A team trained a generative model on 10 antimicrobial peptide scaffolds, asked it to propose chemical edits that would make each one work better, and chemically synthesized 100 of the suggestions. Eighty-five of the 100 cleared the in-vitro bar for antimicrobial activity. Seventy-two of them improved on the template they were derived from. Two of them, tested in mouse models of Acinetobacter baumannii infection, matched or beat a last-resort antibiotic. The system is called ApexGO, published online May 13 in Nature Machine Intelligence ↗.
The premise behind ApexGO is the same one behind every machine-learning antimicrobial paper of the last decade. There is more sequence space than experimental throughput. Most prior approaches handle that by screening fixed libraries (which is fast but ignores chemistry outside the library) or by generating candidates from scratch (which is bold but expensive to validate). ApexGO does neither. It starts from a peptide that already works and proposes edits that should make it work better.
The architecture has two pieces. A transformer variational autoencoder embeds each peptide sequence into a continuous latent space, where similar peptides land near each other. On top of that latent space, a Bayesian optimization loop proposes edits. Pick a point near the current peptide. Decode it back to a sequence. Predict whether the new sequence will be more potent, and if so, queue it for synthesis. The model picks each round's edits with full knowledge of how the prior rounds performed. The chemistry stays inside the architecture the template was already built on, which keeps synthesis tractable.
That is the framing. The numbers from the paper itself are what make the architecture interesting. From 10 template peptides, the team produced a queue of optimized derivatives and chemically synthesized 100 of them. Eighty-five of the 100 cleared the in-vitro hit threshold. Seventy-two improved on their template against Gram-negative pathogens. A typical generative antimicrobial-peptide method publishes hit rates in the 10 to 30 percent range. ApexGO's 85 percent comes from staying close to known scaffolds. The cost of that choice is that the model is not discovering wholly novel architectures.
The in-vivo work is the load-bearing part. The team carried optimized peptides through to two mouse models of Acinetobacter baumannii infection, the Gram-negative pathogen that sits at the top of the WHO's drug-resistance priority list. In both models, the ApexGO-optimized molecules cleared infection at potencies comparable to or exceeding those of the last-resort antibiotic comparator. The same molecules did this at activity levels above their unoptimized templates, which separates the model's contribution from the starting scaffold's intrinsic potency.
The team also did the standard chemical due diligence. Secondary structure. Mechanism of action. Cytotoxicity. The peptides retained the helical content of the templates and showed the same membrane-disrupting mode of action expected of cationic antimicrobial peptides, which suggests the latent-space edits did not push the chemistry across the boundary that separates an antimicrobial from a hemolytic.
What is missing, and the paper acknowledges this, is the next step. A peptide that clears infection in a mouse against a single pathogen is several risk steps short of a clinical candidate. The pharmacokinetics, the immunogenicity, the route of administration, and the manufacturing all still have to clear. The contribution ApexGO makes is upstream of those questions. It shrinks the design loop on the front end of the pipeline.
For peptide chemists running antimicrobial programs, the practical claim is that 100 chemically synthesized molecules can clear an 85 percent hit bar if the generator is constrained to scaffold-local edits and seeded with templates that already work. Whether the framework generalizes outside the cationic-peptide family the paper tested it on, or scales to the more constrained architectures used in oncology or metabolic programs, is the open question. The next paper will tell. For now, the news is that a generative model coupled to Bayesian optimization shipped a usable hit rate at the bench and a translational result in two mouse models, against a pathogen the last-resort antibiotic shelf was built for.