RegVar-Agent: An Intelligent Regulatory Variant Screening System Combining AlphaGenome and DeepSeek

RegVar-Agent is an AI-assisted tool for genomics research that integrates DeepMind's AlphaGenome deep learning model and DeepSeek v4 Pro reasoning capabilities. It can automatically evaluate multi-omics evidence of non-coding region variants and generate experimental validation plans. This project is developed and maintained by raktim-mondol, open-sourced on GitHub (link: https://github.com/raktim-mondol/regvar-agent), and released on June 2, 2026. Its core goal is to address the long-standing challenge in genomics of screening functional regulatory variants from massive genetic variations.

Genome-wide association studies (GWAS) have identified a large number of disease-related genetic loci, but most of them are located in non-coding regions. Interpreting these variants requires integrating multi-omics data (ATAC-seq, RNA-seq, ChIP-seq histone modifications, Hi-C/pcHi-C, etc.). Traditional manual database lookup methods are time-consuming and prone to missing key evidence. RegVar-Agent automates this process through AI.

RegVar-Agent adopts a layered architecture:

1. AlphaGenome Scoring Engine: Encapsulates DeepMind's AlphaGenome API, adding enhanced features such as intelligent caching, exponential backoff retries, and result standardization into DataFrame;
2. DeepSeek Reasoning Agent: Responsible for tool scheduling (calling list_supported_assays/score_regulatory_variant), multi-omics evidence integration, variant priority ranking based on quantile_score, and generating wet experimental validation plans;
3. MCP Server Integration: Supports direct tool calls from IDEs like Claude Code and Cursor, integrating into the development workflow.

Taking non-coding variants near regulatory elements of prostate cancer stromal fibroblasts as an example, the input file candidate_variants.tsv contains information such as chromosome, position, reference/alternative bases, etc. The system output includes:

• Multi-omics scores for each variant;
• Priority list sorted by effect size;
• Regulatory mechanism hypotheses (e.g., disrupting enhancer-promoter looping);
• Validation suggestions (e.g., CRISPR interference to verify the impact of chr8:127401060 on MYC expression).

1. Tool Boundary Design: Strictly distinguishes between deterministic tools (e.g., score_variant) and reasoning agents, improving testability and maintainability;
2. Cross-Model Compatibility: Tool definitions follow OpenAI/Anthropic function calling specifications, supporting switching between multiple model backends;
3. Agent-Free Direct Call: Users can directly obtain variant scores via AlphaGenomeClient (Python example code provided).

RegVar-Agent uses large models as tools for evidence integration and hypothesis generation, accelerating the cycle from GWAS association signals to functional validation. It is suitable for GWAS follow-up validation, disease mechanism research, and drug target discovery. The project is open-source, and the community is encouraged to contribute support for more disease scenarios and detection types.