FLASH-P: Turning decades of biology into accurate causal networks with AI agents

Abstract Mechanistic networks that encode causal regulatory logic can predict the effects of genetic and environmental perturbations but constructing them is a bottleneck in systems biology because the relevant knowledge lies scattered across thousands of resources, untapped for both building and validating such networks. Here we present FLASH-P, a multi-agent framework that autonomously curates this literature into perturbable, signed-directed network models for any trait–species combination in under an hour without much computational power. Twelve FLASH-P networks across seven species predicted the directional outcome of 1,088 published perturbations with a mean accuracy of 90%. This accuracy was driven by the regulatory topology FLASH-P constructs, which is why it outperformed knowledge-graph derived networks. Its merging agent combined six networks into one that preserved single-trait accuracy and recovered pleiotropic effects, and consolidated independent runs of one trait into a comprehensive, high-accuracy network. FLASH-P networks enable applications that require trait models.