Deep learning has made tremendous successes in numerous artificial intelligence applications and is unsurprisingly penetrating into various biomedical domains. High-throughput omics data in the form of molecular profile matrices, such as transcriptomes and metabolomes, have long existed as a valuable resource for facilitating diagnosis of patient statuses/stages. It is timely imperative to compare deep learning neural networks against classical machine learning methods in the setting of matrix-formed omics data in terms of classification accuracy and robustness. Using 37 high throughput omics datasets, covering transcriptomes and metabolomes, we evaluated the classification power of deep learning compared to traditional machine learning methods. Representative deep learning methods, Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN), were deployed and explored in seeking optimal architectures for the best classification performance. Together with five classical supervised classification methods (Linear Discriminant Analysis, Multinomial Logistic Regression, Naïve Bayes, Random Forest, Support Vector Machine), MLP and CNN were comparatively tested on the 37 datasets to predict disease stages or to discriminate diseased samples from normal samples. MLPs achieved the highest overall accuracy among all methods tested. More thorough analyses revealed that single hidden layer MLPs with ample hidden units outperformed deeper MLPs. Furthermore, MLP was one of the most robust methods against imbalanced class composition and inaccurate class labels. Our results concluded that shallow MLPs (of one or two hidden layers) with ample hidden neurons are sufficient to achieve superior and robust classification performance in exploiting numerical matrix-formed omics data for diagnosis purpose. Specific observations regarding optimal network width, class imbalance tolerance, and inaccurate labeling tolerance will inform future improvement of neural network applications on functional genomics data.

Gene expression profiles have been frequently integrated with the human protein interactome to uncover functional modules under specific conditions like disease state. Beyond traditional differential expression analysis, differential co-expression analysis has emerged as a robust approach to reveal condition-specific network modules, with successful applications in a few human disease studies. Hepatocellular carcinoma (HCC), which is often interrelated with the Hepatitis C virus, typically develops through multiple stages. A comprehensive investigation of HCC progression-specific differential co-expression modules may advance our understanding of HCC's pathophysiological mechanisms. Compared with differentially expressed genes, differentially co-expressed genes were found more likely enriched with Hepatitis C virus binding proteins and cancer-mutated genes, and they were clustered more densely in the human reference protein interaction network. These observations indicated that a differential co-expression approach could outperform the standard differential expression network analysis in searching for disease-related modules. We then proposed a differential co-expression network approach to uncover network modules involved in HCC development. Specifically, we discovered subnetworks that enriched differentially co-expressed gene pairs in each HCC transition stage, and further resolved modules with coherent co-expression change patterns over all HCC developmental stages. Our identified network modules were enriched with HCC-related genes and implicated in cancer-related biological functions. In particular, APC and YWHAZ were highlighted as two most remarkable genes in the network modules, and their dynamic interaction partnership was resolved in HCC development. We demonstrated that integration of differential co-expression with the protein interactome could outperform the traditional differential expression approach in discovering network modules of human diseases. In our application of this approach to HCC's gene expression data, we successfully identified subnetworks with marked differential co-expression in individual HCC stage transitions and network modules with coherent co-expression change patterns over all HCC developmental stages. Our results shed light on subtle HCC mechanisms, including temporal activation and dismissal of pivotal functions and dynamic interaction partnerships of key genes.

Non-coding RNAs occupy a significant fraction of the human genome. Their biological significance is backed up by a plethora of emerging evidence. One of the most robust approaches to demonstrate non-coding RNA’s biological relevance is through their prognostic value. Using the rich gene expression data from The Cancer Genome Altas (TCGA), we designed Advanced Expression Survival Analysis (AESA), a web tool which provides several novel survival analysis approaches not offered by previous tools. In addition to the common single-gene approach, AESA computes the gene expression composite score of a set of genes for survival analysis and utilizes permutation test or cross-validation to assess the significance of log-rank statistic and the degree of over-fitting. AESA offers survival feature selection with post-selection inference and utilizes expanded TCGA clinical data including overall, disease-specific, disease-free, and progression-free survival information. Users can analyze either protein-coding or non-coding regions of the transcriptome. We demonstrated the effectiveness of AESA using several empirical examples. Our analyses showed that non-coding RNAs perform as well as messenger RNAs in predicting survival of cancer patients. These results reinforce the potential prognostic value of non-coding RNAs. AESA is developed as a module in the freely accessible analysis suite MutEx.

Somatic mutations are one of the most important factors in tumorigenesis and are the focus of most cancer-sequencing efforts. The co-occurrence of multiple mutations in one tumor has gained increasing attention as a means of identifying cooperating mutations or pathways that contribute to cancer. Using multi-omics, phenotypical, and clinical data from 29,559 cancer subjects and 1747 cancer cell lines covering 78 distinct cancer types, we show that co-mutations are associated with prognosis, drug sensitivity, and disparities in sex, age, and race. Some co-mutation combinations displayed stronger effects than their corresponding single mutations. For example, co-mutation TP53:KRAS in pancreatic adenocarcinoma is significantly associated with disease specific survival (hazard ratio = 2.87, adjusted p-value = 0.0003) and its prognostic predictive power is greater than either TP53 or KRAS as individually mutated genes. Functional analyses revealed that co-mutations with higher prognostic values have higher potential impact and cause greater dysregulation of gene expression. Furthermore, many of the prognostically significant co-mutations caused gains or losses of binding sequences of RNA binding proteins or micro RNAs with known cancer associations. Thus, detailed analyses of co-mutations can identify mechanisms that cooperate in tumorigenesis.

Triple-negative breast cancer (TNBC) is an operational term for breast cancers lacking targetable estrogen receptor expression and HER2 amplifications. TNBC is, therefore, inherently heterogeneous, and is associated with worse prognosis, greater rates of metastasis, and earlier onset. TNBC displays mutational and transcriptional diversity, and distinct mRNA transcriptional subtypes exhibiting unique biology. High-throughput sequencing has extended cancer research far beyond protein coding regions that include non-coding small RNAs, such as miRNA, isomiR, tRNA, snoRNAs, snRNA, yRNA, 7SL, and 7SK. In this study, we performed small RNA profiling of 26 TNBC cell lines, and compared the abundance of non-coding RNAs among the transcriptional subtypes of triple negative breast cancer. We also examined their co-expression pattern with corresponding mRNAs. This study provides a detailed description of small RNA expression in triple-negative breast cancer cell lines that can aid in the development of future biomarker and novel targeted therapies.