AI-Powered Computational Models for Human-Biology-Based Translational Research - SUMMARY Animal models and organoid systems are indispensable for preclinical research, yet their translational accuracy remains limited, with fewer than 10% of therapies effective in mice ultimately succeeding in humans due to fundamental species-specific genetic, molecular, and cellular differences. Organoid systems offer an alternative, but their cellular composition and regulatory programs remain immature relative to human tissues, and a central challenge is the absence of standardized computational frameworks to benchmark the fidelity of these models. At the same time, virtual cell technologies are emerging, but current implementations are fragmented, rely on single data types, lack mechanistic grounding, and often produce inconsistent perturbation predictions, all of which prevent reliable translation from preclinical studies to effective human therapies. This R35 program will establish innovative AI-driven strategies to overcome these challenges through two integrated directions. Direction 1 will develop a fidelity framework to rigorously benchmark organoid and animal models against human references by applying variance decomposition and batch reasoning to disentangle biological signals from technical noise, using heterogeneous graph transformers with contrastive learning to align conserved programs while preserving species-specific divergence. The outcome will be fidelity scorecards that quantify human relevance across regulatory programs, cell–cell communication, and tissue architecture, thereby highlighting deficiencies and providing blueprints for organoid optimization. Direction 2 will construct ensemble virtual cells to predict human-like cellular responses under perturbations by integrating existing gene-expression, protein, and regulatory models into a universal representation across modalities and scales, using hypergraph transformers to harmonize molecular, cellular, and multicellular features, and applying dynamic operator learning to simulate perturbation trajectories with uncertainty while incorporating interactome and metabolome knowledge to capture species-specific drug adaptations. Together, these directions create a closed loop where fidelity benchmarks guide virtual cell grounding and virtual cells provide actionable translational predictions. Applications include accurate synotype characterization in lung disease, where organoids must capture fibroblast heterogeneity and senescence, and cancer drug target prioritization, where virtual cells identify conserved druggable modules while flagging human-specific liabilities. By providing reproducible metrics, interpretable predictions, and uncertainty-aware outputs, this program directly addresses reproducibility and translation in preclinical science, building on my lab’s expertise in graph neural networks, heterogeneous graph transformers, and drug response prediction. Ambitious yet feasible, supported by strong preliminary tools, collaborative datasets, and validated workflows, the proposed research will redefine how preclinical data are evaluated, benchmarked, and applied, ultimately accelerating organoid development, improving drug pipelines, and reducing costly translational failures to improve human health.
