Using Genomic LLMs to Evaluate SV Pathogenicity from Genomic Sequences

This repository contains a code showcase for structural variant (SV) analysis experiments. It integrates ClinVar SV annotations, CADD-SV scores, hosted Evo2 40B scoring, local Evo2 7B embeddings, and AlphaGenome prediction examples.

Repository Layout

• analyses/clinvar_dataset/: Scripts for preparing and building ClinVar SV labels.
• analyses/caddsv/: Input preparation, execution orchestration, and score merging for CADD-SV.
• analyses/evo2_40B_NIM/: Input preparation, scoring, and score integration for the hosted Evo2 API.
• analyses/evo2_7B_local/: Local Evo2 embedding extraction and lightweight probe training.
• analyses/alphagenome/: AlphaGenome panel definition, prediction execution, and variant effect summaries.
• utils/: Common utility helper modules shared across multiple analysis tasks.

Environment Policy

Model and tool environments are not managed in this repository. Please configure each environment following the instructions of the upstream projects you intend to run:

• CADD-SV: Follow the setup guide on the CADD-SV GitHub repository.
• Evo2 40B (Hosted): Follow the instructions on the NVIDIA NIM page.
• Evo2 7B (Local): Follow the setup guide on the Evo2 GitHub repository.
• AlphaGenome: Follow the AlphaGenome API/client documentation.