Systematic discovery of single-cell protein networks in cancer with Shusi

Abstract

Context-specific protein-protein interaction (PPI) drive heterogeneity of primary tumor, forming a formidable challenge to effective cancer therapy. However, systematically mapping and modeling these interactions at single-cell resolution across diverse cancer types remains an unmet need. Here, we present Shusi, a large language model-enhanced variational graph auto-encoder model trained on over single cells across 23 cancer types, to predict context-specific PPIs. Shusi outperforms existing state-of-the-art methods, as validated through orthogonal experimental evidence. Cancer-specific mutations are significantly enriched in Shusi-predicted networks, offering complementary insights to conventional marker gene-based approaches. Through systematic evaluations, we demonstrate strong associations between Shusi-predicted network topologies, genetic vulnerabilities, and therapeutic sensitivity. Finally, in acute myeloid leukemia (AML), a blood cancer where cell-state heterogeneity drives clinical resistance, Shusi pinpointed potential targets as actionable vulnerabilities of resistant leukemia subpopulations, as validated experimentally in primary AML. Shusi offers a deep-learning tool for implementing precision medicine based on context-specific protein network architecture.