A deep learning approach for uncovering lung cancer immunome patterns

Abstract Tumor immune cell infiltration is a well known factor related to survival of cancer patients. This has led to deconvolution approaches that can quantify immune cell proportions for each individual. What is missing, is an approach for modeling joint patterns of different immune cell types. We adapt a deep learning approach, deep Boltzmann machines (DBMs), for modeling immune cell gene expression patterns in lung adenocarcinoma. Specifically, a partially partitioned training approach for dealing with a relatively large number of genes. We also propose a sampling-based approach that smooths the original data according to a trained DBM and can be used for visualization and clustering. The identified clusters can subsequently be judged with respect to association with clinical characteristics, such as tumor stage, providing an external criterion for selecting DBM network architecture and tuning parameters for training. We show that the hidden nodes of the trained networks cannot only be linked to clinical characteristics but also to specific genes, which are the visible nodes of the network. We find that hidden nodes that are linked to tumor stage and survival represent expression of T-cell and mast cell genes among others, probably reflecting specific immune cell infiltration patterns. Thus, DBMs, trained and selected by the proposed approach, might provide a useful tool for extracting immune cell gene expression patterns. In the case of lung adenocarcinomas, these patterns are linked to survival as well as other patient characteristics, which could be useful for uncovering the underlying biology.