ApexOracle

Predicting and generating antibiotics against unseen pathogens

Tianang Leng^† · Fangping Wan^† · Marcelo D. T. Torres^† · Cesar de la Fuente-Nunez
University of Pennsylvania
^† Equal contribution

A modular, multimodal platform that combines molecular structure with pathogen genome and phenotype context to predict antimicrobial activity and guide de novo molecule generation.

Overview · Quick start · Modules · Models & data · Citation

What ApexOracle does

ApexOracle represents molecules with a diffusion language model (DLM), combines them with complementary pathogen representations through cross-attention, and uses task-specific heads for prediction or guidance.

flowchart LR
    M["Molecule structure<br/>DLM"] --> F["Multimodal fusion<br/>cross-attention"]
    G["Pathogen genome<br/>Evo-2"] --> F
    T["Phenotype context<br/>Me-LLaMA"] --> F
    F --> P["MIC · antimicrobial class · synergy"]
    F --> H["Noisy guidance heads"]
    H --> D["de novo candidates<br/>discrete diffusion"]

Quick start

git clone --recurse-submodules https://github.com/DragonDescentZerotsu/ApexOracle.git
cd ApexOracle
python scripts/check_module_locks.py

Already cloned without submodules? Run ./scripts/bootstrap.sh.

[!NOTE] Each scientific module retains its validated environment. ApexOracle intentionally does not force DLM pretraining, Evo-2 extraction, MIC prediction, and guided generation into one Python environment. See the environment guide.

Modular release

The exact gitlink for every module is fixed in manifests/modules.lock.yaml. Scientific implementation commits, release documentation commits, model revisions, and recovery refs are kept distinct in the release provenance.

Models and data

Immutable revisions, file sizes, checksums, licenses, and release scope are recorded in manifests/model_weights.yaml and manifests/data_assets.yaml. Model weights, datasets, embeddings, caches, and raw outputs are not stored as Git objects.

The release distinguishes runnable inference from recomputation of paper numbers. Representative inference-only weights power the public quickstarts; paper-result recomputation uses frozen sample-level predictions, splits, checkpoint provenance, and metric scripts instead of requiring every optimizer-bearing historical checkpoint. The latest Zenodo v5.0.0 retains the full version-series payload, so users do not need to download earlier versions separately. It includes the exact Fig. 1b classification capsule, the post-paper fixed-split MIC reconstruction, the high-confidence synergy replay, and the source-partitioned public model-ready tables. The paper reports 121,265 MIC measurements before token-length filtering; the frozen model-ready table contains 120,955 rows after excluding 310 structures over 1,024 tokens, and all 15,718 private in-house MIC rows are excluded from the public capsule. Immutable v1--v4 DOIs and their incremental release history remain in the release provenance; all versions belong to the same concept DOI 10.5281/zenodo.15612047, not separate Zenodo projects. See the reproducibility scope and compute requirements. Fresh quickstart wall time and peak RAM/VRAM are frozen in manifests/quickstart_benchmarks_2026-08-11.json. The external paper-data capsule plan records which splits are exact, reconstructed, or still unrecovered; the repository hygiene policy prevents those assets from being duplicated into Git.

[!IMPORTANT] The public MIC quickstart demonstrates one inference member. The paper metrics use the frozen seven-member ensemble. Likewise, the compact generation bundle validates the released runtime path; it is not itself an experimental activity result.

ApexOracle/
├── modules/
│   ├── core/             # prediction, evaluation, and reproducibility
│   ├── dlm_pretrain/     # DLM + MTR pretraining producer
│   ├── mdlm/             # embeddings, guidance heads, and scoring
│   ├── evo2/             # genome embedding extraction
│   └── generation/       # guided discrete diffusion
├── quickstarts/          # runnable public examples
├── environments/        # per-module environment policy
├── manifests/           # immutable module and asset records
├── scripts/             # bootstrap, validation, and archive tools
└── docs/                # release status and provenance

python scripts/build_source_archive.py --output ApexOracle-source.tar.gz
python scripts/check_source_archive.py ApexOracle-source.tar.gz

Citation

If ApexOracle is useful in your work, please cite the paper and the software release you used:

@article{leng2025predicting,
  title   = {Predicting and generating antibiotics against future pathogens with ApexOracle},
  author  = {Leng, Tianang and Wan, Fangping and Torres, Marcelo Der Torossian and de la Fuente-Nunez, Cesar},
  journal = {arXiv preprint arXiv:2507.07862},
  year    = {2025}
}

Machine-readable software and dataset citation metadata are available in CITATION.cff.

License and history

The super-repository orchestration layer is released under the MIT License. Each submodule retains its own license; see NOTICE before redistribution. The pre-conversion monorepo remains recoverable from branch legacy-monorepo and annotated tag legacy-monorepo-snapshot-2026-08-10.