整合新抗原免疫原性与肿瘤克隆性以预测免疫治疗应答

Background: Cancer immunotherapy has transformed treatment, yet response rates remain suboptimal. Immunogenic neoantigens-tumor-specific peptides eliciting T-cell responses-represent promising biomarkers, but current predictors have critical limitations: they focus predominantly on MHC-I while neglecting MHC-I/II coordination, and rely on tumor mutation burden (TMB), which lacks immunogenicity specificity and ignores tumor heterogeneity. We developed NeoPrecis, a computational framework integrating immunogenicity prediction with tumor subclonal architecture to improve immunotherapy response prediction. Methods: NeoPrecis comprises two modules capturing mutation-centric and tumor-centric immunogenic contexts. NeoPrecis-Immuno models wild-type to mutant peptide distance to estimate T-cell recognition likelihood, incorporating amino acid embeddings, MHC-binding motifs, positional factors, and peptide sequences. The model was pre-trained on TCR-binding data for peptide-TCR cross-reactivity discrimination, then fine-tuned on T-cell assay data. NeoPrecis-Landscape integrates MHC-I and MHC-II immunogenicity predictions with PyClone-inferred subclonal structure. For each subclone, immunogenicity is computed as the product of its MHC-I and MHC-II scores. Tumor-level immunogenicity is then calculated as the weighted average of all subclonal scores, with weights determined by subclone prevalence. Results: NeoPrecis-Immuno outperformed PRIME, ICERFIRE, and DeepNeo on an independent gastrointestinal cancer dataset with validated CD4+/CD8+ T-cell assays. Its interpretable architecture quantifies allele-specific contributions via allele benefit scores, which showed significant prognostic associations in melanoma (p=0.04) and NSCLC (p=0.01) independent of specific mutations. Across five melanoma and three NSCLC cohorts, NeoPrecis-Landscape outperformed TMB in stratifying ICI responders, particularly in melanoma and heterogeneous NSCLCs. Homogeneous, heavily immunoedited NSCLCs (predominantly smoker tumors) showed reduced neoantigen-based predictive power. Conclusion: NeoPrecis provides an interpretable framework for neoantigen immunogenicity assessment. By integrating tumor subclonal structure, it outperforms TMB in predicting ICI response, especially in melanoma and heterogeneous NSCLCs with low immunoediting. Poor performance in immunoedited tumors suggests immune evasion mechanisms may dominate ICI response in these contexts.