An Analysis Graph for Statistical Genetics Agents

Stephen Dorn, Jason Mohoney, Tim Kraska, Samuel Madden, Noah Zaitlen, Margaux Hujoel

Submission Type: Project Talk Proposal

Keywords: Genetics, Knowledge Graph, Agents

TL;DR: We created a knowledge graph used by coding agents for statistical genetics analysis tasks and provenance tracking.

Abstract:

Statistical genetics workflows have grown more complex in recent years, with new methods,
software, genome annotations, and publicly available GWAS summary statistics
outpacing what any one researcher can keep track of.
The problem is sharpest for clinical investigators, who often have deep
disease expertise but limited bandwidth to track an expanding post-GWAS
toolkit. Coding agents can run the software, but command execution
alone does not produce a reviewable, reusable analysis. We
propose a graph-backed analysis agent for statistical genetics.
Skills describe each analysis task; the agent executes them and
uses the graph to pick methods, reconcile mismatched inputs, and
route results between stages. The graph also records typed claims
about every artifact, command, reference, and result, including
dataset identity, genome build, linkage-disequilibrium (LD)
reference panel, software version, quality-control summaries, and
unresolved expert-review decisions. In a preliminary case study, the agent
ran a nine-stage post-GWAS pipeline on a public inflammatory
bowel disease (IBD) GWAS, finishing in 94.7 minutes on a single
laptop CPU. The agent recovered the canonical IBD genetic architecture and left a queryable, reviewable record of every method, reference, and decision.

Submission Number: 12