MOTIVATION: Traditional genome-wide association studies (GWAS) aim to uncover the genetic variants associated with a single phenotype of interest (typically a disease), and to elucidate its genotypic architecture. However, many of today's GWAS simultaneously measure multiple related phenotypes, leading to the possibility of pursuing the reverse aim of elucidating the "phenotypic architecture" of a single genetic variant. In other words, we may ask what combination of measured phenotypes is associated with a given genotypic variant. ReverseGWAS is an algorithmic platform for answering such questions in the context of large-scale multi-phenotype GWAS. RESULTS: We demonstrate the effectiveness of ReverseGWAS on simulated data, showing its ability to identify logical combinations of phenotypes with a reasonable amount of noise. We then apply it to a selection of combined phenotypes from the UK Biobank, obtaining 719 candidate associations using autoimmune diseases and 205 using common ICD10 codes. We find that the majority of these associations (546/719 and 111/205, respectively) successfully replicate in an independent cohort, FinnGen. AVAILABILITY AND IMPLEMENTATION: The source code of ReverseGWAS is freely available to non-commercial users as an installable R package at https://github.com/Leonardini/rgwas.

BACKGROUND: The ability to confidently predict health outcomes from gene expression would catalyze a revolution in molecular diagnostics. Yet, the goal of developing actionable, robust, and reproducible predictive signatures of phenotypes such as clinical outcome has not been attained in almost any disease area. Here, we report a comprehensive analysis spanning prediction tasks from ulcerative colitis, atopic dermatitis, diabetes, to many cancer subtypes for a total of 24 binary and multiclass prediction problems and 26 survival analysis tasks. We systematically investigate the influence of gene subsets, normalization methods and prediction algorithms. Crucially, we also explore the novel use of deep representation learning methods on large transcriptomics compendia, such as GTEx and TCGA, to boost the performance of state-of-the-art methods. The resources and findings in this work should serve as both an up-to-date reference on attainable performance, and as a benchmarking resource for further research. RESULTS: -regularized regression methods applied to centered log-ratio transformed transcript abundances provide the best predictive analyses overall. CONCLUSIONS: Transcriptomics-based phenotype prediction benefits from proper normalization techniques and state-of-the-art regularized regression approaches. In our view, breakthrough performance is likely contingent on factors which are independent of normalization and general modeling techniques; these factors might include reduction of systematic errors in sequencing data, incorporation of other data types such as single-cell sequencing and proteomics, and improved use of prior knowledge.

MOTIVATION: Patient stratification methods are key to the vision of precision medicine. Here, we consider transcriptional data to segment the patient population into subsets relevant to a given phenotype. Whereas most existing patient stratification methods focus either on predictive performance or interpretable features, we developed a method striking a balance between these two important goals. RESULTS: We introduce a Bayesian method called SUBSTRA that uses regularized biclustering to identify patient subtypes and interpretable subtype-specific transcript clusters. The method iteratively re-weights feature importance to optimize phenotype prediction performance by producing more phenotype-relevant patient subtypes. We investigate the performance of SUBSTRA in finding relevant features using simulated data and successfully benchmark it against state-of-the-art unsupervised stratification methods and supervised alternatives. Moreover, SUBSTRA achieves predictive performance competitive with the supervised benchmark methods and provides interpretable transcriptional features in diverse biological settings, such as drug response prediction, cancer diagnosis, or kidney transplant rejection. AVAILABILITY AND IMPLEMENTATION: The R code of SUBSTRA is available at https://github.com/sahandk/SUBSTRA. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Stratification of patient subpopulations that respond favorably to treatment or experience and adverse reaction is an essential step toward development of new personalized therapies and diagnostics. It is currently feasible to generate omic-scale biological measurements for all patients in a study, providing an opportunity for machine learning models to identify molecular markers for disease diagnosis and progression. However, the high variability of genetic background in human populations hampers the reproducibility of omic-scale markers. In this paper, we develop a biological network-based regularized artificial neural network model for prediction of phenotype from transcriptomic measurements in clinical trials. To improve model sparsity and the overall reproducibility of the model, we incorporate regularization for simultaneous shrinkage of gene sets based on active upstream regulatory mechanisms into the model. RESULTS: We benchmark our method against various regression, support vector machines and artificial neural network models and demonstrate the ability of our method in predicting the clinical outcomes using clinical trial data on acute rejection in kidney transplantation and response to Infliximab in ulcerative colitis. We show that integration of prior biological knowledge into the classification as developed in this paper, significantly improves the robustness and generalizability of predictions to independent datasets. We provide a Java code of our algorithm along with a parsed version of the STRING DB database. CONCLUSION: In summary, we present a method for prediction of clinical phenotypes using baseline genome-wide expression data that makes use of prior biological knowledge on gene-regulatory interactions in order to increase robustness and reproducibility of omic-scale markers. The integrated group-wise regularization methods increases the interpretability of biological signatures and gives stable performance estimates across independent test sets.

