Interpretable deep learning quantifies the impact of genomic variations on CD8+ T cell infiltration

Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, but challenges remain in identifying molecular determinants of clinical response. To address this, we developed a Transformer-based model, ImmRegVar, to predict CD8+ T cell infiltration using somatic genomic variations from 7972 solid tumor samples. ImmRegVar outperforms traditional regression methods in predictive accuracy and model performance. We introduce an innovative explanatory factor that combines Transformer attention weights with feature ablation techniques, allowing precise quantification of each gene mutation’s contribution. This factor offers an interpretable metric for understanding how specific mutations affect CD8+ T cell infiltration, shedding light on immune regulatory mechanisms. Validation through random permutation tests confirms its robustness and stability in high-dimensional sparse data. Our analysis also aligns with known immune regulatory genes, supporting the biological credibility of ImmRegVar. This research advances the understanding of key factors in immunotherapy and opens new pathways for exploring immune regulation and clinical applications. ImmRegVar and the explanatory factor are available at //github.com/Liwen-Liberty/ImmRegVar.