Heterogeneity-optimized method for predicting immune checkpoint blockade response

Interpatient tumor heterogeneity manifests as multimodal distributions across genomic, transcriptomic, and microenvironmental profiles. This fundamentally violates the unimodal assumption of conventional machine learning models, impairing immune checkpoint blockade (ICB) response prediction. To resolve this limitation, we propose a heterogeneity-optimized framework that applies K-means clustering to stratify patients into biologically distinct hot-tumor and cold-tumor subgroups, demonstrating superiority over hierarchical/DBSCAN clustering. Subsequently, heterogeneity-optimized predictive models-a support vector machine for hot-tumor subtypes and a random forest for cold-tumor subtypes-were developed utilizing seven heterogeneity-associated biomarkers to circumvent unimodal constraints. The proposed model significantly enhances ICB response prediction in melanoma, NSCLC, other cancer types, and pan-cancer datasets, achieving a mean accuracy gain of at least 1.24% compared to 11 baseline methods. This performance improvement was consistently validated in an independent external cohort. Furthermore, the approach enables biologically interpretable precision immunotherapy by explicitly modeling multimodal heterogeneity.