Virtual Screening on Cellular Systems

Caleb Ellington, Sohan Addagudi, Jiaqi Wang, Ben Lengerich, Eric P. Xing

Keywords: virtual screening, drug repurposing, virtual cells, perturbation modeling

TL;DR: We develop (i) two new benchmarks for evaluating the clinical relevance of virtual screening methods, such as virtual cells, and (ii) a framework for virtual screening against entire cellular systems.

Abstract:

Virtual screening has traditionally focused on molecular targets, often failing to anticipate the complex, system-level failures that arise during clinical trials.
To address this, we propose (i) two new benchmarks for evaluating the clinical relevance of virtual screening methods, such as virtual cells, and (ii) a framework for virtual screening against entire cellular systems.
Our framework uses contextualized modeling, a multi-task learning approach for inferring context-specific network models, to infer perturbation-specific coexpression networks from large-scale screening datasets, enabling accurate prediction of network restructuring under diverse cellular and therapeutic contexts.
We demonstrate that context-adaptive models outperform even observed expression profiles for predicting disease and drug mechanisms, suggesting low-cost improvements to common virtual cell objectives.
At test-time, contextualized networks generate accurate models of gene network reorganization on-demand for completely unseen cell types and therapies.
Across multiple independent runs, networks provide a standard, cohesive, and constrained latent space to compare therapeutic effects from different perturbation modalities (knockout, overexpression, small molecule).
Comparing perturbations in terms of cell-level effects leads to a principled approach to drug repurposing, safety profiling, and interpreting mechanism of action.
Rethinking virtual cell benchmarks to target clinical relevance and drug repurposing opens a path to hill-climbing on prelinical screening.

Submission Number: 92