BACKGROUND: Inference of active regulatory cascades under specific molecular and environmental perturbations is a recurring task in transcriptional data analysis. Commercial tools based on large, manually curated networks of causal relationships offering such functionality have been used in thousands of articles in the biomedical literature. The adoption and extension of such methods in the academic community has been hampered by the lack of freely available, efficient algorithms and an accompanying demonstration of their applicability using current public networks. RESULTS: In this article, we propose a new statistical method that will infer likely upstream regulators based on observed patterns of up- and down-regulated transcripts. The method is suitable for use with public interaction networks with a mix of signed and unsigned causal edges. It subsumes and extends two previously published approaches and we provide a novel algorithmic method for efficient statistical inference. Notably, we demonstrate the feasibility of using the approach to generate biological insights given current public networks in the context of controlled in-vitro overexpression experiments, stem-cell differentiation data and animal disease models. We also provide an efficient implementation of our method in the R package QuaternaryProd available to download from Bioconductor. CONCLUSIONS: In this work, we have closed an important gap in utilizing causal networks to analyze differentially expressed genes. Our proposed Quaternary test statistic incorporates all available evidence on the potential relevance of an upstream regulator. The new approach broadens the use of these types of statistics for highly curated signed networks in which ambiguities arise but also enables the use of networks with unsigned edges. We design and implement a novel computational method that can efficiently estimate p-values for upstream regulators in current biological settings. We demonstrate the ready applicability of the implemented method to analyze differentially expressed genes using the publicly available networks.

MOTIVATION: Understanding and predicting an individual's response in a clinical trial is the key to better treatments and cost-: effective medicine. Over the coming years, more and more large-scale omics datasets will become available to characterize patients with complex and heterogeneous diseases at a molecular level. Unfortunately, genetic, phenotypical and environmental variation is much higher in a human trial population than currently modeled or measured in most animal studies. In our experience, this high variability can lead to failure of trained predictors in independent studies and undermines the credibility and utility of promising high-dimensional datasets. METHODS: We propose a method that utilizes patient-level genome-wide expression data in conjunction with causal networks based on prior knowledge. Our approach determines a differential expression profile for each patient and uses a Bayesian approach to infer corresponding upstream regulators. These regulators and their corresponding posterior probabilities of activity are used in a regularized regression framework to predict response. RESULTS: We validated our approach using two clinically relevant phenotypes, namely acute rejection in kidney transplantation and response to Infliximab in ulcerative colitis. To demonstrate pitfalls in translating trained predictors across independent trials, we analyze performance characteristics of our approach as well as alternative feature sets in the regression on two independent datasets for each phenotype. We show that the proposed approach is able to successfully incorporate causal prior knowledge to give robust performance estimates.

MOTIVATION: The abundance of many transcripts changes significantly in response to a variety of molecular and environmental perturbations. A key question in this setting is as follows: what intermediate molecular perturbations gave rise to the observed transcriptional changes? Regulatory programs are not exclusively governed by transcriptional changes but also by protein abundance and post-translational modifications making direct causal inference from data difficult. However, biomedical research over the last decades has uncovered a plethora of causal signaling cascades that can be used to identify good candidates explaining a specific set of transcriptional changes. METHODS: We take a Bayesian approach to integrate gene expression profiling with a causal graph of molecular interactions constructed from prior biological knowledge. In addition, we define the biological context of a specific interaction by the corresponding Medical Subject Headings terms. The Bayesian network can be queried to suggest upstream regulators that can be causally linked to the altered expression profile. RESULTS: Our approach will treat candidate regulators in the right biological context preferentially, enables hierarchical exploration of resulting hypotheses and takes the complete network of causal relationships into account to arrive at the best set of upstream regulators. We demonstrate the power of our method on distinct biological datasets, namely response to dexamethasone treatment, stem cell differentiation and a neuropathic pain model. In all cases relevant biological insights could be validated. AVAILABILITY AND IMPLEMENTATION: Source code for the method is available upon request.

