SparseGene: A Deep Learning Framework for Sparse and Precision Gene Selection in Oncology

Abstract

In precision oncology, paths of the decision genes of interest are especially crucial to the advancement of stable diagnosis, individual therapeutic interventions. The problem of interpretation or overfitting is one of the inadequacies of the traditional method of LASSO regression, and also the DL model. We redress this through the paradigm of unsupervised representation learning through sparse supervised classification by a newly designed deep learning strategy, SparseGene, which investigates biologically meaningful possibilities on a subset of the genes in high-dimensional RNA-Seq data. SparseGene fits a deep autoencoder to the dimension reduction and a layer constrained in sparsity of classification, and detects a small, precise predictive set of genes. On a non-small cell lung cancer (NSCLC) dataset (GSE89843), SparseGene is able to distinguish between predictive models at 94% accuracy, 0.93 F1-score, with only 28 key genes selected, many of which correspond to other known predictive models in NSCLC that belong to other well-discovered oncogenic circuits such as PI3K-Akt and p53 signaling, as well as EGFR resistance. The robustness of the applicability and translatability to clinical practice, as well as its understandability, is proven by the further pathway enrichment and ROC analysis. The SparseGene provides a very good path to scalable, interpretable, and biologically-aware marker discovery of cancer genomics.