ICIsc: A Deep Learning Framework for Predicting Immune Checkpoint Inhibitor Response by Integrating scRNA-Seq and Protein Language Models

Immune checkpoint inhibitors (ICIs) targeting PD-1/PD-L1 and CTLA-4 are widely used in the treatment of several cancers and have significantly improved survival outcomes in responsive patients. However, a substantial proportion of patients fail to benefit from these therapies, underscoring the urgent need for accurate prediction of ICI response. We propose a deep learning framework, ICIsc, to accurately predict ICI response by integrating single-cell RNA sequencing (scRNA-seq) data with protein large language models. Specifically, patient representations are constructed using transcriptomic profiles and immune-related gene set scores as latent embedding features, while drug representations are derived from amino acid sequences of ICI encoded by the Evolutionary Scale Modeling 2 (ESM2). For bulk data, ICIsc employs a bilinear attention module to fuse patient and drug embeddings for response prediction. For scRNA-seq data, ICIsc infers cell-cell interactions using a single-sample network (SSN) approach and applies GATv2 to model immune microenvironment heterogeneity at the single-cell level. Benchmark evaluations and independent validation demonstrate that ICIsc consistently outperforms baseline models and exhibits robust generalization performance. SHAP-based interpretability analysis further identifies key genes (e.g., <i>GAPDH</i>) associated with immunotherapy response and patient prognosis. Overall, ICIsc provides an accurate and interpretable framework for predicting immunotherapy outcomes and elucidating underlying mechanisms.