractable. The problem was accuracy at short lengths. Traditional prediction models were trained on peptides of 80 to 255 amino acids. When applied to the shorter sequences that are cheaper to synthesize and easier to optimize, their accuracy dropped from above 90% to below 75%. The short sequences were hiding in plain sight; the tools to find them had not yet been developed.
Deep-AmPEP30, developed to address exactly this gap, focused exclusively on peptides of 30 amino acids or fewer. Its key innovation was to cluster amino acids by their evolutionary and physicochemical similarities rather than treating all 20 as independent. This compressionsion that reduced noise and allowed the model to extract conserved features from sequences that had previously looked like random noise. It was also the first predictor designed to scan fully assembled genomic DNA rather than individual sequences, a capability that foreshadowed the genome-mining approaches that would come later.
Architecture, however, was only the first frontier. AMPlify, published in 2022, broke the ceiling on what prediction accuracy was possible by becoming the first AMP model to apply attention mechanisms. Its bidirectional architecture allowed the model to weigh every residue against every other, regardless of their distance in the sequence, capturing the long-range structural dependencies that govern how peptides fold and function. An ensemble of five sub-models trained on different data subsets provided robustness that any single model lacked. AMPlify achieved an accuracy of 93.71%, but more significantly, it was used to mine the genome of the North American Bullfrog and identify novel peptides with confirmed bioactivity against WHO priority pathogens. The transition from filtering to discovering had begun.
TransImbAMP, published the same year, addressed a different structural flaw: the severe imbalance between positive and negative training samples. In most peptide datasets, confirmed antimicrobials are vastly outnumbered by inactive sequences. A model trained on this data achieves apparent high accuracy simply by predicting “inactive” for everything technically correct, biologically useless. TransImbAMP’s solution was to borrow a framework from natural language processing. A Transformer backbone pretrained on massive protein databases learned the underlying “grammar” of amino acid sequences, while an asymmetric loss function forced the model to pay disproportionate attention to the rare positive samples. The approach extended across seven distinct target categories: gram-negative bacteria, fungi, cancer cells, and others, achieving 96.85% balanced accuracy in identification and 79.83% in functional annotation.
The subsequent generation of models turned from sequence to structure. sAMPpred-GAT, published in 2023, was the first predictor to incorporate three-dimensional spatial information. Using predicted protein folding data, it constructed graphs in which amino acids were nodes and their physical proximities were edges, then applied a Graph Attention Network to learn which spatial relationships determined antimicrobial activity. The architecture required careful calibration, with too few layers and the model missing complex structural patterns, too many, and individual node features blurred into indistinguishable noise, but at exactly three layers, it outperformed every prior approach on novel sequences.
Parallel to this work on contemporary biology, a more unexpected direction was emerging: extinction. The APEX platform, published in Nature Biomedical Engineering in 2024, operated on the premise that evolution had already solved many of the problems researchers were struggling with and that those solutions were archived in the proteins of species that no longer exist. By mining the proteomes of 208 extinct organisms, including the woolly mammoth, the giant sloth, and the ancient sea cow, APEX identified over 11,000 archaic sequences with no counterpart in any living species. These compounds proved effective in mouse models against modern drug-resistant pathogens, and their mechanism was distinct: rather than targeting the outer bacterial membrane, they depolarize the inner cytoplasmic membrane, the bacterial equivalent of draining a battery rather than smashing a wall.
Simultaneously, generative models were beginning to push past predictions entirely. G-SchNet revealed the depth of a persistent challenge: when researchers analyzed what generative models actually produced, they found that systematic bias molecules consistently had less saturated, more heteroatom-rich, and chemically distinct from training data in ways that invalidated downstream property predictions. SiMGen addressed this by abandoning global molecular representations and building structures atom by atom, guided by local atomic similarities to a small reference set. The approach was both data-efficient, matching trained models with as few as 256 reference molecules, and geometrically flexible, allowing researchers to specify desired molecular shapes before the AI filled them with valid chemistry.
The clinical dimension of the problem was not neglected. UTI SmartSet demonstrated that machine learning could predict individual patient resistance risk the moment a culture was collected before the 48-to-72-hour wait for laboratory confirmation sing only electronic health records, vital signs, and prior infection history. MALDI-TOF ML tackled a methodological problem lurking inside resistance prediction itself: the assumption that training and test data are independently sampled. By inferring latent phylogenetic structure directly from mass spectrometry data, it produced more accurate and reproducible resistance estimates for major pathogens.
What this decade of work produced was not a solution, but a set of tools: models that could predict, classify, mine, and, in limited ways, generate. They were sophisticated, genuinely useful, and still fundamentally constrained, each one working at the edges of what humans had already explored. The revolutionary shift that followed them came from a different question entirely.
