MDRAMP model assets
Model assets for the MDRAMP peptide modeling project. Source code: https://github.com/SiyiPeng66/MDRAMP (the matching engineering release must accompany these assets).
Contents
- BaseModel: public-data-trained experts and autoregressive generator. It has no quantitative MIC head; experimental calibration initializes one.
- FinalModel: preference-trained MIC head plus the existing prospective generator. ESM3, expert weights, feature scaling and a frozen 575-protein panel are shared. Generation samples the policy and reranks candidates with the current MIC head; it does not optimize the policy on each request.
- shared/: expert weights, descriptor scaler and the fixed scoring panel. The small panel embedding file is included to preserve the exact frozen model binding. Large public-training feature caches are not included.
FinalModel uses user-adjusted best-MIC labels and explicit ranking preferences. Outputs are not new experimental MIC measurements or independently calibrated efficacy probabilities. Requested ranks were fitted within fixed candidate sets; they do not transfer to arbitrary inputs. No clinical efficacy or safety claims are made.
Download
Download this repository into assets/ in the matching MDRAMP project. Pin the repository commit for reproducibility. Install the project's dependencies first.
from huggingface_hub import snapshot_download
snapshot_download("BillyLiang/MDRAMP", local_dir="assets", revision="<commit>")
snapshot_download(
"biohub/esm3-sm-open-v1",
revision="47f0545b2b6daf26a93439a3cd610f4f7f3d5478",
local_dir="assets/shared/esm3",
)
ESM3 is not redistributed here. Access and use are subject to its upstream terms; authenticate and obtain access if required. external_assets.json records the exact upstream revision and main-weight SHA-256. The full upstream snapshot is used because the current ESM SDK expects companion resources.
python main.py predict --input demo/predict.fasta
python main.py calibrate --input demo/experiments.csv
python main.py generate --count 100
Custom weights require a companion scorer.json. Configurations use relative paths. checksums.json records downloadable asset hashes. The complete project, experimental data and candidate workbook are separate deliverables, not part of this model repository.
Provenance
Public experts were trained using public AMP/background and peptide–protein inputs. The final head was fitted using private experimental observations with explicit target transformations and preference constraints. Original training provenance is retained in checkpoints; historical local path strings are metadata, not required runtime paths. Shared assets are reused by both model roles.
This repository does not assign a new license to upstream ESM3 or source datasets. Consult their original terms and the project owner regarding redistribution and downstream use.
Negative-constraint update
FinalModel now fits 27 experimental negatives and 57 public conditional negatives above 256 ug/mL. The other 203 original predictions changed by at most 1.05%; fixed-set DN01/DN05/LO18 ranks remain 1/5/8. Frozen internal MIC-head holdout: 7/12 above 256, with prior AMP-pretraining exposure. Public source primary review is incomplete. Very high MIC outputs are not validated quantitative estimates. See FinalModel/validation.json.