Multi-scale and multi-modal computational modeling of cell state plasticity and drug response

The National Institute of General Medical Sciences awarded Oregon Health & Science University $390,000 on May 1, 2026, under the Biomedical Research and Research Training program (CFDA 93.859) to develop computational approaches for predicting drug response in heterogeneous patient cell populations at single-cell resolution.

The funded research addresses the translational challenge of predicting how primary patient samples respond to drug perturbations by integrating single-cell high-throughput measurements with machine learning and causal modeling. The project proposes a principled computational approach combining cell state prediction, causal regularization, and distributional alignment to forecast responses to previously unseen mono- and combination therapies in new patients. The work aims to generalize across different drugs, patient cohorts, and measurement scales—from bulk to single-cell assays—while incorporating paired perturbed and non-perturbed cellular measurements to model cell state transitions and population dynamics.

The award is a Project Grant, a standard NIGMS assistance mechanism. Performance occurs in Portland, Oregon, with a period of performance extending through March 31, 2031. The research builds on OHSU's substantial NIH grant portfolio in biomedical research and its embedded research infrastructure, including research cores that support investigator-initiated basic and translational science across multiple disease areas.

Abstract

Predicting drug perturbation effects on primary patient samples is a core translational problem critical for effective patient treatment. Single-cell high-throughput technologies enable the study of drug effects on heterogeneous cell populations, capturing patient-specific variability in disease mechanisms and treatment response. While unimodal and multiomic single-cell platforms offer unprecedented resolution, they introduce new challenges for AI/ML models. High cell input requirements and destructive measurement protocols limit patient profiling, especially in the context of combination therapies, where the number of possible drug regimens grows exponentially. Most datasets also lack paired measurements of pre- and post-treatment states, complicating efforts to model cellular heterogeneity and subpopulation dynamics under perturbation. Furthermore, clinically relevant datasets with long-term treatment effects remain scarce. There is a critical need for computational approaches that generalize across unseen drugs and patients, and that can integrate across assays, cohorts, and scales-from bulk to single-cell. We propose to address these needs through novel computational approaches that span two independent research directions. In the first, we propose a principled approach that leverages emerging technologies to predict drug response at a single cell resolution, including of previously unseen mono- and combination therapies and in new patients. Our model combines cell state prediction, causal regularization, and distributional alignment that aims to blend data-driven learning with known mechanisms. In contrast to current methods, the proposed approach will be grounded with a reference dataset that will (i) provide paired perturbed and non-perturbed cellular measurements to precisely establish cell state transitions including cell death and persistent states, and (II) will measure long term drug treatment effects more closely resembling clinical regimens. In the second direction, we propose a unifying transfer learning framework for drug response prediction, that enables the integration of data across multiple patient cohorts, modalities, and scale. This proposal aims to uncover drug response mechanisms at single-cell resolution to advance the development of combination therapies that overcome resistance. We will deliver a validated computational framework that models cell state transitions under perturbation, using novel AI-based methods and a time- resolved EX vivo reference dataset. Our approach maximizes the value of limited patient material and enables scalable, platform-independent analysis across diseases and treatments. Importantly, the framework supports federated learning across institutions, allowing model training on decentralized data while preserving patient privacy. This work is in strong support of the mission of nigms for advances in disease diagnosis, treatment, and prevention.