Transcriptome-wide prediction of prostate cancer gene expression from histopathology images using co-expression-based convolutional neural networks

• DOI: 10.1093/bioinformatics/btac343
• Publication Date: 2022-05-20T23:36:17Z
• Authors: Philippe Weitz; Yinxi Wang; Kimmo Kartasalo; Lars Egevad; Johan Lindberg; Henrik Grönberg; Martin Eklund; Mattias Rantalainen
• Keywords: Histopathology; Expression (computer science)

Molecular phenotyping by gene expression profiling is central in contemporary cancer research and molecular diagnostics but remains resource intense to implement. Changes occurring tumours cause morphological changes tissue, which can be observed on the microscopic level. The relationship between patterns some of phenotypes exploited predict from routine haematoxylin eosin-stained whole slide images (WSIs) using convolutional neural networks (CNNs). In this study, we propose a new, computationally efficient approach model relationships morphology expression.We conducted first transcriptome-wide analysis prostate cancer, CNNs bulk RNA-sequencing estimates WSIs for 370 patients TCGA PRAD study. Out 15 586 protein coding transcripts, 6618 had predicted significantly associated with RNA-seq (FDR-adjusted P-value <1×10-4) cross-validation 5419 (81.9%) these associations were subsequently validated held-out test set. We furthermore prognostic cell-cycle progression score directly WSIs. These findings suggest that computer vision models offer an inexpensive scalable solution prediction WSIs, providing opportunity cost-effective large-scale studies diagnostics.A self-contained example available http://github.com/phiwei/prostate_coexpression. Model predictions metrics are doi.org/10.5281/zenodo.4739097.Supplementary data at Bioinformatics online.