Deep Genomic Signature for early metastasis prediction in prostate cancer

Motivation: For Prostate Cancer (PCa) patients, timing and intensity of the therapy is adjusted based on their prognosis. This can be predicted from clinical/pathological information and, recently, gene expression signatures. One major challenge in developing such signatures is that all of them are based on cohorts which have limited number of patients with complete clinical outcomes (labelled), especially for slow progressing cancers such as PCa. This poses a challenge to the model development in conjunction with high dimensionality of the transcriptomic data. Results: In this study we aim to exploit the previously untapped potential of a large cohort (n=15,136) with genomic data but no clinical outcome (unlabelled), to improve the performance of the genomic classifiers for predicting PCa metastasis. We propose, Deep Genomic Signature (DGS), based on Denoising Auto-Encoder (DAE) for feature extraction and selection. In order to capture information from the unlabelled and labelled data we train two DAEs separately and apply transfer learning to bridge the gap between them. We show that DGS captures information from these cohorts that can be utilized to build a logistic regression model to predict metastasis. Results on five validation cohorts indicate that this classifier, which is based on high weight genes in the DAEs, outperforms state-of-the-art signatures for metastatic PCa in terms of prediction accuracy. Survival analysis demonstrate the clinical utility of our signature which adds information to the well-established clinical factors and state-of-the-art signatures. Furthermore, pathway analysis reveals that the signature discovered by our DGS captures the hallmarks of PCa metastasis.