Combinatorial prediction of therapeutic perturbations using causally-inspired neural networks

Abstract

Phenotype-driven drug discovery, as an emerging alternative to target-driven strategies, identifies compounds that counteract the overall effects of diseases by analyzing phenotypic signatures. Our study introduces a novel approach to this field, aiming to expand the search space for new therapeutic agents. We introduce PDGrapher, a causally-inspired graph neural network model designed to predict arbitrary perturbagens – a set of therapeutic targets – capable of reversing disease effects. Unlike current methods, which are limited by their reliance on predefined compound libraries, PDGrapheremploys a novel combinatorial prediction framework to widen the search scope. PDGrapherhas demonstrated significant improvements in predicting effective perturbagens, as evidenced by our evaluation across four datasets of genetic and chemical perturbations. PDGraphersuccessfully predicted effective perturbagens in up to 10% additional test samples and ranked known therapeutic targets up to 35% higher than competing methods. A key innovation of PDGrapheris its direct prediction capability, which contrasts with the indirect, computationally-intensive models traditionally used in phenotype-driven drug discovery that only predict changes in phenotype that occur as a result of a perturbation. The direct approach enables PDGrapherto train up to 30 times faster, representing a significant leap in efficiency. Our results suggest that PDGraphercan advance phenotype-driven drug discovery, offering a fast and comprehensive approach to identifying therapeutically useful perturbations.

scPRINT: pre-training on 50 million cells allows robust gene network predictions

“…Understanding the cellular mechanism is considered a milestone in biology, allowing us to predict cell behavior and the impact of drugs and gene knock-outs [1][2][3] . A cell is regulated by a complex interplay of myriads of macromolecules that define its state.…”