A Multi-modal LLM-Knowledge Fusion Framework for Predicting Single-cell Genetic Perturbation Effects

Abstract

Understanding cellular responses to genetic perturbations is fundamental for drug discovery, yet experimental approaches face significant limitations in coverage and cost that prevent comprehensive mapping of cellular behavior. This has motivated the development of virtual cells--computational models that learn the relationship between cell state and function to predict the consequences of perturbations across diverse contexts. However, current computational methods suffer from limited accuracy in complex genetic interactions, poor biological interpretability, and inadequate generalization to unseen genes, severely constraining virtual cell capabilities. We present scPert, a multi-modal framework based on Transformer architecture that integrates large language model embeddings with structured biological knowledge to predict single-cell transcriptomic responses to genetic perturbations. Through hierarchical fusion of knowledge graph representations, contextual embeddings from foundation models, and gene-specific encodings, scPert achieves significant performance improvements in both single-gene and combinatorial perturbations over existing methods. In cancer-relevant applications, scPert demonstrates the capability to reveal p53 pathway dynamics and immune checkpoint regulatory mechanisms. Systematic evaluation on 42 cancer dependency genes demonstrates scPerts ability to identify critical potential therapeutic targets. Our framework establishes a powerful computational foundation for virtual cell construction and accelerates drug target discovery.