The Revolutionary Shift When AI Stopped Helping and Started Leading
To understand how revolutionary the next step was, it helps to understand the assumption that had been concrete in drug discovery for the previous models. That assumption was that a molecule worth investigating should look like something we already know works. This is called chemical similarity, and it was not an unreasonable heuristic. It reduced the search space from billions of possible structures to a manageable set of known-adjacent candidates. It worked for decades. It also quietly left the vast majority of chemistry unexplored.
The MIT team that changed this, led by James Collins and Regina Barzilay, with biologist Jonathan Stokes as the experimental bridge between digital prediction and physical reality, did not set out to make a philosophical argument about chemical space. They set out to find new antibiotics. But the model they built to do so was trained not on what molecules looked like, but exclusively on what they did: did this compound kill bacteria, and how effectively? When they applied it to a library of over 6,000 compounds in 2020, it returned a molecule that had originally been designed to treat diabetes. It bore no structural resemblance to any known antibiotic. A human chemist would never have selected it as a candidate. They named it Halicin after HAL 9000, the artificial intelligence of 2001: A Space Odyssey, and it proved effective against a broad spectrum of drug-resistant pathogens in both cell cultures and animal models.
The 2023 follow-up paper, from Wong, Zheng, Collins, and collaborators at MIT and Harvard, extended this in a direction that proved equally significant. Previous AI models, however accurate, were black boxes: they returned predictions without explanation, which made it impossible for human chemists to learn from them or to extend their logic. The 2023 work integrated explainability directly into the framework, allowing researchers to identify the specific chemical substructures that the model was using to make its decisions. What they found inside the model’s reasoning was not a refinement of existing antibiotic logic but the blueprint for an entirely new structural class of compounds that killed MRSA by disrupting the proton motive force, the electrochemical gradient that functions as the bacterium’s internal battery. The AI had not just found a new drug. It had found a new mechanism and made that mechanism accessible to human researchers.
The subsequent generative work from the Jameel Clinic completed the arc. Rather than screening existing libraries or even mining evolutionary history, these models operate as architects: given a target vulnerability and a set of structural constraints, they design molecules from scratch that have never existed anywhere in nature. Crucially, they do so within the boundaries of the Enamine REAL Space, a library of approximately 48 million commercially available chemical building blocks combinable through validated reactions, ensuring that every designed compound comes with a viable synthesis route.
The following table captures how each of the four core papers shifted the paradigm:
| paper | what changed | historical approach | new approach |
| Halicin (2020) Stokes, Collins et al. MIT/Broad | The search logic | Similarity, Find molecules that look like known antibiotics | Function-first, Find molecules that kill bacteria, regardless of structure |
| MRSA class (2023) Wong, Zheng, Collins et al. MIT/Harvard | Model transparency | Black box, AI says yes/no; humans can’t see why | Explainable, AI highlights the exact substructures driving the decision |
| Generative design (2024) Jameel Clinic, Collins Lab MIT | The creative act | Mining, Search through what already exists | Generative Design, entirely new structures from scratch |
| Mechanism (all four), Proton motive force targeting | Bacterial target | Outer membrane, Target cell wall; bacteria evolve thicker walls | Inner battery Disrupt PMF; fundamental to life, hard to evolve around |
The People
Scientific revolutions are rarely the product of a single mind, and the antibiotic work emerging from MIT’s Abdul Latif Jameel Clinic is no exception. Behind the models and the metrics, three researchers shaped the paradigm in ways that no algorithm could have done alone.
James Collins arrived at the problem with decades of biological intuition. Often credited as one of the founding figures of synthetic biology, Collins had watched the antibiotic pipeline dry up across his career, not because bacteria were invincible, but because the industry kept searching in the same places and expecting different results. His conviction was that the “dark space” of chemistry, the vast landscape of molecules no one had thought to examine, was not barren. It simply hadn’t been explored with the right tools.
Regina Barzilay provided those tools. One of the world’s foremost AI researchers, her turn toward medicine was not purely academic. After surviving breast cancer, Barzilay grew impatient with the gap between what machine learning could theoretically do and what it was actually being used for. She brought to the collaboration something the field had lacked: a rigorous framework for building models that didn’t just classify molecules but learned the underlying logic of why certain structures interact with biological systems in the ways they do.
Jonathan Stokes occupied the space between them. A biologist who had deliberately immersed himself in machine learning, Stokes functioned as the project’s translator, someone who could interrogate a model’s predictions with the skepticism of an experimentalist and then carry those predictions into a laboratory to find out whether they survived contact with reality. It is one thing for an AI to rank a molecule highly on a screen. It is another to watch it kill a drug-resistant bacterium in a petri dish.
