Pipette: Encoding scientific literature into an executable Skill Graph for multi-agent bioinformatics

Abstract

The cost of genomic sequencing has fallen by several orders of magnitude, yet data analysis remains a bottleneck concentrated among researchers with specialized computational expertise. While Large Language Models can generate bioinformatics code, they frequently produce incoherent multi-step workflows due to the absence of domain-specific analytical constraints. Here, we present Pipette, a multi-agent AI framework that orchestrates end-to-end bioinformatics workflows through natural language interaction, guided by a literature-derived Skill Graph. This directed, edge-weighted knowledge graph, extracted from over 20,000 peer-reviewed publications, constrains workflow generation to biologically valid analytical transitions, preventing incomplete or incoherent workflows. We benchmarked Pipette across four biological domains using published datasets: single-cell RNA-seq analysis of peripheral blood mononuclear cells and a human pancreas atlas, bulk RNA-seq differential expression in rice under environmental stress, and two structure-based computational drug design workflows. In ablation against two LLMs operating without Skill Graph constraints, Pipette matched or exceeded both baselines across all quantitative metrics while uniquely completing multi-step cross-domain transitions. We further evaluated Pipette on a clinical genomics task, where it executed an ACMG/AMP-compliant variant classification on a reference human genome. In all cases, Pipette recapitulated established biological and clinical findings while generating a fully reproducible, machine-readable provenance record. By reducing the computational expertise required to execute standard genomic analyses, Pipette lowers the barrier between sequencing data and biological insight for bench scientists. Pipette is available at https://pipette.bio.