Biologically explainable multi-omics feature demonstrates greater learning potential by identifying tissue of origin, stages, and subtypes for pan-cancer classification.

Cancer is a complex disease characterized by uncontrolled cell growth, which can invade surrounding tissues and spread to distant organs. Most of the conventional methods of diagnosis fails to identify the primary organ when cancer spreads to other organs, thereby adding another level of complexity for cancer detection. It is also critical to determine the stages and subtypes of cancer, and develop a clinically applicable model for precision therapy. With a dataset of 7632 samples from 30 different cancer originating from distinct organs, we have constructed a deep learning framework to solve all of these challenges. We have applied a hybrid feature selection method to identify cancer-associated features in the transcriptome, methylome, and microRNA datasets. This was achieved by combining both gene set enrichment analysis and Cox regression analysis to build an explainable AI model. We performed the early integration using an autoencoder to embed the cancer-associated multi-omics data into a lower-dimensional space; an ANN classifier was constructed using the latent features. In addition to correctly classifying 30 different cancer types by their tissue of origin, our framework also identifies individual subtypes and stages of cancer with an accuracy ranging from 87.31% to 94.0% and 83.33% to 93.64%, respectively. The current model demonstrates higher accuracy even when tested with external datasets, and shows better stability and accuracy in making predictions compared to the existing models. This approach offers explainable strategies for selecting features in AI-based prediction of tumor types for personalized therapy, aiding clinicians in making real-time treatment decisions.