Lab Newsletter — September 10, 2026: An Antibiotic in the Machine

Antibiotic resistance is one of those slow emergencies that never makes the front page until it’s personal. The uncomfortable arithmetic: resistant bacteria are evolving faster than we discover drugs to stop them, and the traditional discovery pipeline — culture, screen, medicinal chemistry — has been sputtering for decades. This week we’ve watched AI design proteins, evolve them, and compute the forces that hold molecules together. Today those same tools meet a problem with a body count. The move throughout is one this digest keeps returning to: when the search space is too large for people to comb, teach a machine to comb it — and point it somewhere new.

💊 The landmark: predict activity, then go looking

The paper that showed this could work is Stokes et al. (Cell, 2020). Its premise is stark: “due to the rapid emergence of antibiotic-resistant bacteria, there is a growing need to discover new antibiotics. To address this challenge, we trained a deep neural network capable of predicting molecules with antibacterial activity.” Rather than design a molecule, they used the network as a filter over libraries humans had already made — and it found something odd and wonderful: “a molecule from the Drug Repurposing Hub—halicin—that is structurally divergent from conventional antibiotics and displays bactericidal activity against a wide phylogenetic spectrum of pathogens including Mycobacterium tuberculosis and carbapenem-resistant Enterobacteriaceae.” Then they aimed it at the vast unknown: “from a discrete set of 23 empirically tested predictions from >107 million molecules curated from the ZINC15 database, our model identified eight antibacterial compounds that are structurally distant from known antibiotics.” A model looking where chemists’ intuitions don’t.

🎯 Aiming at the hardest targets

Broad-spectrum discovery is one thing; hitting a specific, nightmarish pathogen is another. Liu et al. (Nature Chemical Biology, 2023) went after one of the worst — “Acinetobacter baumannii … a nosocomial Gram-negative pathogen that often displays multidrug resistance,” a WHO priority target notoriously resistant to new drugs. Their pitch for ML is the efficiency: it “allow[s] for the rapid exploration of chemical space, increasing the probability of discovering new antibacterial molecules.” From a modest, focused experiment — “we screened ~7,500 molecules for those that inhibited the growth of A. baumannii in vitro” — they trained a network and predicted new structures, arriving at “abaucin, an antibacterial compound with narrow-spectrum activity against A. baumannii.” Narrow-spectrum is a feature, not a bug: a drug that spares the rest of your microbiome. And it worked in a living host — abaucin “could control an A. baumannii infection in a mouse wound model.”

🔍 Making the black box explain itself

A recurring worry with deep learning in discovery is that it’s an oracle: it says yes without saying why. Wong et al. (Nature, 2024) set out to fix that, noting that such approaches “typically use black box models and do not provide chemical insights.” Their answer married scale to interpretability. They measured “the antibiotic activities and human cell cytotoxicity profiles of 39,312 compounds and applied ensembles of graph neural networks to predict antibiotic activity and cytotoxicity for 12,076,365 compounds” — then used explainable graph algorithms to extract the substructures driving the prediction. The result was not one molecule but a class: one “selective against methicillin-resistant S. aureus (MRSA) and vancomycin-resistant enterococci, evades substantial resistance, and reduces bacterial titres in mouse models.” The headline for us is the methodological one: “machine learning models in drug discovery can be explainable, providing insights into the chemical substructures that underlie selective antibiotic activity.”

🧪 From screening to designing

Every model so far ranks molecules that already exist. The next step is to invent them. Das et al. (Nature Biomedical Engineering, 2021) did, tackling the fact that “the de novo design of antimicrobial therapeutics involves the exploration of a vast chemical repertoire to find compounds with broad-spectrum potency and low toxicity.” Their pipeline is a nice echo of yesterday’s physics: it “leverages guidance from classifiers trained on an informative latent space of molecules modelled using a deep generative autoencoder, and screens the generated molecules using deep-learning classifiers as well as physicochemical features derived from high-throughput molecular dynamics simulations.” Generation, then a simulation-based sanity check. The payoff was fast and real: “within 48 days, we identified, synthesized and experimentally tested 20 candidate antimicrobial peptides, of which two displayed high potency against diverse Gram-positive and Gram-negative pathogens … and a low propensity to induce drug resistance.”

🦠 Mining nature’s own arsenal

Here’s the twist the field didn’t see coming: we don’t only have to invent antibiotics — the microbial world is already full of them, waiting to be read out of sequence. Ma et al. (Nature Biotechnology, 2022) turned language models loose on the gut. The catch is size: “the human gut microbiome encodes a large variety of antimicrobial peptides (AMPs), but the short lengths of AMPs pose a challenge for computational prediction.” Their fix borrowed straight from NLP — they “combined multiple natural language processing neural network models, including LSTM, Attention and BERT, to form a unified pipeline for candidate AMP identification from human gut microbiome data.” The hit rate is what stuns: “of 2,349 sequences identified as candidate AMPs, 216 were chemically synthesized, with 181 showing antimicrobial activity (a positive rate of >83%).” Reading DNA as text to find drugs written into the microbiome.

🌍 At planetary scale, and open to all

And then the scale went global. Santos-Júnior et al. (Cell, 2024) mined essentially the whole known microbial world. From “a vast dataset of 63,410 metagenomes and 87,920 prokaryotic genomes from environmental and host-associated habitats,” their machine-learning approach built “the AMPSphere, a comprehensive catalog comprising 863,498 non-redundant peptides, few of which match existing databases.” Then they proved it: “we synthesized and tested 100 AMPs against clinically relevant drug-resistant pathogens … A total of 79 peptides were active, with 63 targeting pathogens.” Nearly a million candidate antibiotics — and, in the detail that makes it our kind of science, “an open-access resource for antibiotic discovery.” Discovery released as public infrastructure, not locked in a pipeline.

🧬 Why it’s our kind of problem

The thread tying these six papers together is a single verb: AI widens where we look for medicine. Old drug libraries (halicin), a focused in-house screen (abaucin), twelve million virtual compounds (Wong), a generative latent space (Das), and the microbiome’s own encrypted arsenal (Ma, AMPSphere) — each is a search space too vast for intuition, made tractable by a model. It’s the same pattern behind reading spectra as language or genomes as text: point a general learner at biology’s haystacks. It closes a loop we care about, too — predict, synthesize, assay in a dish and a mouse, refine — exactly the design–build–test–learn cycle a self-driving lab is built to turn. And the ethos is ours to the core: AMPSphere is open; the models and growth-inhibition datasets are shared yardsticks — the same publish-the-model-and-the-data spirit behind the BioImage Model Zoo and BioEngine. A drug-discovery model you can call like a service, benchmarked on shared data, aimed at one of medicine’s most urgent problems: that’s AI for life science doing exactly what we hope it will.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.