Deep learning-based pan-cancer classification model reveals cancer-specific gene expression signatures

Abstract

The identification of cancer-specific biomarkers and therapeutic targets is one of the primary goals of cancer genomics. Thousands of cancer genomes, exomes, and transcriptomes have been sequenced to date. In this study, we conducted a pan-cancer analysis of transcriptome datasets from 37 cancer types provided by The Cancer Genome Atlas (TCGA) in an effort to identify cancer-specific gene expression signatures. We employed deep neural networks to train a model on the transcriptome profile datasets for all cancer types. The model was validated, and its predictive accuracy was determined using an independent dataset, achieving > 97% prediction accuracy across cancer types. This strongly suggests that there are distinct gene expression signatures associated with various cancer types. We interpreted the model using SHapley Additive exPlanations (SHAP) to identify specific gene signatures that significantly contributed to the classification of cancer types. In addition to known biomarkers, we identified several novel biomarkers in different cancer types. These cancer-specific gene signatures are valuable candidates for future studies of their potential utility as cancer biomarkers and putative therapeutic targets.

Deep Learning-Based Pan-Cancer Classification Model Reveals Tissue-of-Origin Specific Gene Expression Signatures

“…There are some limitations associated with our pan-cancer analysis-based assessment of cancer tissue-of-origin specific gene expression signatures [ 31 ]. For example, we only considered genes that were expressed at sufficiently high levels (≥5 FPKM) in at least 50% of samples within a cancer type.…”

Machine learning for pan-cancer classification based on RNA sequencing data

“…The prediction accuracy varies by tumor type, with some tumor types being more frequently mispredicted. Patterns of more frequent misclassifications among groups of cancers arising from the same organ (e.g., kidney renal clear cell carcinoma, kidney renal papillary cell carcinoma and kidney chromophobe carcinoma, or lung adenocarcinoma and lung squamous cell carcinoma), and/or among cancers represented by a small number of samples in the training set (e.g., cholangiocarcinoma, which is frequently predicted as liver hepatocellular carcinoma and vice versa), as noted in multiple studies (Bagge et al, 2018;Lyu and Haque, 2018;Bavafaye Haghighi et al, 2019;De Guia et al, 2019;Grewal et al, 2019;Zhao et al, 2020;Vibert et al, 2021;Divate et al, 2022;Jones et al, 2022;Moiso et al, 2022). All of this implies that the distribution of different cancer types in the training set is one of the key factors contributing to the prediction accuracy of the model.…”

“…The majority of studies were based on deep learning, using neural networks of different architectures (Lyu and Haque, 2018;Azarkhalili et al, 2019;De Guia et al, 2019;He et al, 2020b;Mostavi et al, 2020;Zhao et al, 2020;Vibert et al, 2021;Divate et al, 2022;Hong et al, 2022;Jones et al, 2022;Moiso et al, 2022). Several studies utilized ensemble learning methods, in which the final prediction is a combination of multiple predictors (Grewal et al, 2019;He et al, 2020a;Ramroach et al, 2020;Chen et al, 2021;Liu et al, 2021).…”

“…The remaining studies used independent test sets consisting of 5-18 cancer types and showed a reduction in prediction accuracy ranging from 5.8% to 26.47% compared to cross-validation (with a mean reduction in accuracy of 13.32%). Out of these, two studies used metastatic samples (Bavafaye Haghighi et al, 2019;Zhao et al, 2020), two used a mixture of primary and metastatic samples (Bagge et al, 2018;Divate et al, 2022) and one did not report the exact source of the independent test set (Chen et al, 2021).…”

“…Interestingly, when testing the accuracy of the same model on test sets composed of both primary and metastatic samples, metastatic samples showed a lower prediction accuracy. For example, Divate et al (2022) reported 88.10% accuracy for metastatic samples compared to 92.13% for primary samples. Bagge et al (2018) found 53.84% accuracy for metastatic samples compared to 96.67% for patient-derived xenografts of primary cancer and 100% for primary cancer.…”