Artificial neural networks have recently gained significant attention in biomedical research. However, their utility in survival analysis still faces many challenges. In addition to designing models for high accuracy, it is essential to optimize models that provide biologically meaningful insights. With these considerations in mind, we developed a deep neural network model, MaskedNet, to identify genes and pathways whose expression at the time of diagnosis is associated with overall survival. MaskedNet was trained using TCGA breast cancer transcriptome and clinical data, and the model’s final output was the predicted logarithm of the hazard ratio for death. The trained model was interpreted using SHapley Additive exPlanations (SHAP), a technique grounded in robust mathematical principles that assigns importance scores to input features. Compared to traditional Cox proportional hazards regression, MaskedNet had higher accuracy, as measured by Harrell’s C-index. We also found that aggregating outputs from several model runs identified multiple genes and pathways associated with overall survival, including IFNG and PIK3CA genes , along with their related pathways. To further elucidate the role of the IFNG gene, tumors were partitioned into two groups based on low and high IFNG SHAP values, respectively. Tumors with lower IFNG SHAP values exhibited higher IFNG expression and better overall survival, which were linked to more abundant presence of M1 macrophages and activated CD4+ and CD8+ T cells in the tumor microenvironment. The association of the IFNG pathway with overall survival was validated in the trastuzumab arm of the NCCTG-N9831 trial, an independent breast cancer study.

[This corrects the article PMC11643596.].

MOTIVATION: Functions of cancer driver genes vary substantially across tissues and organs. Distinguishing passenger genes, oncogenes (OGs) and tumor-suppressor genes (TSGs) for each cancer type is critical for understanding tumor biology and identifying clinically actionable targets. Although many computational tools are available to predict putative cancer driver genes, resources for context-aware classifications of OGs and TSGs are limited. RESULTS: We show that the direction and magnitude of somatic selection of protein-coding mutations are significantly different for passenger genes, OGs and TSGs. Based on these patterns, we develop a new method (genes under selection in tumors) to discover OGs and TSGs in a cancer-type specific manner. Genes under selection in tumors shows a high accuracy (92%) when evaluated via strict cross-validations. Its application to 10 172 tumor exomes found known and novel cancer drivers with high tissue-specificities. In 11 out of 13 OGs shared among multiple cancer types, we found functional domains selectively engaged in different cancers, suggesting differences in disease mechanisms. AVAILABILITY AND IMPLEMENTATION: An R implementation of the GUST algorithm is available at https://github.com/liliulab/gust. A database with pre-computed results is available at https://liliulab.shinyapps.io/gust. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: After decades of identifying risk factors using array-based genome-wide association studies (GWAS), genetic research of complex diseases has shifted to sequencing-based rare variants discovery. This requires large sample sizes for statistical power and has brought up questions about whether the current variant calling practices are adequate for large cohorts. It is well-known that there are discrepancies between variants called by different pipelines, and that using a single pipeline always misses true variants exclusively identifiable by other pipelines. Nonetheless, it is common practice today to call variants by one pipeline due to computational cost and assume that false negative calls are a small percent of total. RESULTS: We analyzed 10,000 exomes from the Alzheimer's Disease Sequencing Project (ADSP) using multiple analytic pipelines consisting of different read aligners and variant calling strategies. We compared variants identified by using two aligners in 50,100, 200, 500, 1000, and 1952 samples; and compared variants identified by adding single-sample genotyping to the default multi-sample joint genotyping in 50,100, 500, 2000, 5000 and 10,000 samples. We found that using a single pipeline missed increasing numbers of high-quality variants correlated with sample sizes. By combining two read aligners and two variant calling strategies, we rescued 30% of pass-QC variants at sample size of 2000, and 56% at 10,000 samples. The rescued variants had higher proportions of low frequency (minor allele frequency [MAF] 1-5%) and rare (MAF < 1%) variants, which are the very type of variants of interest. In 660 Alzheimer's disease cases with earlier onset ages of ≤65, 4 out of 13 (31%) previously-published rare pathogenic and protective mutations in APP, PSEN1, and PSEN2 genes were undetected by the default one-pipeline approach but recovered by the multi-pipeline approach. CONCLUSIONS: Identification of the complete variant set from sequencing data is the prerequisite of genetic association analyses. The current analytic practice of calling genetic variants from sequencing data using a single bioinformatics pipeline is no longer adequate with the increasingly large projects. The number and percentage of quality variants that passed quality filters but are missed by the one-pipeline approach rapidly increased with sample size.

