Protein-ML: Peptide Modeling Pipeline

This repository contains a multi-phase pipeline for peptide and protein modeling, starting from hemolysis prediction and extending toward generative design and closed-loop optimization. It is designed as a realistic end-to-end ML project rather than a single benchmark model.

Phase roadmap

• Phase 0 – Setup & Data

  • Source peptide datasets (AMPDeep-style hemolysis), engineer basic physicochemical features, and run exploratory data analysis.
  • Export CPU-friendly sampled datasets for rapid experimentation.

• Phase 1 – Transformer Predictor

  • Fine-tune a transformer model (ESM-2) to classify hemolytic vs non-hemolytic peptides.
  • Save model weights, metrics, and reusable sequence embeddings for downstream tasks.

• Phase 2 – Transformer LM (planned)

  • Conditional sequence generation conditioned on properties (e.g., hemolysis/solubility labels).
  • Use pretrained protein language models as a base for guided peptide generation.

• Phase 3 – Diffusion Model (planned)

  • Discrete diffusion / denoising model over amino acid tokens.
  • Sample valid, diverse peptide sequences under structural and property constraints.

• Phase 4 – Property Scorers (planned)

  • Ensemble of ML models for multiple peptide properties (e.g., hemolysis, solubility, other developability metrics).
  • Fast scoring interface to plug into generators and screening loops.

• Phase 5 – Closed-Loop Pipeline (planned)

  • Active learning cycles that propose, score, and (in a real setting) test new sequences.
  • Track hit rate improvements over multiple design–evaluate iterations.

• Phase 6 – Analysis & Scaling (planned)

  • Compare model families (classifier, LM, diffusion, ensemble).
  • Perform motif analysis, interpretability, and plan for larger-scale training.

Current implementation status

• ✅ Phase 0: data sourcing, feature engineering, rule-based solubility heuristic, and EDA utilities (phase0_peptide_data_prep_eda.py).
• ✅ Phase 1: ESM-2-based hemolysis classifier with full training/evaluation loop and saved artifacts (phase1_esm2_hemolysis_classifier.py).
• 🔜 Phases 2–6: generative modeling, property scoring, and closed-loop optimization.

Environment

The project uses a conda environment tailored for PyTorch, Hugging Face, and scientific Python.

To create it:

conda env create -f environment.yml
conda activate peptide-design