PeptideLocator2

Endogenous peptide region prediction from protein sequences using frozen PLM embeddings.

Amritansh Tiwari & Raúl Fernández-Díaz — UCD Shields Lab, University College Dublin

---

Overview

PeptideLocator2 predicts endogenous bioactive peptide regions directly from protein sequences. It uses a MCC-weighted ensemble of 20 models (4 PLM backbones × 5 folds) trained on the DeepPeptide dataset with GraphPart cross-validation splits.

Backbones

• ProtT5-XL (Elnaggar et al., 2021)
• ESM2-3B (Lin et al., 2023)
• ESMC-300M (Hayes et al., 2024)
• ESMC-600M (Hayes et al., 2024)

Usage

git clone https://github.com/Amt11cdr/PepLoc2
cd PepLoc2
pip install -r requirements.txt
python predict.py --fasta input.fasta --output results.json

Arguments

Argument	Default	Description
--fasta	required	Input FASTA file
--output	required	Output JSON file
--threshold	0.5	Prediction probability threshold
--min_len	5	Minimum peptide length (AA)

Output format

{
  "PROTEIN_ID": {
    "sequence": "MKTL...",
    "length": 120,
    "peptides": [
      {"start": 45, "end": 62, "length": 18}
    ],
    "probabilities": [0.02, 0.03, ...]
  }
}

License

Apache 2.0