MOTIVATION: RNA-seq has become the method of choice to quantify genes and exons, discover novel transcripts and detect fusion genes. However, reliable variant identification from RNA-seq data remains challenging because of the complexities of the transcriptome, the challenges of accurately mapping exon boundary spanning reads and the bias introduced during the sequencing library preparation. METHOD: We developed RVboost, a novel method specific for RNA variant prioritization. RVboost uses several attributes unique in the process of RNA library preparation, sequencing and RNA-seq data analyses. It uses a boosting method to train a model of 'good quality' variants using common variants from HapMap, and prioritizes and calls the RNA variants based on the trained model. We packaged RVboost in a comprehensive workflow, which integrates tools of variant calling, annotation and filtering. RESULTS: RVboost consistently outperforms the variant quality score recalibration from the Genome Analysis Tool Kit and the RNA-seq variant-calling pipeline SNPiR in 12 RNA-seq samples using ground-truth variants from paired exome sequencing data. Several RNA-seq-specific attributes were identified as critical to differentiate true and false variants, including the distance of the variant positions to exon boundaries, and the percent of the reads supporting the variant in the first six base pairs. The latter identifies false variants introduced by the random hexamer priming during the library construction. AVAILABILITY AND IMPLEMENTATION: The RVboost package is implemented to readily run in Mac or Linux environments. The software and user manual are available at http://bioinformaticstools.mayo.edu/research/rvboost/.

MOTIVATION: Exome sequencing (exome-seq) data, which are typically used for calling exonic mutations, have also been utilized in detecting DNA copy number variations (CNVs). Despite the existence of several CNV detection tools, there is still a great need for a sensitive and an accurate CNV-calling algorithm with built-in QC steps, and does not require a paired reference for each sample. RESULTS: We developed a novel method named PatternCNV, which (i) accounts for the read coverage variations between exons while leveraging the consistencies of this variability across different samples; (ii) reduces alignment BAM files to WIG format and therefore greatly accelerates computation; (iii) incorporates multiple QC measures designed to identify outlier samples and batch effects; and (iv) provides a variety of visualization options including chromosome, gene and exon-level views of CNVs, along with a tabular summarization of the exon-level CNVs. Compared with other CNV-calling algorithms using data from a lymphoma exome-seq study, PatternCNV has higher sensitivity and specificity. AVAILABILITY AND IMPLEMENTATION: The software for PatternCNV is implemented using Perl and R, and can be used in Mac or Linux environments. Software and user manual are available at http://bioinformaticstools.mayo.edu/research/patterncnv/, and R package at https://github.com/topsoil/patternCNV/.

