Biocomputing 2017

The following sections are included:

• Preface
• PACIFIC SYMPOSIUM ON BIOCOMPUTING 2017

The following sections are included:

• Introduction
• Overview of Contributions
• References

With continued rapid growth in the number and quality of fully sequenced and accurately annotated bacterial genomes, we have unprecedented opportunities to understand metabolic diversity. We selected 101 diverse and representative completely sequenced bacteria and implemented a manual curation effort to identify 846 unique metabolic variants present in these bacteria. The presence or absence of these variants act as a metabolic signature for each of the bacteria, which can then be used to understand similarities and differences between and across bacterial groups. We propose a novel and robust method of summarizing metabolic diversity using metabolic signatures and use this method to generate a metabolic tree, clustering metabolically similar organisms. Resulting analysis of the metabolic tree confirms strong associations with well-established biological results along with direct insight into particular metabolic variants which are most predictive of metabolic diversity. The positive results of this manual curation effort and novel method development suggest that future work is needed to further expand the set of bacteria to which this approach is applied and use the resulting tree to test broad questions about metabolic diversity and complexity across the bacterial tree of life.

Current automated computational methods to assign functional labels to unstudied genes often involve transferring annotation from orthologous or paralogous genes, however such genes can evolve divergent functions, making such transfer inappropriate. We consider the problem of determining when it is correct to make such an assignment between paralogs. We construct a benchmark dataset of two types of similar paralogous pairs of genes in the well-studied model organism S. cerevisiae: one set of pairs where single deletion mutants have very similar phenotypes (implying similar functions), and another set of pairs where single deletion mutants have very divergent phenotypes (implying different functions). State of the art methods for this problem will determine the evolutionary history of the paralogs with references to multiple related species. Here, we ask a first and simpler question: we explore to what extent any computational method with access only to data from a single species can solve this problem.

We consider divergence data (at both the amino acid and nucleotide levels), and network data (based on the yeast protein-protein interaction network, as captured in BioGRID), and ask if we can extract features from these data that can distinguish between these sets of paralogous gene pairs. We find that the best features come from measures of sequence divergence, however, simple network measures based on degree or centrality or shortest path or diffusion state distance (DSD), or shared neighborhood in the yeast protein-protein interaction (PPI) network also contain some signal. One should, in general, not transfer function if sequence divergence is too high. Further improvements in classification will need to come from more computationally expensive but much more powerful evolutionary methods that incorporate ancestral states and measure evolutionary divergence over multiple species based on evolutionary trees.

Automated annotation of protein function has become a critical task in the post-genomic era. Network-based approaches and homology-based approaches have been widely used and recently tested in large-scale community-wide assessment experiments. It is natural to integrate network data with homology information to further improve the predictive performance. However, integrating these two heterogeneous, high-dimensional and noisy datasets is non-trivial. In this work, we introduce a novel protein function prediction algorithm ProSNet. An integrated heterogeneous network is first built to include molecular networks of multiple species and link together homologous proteins across multiple species. Based on this integrated network, a dimensionality reduction algorithm is introduced to obtain compact low-dimensional vectors to encode proteins in the network. Finally, we develop machine learning classification algorithms that take the vectors as input and make predictions by transferring annotations both within each species and across different species. Extensive experiments on five major species demonstrate that our integration of homology with molecular networks substantially improves the predictive performance over existing approaches.

Over the last decades, we have observed an ongoing tremendous growth of available sequencing data fueled by the advancements in wet-lab technology. The sequencing information is only the beginning of the actual understanding of how organisms survive and prosper. It is, for instance, equally important to also unravel the proteomic repertoire of an organism. A classical computational approach for detecting protein families is a sequence-based similarity calculation coupled with a subsequent cluster analysis. In this work we have intensively analyzed various clustering tools on a large scale. We used the data to investigate the behavior of the tools' parameters underlining the diversity of the protein families. Furthermore, we trained regression models for predicting the expected performance of a clustering tool for an unknown data set and aimed to also suggest optimal parameters in an automated fashion. Our analysis demonstrates the benefits and limitations of the clustering of proteins with low sequence similarity indicating that each protein family requires its own distinct set of tools and parameters. All results, a tool prediction service, and additional supporting material is also available online under http://proteinclustering.compbio.sdu.dk.

