How djcrocker/caliva works
CALIVA is a missense-variant pathogenicity predictor that turns ClinVar-labeled variants and protein data into structural and sequence features, trains graph and boosted-tree models, and produces calibrated probability predictions. The documented rebuild workflow runs from local cohort and structure preparation through feature generation and graph assembly; training and scoring consume those representations. Other product branches fetch and compare reference-predictor scores, calibrate and evaluate predictions, and analyze model interpretation and errors. Some links are documented in the README rather than shown in sampled source; uncertain implementation links are omitted.
Main entry points: score_full_cohort.py Cohort scoring · graph_dataset.py Graph dataset · generate_graphs.py Graph assembly · train_gnn.py GNN training
Components 22
Cohort preparation
- Cohort builder
src/local/build_cohort.py - Structure and FoldX prep
src/local/run_foldx.py - Integrity checks
src/local/verify_integrity.py
Protein features
- ESM embeddings
src/features/generate_esm_embeddings.py - MSA conservation
src/features/conservation_features.py - Variant features
src/features/extract_features.py - Interface features
src/features/interface_features.py - Graph assembly
src/features/generate_graphs.py
Modeling
- Graph dataset
src/ml/graph_dataset.py - GNN architectures
src/ml/gnn_model.py - GNN training
src/ml/train_gnn.py - Boosted-tree training
src/ml/train_ensemble_gbdt.py - Ensemble selection
src/ml/find_best_ensemble.py - Cohort scoring
src/ml/score_full_cohort.py
Calibration and analysis
- Reference score collection
src/external/fetch_alphamissense.py - External comparison
src/external/external_comparison.py - Probability calibration
src/external/temperature_scale.py - ACMG calibration
src/external/acmg_calibration.py - Model evaluation
src/ml/evaluate_model.py - SHAP interpretation
src/stats/shap_analysis.py - Error analysis
src/stats/error_analysis.py
Other components
- Researcher
Connections 25
- Researcher → Cohort builder: prepares cohort
README.md - Cohort builder → Structure and FoldX prep: precedes prep
README.md - Structure and FoldX prep → Variant features: supplies structure features
README.md - Cohort builder → Variant features: supplies cohort data
README.md - Variant features → Graph assembly: provides variant features
src/features/generate_graphs.py - ESM embeddings → Graph assembly: adds embeddings
src/features/generate_graphs.py - MSA conservation → Graph assembly: adds conservation
src/features/generate_graphs.py - Interface features → Graph assembly: adds annotations
README.md - Graph assembly → Graph dataset: loads graph files
src/ml/graph_dataset.py - Graph dataset → GNN training: provides training data
src/ml/train_gnn.py - Graph dataset → Boosted-tree training: provides training data
src/ml/train_ensemble_gbdt.py - GNN training → GNN architectures: trains architecture
src/ml/train_gnn.py - GNN training → Graph dataset: uses dataset
src/ml/train_gnn.py - Boosted-tree training → Graph dataset: uses dataset
src/ml/train_ensemble_gbdt.py - Ensemble selection → GNN training: selects ensemble members
src/ml/find_best_ensemble.py - Ensemble selection → Boosted-tree training: selects ensemble members
src/ml/find_best_ensemble.py - Ensemble selection → Cohort scoring: supports deployed scoring
- Cohort scoring → External comparison: supplies CALIVA predictions
src/external/external_comparison.py - Reference score collection → External comparison: supplies reference scores
- Cohort scoring → Probability calibration: temperature scales scores
README.md - Probability calibration → ACMG calibration: calibrates probabilities
src/external/acmg_calibration.py - Cohort scoring → Model evaluation: evaluates predictions
- Boosted-tree training → SHAP interpretation: explains tree model
src/stats/shap_analysis.py - Cohort scoring → Error analysis: supports error review
- Structure and FoldX prep → Integrity checks: checks outputs