DeepTI: A deep learning-based framework decoding tumor-immune interactions for precision immunotherapy in oncology

Background: Increasing evidence suggests the immunomodulatory potential of genes in oncology. But the identification of immune attributes of genes is costly and time-consuming, which leads to an urgent demand to develop a prediction model. Method: We developed a deep learning-based model to predict the immune properties of genes. This model is trained in 70% of samples and evaluated in 30% of samples. Furthermore, it uncovers 60 new immune-related genes. We analyzed the expression perturbation and prognostic value of these genes in gastric cancer. Finally, we validated these genes in immunotherapy-related datasets to check the predictive potential of immunotherapeutic sensitivity. Result: This model classifies genes as immune-promoted or immune-inhibited based on the human PPI network and it achieves an accuracy of 0.68 on the test set. It uncovers 60 new immune-related genes, most of which are validated in the published literature. These genes are found to be downregulated in gastric cancer and significantly associated with the immune microenvironment in gastric cancer. Analysis of immunotherapy shows that these genes can discriminate between responder and non-responder. Conclusion: This model can facilitate the identification of immune properties of genes, decoding tumor-immune interactions for precision immunotherapy in oncology.