Automated variant network assembly and prediction framework

This repository implements:

• An approach for deriving variant-specific networks from the INDRA Database, connecting genetic variants to their downstream biological and disease processes potentially via molecular intermediaries. (indra_variant.generate module)
• A web application that allows browsing and interacting with these networks, deployed at https://variants.indra.bio. (indra_variant.app module)
• An integrative, transformer-based prediction framework that makes use of the variant-effect networks as well as protein language model embeddings and other features to predict variant-effect mechanisms. (indra_variant.predict module)

Requirements

• python >= 3.9

Data availability:

Place the following files under data/:

• training_feature_input.tsv - input features for training
• genes_to_pmids.tsv - mapping of proteins to relevant publications
• label_classified.tsv - mapping of 1085 bp/disease labels to 30 categories
• human_domains.tsv - domain info from UniProt uniprot_sprot.dat.gz (downloaded May 13, 2025)
• clinvar_patho_subset.tsv.gz - extract the fields from the ClinVar clinvar_variant_summary.txt (.gz)

Usage

Generate protein embeddings

Use predict/seq_embedding/embedding.py to convert protein features into ESM-2 embeddings for downstream tasks.

• input: data/training_feature_input.tsv
• output: training_feature_esm2.tsv

Extract knowledge-graph paths

Use predict/gnn_pretraining/extract_triples.py to extract triples (subject, rel, object) from literature-derived causal paths.

• input: data/training_feature_esm2.tsv
• output: triples.csv; triples_unique.csv; variant_paths.tsv

Generate node embeddings using GNN

Use predict/gnn_pretraining/rgcn_pretrain.py to learn structural representations of nodes in the knowledge graph.

• input: triples_unique.csv
• output: node_embeds.pt (graph node embedding weights)

Split dataset

Use predict/train/dataset_split.py to split the dataset randomly for model training and evaluation.

• input: data/label_clssified.tsv; variant_paths.tsv
• output: splits/(splits_index)

Build training dataset

Use predict/train/build_train_dataset.py to assemble complete training dataset by combining embeddings, paths, and label mappings.

• input: training_feature_esm2.tsv; variant_paths.tsv; data/label_classified.tsv; node_embeds.pt
• output: path_dataset_bag_full.pt; W_var.pt

Train model

Use predict/train/training.py.

• input: path_dataset_bag_full.pt

Run the script with the -h argument to see additional arguments.

Predict

Use predict/train/prediction.py.

Run the script with the -h argument to see additional arguments.