How izgys/enzyme-llm-platform works
This is an enzyme EC-subclass classifier built around ESM-2 fine-tuning with LoRA and homology-aware evaluation. Its principal workflow downloads and curates Swiss-Prot sequences, creates cluster-separated splits, trains and evaluates a classifier, then calibrates confidence for prediction. Interpretability, structure mapping, visualization, and experiment tracking are additional documented capabilities. Source excerpts substantiate the data pipeline and dataset layer; several model-serving and analysis modules are represented from README claims and tree paths, so their internal wiring is intentionally limited.
Main entry points: interpretability Integrated gradients · model ESM-2 classifier · clustering.py Homology splitting · training LoRA fine-tuning
Components 20
Data preparation
- Fetch and filter
src/data/download.py - Raw records
data/raw - Clean and label
src/data/preprocess.py - Processed records
data/processed - Homology splitting
src/data/clustering.py - Train val test
data/processed - Tokenized batches
src/data/dataset.py
Model learning
- LoRA fine-tuning
src/training - ESM-2 classifier
src/model - Zero-shot baseline
experiments/baseline_zeroshot/run_baseline.py - Confidence calibration
src/calibration
Prediction
- Inference endpoint
src/inference
Interpretation
- Integrated gradients
src/interpretability - MSA conservation
- PDB mapping
src/interpretability - Result visualization
src/visualisation
Experiment operations
- MLflow tracking
Other components
- Swiss-Prot
- MMseqs2
- Prediction client
Connections 20
- Swiss-Prot → Fetch and filter: provides entries
- Fetch and filter → Raw records: writes records
- Raw records → Clean and label: loads records
- Clean and label → Processed records: writes labels
- Processed records → Homology splitting: loads records
- Homology splitting → MMseqs2: runs clustering
- Homology splitting → Train val test: writes splits
- Train val test → Tokenized batches: loads sequences
- Tokenized batches → LoRA fine-tuning: supplies batches
- LoRA fine-tuning → ESM-2 classifier: fine-tunes
- ESM-2 classifier → Confidence calibration: supplies logits
- Confidence calibration → Inference endpoint: calibrates scores
- Prediction client → Inference endpoint: requests prediction
- Inference endpoint → ESM-2 classifier: predicts with
- ESM-2 classifier → Integrated gradients: attributes predictions
- Integrated gradients → MSA conservation: cross-checks conservation
- Integrated gradients → PDB mapping: maps scores
- Integrated gradients → Result visualization: provides attributions
- LoRA fine-tuning → MLflow tracking: tracks runs
- Zero-shot baseline → Train val test: evaluates embeddings