When Machines Learn to Kill Bacteria: The AI Revolution in Antibiotic Discovery

The last truly novel class of antibiotic was discovered in the 1980s. Since then, humanity has relied on variations of the same chemical themes — reshuffling known scaffolds while resistance outpaces us. But in the past three years, artificial intelligence has done something no human chemist could: designed entirely new antibiotic molecules from scratch, with mechanisms of action never seen in nature or the lab. We are witnessing the shift from finding needles in haystacks to manufacturing new haystacks entirely.

Two Waves of AI

The story of AI in antibiotic discovery has two distinct chapters, and understanding the difference between them matters.

Wave 1 was screening. In 2020, James Collins's lab at MIT used a graph neural network to screen the Drug Repurposing Hub — about 6,000 existing compounds — for molecules that could kill Escherichia coli. The model identified halicin, a compound originally investigated for diabetes, as a potent broad-spectrum antibiotic. It worked by disrupting the electrochemical gradient across the bacterial membrane — a mechanism unlike any existing drug class. It was the first antibiotic discovered by AI.

In 2023, the same group narrowed the approach. They trained a model on Acinetobacter baumannii specifically and discovered abaucin, a narrow-spectrum agent that inhibits LolE, a lipoprotein trafficking protein. Abaucin killed A. baumannii — one of the WHO's most critical priority pathogens — while leaving other bacterial species largely untouched. Narrow-spectrum antibiotics that spare the microbiome are the holy grail. AI found one.

But both halicin and abaucin came from existing chemical libraries. AI was used as a smarter search engine — extraordinarily powerful, but still limited to what humans had already synthesized.

Wave 2 is generation. And this is where the revolution truly begins.

Creating Molecules That Never Existed

In August 2025, the Collins lab published a landmark paper in Cell that changed the game. Instead of screening existing compounds, they built generative AI models — graph neural networks combined with CReM and variational autoencoder (VAE) architectures — that could design entirely new molecules optimized for antibacterial activity, synthesizability, and safety.

The scale was staggering. The models generated over 36 million candidate compounds. From these, 24 were selected and synthesized. Seven showed validated antibacterial activity. Two stood out.

NG1 is particularly noteworthy. It targets LptA, a protein in the lipopolysaccharide (LPS) transport pathway that shuttles LPS molecules from the inner membrane to the outer membrane of Gram-negative bacteria. Without LPS in its outer membrane, a Gram-negative bacterium loses the very structure that makes it so hard to kill. No existing antibiotic class targets LptA.

And here is a connection that may prove important: zosurabalpin, Roche's tethered macrocyclic peptide now entering Phase 3 clinical trials against carbapenem-resistant A. baumannii (CRAB), targets a different step in the same LPS transport pathway — the LptB2FGC complex. Two independently discovered molecules, one by AI and one by traditional screening, both converging on the same machinery. The LPS transport pathway is emerging as the most promising new target space for Gram-negative infections — potentially the first new class since 1968.

Screening at Planetary Scale

While the Collins lab was generating new molecules, a team at Genentech, NVIDIA, and MILA (the Montreal AI institute) was pushing the screening approach to its theoretical limits.

In November 2025, they published GNEprop in Nature Biotechnology — a deep-learning virtual screening platform trained on approximately 2 million molecules from a high-throughput screen against a sensitized E. coli strain. They then applied the model to screen 1.4 billion synthetically accessible compounds from the Enamine REAL library.

The numbers alone are impressive, but what matters more is the novelty of the hits. Many of the 82 validated antibacterial compounds were structurally dissimilar to any known antibiotic — they came from chemical space that no human chemist would have thought to explore. The 90-fold improvement in hit rate means AI is not just faster than traditional screening; it is finding things in regions of chemical space that traditional approaches would never reach.

And crucially, Genentech open-sourced the entire platform on GitHub. Anyone with the technical capability can now screen billions of compounds for antibacterial activity. In a field plagued by market failure, this matters.

From Molecules to Medicine: The Translation Problem

Discovering a molecule is not the same as delivering a drug. The gap between a promising hit and an FDA-approved antibiotic costs approximately $1 billion and takes 10–15 years. Most candidates fail. This is where the economics of antibiotic development have historically broken down — pharmaceutical companies cannot recoup their investment on drugs designed to be used sparingly and briefly.

AI accelerates discovery, but it does not fix the business model. So a new kind of organization has emerged to bridge the gap.

Phare Bio: A New Model for a Broken Market

Phare Bio is a nonprofit advancing the Collins lab's AI-designed antibiotic candidates toward the clinic. Their model is deliberately different from traditional pharmaceutical development:

• Nonprofit structure — No shareholders demanding returns. The mission is the product.
• Government-funded — $27 million ARPA-H TARGET grant (with the Collins Lab and Wyss Institute at Harvard) to screen 107 million molecules and generate new candidates. Goal: 15 preclinical candidates by 2030.
• Open-source AI — They are open-sourcing their generative AI models so other researchers can use them. In a market-failure disease, proprietary hoarding is a luxury the world cannot afford.
• Industry partnership for development — In December 2025, Phare Bio signed a partnership with Basilea Pharmaceutica — the first commercial agreement for AI-designed antibiotics. Phare designs molecules to a target product profile; Basilea brings manufacturing expertise and clinical development capability. Success-based payments align incentives.