Imaging genomics is an emerging research field, where integrative analysis of imaging and omics data is performed to provide new insights into the phenotypic characteristics and genetic mechanisms of normal and/or disordered biological structures and functions, and to impact the development of new diagnostic, therapeutic and preventive approaches. The Imaging Genomics Session at PSB 2017 aims to encourage discussion on fundamental concepts, new methods and innovative applications in this young and rapidly evolving field.

Due to its high dimensionality and high noise levels, analysis of a large brain functional network may not be powerful and easy to interpret; instead, decomposition of a large network into smaller subcomponents called modules may be more promising as suggested by some empirical evidence. For example, alteration of brain modularity is observed in patients suffering from various types of brain malfunctions. Although several methods exist for estimating brain functional networks, such as the sample correlation matrix or graphical lasso for a sparse precision matrix, it is still difficult to extract modules from such network estimates. Motivated by these considerations, we adapt a weighted gene co-expression network analysis (WGCNA) framework to resting-state fMRI (rs-fMRI) data to identify modular structures in brain functional networks. Modular structures are identified by using topological overlap matrix (TOM) elements in hierarchical clustering. We propose applying a new adaptive test built on the proportional odds model (POM) that can be applied to a high-dimensional setting, where the number of variables (p) can exceed the sample size (n) in addition to the usual p < n setting. We applied our proposed methods to the ADNI data to test for associations between a genetic variant and either the whole brain functional network or its various subcomponents using various connectivity measures. We uncovered several modules based on the control cohort, and some of them were marginally associated with the APOE4 variant and several other SNPs; however, due to the small sample size of the ADNI data, larger studies are needed.

Characterizing the transcriptome architecture of the human brain is fundamental in gaining an understanding of brain function and disease. A number of recent studies have investigated patterns of brain gene expression obtained from an extensive anatomical coverage across the entire human brain using experimental data generated by the Allen Human Brain Atlas (AHBA) project. In this paper, we propose a new representation of a gene's transcription activity that explicitly captures the pattern of spatial co-expression across different anatomical brain regions. For each gene, we define a Spatial Expression Network (SEN), a network quantifying co-expression patterns amongst several anatomical locations. Network similarity measures are then employed to quantify the topological resemblance between pairs of SENs and identify naturally occurring clusters. Using network-theoretical measures, three large clusters have been detected featuring distinct topological properties. We then evaluate whether topological diversity of the SENs reects significant differences in biological function through a gene ontology analysis. We report on evidence suggesting that one of the three SEN clusters consists of genes specifically involved in the nervous system, including genes related to brain disorders, while the remaining two clusters are representative of immunity, transcription and translation. These findings are consistent with previous studies showing that brain gene clusters are generally associated with one of these three major biological processes.

Lung cancer is one of the most deadly cancers and lung adenocarcinoma (LUAD) is the most common histological type of lung cancer. However, LUAD is highly heterogeneous due to genetic difference as well as phenotypic differences such as cellular and tissue morphology. In this paper, we systematically examine the relationships between histological features and gene transcription. Specifically, we calculated 283 morphological features from histology images for 201 LUAD patients from TCGA project and identified the morphological feature with strong correlation with patient outcome. We then modeled the morphology feature using multiple co-expressed gene clusters using Lasso-regression. Many of the gene clusters are highly associated with genetic variations, specifically DNA copy number variations, implying that genetic variations play important roles in the development cancer morphology. As far as we know, our finding is the first to directly link the genetic variations and functional genomics to LUAD histology. These observations will lead to new insight on lung cancer development and potential new integrative biomarkers for prediction patient prognosis and response to treatments.

