AI AMP Search
Trained models for antimicrobial peptide (AMP) prediction, part of the ai-amp-search project.
Models in this repo
| File | What it is | Size | Verified |
|---|---|---|---|
esm_amp.joblib |
ESM-2 embeddings + logistic regression (AMP classifier) | 3 KB | acc 0.992, AUC 1.000 (naive split) / 0.971 (homology-controlled) |
hemolysis.joblib |
ESM-2 embeddings + logistic regression (toxicity) | 3 KB | acc 0.994, AUC 0.998 |
amp_gpt.pt |
Character-level GPT trained on AMP sequences (2.69M params) | 10 MB | val perplexity 10.13 (best epoch) |
amp_scan.joblib |
Physicochemical RF (AMP classifier + log10 MIC regressor) | 141 MB | acc 0.986, AUC 0.999, PR-AUC 0.999 |
esm_finetuned.pt |
ESM-2 35M fine-tuned end-to-end with classification head | 134 MB | acc 0.956, AUC 0.990 |
Usage
import joblib, torch
from transformers import AutoModel, AutoTokenizer
# AMP classifier (frozen ESM-2 embeddings + logistic regression)
clf = joblib.load("esm_amp.joblib")["clf"]
tok = AutoTokenizer.from_pretrained("facebook/esm2_t6_8M_UR50D")
esm = AutoModel.from_pretrained("facebook/esm2_t6_8M_UR50D")
# ... embed with mean pooling, then clf.predict_proba(X)
Full pipeline (train, score, generate, evaluate) in the GitHub repo:
https://github.com/tomdabro/ai-amp-search
Data
- GRAMPA (Witten & Witten 2019, bioRxiv): 6,642 unique AMP sequences with log₁₀ MIC against E. coli. DOI: 10.1101/692681. Source databases: APD (Wang 2004, DOI: 10.1093/nar/gkh025), DRAMP (Fan et al. 2016, DOI: 10.1038/srep24482), DBAASP (Gogoladze et al. 2014, DOI: 10.1111/1574-6968.12489).
- ConsAMPHemo (Xie et al. 2025, Protein Science): 884 peptides, hemolytic vs non-hemolytic. DOI: 10.1002/pro.70087.
- UniProt (CC-BY-4.0): non-AMP negatives. The UniProt Consortium 2017, Nucleic Acids Research 45(D1):D158-D169. DOI: 10.1093/nar/gkw1099.
- DBAASP (Gogoladze et al. 2014, FEMS Microbiology Letters 357(1):63-68): clinical-isolate AMP-MIC validation data. DOI: 10.1111/1574-6968.12489.
Honest evaluation
The naive random split gives AUC 1.000 — inflated by homology leakage (near-identical family variants in both train and test). With a homology-controlled split (k-mer Jaccard >= 0.8 clustering, split by cluster), the honest AUC is 0.971. Length-binned AUC: [5,20] 0.959, [21,50] 0.972 — the model is not just predicting length.
License
MIT (code). Data is third-party with its own terms.