EPCO-60. FOUNDATIONAL MODEL FOR VIRTUAL EXPERIMENTAL FRAMEWORK (VEF) FOR BIOMEDICAL APPLICATIONS PREDICTS MASTER REGULATORS IN HUMAN GLIOBLASTOMA

Abstract BACKGROUND Mapping causal effects of genetic perturbations is hampered by the combinatorial complexity of multigene interactions and the limits of wet-lab screens. Current deep-learning methods treat perturbations as binary, ignore dosage effects, lack support for high-order multigene edits. To that end, we developed VEF, a foundational model that predicts transcriptomic, regulatory outcomes under arbitrary continuous, bidirectional gene perturbations, including unseen multigene combinations and applied it to identify master regulators (MR) in glioblastoma. METHODS VEF integrates >200K bulk RNA-seq samples, ~70 Perturb-seq single-cell screens, 20 million unperturbed scRNA-seq profiles. It combines a variational autoencoder for shared biology, graph neural network for gene interactions, transformer for transcription factor (TF) states, gated cross-attention for TF→gene regulation. By injecting motif and pathway priors via large language model-derived gene embeddings and encoding continuous, bidirectional perturbation magnitudes, VEF trains on both paired and unpaired cells for robust generalization. RESULTS VEF outperforms prior models on a standard Norman-et-al Perturb-seq benchmark––achieving a Pearson correlation r=0.71 between predicted and observed differential gene expression, compared to <0.60 with current best model, a >15% gain. Ablation studies confirm that each key innovation contribute substantial accuracy improvements. Applying VEF to a glioblastoma→dendritic cell (DC) RNA-seq dataset identified four transcription factors (Pu.1/Irf8/Batf3/Id2) that convert glioblastoma cells into functional DCs in situ including cross-presenting tumor antigens, inducing local immunization, and extending survival. Finally, VEF analysis of 29 patient-derived glioma stem-like (GSC) lines, six neural/astrocyte lines, and 167 TCGA GBM samples identified nine MRs of the GBM state. Enforcing this program in healthy astrocytes generated GSC-like cells, while its depletion blunted GSCs’ in-vivo tumorigenesis, validating VEF’s predictive power. CONCLUSIONS VEF is a scalable platform that accurately predicts genetic-edit effects, prioritizes wet-lab experiments, accelerates causal cell-state discovery, and speeds drug-target identification demonstrated by its pinpointing glioblastoma MRs and iDC-conversion drivers.