Aarthi Venkat, PhD

Postdoctoral Fellow<br>Broad Institute of MIT and Harvard

Geometric machine learning enables multiscale perturbational reasoning

Genetic association studies have identified candidate causal genes for thousands of traits and diseases, but how affected genes co-regulate to mediate trait association through intermediate phenotypes is far from resolved. Modeling effects of genetic perturbations, including from genome-wide CRISPR screens, mouse knockout models, and, more recently, single-cell approaches such as Perturb-seq and OPS, remains a promising tool for guiding phenotypic interpretation and target discovery efforts. Despite this, there exist few approaches to systematically integrate perturbation datasets across scales and reason over complex relational paths to transfer insights to underpowered or unphenotyped entities.

To this end, we constructed a perturbational knowledge graph integrating genetic associations and molecular quantitative trait loci evidence with transcriptional, morphological, functional, and organismal perturbation readouts, producing a network consisting of 45,159 nodes and over 22 million edges. This network introduces functional gene-gene relationships defined by single-cell Perturb-seq and OPS data, which we show capture protein complex formation, protein-protein interaction, and pathway association. Additionally, we present a graph-based machine learning approach to query multiple entities and relations simultaneously, enabling multi-hop reasoning where standard approaches fail. This achieves strong performance over single-cell transcriptomic and perturbation foundation models on tasks such as disease-specific gene-gene interaction and context-aware disease-gene prioritization.