Brain imaging and protein expression, from both cerebrospinal fluid and blood plasma, have been found to provide complementary information in predicting the clinical outcomes of Alzheimer’s disease (AD). But the underlying associations that contribute to such a complementary relationship have not been previously studied yet. In this work, we will perform an imaging proteomics association analysis to explore how they are related with each other. While traditional association models, such as Sparse Canonical Correlation Analysis (SCCA), can not guarantee the selection of only disease-relevant biomarkers and associations, we propose a novel discriminative SCCA (denoted as DSCCA) model with new penalty terms to account for the disease status information. Given brain imaging, proteomic and diagnostic data, the proposed model can perform a joint association and multi-class discrimination analysis, such that we can not only identify disease-relevant multimodal biomarkers, but also reveal strong associations between them. Based on a real imaging proteomic data set, the empirical results show that DSCCA and traditional SCCA have comparable association performances. But in a further classification analysis, canonical variables of imaging and proteomic data obtained in DSCCA demonstrate much more discrimination power toward multiple pairs of diagnosis groups than those obtained in SCCA.

We consider the problem of multimodal data integration for the study of complex neurological diseases (e.g. schizophrenia). Among the challenges arising in such situation, estimating the link between genetic and neurological variability within a population sample has been a promising direction. A wide variety of statistical models arose from such applications. For example, Lasso regression and its multitask extension are often used to fit a multivariate linear relationship between given phenotype(s) and associated observations. Other approaches, such as canonical correlation analysis (CCA), are widely used to extract relationships between sets of variables from different modalities. In this paper, we propose an exploratory multivariate method combining these two methods. More Specifically, we rely on a ’CCA-type’ formulation in order to regularize the classical multimodal Lasso regression problem. The underlying motivation is to extract discriminative variables that display are also co-expressed across modalities. We first evaluate the method on a simulated dataset, and further validate it using Single Nucleotide Polymorphisms (SNP) and functional Magnetic Resonance Imaging (fMRI) data for the study of schizophrenia.

Science is not done in a vacuum – across fields of biomedicine, scientists have built on previous research and used data published in previous papers. A mainstay of scientific inquiry is the publication of one’s research and recognition for this work is given in the form of citations and notoriety – ideally given in proportion to the quality of the work. Academic incentives, however, may encourage individual researchers to prioritize career ambitions over scientific truth. Recently, the New England Journal of Medicine published a commentary calling scientists who repurpose data “research parasites” who misuse data generated by others to demonstrate alternative hypotheses. In our opinion, the concept of data hoarding not only runs contrary to the spirit of, but also hinders scientific progress. Scientific research is meant to seek objective truth, rather than promote a personal agenda, and the only way to do so is through maximum transparency and reproducibility, no matter who is using the data…

Network reconstruction algorithms are increasingly being employed in biomedical and life sciences research to integrate large-scale, high-dimensional data informing on living systems. One particular class of probabilistic causal networks being applied to model the complexity and causal structure of biological data is Bayesian networks (BNs). BNs provide an elegant mathematical framework for not only inferring causal relationships among many different molecular and higher order phenotypes, but also for incorporating highly diverse priors that provide an efficient path for incorporating existing knowledge. While significant methodological developments have broadly enabled the application of BNs to generate and validate meaningful biological hypotheses, the reproducibility of BNs in this context has not been systematically explored. In this study, we aim to determine the criteria for generating reproducible BNs in the context of transcription-based regulatory networks. We utilize two unique tissues from independent datasets, whole blood from the GTEx Consortium and liver from the Stockholm-Tartu Atherosclerosis Reverse Network Engineering Team (STARNET) study. We evaluated the reproducibility of the BNs by creating networks on data subsampled at different levels from each cohort and comparing these networks to the BNs constructed using the complete data. To help validate our results, we used simulated networks at varying sample sizes. Our study indicates that reproducibility of BNs in biological research is an issue worthy of further consideration, especially in light of the many publications that now employ findings from such constructs without appropriate attention paid to reproducibility. We find that while edge-to-edge reproducibility is strongly dependent on sample size, identification of more highly connected key driver nodes in BNs can be carried out with high confidence across a range of sample sizes.

