ApexGO

A peptide that inhibits Acinetobacter baumannii at 32 μmol l⁻¹ is not a drug, but it is a lead. Improving it means asking which of the roughly 3 × 10¹² sequences within five edits of a 20-residue scaffold is more potent — a question no fixed library can answer, because the answer is not in any library. Lead optimization needs a search that starts at a given sequence and works outward under a similarity budget.

ApexGO (APEX generative optimization) is the University of Pennsylvania answer to that setting, from the Machine Biology Group and Jacob Gardner's machine learning group. It reframes discrete sequence editing as continuous optimization: a transformer variational autoencoder maps peptides into a 256-dimensional latent space, and Bayesian optimization searches that space for points whose decoded sequences score well under APEX, the lab's minimum inhibitory concentration (MIC) predictor, used here as a frozen black-box oracle. A hard constraint keeps every proposal at least 75% similar to the template by edit distance, so the search returns derivatives rather than unrelated peptides. The work appeared as a preprint in November 2024 and in Nature Machine Intelligence in May 2026.

The feedback loop is what separates it from one-shot generative AMP models: the optimizer sees each candidate's oracle score before proposing the next, so it can navigate neighbourhoods the generative model never encountered in training. The paper reports experimentally active derivatives for all ten templates, including two whose motifs fall outside curated AMP databases.

#Key Features

• A pretrained peptide VAE, reused across every run: six-layer transformer encoder and decoder with d_model 128, bottlenecking to two 128-dimensional latent tokens. One self-supervised checkpoint starts every optimization run in the paper.
• Constrained latent-space search: a second Gaussian process models the similarity constraint, so Thompson sampling discards candidates predicted to violate the 75% threshold before the oracle is called.
• Multiple trust regions: twenty adaptive hyper-rectangles, each recentred on one of the twenty best unique sequences so far, return a diverse design set rather than one optimum.
• Joint VAE–surrogate updates: following the LOL-BO algorithm, encoder and Gaussian process are updated together every ten steps, reorganizing the latent space so peptides with similar MICs cluster.
• Per-template adaptation without labels: 20,000 synthetic derivatives from random mutagenesis within the similarity budget continue training the pretrained VAE toward a template's neighbourhood, raising the feasible-proposal rate.

#Technical Details

The VAE was pretrained on 4.5 million UniRef sequences under 50 residues, randomly cropped, for 118 epochs on a single NVIDIA RTX A6000, reaching 99.94% test-set reconstruction accuracy. Optimization uses a parametric Gaussian process regressor with 1,024 inducing points, an acquisition batch size of 50, and initialization from 10,000 sequences randomly mutated from the template.

Ten peptides mined from extinct proteomes served as templates, each with mid-micromolar activity chosen to leave headroom. ApexGO produced 100 designs — five Gram-negative-targeted and five broad-spectrum per template — all synthesized and assayed against 11 clinically relevant strains. Of these, 86 inhibited at least one strain at ≤64 μmol l⁻¹, and 68% improved on their parent's mean MIC, rising to a 72% improvement rate and an 85% hit rate on Gram-negative pathogens. Pearson and Spearman correlations between APEX predictions and measured MICs were 0.463 and 0.462. For mammuthusin-3 and arctoterin-1, short enough to enumerate exhaustively, ApexGO recovered all twenty of the top-twenty derivatives the oracle scores best. On the same constrained task, HydrAMP satisfied the similarity threshold for only 22.2% of proposals at its most conservative temperature, and PepDiffusion for none of a million samples.

#Applications

The intended use is scaffold optimization: supply a peptide with measurable but insufficient activity, a similarity budget and a target panel, and receive derivatives predicted to be more potent. Relaxing the constraint to zero turns the same machinery into unconstrained de novo design. In two preclinical mouse models of A. baumannii infection, mammuthusin-3 and mylodonin-2 derivatives outperformed their template controls and matched or exceeded polymyxin B.

#Impact

ApexGO supplies the optimization step a discovery pipeline built on prediction alone was missing, and it is the first ground-truth in vitro and in vivo validation of generative Bayesian optimization in any domain — a loop that applies wherever a cheap oracle and an expensive assay coexist. Its ceiling is the oracle's: a rank correlation near 0.46 bounds how well any search over APEX can track reality. The objective covers antimicrobial potency alone, leaving stability, pharmacokinetics and cytotoxicity outside the loop, and each template requires its own fine-tuning pass first. Code is MIT-licensed and ships the pretrained VAE checkpoint alongside the APEX oracle weights, archived on Zenodo and as a Docker image; the paper and its experimental data are CC BY. The same lab's later ApexOracle and ApexAmphion extend the line toward pathogen-conditioned prediction and reinforcement-learning-driven design.