Widespread redundancy in -omics profiles of cancer mutation states

Abstract

In studies of cellular function in cancer, researchers are increasingly able to choose from many -omics assays as functional readouts. Choosing the correct readout for a given study can be difficult, and which layer of cellular function is most suitable to capture the relevant signal may be unclear. In this study, we consider prediction of cancer mutation status (presence or absence) from functional -omics data as a representative problem. Since functional signatures of cancer mutation have been identified across many data types, this problem presents an opportunity to quantify and compare the ability of different -omics readouts to capture signals of dysregulation in cancer. The TCGA Pan-Cancer Atlas contains genetic alteration data including somatic mutations and copy number variants (CNVs), as well as several -omics data types. From TCGA, we focus on RNA sequencing, DNA methylation arrays, reverse phase protein arrays (RPPA), microRNA, and somatic mutational signatures as -omics readouts.Across a collection of genes recurrently mutated in cancer, RNA sequencing and DNA methylation were the most effective predictors of mutation state. Surprisingly, we found that for most mutated genes, they were approximately equally effective predictors. Performance was more variable between mutations than it was between data types for the same mutation, and there was little difference between the top data types for the majority of genes. We also found that combining data types into a single multi-omics model provided little or no improvement in predictive ability over the best individual data type. Based on our results, for the design of studies focused on the functional outcomes of cancer mutations, we recommend focusing on gene expression or DNA methylation as first-line readouts.

wenda_gpu: fast domain adaptation for genomic data

“…One example use case is building classifiers on transcriptomic data to identify mutations in common cancer driver genes, which has been shown to be possible in a pan-cancer context ( Knijnenburg et al , 2018 ; Way et al , 2018 ). Mutation classifiers trained on transcriptomic data from one cancer type can even be predictive of mutation status in other cancer types ( Crawford et al , 2022 ). We postulated that wenda would allow prediction models trained on one cancer type to perform better on another cancer type, opening the possibility of these models being used on rare cancers for which sample sizes are limited.…”

Deep learning can predict multi-omic biomarkers from routine pathology images: A systematic large-scale study

“…Our study showed that genetic alterations are detectable at different levels of the omics landscape for many genes. However, the benefit of integrating multi-omic data for predicting mutations is still not fully understood and further analysis is required to explore the impact of combining multiple omics on predictability 58 . The questions around the specific mechanism of biomarker detectability, such as the predictive pattern and how it is conserved across the different landscapes of the molecular landscape, merit more investigation.…”

Cross-platform normalization enables machine learning model training on microarray and RNA-seq data simultaneously

“…In addition, both BRCA and GBM have well-defined molecular subtypes that are suitable for use as labels/classes for supervised machine learning approaches we describe below. BRCA and GBM also show recurrent TP53 and PIK3CA gene mutations, allowing for the prediction of mutation status based on gene expression profiles [32][33][34] . For BRCA (520 pairs of matched samples), we used log 2 -transformed, lowess normalized Agilent 244 K microarray data 29 and RSEM (RNA-seq by Expectation Maximization) gene-level count RNA-seq data 35 .…”