ApexOracle

Antimicrobial resistance is implicated in roughly 4.95 million deaths per year, and the antibiotic pipeline has not kept pace. Machine learning has helped, but most antimicrobial activity predictors are built around a fixed panel of target strains and need a bespoke, expensively assayed dataset for every new one. That design fails precisely when it matters most — against a newly emerged pathogen or a resistant clinical isolate for which no minimum inhibitory concentration (MIC) data exist at all.

ApexOracle, from the Machine Biology Group at the University of Pennsylvania and released as a preprint in July 2025, attacks that gap by conditioning on the pathogen rather than memorizing it. It takes three inputs — a strain's complete genome, a short textual description of its traits, and a candidate molecule — and returns a predicted potency, a generated molecule, or both. Genomic context comes from Evo 2, whose per-base hidden states are pooled into a whole-genome representation; textual context comes from a Me-LLaMA3-8B encoding of strain-trait passages covering taxonomy, Gram status, characteristic mutations, and documented resistance profiles.

The molecular side is a masked discrete diffusion language model over SELFIES strings, serving simultaneously as the representation extractor for prediction and the generative engine for design. Because SELFIES describes chemistry at the atom level rather than as a residue string, one model spans small-molecule antibiotics and peptides alike, including the 71.5% of antimicrobial peptides in DBAASP that carry non-canonical residues, terminal modifications, or intrachain bonds — a population residue-level predictors either simplify away or cannot encode.

#Key Features

• Pathogen-conditioned activity prediction: Cross-attention fuses a molecule query against two strain-knowledge banks, genomic and textual, so the model can score a compound against an organism absent from training.
• Unified prediction and generation: The same diffusion backbone that produces molecular embeddings also samples new molecules, with the internal MIC regressor steering generation via discrete predictor-based guidance.
• One chemistry for peptides and small molecules: A companion converter, PepLink, maps peptide sequences to SELFIES and back, supporting 404 non-canonical amino acids, 11 intrachain bond types, and 298 terminal modifications.
• Transfer without strain-specific labels: The model predicts small-molecule antibiotic activity zero-shot for strains carrying no small-molecule training labels, and beats strain-specific baselines once fine-tuned.
• Synergy prediction: Concatenating two fused molecule–strain embeddings feeds a head that classifies fractional inhibitory concentration index outcomes for drug combinations.

#Technical Details

The diffusion language model is a 12-layer Diffusion Transformer with a latent dimension of 768, trained with the MDLM absorbing-state objective on 121.6 million molecules of up to 1,024 SELFIES tokens: PubChem compounds obtained via MoLFormer-XL (111.4M), SmProt v2.0 peptides (826K), short UniRef and UniProt entries (7.0M and 3.7M), and 10M modification-rich peptides from CycloPS. An auxiliary multi-task regression loss trains the <CLS> embedding to predict 209 RDKit descriptors. Under a strain-unaware five-fold protocol on 19 well-covered DBAASP strains, the encoder exceeded the next-best chemical language model by 27.1% in R². Pretraining ran on eight A100 80GB GPUs to a validation perplexity of 1.58.

Downstream training used 121,265 MIC measurements spanning 17,988 peptides and 5,632 strains, plus 49,331 molecule–strain pairs with binary activity labels against S. aureus RN4220, E. coli BW25113, and A. baumannii ATCC 17978. Held-out strain-wise evaluation gave an average R² of 0.5793 with a seven-model ensemble and 0.5032 without, degrading gracefully with taxonomic distance to 0.4337 across an eleven-cluster species-wise split and 0.3744 under the hardest three-cluster split. Fine-tuned, ApexOracle improved on published strain-specific small-molecule classifiers by 8.3% AUROC and 37.7% AUPRC on average, and the synergy head reached 0.7539 AUROC on 2,732 curated combination pairs. Ablations show genomic embeddings contribute most, with textual traits close behind.

#Applications

The intended workflow is genome-in, candidate-out: sequence a clinical or environmental isolate, assemble a trait description, and either rank existing compounds by predicted MIC or sample new chemotypes tailored to that organism. In the paper's demonstration, the model designed molecules against two resistant strains it had never seen — colistin-resistant E. coli BAA-3170 and multidrug-resistant P. aeruginosa BAA-3197 — and guided sampling shifted predicted MICs downward while lowering maximum Tanimoto similarity to DBAASP compounds, indicating designs distinct from the training chemistry.

#Impact

ApexOracle reframes antimicrobial modeling as a pathogen-conditioned problem, removing the per-target dataset requirement that has confined earlier predictors to a handful of laboratory strains, and its pairing of a genome language model with a medical LLM as context encoders is a template other target-conditioned design tasks can borrow. The results remain in silico: the preprint reports no wet-lab confirmation of generated molecules, generation optimizes potency without modeling toxicity, metabolic stability, or synthetic accessibility, and accuracy falls away for clades whose biology is unlike the training data — notably the cell-wall-deficient Mycoplasmatota. The work has not yet appeared in a peer-reviewed journal. Code is MIT-licensed, the trained checkpoint and precomputed embeddings are deposited on Zenodo under CC BY 4.0, and an embedding-extraction model is published on HuggingFace.