Repurposing existing drugs for new uses has attracted considerable attention over the past years. To identify potential candidates that could be repositioned for a new indication, many studies make use of chemical, target, and side effect similarity between drugs to train classifiers. Despite promising prediction accuracies of these supervised computational models, their use in practice, such as for rare diseases, is hindered by the assumption that there are already known and similar drugs for a given condition of interest. In this study, using publicly available data sets, we question the prediction accuracies of supervised approaches based on drug similarity when the drugs in the training and the test set are completely disjoint. We first build a Python platform to generate reproducible similarity-based drug repurposing models. Next, we show that, while a simple chemical, target, and side effect similarity based machine learning method can achieve good performance on the benchmark data set, the prediction performance drops sharply when the drugs in the folds of the cross validation are not overlapping and the similarity information within the training and test sets are used independently. These intriguing results suggest revisiting the assumptions underlying the validation scenarios of similarity-based methods and underline the need for unsupervised approaches to identify novel drug uses inside the unexplored pharmacological space. We make the digital notebook containing the Python code to replicate our analysis that involves the drug repurposing platform based on machine learning models and the proposed disjoint cross fold generation method freely available at github.com/emreg00/repurpose.

A major contributor to the scientific reproducibility crisis has been that the results from homogeneous, single-center studies do not generalize to heterogeneous, real world populations. Multi-cohort gene expression analysis has helped to increase reproducibility by aggregating data from diverse populations into a single analysis. To make the multi-cohort analysis process more feasible, we have assembled an analysis pipeline which implements rigorously studied meta-analysis best practices. We have compiled and made publicly available the results of our own multi-cohort gene expression analysis of 103 diseases, spanning 615 studies and 36,915 samples, through a novel and interactive web application. As a result, we have made both the process of and the results from multi-cohort gene expression analysis more approachable for non-technical users.

As biomedical data has become increasingly easy to generate in large quantities, the methods used to analyze it have proliferated rapidly. Reproducible and reusable methods are required to learn from large volumes of data reliably. To address this issue, numerous groups have developed workflow specifications or execution engines, which provide a framework with which to perform a sequence of analyses. One such specification is the Common Workflow Language, an emerging standard which provides a robust and flexible framework for describing data analysis tools and workflows. In addition, reproducibility can be furthered by executors or workflow engines which interpret the specification and enable additional features, such as error logging, file organization, optim1izations to computation and job scheduling, and allow for easy computing on large volumes of data. To this end, we have developed the Rabix Executor, an open-source workflow engine for the purposes of improving reproducibility through reusability and interoperability of workflow descriptions.

Open sharing of clinical genetic data promises to both monitor and eventually improve the reproducibility of variant interpretation among clinical testing laboratories. A significant public data resource has been developed by the NIH ClinVar initiative, which includes submissions from hundreds of laboratories and clinics worldwide. We analyzed a subset of ClinVar data focused on specific clinical areas and we find high reproducibility (>90% concordance) among labs, although challenges for the community are clearly identified in this dataset. We further review results for the commonly tested BRCA1 and BRCA2 genes, which show even higher concordance, although the significant fragmentation of data into different silos presents an ongoing challenge now being addressed by the BRCA Exchange. We encourage all laboratories and clinics to contribute to these important resources.

Given the exponential growth of biomedical data, researchers are faced with numerous challenges in extracting and interpreting information from these large, high-dimensional, incomplete, and often noisy data. To facilitate addressing this growing concern, the “Patterns in Biomedical Data-How do we find them?” session of the 2017 Pacific Symposium on Biocomputing (PSB) is devoted to exploring pattern recognition using data-driven approaches for biomedical and precision medicine applications. The papers selected for this session focus on novel machine learning techniques as well as applications of established methods to heterogeneous data. We also feature manuscripts aimed at addressing the current challenges associated with the analysis of biomedical data.