This is a genuinely new model: a nonprofit creating drugs with public money and open-source AI, then partnering with industry for the expensive late-stage work. Phare Bio's CEO called the Basilea deal a "watershed moment." She may be right. If AI can compress the discovery phase from years to months while a nonprofit-industry hybrid handles development, the economic equation changes.

The Parallel Track: Structural Biology Fights Back

AI is not the only source of genuinely novel antibiotics. Two discoveries from structural biology deserve attention because they use a related but distinct strategy: exploiting deep knowledge of molecular structure to design drugs that are inherently harder to resist.

Lariocidin: The Lasso That Ties Up Resistance

In March 2025, Gerry Wright's lab at McMaster University published in Nature the discovery of lariocidin — a lasso peptide produced by a Paenibacillus soil bacterium that had been cultured for nearly a year to coax it into revealing its chemical secrets.

Lariocidin binds to a site on the small ribosomal subunit that no existing antibiotic targets. But the truly clever feature is how it binds: it interacts primarily with the RNA backbone rather than the nucleobases. Why does this matter? Most ribosome-targeting antibiotics bind to specific nucleotides. Bacteria can develop resistance through point mutations that change those nucleotides. But the RNA backbone — the sugar-phosphate chain that holds the nucleotides together — is structurally constrained. Mutating it would compromise the ribosome itself. Resistance through mutation becomes far harder.

And lariocidin doesn't stop at binding. It has a dual mechanism: it inhibits ribosomal translocation (the movement of the ribosome along mRNA) and induces miscoding (causing the ribosome to insert wrong amino acids). Two killing mechanisms from one molecule. In mouse models of CRAB infection, lariocidin achieved 100% survival versus 0% in untreated controls.

Cresomycin: Pre-Organized for Killing

Andrew Myers's lab at Harvard took a different approach. Published in Science in February 2024, cresomycin is a fully synthetic antibiotic designed using crystal structures of the ribosome. The key insight was pre-organization: the molecule is rigidified so that its shape in solution already matches its target-bound conformation. This eliminates the entropic penalty of binding — the molecule doesn't need to rearrange itself to fit, so it binds tighter than flexible competitors.

Cresomycin overcomes resistance mechanisms that defeat existing ribosome-targeting drugs. It is active against MRSA, resistant E. coli, and P. aeruginosa. Kinvard Bio is now commercializing the oxepanoprolinamide (OPP) platform with CARB-X funding. Both IV and oral formulations are planned, enabling step-down therapy — start patients on IV in the hospital, switch to oral for outpatient completion.

The Missing Piece

AI can now find and create antibiotic molecules faster than at any point in human history. Structural biology is yielding compounds with built-in resistance resilience. The science is working.

But the economics remain broken.

Clinical trials still cost upward of $1 billion. The FDA approval process takes a decade. And once approved, antibiotics generate modest revenue because — by design — they should be used sparingly. The PASTEUR Act, which would create a subscription model for novel antibiotics ($75–300 million per year, delinked from volume), has been reintroduced four times since 2020 and has never received a floor vote in either chamber. The UK's subscription pilot is expanding to £100 million per year in April 2026, but evidence of its impact on stimulating R&D remains thin.

AI compresses the front end of the pipeline. It reduces the time from "what should we look for" to "here is a validated hit" from years to weeks. But it does not compress the back end — manufacturing, toxicology, Phase I through III trials, regulatory review. Until the economic incentives for the back end are fixed, we will continue to discover molecules faster than we can develop them.

What This Means

We are at a genuine inflection point. For the first time, machines can create molecules that humans never imagined, test them virtually against billions of alternatives, and hand off the best candidates to organizations built specifically to navigate the valley of death between discovery and clinic.

The convergence is remarkable. AI-generated NG1 and Roche's zosurabalpin independently identified the same target pathway — LPS transport — as the Achilles' heel of Gram-negative bacteria. McMaster's lariocidin and Harvard's cresomycin both exploit structural biology to build resistance-proof ribosome binders. GNEprop and the Collins lab's generative models both push beyond known chemical space into territories no human intuition would have explored.

Multiple independent approaches are converging on the same conclusion: there are still antibiotics to be found. The chemical space is vast. The targets exist. Evolution has not closed every door.

But AI discovers. It does not deliver. The pipeline from molecule to medicine still runs through a system that was designed for blockbuster drugs, not antibiotics that save lives by being used as little as possible. Until the incentives match the science, we will keep finding cures faster than we can bring them to patients.

The machines have learned to kill bacteria. Now we need to learn to bring their discoveries to the bedside.

Sources: Stokes et al., "A Deep Learning Approach to Antibiotic Discovery," Cell (2020); Liu et al., "An Exploration Strategy Improves the Diversity of de novo Ligands," Cell (Aug 2025); Wong et al., "Discovery of a structural class of antibiotics with explainable deep learning," Nature (2023); Lam et al., "GNEprop: Generative Neural Network-based Propensity Scoring for Virtual Screening," Nature Biotechnology (Nov 2025); Copp et al., "Lariocidin, a lasso peptide antibiotic with a novel mechanism of action," Nature (Mar 2025); Mitcheltree et al., "A synthetic antibiotic class overcoming bacterial multidrug resistance," Science (Feb 2024); Phare Bio press releases (ARPA-H TARGET grant, Basilea partnership, Dec 2025); WHO Antibacterial Pipeline Database (2025); Roche/Genentech zosurabalpin Phase 3 announcement (2026).