Tianhao Wang; Haoran Li; Jie Liao · 2026 · EXO
Paper
Predicting developmental trajectories and disease progression is a central goal of virtual biology. While conventional biological network models struggle with noise and computational complexity, black-box deep learning obscures mechanistic interpretability. In this Perspective, we introduce “gene programs” (GPs) as the essential intermediate layer to close this gap. Extracted from massive single-cell datasets, GPs provide relatively robust and interpretable modules that capture coherent biological functions. Supported by recent multi-modal and large-scale perturbation advances, GPs may reflect underlying cross-layer biophysical structures. We propose a framework where next-generation virtual cells encode GPs as latent variables in neural networks. By incorporating spatial multicellular contexts and multimodal GP extraction, this GP-centric framework will reduce data noise and allow direct perturbation-to-phenotype mapping, enabling high-fidelity simulations of complex tissue dynamics. Anchoring virtual models on GPs paves the way for highly personalized “Patient Digital Twins”, empowering predictive in silico clinical simulations and advancing mechanism-driven precision medicine.
Analysis
Preparing paper insights from the available abstract and paper details.
Discovery
Leonard Azamfirei; Dorin Bica; Andrei Calin Dragomir; Emoke Almasy
Dr. Kunal Mehta; Dr. Farah Ali
Rasit Dinc; Nurittin Ardic
Lorenzo Cipriano
Sandra Saade; Susanna Nordin; Johan Borg
Adil Adam; Mussaab Alshbib
Source record