BACKGROUND: Although the costs of next generation sequencing technology have decreased over the past years, there is still a lack of simple-to-use applications, for a comprehensive analysis of RNA sequencing data. There is no one-stop shop for transcriptomic genomics. We have developed MAP-RSeq, a comprehensive computational workflow that can be used for obtaining genomic features from transcriptomic sequencing data, for any genome. RESULTS: For optimization of tools and parameters, MAP-RSeq was validated using both simulated and real datasets. MAP-RSeq workflow consists of six major modules such as alignment of reads, quality assessment of reads, gene expression assessment and exon read counting, identification of expressed single nucleotide variants (SNVs), detection of fusion transcripts, summarization of transcriptomics data and final report. This workflow is available for Human transcriptome analysis and can be easily adapted and used for other genomes. Several clinical and research projects at the Mayo Clinic have applied the MAP-RSeq workflow for RNA-Seq studies. The results from MAP-RSeq have thus far enabled clinicians and researchers to understand the transcriptomic landscape of diseases for better diagnosis and treatment of patients. CONCLUSIONS: Our software provides gene counts, exon counts, fusion candidates, expressed single nucleotide variants, mapping statistics, visualizations, and a detailed research data report for RNA-Seq. The workflow can be executed on a standalone virtual machine or on a parallel Sun Grid Engine cluster. The software can be downloaded from http://bioinformaticstools.mayo.edu/research/maprseq/.

UNLABELLED: TREAT (Targeted RE-sequencing Annotation Tool) is a tool for facile navigation and mining of the variants from both targeted resequencing and whole exome sequencing. It provides a rich integration of publicly available as well as in-house developed annotations and visualizations for variants, variant-hosting genes and host-gene pathways. AVAILABILITY AND IMPLEMENTATION: TREAT is freely available to non-commercial users as either a stand-alone annotation and visualization tool, or as a comprehensive workflow integrating sequencing alignment and variant calling. The executables, instructions and the Amazon Cloud Images of TREAT can be downloaded at the website: http://ndc.mayo.edu/mayo/research/biostat/stand-alone-packages.cfm.

UNLABELLED: Reduced representation bisulfite sequencing (RRBS) is a cost-effective approach for genome-wide methylation pattern profiling. Analyzing RRBS sequencing data is challenging and specialized alignment/mapping programs are needed. Although such programs have been developed, a comprehensive solution that provides researchers with good quality and analyzable data is still lacking. To address this need, we have developed a Streamlined Analysis and Annotation Pipeline for RRBS data (SAAP-RRBS) that integrates read quality assessment/clean-up, alignment, methylation data extraction, annotation, reporting and visualization. This package facilitates a rapid transition from sequencing reads to a fully annotated CpG methylation report to biological interpretation. AVAILABILITY AND IMPLEMENTATION: SAAP-RRBS is freely available to non-commercial users at the web site http://ndc.mayo.edu/mayo/research/biostat/stand-alone-packages.cfm.

BACKGROUND: The developments of high-throughput genotyping technologies, which enable the simultaneous genotyping of hundreds of thousands of single nucleotide polymorphisms (SNP) have the potential to increase the benefits of genetic epidemiology studies. Although the enhanced resolution of these platforms increases the chance of interrogating functional SNPs that are themselves causative or in linkage disequilibrium with causal SNPs, commonly used single SNP-association approaches suffer from serious multiple hypothesis testing problems and provide limited insights into combinations of loci that may contribute to complex diseases. Drawing inspiration from Gene Set Enrichment Analysis developed for gene expression data, we have developed a method, named GLOSSI (Gene-loci Set Analysis), that integrates prior biological knowledge into the statistical analysis of genotyping data to test the association of a group of SNPs (loci-set) with complex disease phenotypes. The most significant loci-sets can be used to formulate hypotheses from a functional viewpoint that can be validated experimentally. RESULTS: In a simulation study, GLOSSI showed sufficient power to detect loci-sets with less than 10% of SNPs having moderate-to-large effect sizes and intermediate minor allele frequency values. When applied to a biological dataset where no single SNP-association was found in a previous study, GLOSSI was able to identify several loci-sets that are significantly related to blood pressure response to an antihypertensive drug. CONCLUSION: GLOSSI is valuable for association of SNPs at multiple genetic loci with complex disease phenotypes. In contrast to methods based on the Kolmogorov-Smirnov statistic, the approach is parametric and only utilizes information from within the interrogated loci-set. It properly accounts for dependency among SNPs and allows the testing of loci-sets of any size.