There is heterogeneity in the manifestation of diseases, therefore it is essential to understand the patterns of progression of a disease in a given population for disease management as well as for clinical research. Disease status is often summarized by repeated recordings of one or more physiological measures. As a result, historical values of these physiological measures for a population sample can be used to characterize disease progression patterns. We use a method for clustering sparse functional data for identifying sub-groups within a cohort of patients with chronic kidney disease (CKD), based on the trajectories of their Creatinine measurements. We demonstrate through a proof-of-principle study how the two sub-groups that display distinct patterns of disease progression may be compared on clinical attributes that correspond to the maximum difference in progression patterns. The key attributes that distinguish the two sub-groups appear to have support in published literature clinical practice related to CKD.

Osteosarcoma is one of the most common types of bone cancer in children. To gauge the extent of cancer treatment response in the patient after surgical resection, the H&E stained image slides are manually evaluated by pathologists to estimate the percentage of necrosis, a time consuming process prone to observer bias and inaccuracy. Digital image analysis is a potential method to automate this process, thus saving time and providing a more accurate evaluation. The slides are scanned in Aperio Scanscope, converted to digital Whole Slide Images (WSIs) and stored in SVS format. These are high resolution images, of the order of 10⁹ pixels, allowing up to 40X magnification factor. This paper proposes an image segmentation and analysis technique for segmenting tumor and non-tumor regions in histopathological WSIs of osteosarcoma datasets. Our approach is a combination of pixel-based and object-based methods which utilize tumor properties such as nuclei cluster, density, and circularity to classify tumor regions as viable and non-viable. A K-Means clustering technique is used for tumor isolation using color normalization, followed by multi-threshold Otsu segmentation technique to further classify tumor region as viable and non-viable. Then a Flood-fill algorithm is applied to cluster similar pixels into cellular objects and compute cluster data for further analysis of regions under study. To the best of our knowledge this is the first comprehensive solution that is able to produce such a classification for Osteosarcoma cancer. The results are very conclusive in identifying viable and non-viable tumor regions. In our experiments, the accuracy of the discussed approach is 100% in viable tumor and coagulative necrosis identification while it is around 90% for fibrosis and acellular/hypocellular tumor osteoid, for all the sampled datasets used. We expect the developed software to lead to a significant increase in accuracy and decrease in inter-observer variability in assessment of necrosis by the pathologists and a reduction in the time spent by the pathologists in such assessments.

Electronic health records (EHRs) have become a vital source of patient outcome data but the widespread prevalence of missing data presents a major challenge. Different causes of missing data in the EHR data may introduce unintentional bias. Here, we compare the effectiveness of popular multiple imputation strategies with a deeply learned autoencoder using the Pooled Resource Open-Access ALS Clinical Trials Database (PRO-ACT). To evaluate performance, we examined imputation accuracy for known values simulated to be either missing completely at random or missing not at random. We also compared ALS disease progression prediction across different imputation models. Autoencoders showed strong performance for imputation accuracy and contributed to the strongest disease progression predictor. Finally, we show that despite clinical heterogeneity, ALS disease progression appears homogenous with time from onset being the most important predictor.

Cancer detection from gene expression data continues to pose a challenge due to the high dimensionality and complexity of these data. After decades of research there is still uncertainty in the clinical diagnosis of cancer and the identification of tumor-specific markers. Here we present a deep learning approach to cancer detection, and to the identification of genes critical for the diagnosis of breast cancer. First, we used Stacked Denoising Autoencoder (SDAE) to deeply extract functional features from high dimensional gene expression profiles. Next, we evaluated the performance of the extracted representation through supervised classification models to verify the usefulness of the new features in cancer detection. Lastly, we identified a set of highly interactive genes by analyzing the SDAE connectivity matrices. Our results and analysis illustrate that these highly interactive genes could be useful cancer biomarkers for the detection of breast cancer that deserve further studies.