The human dimension of this work is perhaps most legible in a single naming decision. When the model surfaced its first major hit, a failed diabetes drug that the AI had identified as a potent bacterial killer, the team did not assign it a catalog number. They named it Halicin, after HAL 9000, the fictional artificial intelligence at the center of 2001: A Space Odyssey. The choice was not frivolous. It was a deliberate signal: this discovery belonged to a new era of medicine, one where the intelligence doing the searching was not entirely human.
Closing the Loop
Every generative AI system in the history of drug discovery has faced the same quiet embarrassment. The molecules it produces look extraordinary on a computer screen, optimized, novel, theoretically lethal to bacteria. Then a chemist examines them and explains, with varying degrees of patience, that several of the proposed structures violate basic rules of chemical stability, that the required precursors don’t exist commercially, or that the synthesis would require conditions no laboratory can sustain. The molecule is discarded. The AI moves on, undeterred, to produce more impossibilities.
This is the synthesizability gap, and for years, it was the most honest argument against taking generative drug discovery seriously. The 2024 and 2025 work from the Collins lab closed it through a constraint that is elegant in its simplicity: the AI is not permitted to dream freely. Instead, it works exclusively within the Enamine REAL Space, a library of approximately 48 million commercially available chemical building blocks, combinable through a defined set of validated reactions. Every structure the model proposes is, by construction, a structure that can be ordered, assembled, and tested within weeks.
The shift this represents is easier to grasp with an analogy. Imagine giving a chef unlimited imagination but constraining them to the ingredients available at a single grocery store. The creative space is still vast, far larger than any single chef could exhaust in a lifetime, but every dish they design can actually be cooked. The constraint does not diminish the creativity. It redeems it by ensuring that creativity produces something real.
This is what separates the current generation of generative models from everything that preceded them. G-SchNet could imagine molecules. SiMGen could build them to specification. SyntheMol could guarantee a synthesis route within a narrow chemical library. The synthesis-aware frameworks of 2024 and 2025 have combined all three capabilities at a scale of 48 million building blocks, a virtually unbounded combinatorial space that makes “we can’t build it” an answer that is increasingly difficult to give.
Two Ways to See It
For readers without a background in chemistry, the significance of this shift can be difficult to feel intuitively. Two analogies make it concrete.
The first is the distinction between a filter and an artist. The historical models described in this blog, Deep-AmPEP30, AMPlify, TransImbAMP, and sAMPpred-GAT were, at their core, filters. Sophisticated, powerful, genuinely useful filters, but filters nonetheless. They sorted through what already existed and identified the most promising candidates. The new generative models are something categorically different. They do not sort through a pile; they pick up a paintbrush. In the same way that image generation models produce images that have never existed by learning the deep structure of visual information, these models produce molecules that have never existed by learning the deep structure of chemical space.
The second is the “Halicin metric.” In chemistry, structural similarity is measured by a value called the Tanimoto coefficient, where 1.0 indicates identical structures and 0.0 indicates no shared features. Most drugs brought to market are close variations of existing compounds; their Tanimoto similarity to a known drug typically exceeds 0.9. When the MIT team analyzed Halicin against the full library of known antibiotics, they found a structural divergence of roughly 80%. If most antibiotics are different models of the same vehicle, Halicin was a different mode of transport entirely. Bacteria evolve resistance by learning to recognize and neutralize the things that attack them. A molecule they have never encountered, one that looks nothing like anything in their evolutionary experience, arrives without a target on its back.
Summary
The arc traced across this blog is, at its deepest level, a story about the expansion of the imaginable. The historical section described researchers learning to search more cleverly, mining extinct proteomes, reading peptide structures in three dimensions, and predicting clinical resistance from patient records. The revolutionary shift section describes researchers learning to stop searching entirely and start designing. The people and technical sections described the specific individuals and innovations that made that transition real.
We began with a crisis: drug-resistant infections killing over a million people annually, a number projected to exceed ten million by 2050, and a pharmaceutical pipeline that had been running dry for decades. We end with something that, a decade ago, would have sounded like science fiction: an AI that designs molecules that have never existed, constrained only by what chemists can actually build, targeting bacterial mechanisms that bacteria cannot easily evolve around.
This is not a solved problem. The gap between a promising mouse model result and a drug approved for human use remains wide, expensive, and full of failure. The synthesizability constraint narrows the distance but does not close it. Resistance will continue to evolve. The bacteria will not stop.
But for the first time in seventy years, the scientists are moving faster. The algorithms are learning more quickly than the pathogens are adapting. And the question is no longer whether AI will transform antibiotic discovery; it already has, but how far that transformation will carry us, and how quickly we choose to follow it.
We went from mining the past, to predicting the present, to architecting the future. The Antibiotic Winter is ending. What comes next is up to us.
