APEX

Antimicrobial potency is not one number. A 20-residue peptide can clear Acinetobacter baumannii at a few micromolar and leave Staphylococcus aureus untouched, and the scoring functions that dominated peptide-antibiotic mining — net charge plus normalized hydrophobicity, collapsed into a single "antimicrobial" score — cannot express that distinction. They rank sequences by resemblance to a canonical cationic amphipathic peptide, exactly the prior a search for new chemistry should not carry.

APEX, short for antibiotic peptide de-extinction, replaces that score with 34 of them. Built by the Machine Biology Group at the University of Pennsylvania and released as a preprint in October 2023 before appearing in Nature Biomedical Engineering in June 2024, it reads a raw sequence of 8–50 residues and regresses a separate minimum inhibitory concentration (MIC) for each of 34 bacterial strains, from ESKAPEE pathogens and resistant clinical isolates to gut commensals. A multitask constraint on the final layer ties those 34 outputs to the bacterial phylogeny, so strains close on the Genome Taxonomy Database tree are pushed toward consistent predictions rather than learned as 34 unrelated problems.

The de-extinction in the name is what the trained ensemble was pointed at: 12,860 proteins from 208 extinct species, reduced to 5,190 non-redundant sequences, yielded 10,311,899 substrings of 8–50 residues, none of them in the training set. APEX flagged 37,176 as broadly active (predicted median MIC ≤ 80 μmol l⁻¹), 11,035 of which appear in no extant proteome.

#Key Features

• Strain-resolved MIC regression: One forward pass returns 34 concentrations, so a candidate is triaged by which pathogen it hits rather than by a generic activity score.
• Phylogeny-coupled multitask head: Weights in the final regression layer are regularized toward similarity for taxonomically close strains, using a distance matrix derived from the GTDB bacterial tree.
• Auxiliary classifier as data augmentation: A second head trained on public AMP and non-AMP labels shares the encoder, letting database-scale binary labels improve a regressor fit to a much smaller quantitative dataset.
• Physicochemical sequence encoding: Each residue is represented by its 566-dimensional AAindex property vector rather than a pretrained language-model embedding, making the model small enough to screen ten million sequences.
• Forty-model ensemble, shipped whole: The released predictor averages 8 architecture and training configurations × 5 random seeds, all 40 checkpoints included.

#Technical Details

The encoder is a bidirectional gated recurrent unit followed by layer normalization and a two-stage attention block: the first attention layer mixes residue positions using a residual concatenation of the GRU states with the raw AAindex input, and the second pools the result to a single peptide vector. Two fully connected heads sit on that shared representation — the 34-way MIC regressor and the AMP/non-AMP classifier. Training used 988 in-house peptides carrying 14,738 measured activity values across the 34 strains, plus 5,093 AMPs and 5,500 non-AMPs from DBAASP with overlapping sequences removed; 790 of the in-house peptides went to five-fold cross-validated tuning and 198 were held out.

On that held-out set the best single model reached R² 0.369 and Pearson correlation 0.621, an eight-model ensemble 0.473 and 0.669, and the 40-model ensemble that became the released APEX reached R² 0.546, Pearson 0.728 and Spearman 0.607 — ahead of elastic net, linear SVR, extra-trees, random forest and gradient-boosted baselines on most pathogens. Against the 69 de-extinct peptides later synthesized and assayed, predicted and measured log₂ MICs correlated at Pearson 0.448.

#Applications

The model's practical role is triage: it converts a proteome or a generative model's output into a ranked shortlist small enough to synthesize. Because the checkpoint is fixed and inference is cheap, later antimicrobial work reuses it as a scoring oracle — ApexGO treats predicted MIC as the objective in latent-space Bayesian optimization, AMP-Diffusion and AMPGPT rank generated candidates with it, and the same lab's ApexOracle benchmarks against it. The lab hosts a web server that runs the predictor on user-supplied sequences.

#Impact

APEX carried a search from an extinct proteome through to a preclinical result without retraining at any step. Of 69 synthesized candidates, 41 were active at ≤128 μmol l⁻¹ against at least one strain — a 59% hit rate against 24% for the charge-and-hydrophobicity scoring function on the same proteomes — and most killed bacteria by depolarizing the cytoplasmic membrane rather than permeabilizing the outer membrane, the mechanism typical of known AMPs. In a neutropenic mouse thigh infection model, single doses of elephasin-2, megalocerin-1, mammuthusin-2 and mylodonin-2 cut A. baumannii burden by two to four orders of magnitude, with mylodonin-2 matching polymyxin B. The selectivity story is one-sided: predicted Gram-negative selectivity held up in vitro (Pearson 0.295), while peptides predicted to be Gram-positive-selective were not (Pearson 0.11, mean selectivity score 1.02). The training MICs are available only on request, and the code and checkpoints carry a Penn academic non-commercial licence.