Socioeconomic status (SES) is a fundamental contributor to health, and a key factor underlying racial disparities in disease. However, SES data are rarely included in genetic studies due in part to the difficultly of collecting these data when studies were not originally designed for that purpose. The emergence of large clinic-based biobanks linked to electronic health records (EHRs) provides research access to large patient populations with longitudinal phenotype data captured in structured fields as billing codes, procedure codes, and prescriptions. SES data however, are often not explicitly recorded in structured fields, but rather recorded in the free text of clinical notes and communications. The content and completeness of these data vary widely by practitioner. To enable gene-environment studies that consider SES as an exposure, we sought to extract SES variables from racial/ethnic minority adult patients (n=9,977) in BioVU, the Vanderbilt University Medical Center biorepository linked to de-identified EHRs. We developed several measures of SES using information available within the de-identified EHR, including broad categories of occupation, education, insurance status, and homelessness. Two hundred patients were randomly selected for manual review to develop a set of seven algorithms for extracting SES information from de-identified EHRs. The algorithms consist of 15 categories of information, with 830 unique search terms. SES data extracted from manual review of 50 randomly selected records were compared to data produced by the algorithm, resulting in positive predictive values of 80.0% (education), 85.4% (occupation), 87.5% (unemployment), 63.6% (retirement), 23.1% (uninsured), 81.8% (Medicaid), and 33.3% (homelessness), suggesting some categories of SES data are easier to extract in this EHR than others. The SES data extraction approach developed here will enable future EHR-based genetic studies to integrate SES information into statistical analyses. Ultimately, incorporation of measures of SES into genetic studies will help elucidate the impact of the social environment on disease risk and outcomes.

Type 2 diabetes (T2D) is the result of metabolic defects in insulin secretion and insulin sensitivity, yet most T2D loci identified to date influence insulin secretion. We hypothesized that T2D loci, particularly those affecting insulin sensitivity, can be identified through interaction with known T2D loci implicated in insulin secretion. To test this hypothesis, single nucleotide polymorphisms (SNPs) nominally associated with acute insulin response to glucose (AIR_g), a dynamic measure of first-phase insulin secretion, and previously associated with T2D in genome-wide association studies (GWAS) were identified in African Americans from the Insulin Resistance Atherosclerosis Family Study (IRASFS; n=492 subjects). These SNPs were tested for interaction, individually and jointly as a genetic risk score (GRS), using GWAS data from five cohorts (ARIC, CARDIA, JHS, MESA, WFSM; n=2,725 cases, 4,167 controls) with T2D as the outcome. In single variant analyses, suggestively significant (P_interaction < 5×10^-6) interactions were observed at several loci including DGKB (rs978989), CDK18 (rs12126276), CXCL12 (rs7921850), HCN1 (rs6895191), FAM98A (rs1900780), and MGMT (rs568530). Notable beta-cell GRS interactions included two SNPs at the DGKB locus (rs6976381; rs6962498). These data support the hypothesis that additional genetic factors contributing to T2D risk can be identified by interactions with insulin secretion loci.

Deep neural network (DNN) models have recently obtained state-of-the-art prediction accuracy for the transcription factor binding (TFBS) site classification task. However, it remains unclear how these approaches identify meaningful DNA sequence signals and give insights as to why TFs bind to certain locations. In this paper, we propose a toolkit called the Deep Motif Dashboard (DeMo Dashboard) which provides a suite of visualization strategies to extract motifs, or sequence patterns from deep neural network models for TFBS classification. We demonstrate how to visualize and understand three important DNN models: convolutional, recurrent, and convolutional-recurrent networks. Our first visualization method is finding a test sequence's saliency map which uses first-order derivatives to describe the importance of each nucleotide in making the final prediction. Second, considering recurrent models make predictions in a temporal manner (from one end of a TFBS sequence to the other), we introduce temporal output scores, indicating the prediction score of a model over time for a sequential input. Lastly, a class-specific visualization strategy finds the optimal input sequence for a given TFBS positive class via stochastic gradient optimization. Our experimental results indicate that a convolutional-recurrent architecture performs the best among the three architectures. The visualization techniques indicate that CNN-RNN makes predictions by modeling both motifs as well as dependencies among them.