MOTIVATION: The interpretation of high-throughput datasets has remained one of the central challenges of computational biology over the past decade. Furthermore, as the amount of biological knowledge increases, it becomes more and more difficult to integrate this large body of knowledge in a meaningful manner. In this article, we propose a particular solution to both of these challenges. METHODS: We integrate available biological knowledge by constructing a network of molecular interactions of a specific kind: causal interactions. The resulting causal graph can be queried to suggest molecular hypotheses that explain the variations observed in a high-throughput gene expression experiment. We show that a simple scoring function can discriminate between a large number of competing molecular hypotheses about the upstream cause of the changes observed in a gene expression profile. We then develop an analytical method for computing the statistical significance of each score. This analytical method also helps assess the effects of random or adversarial noise on the predictive power of our model. RESULTS: Our results show that the causal graph we constructed from known biological literature is extremely robust to random noise and to missing or spurious information. We demonstrate the power of our causal reasoning model on two specific examples, one from a cancer dataset and the other from a cardiac hypertrophy experiment. We conclude that causal reasoning models provide a valuable addition to the biologist's toolkit for the interpretation of gene expression data. AVAILABILITY AND IMPLEMENTATION: R source code for the method is available upon request.

Abstract Background Causal graphs are an increasingly popular tool for the analysis of biological datasets. In particular, signed causal graphs--directed graphs whose edges additionally have a sign denoting upregulation or downregulation--can be used to model regulatory networks within a cell. Such models allow prediction of downstream effects of regulation of biological entities; conversely, they also enable inference of causative agents behind observed expression changes. However, due to their complex nature, signed causal graph models present special challenges with respect to assessing statistical significance. In this paper we frame and solve two fundamental computational problems that arise in practice when computing appropriate null distributions for hypothesis testing. Results First, we show how to compute a p-value for agreement between observed and model-predicted classifications of gene transcripts as upregulated, downregulated, or neither. Specifically, how likely are the classifications to agree to the same extent under the null distribution of the observed classification being randomized? This problem, which we call "Ternary Dot Product Distribution" owing to its mathematical form, can be viewed as a generalization of Fisher's exact test to ternary variables. We present two computationally efficient algorithms for computing the Ternary Dot Product Distribution and investigate its combinatorial structure analytically and numerically to establish computational complexity bounds. Second, we develop an algorithm for efficiently performing random sampling of causal graphs. This enables p-value computation under a different, equally important null distribution obtained by randomizing the graph topology but keeping fixed its basic structure: connectedness and the positive and negative in- and out-degrees of each vertex. We provide an algorithm for sampling a graph from this distribution uniformly at random. We also highlight theoretical challenges unique to signed causal graphs; previous work on graph randomization has studied undirected graphs and directed but unsigned graphs. Conclusion We present algorithmic solutions to two statistical significance questions necessary to apply the causal graph methodology, a powerful tool for biological network analysis. The algorithms we present are both fast and provably correct. Our work may be of independent interest in non-biological contexts as well, as it generalizes mathematical results that have been studied extensively in other fields.

BACKGROUND: Identification of active causal regulators is a crucial problem in understanding mechanism of diseases or finding drug targets. Methods that infer causal regulators directly from primary data have been proposed and successfully validated in some cases. These methods necessarily require very large sample sizes or a mix of different data types. Recent studies have shown that prior biological knowledge can successfully boost a method's ability to find regulators. RESULTS: We present a simple data-driven method, Correlation Set Analysis (CSA), for comprehensively detecting active regulators in disease populations by integrating co-expression analysis and a specific type of literature-derived causal relationships. Instead of investigating the co-expression level between regulators and their regulatees, we focus on coherence of regulatees of a regulator. Using simulated datasets we show that our method performs very well at recovering even weak regulatory relationships with a low false discovery rate. Using three separate real biological datasets we were able to recover well known and as yet undescribed, active regulators for each disease population. The results are represented as a rank-ordered list of regulators, and reveals both single and higher-order regulatory relationships. CONCLUSIONS: CSA is an intuitive data-driven way of selecting directed perturbation experiments that are relevant to a disease population of interest and represent a starting point for further investigation. Our findings demonstrate that combining co-expression analysis on regulatee sets with a literature-derived network can successfully identify causal regulators and help develop possible hypothesis to explain disease progression.