Integrative Modeling of Gene Expression and Immune Response using Deep Learning for Cervical Cancer Treatment Stratification

Immunotherapy using PD-1 checkpoint inhibitors is good in treating cervical cancer, predicting which patients will respond remains a challenge. In this research, proposed a comprehensive deep learning framework to classify immunotherapy responders based on peripheral blood gene expression profiles. Using differentially expressed gene data from a publicly available dataset (GSE205247), extracted post-treatment minus pre-treatment expression vectors for one partial responder (PR) and three progressive disease (PD) patients. A feedforward neural network classifier is trained using Leave-One-Out Cross-Validation, achieving 100% classification accuracy. Gradient saliency and SHAP explain ability is employed to find key immune related genes influencing the model’s predictions. Dimensionality reduction by PCA, t-SNE, and UMAP confirmed clear separation between PR and PD patients in both original and latent feature spaces. An auto encoder further reduced 100 gene input profiles into a 16-dimensional latent space, preserving biological variance while enabling compact visualization. These latent representations revealed distinct clustering of the PR patient. Our integrated AI pipeline not only demonstrates robust prediction of PD-1 therapy response but also identifies interpretable molecular biomarkers. This work presents a scalable, explainable, and non-invasive approach toward personalized immunotherapy strategies in cervical cancer.