The utility of multi-cohort two-class meta-analysis to identify robust differentially expressed gene signatures has been well established. However, many biomedical applications, such as gene signatures of disease progression, require one-class analysis. Here we describe an R package, MetaCorrelator, that can identify a reproducible transcriptional signature that is correlated with a continuous disease phenotype across multiple datasets. We successfully applied this framework to extract a pattern of gene expression that can predict lung function in patients with chronic obstructive pulmonary disease (COPD) in both peripheral blood mononuclear cells (PBMCs) and tissue. Our results point to a disregulation in the oxidation state of the lungs of patients with COPD, as well as underscore the classically recognized inammatory state that underlies this disease.

Reduction of preventable hospital readmissions that result from chronic or acute conditions like stroke, heart failure, myocardial infarction and pneumonia remains a significant challenge for improving the outcomes and decreasing the cost of healthcare delivery in the United States. Patient readmission rates are relatively high for conditions like heart failure (HF) despite the implementation of high-quality healthcare delivery operation guidelines created by regulatory authorities. Multiple predictive models are currently available to evaluate potential 30-day readmission rates of patients. Most of these models are hypothesis driven and repetitively assess the predictive abilities of the same set of biomarkers as predictive features. In this manuscript, we discuss our attempt to develop a data-driven, electronic-medical record-wide (EMR-wide) feature selection approach and subsequent machine learning to predict readmission probabilities. We have assessed a large repertoire of variables from electronic medical records of heart failure patients in a single center. The cohort included 1,068 patients with 178 patients were readmitted within a 30-day interval (16.66% readmission rate). A total of 4,205 variables were extracted from EMR including diagnosis codes (n=1,763), medications (n=1,028), laboratory measurements (n=846), surgical procedures (n=564) and vital signs (n=4). We designed a multistep modeling strategy using the Naïve Bayes algorithm. In the first step, we created individual models to classify the cases (readmitted) and controls (non-readmitted). In the second step, features contributing to predictive risk from independent models were combined into a composite model using a correlation-based feature selection (CFS) method. All models were trained and tested using a 5-fold cross-validation method, with 70% of the cohort used for training and the remaining 30% for testing. Compared to existing predictive models for HF readmission rates (AUCs in the range of 0.6-0.7), results from our EMR-wide predictive model (AUC=0.78; Accuracy=83.19%) and phenome-wide feature selection strategies are encouraging and reveal the utility of such datadriven machine learning. Fine tuning of the model, replication using multi-center cohorts and prospective clinical trial to evaluate the clinical utility would help the adoption of the model as a clinical decision system for evaluating readmission status.

Prediction problems in biomedical sciences are generally quite difficult, partially due to incomplete knowledge of how the phenomenon of interest is influenced by the variables and measurements used for prediction, as well as a lack of consensus regarding the ideal predictor(s) for specific problems. In these situations, a powerful approach to improving prediction performance is to construct ensembles that combine the outputs of many individual base predictors, which have been successful for many biomedical prediction tasks. Moreover, selecting a parsimonious ensemble can be of even greater value for biomedical sciences, where it is not only important to learn an accurate predictor, but also to interpret what novel knowledge it can provide about the target problem. Ensemble selection is a promising approach for this task because of its ability to select a collectively predictive subset, often a relatively small one, of all input base predictors. One of the most well-known algorithms for ensemble selection, CES (Caruana et al.’s Ensemble Selection), generally performs well in practice, but faces several challenges due to the difficulty of choosing the right values of its various parameters. Since the choices made for these parameters are usually ad-hoc, good performance of CES is difficult to guarantee for a variety of problems or datasets. To address these challenges with CES and other such algorithms, we propose a novel heterogeneous ensemble selection approach based on the paradigm of reinforcement learning (RL), which offers a more systematic and mathematically sound methodology for exploring the many possible combinations of base predictors that can be selected into an ensemble. We develop three RL-based strategies for constructing ensembles and analyze their results on two unbalanced computational genomics problems, namely the prediction of protein function and splice sites in eukaryotic genomes. We show that the resultant ensembles are indeed substantially more parsimonious as compared to the full set of base predictors, yet still offer almost the same classification power, especially for larger datasets. The RL ensembles also yield a better combination of parsimony and predictive performance as compared to CES.

