通过免疫原性张力图谱量化肿瘤免疫逃逸

Predicting response to immunotherapy remains challenging because biomarkers such as TMB, PD-L1 expression, and bulk immune infiltration offer only static snapshots of tumor biology and fail to capture the cumulative influence of immune pressure over time. Immune pressure leaves discernible footprints on tumor evolution, and characterizing these signatures may help distinguish tumors shaped by immune editing from those that have escaped immune surveillance. Comparing observed tumor genomes with patient-specific neutral expectations provides a direct framework for revealing such evolutionary deviations. Using TCGA colorectal cancers, we generated mutational-signature-matched neutral simulations and quantified how far each tumor diverged across multiple evolutionary metrics. Observed coding mutations showed substantially greater dispersion than neutral simulations. Although the cohort appeared near-neutral overall, tumors showing strong deviation in one metric consistently deviated across others, revealing a coherent subset under stronger selective pressure. These strongly deviated tumors aligned with well-established hypermutated categories-including POLE-mutated hypermutated tumors and MSI-H/indel-rich tumors-and displayed steep TMB gradients (44.8 vs 5.3 mut/Mb across quartiles). They also exhibited stronger immune-associated features, including significant higher CD8-related infiltration and elevated CD103 ratios in extreme cases, suggesting that immune activity contributes to the observed evolutionary divergence. To evaluate whether broader mutation patterns capture deeper evolutionary structure, we embedded observed and simulated mutation catalogs into a genomic foundation model. The resulting latent space clearly distinguished strongly deviated tumors from near-neutral cases, and also separated molecular subtypes and each sample's mutational-signature composition. These findings suggest that zero-shot sequence representations capture contextual patterns linked to immune-driven tumor evolution. Because these embeddings encode mutation context at a finer resolution than classical summaries, they can recover evolutionary trajectories and underlying mechanisms that remain obscured in low-dimensional metrics. Together, these results show that immune-linked evolutionary divergence is detectable even when cohort-level averages appear neutral. Tumors that diverge most strongly from neutral expectation correspond to hypermutated biology and increased CD8-driven immune activity. Evolution-based approaches-especially when combined with foundation-model embeddings-offer a scalable framework for quantifying historical immune pressure and refining immunologic stratification beyond conventional biomarkers.