In our recent Asthma Mobile Health Study (AMHS), thousands of asthma patients across the country contributed medical data through the iPhone Asthma Health App on a daily basis for an extended period of time. The collected data included daily self-reported asthma symptoms, symptom triggers, and real time geographic location information. The AMHS is just one of many studies occurring in the context of now many thousands of mobile health apps aimed at improving wellness and better managing chronic disease conditions, leveraging the passive and active collection of data from mobile, handheld smart devices. The ability to identify patient groups or patterns of symptoms that might predict adverse outcomes such as asthma exacerbations or hospitalizations from these types of large, prospectively collected data sets, would be of significant general interest. However, conventional clustering methods cannot be applied to these types of longitudinally collected data, especially survey data actively collected from app users, given heterogeneous patterns of missing values due to: 1) varying survey response rates among different users, 2) varying survey response rates over time of each user, and 3) non-overlapping periods of enrollment among different users. To handle such complicated missing data structure, we proposed a probability imputation model to infer missing data. We also employed a consensus clustering strategy in tandem with the multiple imputation procedure. Through simulation studies under a range of scenarios reflecting real data conditions, we identified favorable performance of the proposed method over other strategies that impute the missing value through low-rank matrix completion. When applying the proposed new method to study asthma triggers and symptoms collected as part of the AMHS, we identified several patient groups with distinct phenotype patterns. Further validation of the methods described in this paper might be used to identify clinically important patterns in large data sets with complicated missing data structure, improving the ability to use such data sets to identify at-risk populations for potential intervention.

A new computational method is presented to extract disease patterns from heterogeneous and text-based data. For this study, 22 million PubMed records were mined for co-occurrences of gene name synonyms and disease MeSH terms. The resulting publication counts were transferred into a matrix M_data. In this matrix, a disease was represented by a row and a gene by a column. Each field in the matrix represented the publication count for a co-occurring disease–gene pair. A second matrix with identical dimensions M_relevance was derived from M_data. To create M_relevance the values from M_data were normalized. The normalized values were multiplied by the column-wise calculated Gini coefficient. This multiplication resulted in a relevance estimator for every gene in relation to a disease. From M_relevance the similarities between all row vectors were calculated. The resulting similarity matrix Srelevance related 5,000 diseases by the relevance estimators calculated for 15,000 genes. Three diseases were analyzed in detail for the validation of the disease patterns and the relevant genes. Cytoscape was used to visualize and to analyze M_relevance and S_relevance together with the genes and diseases. Summarizing the results, it can be stated that the relevance estimator introduced here was able to detect valid disease patterns and to identify genes that encoded key proteins and potential targets for drug discovery projects.

Intimate partner violence (IPV) is a serious problem with devastating health consequences. Screening procedures may overlook relationships between IPV and negative health effects. To identify IPV-associated women’s health issues, we mined national, aggregated de-identified electronic health record data and compared female health issues of domestic abuse (DA) versus non-DA records, identifying terms significantly more frequent for the DA group. After coding these terms into 28 broad categories, we developed a network map to determine strength of relationships between categories in the context of DA, finding that acute conditions are strongly connected to cardiovascular, gastrointestinal, gynecological, and neurological conditions among victims.

Advances in cellular, molecular, and disease biology depend on the comprehensive characterization of gene interactions and pathways. Traditionally, these pathways are curated manually, limiting their efficient annotation and, potentially, reinforcing field-specific bias. Here, in order to test objective and automated identification of functionally cooperative genes, we compared a novel algorithm with three established methods to search for communities within gene interaction networks. Communities identified by the novel approach and by one of the established method overlapped significantly (q < 0.1) with control pathways. With respect to disease, these communities were biased to genes with pathogenic variants in ClinVar (p ≪ 0.01), and often genes from the same community were co-expressed, including in breast cancers. The interesting subset of novel communities, defined by poor overlap to control pathways also contained co-expressed genes, consistent with a possible functional role. This work shows that community detection based on topological features of networks suggests new, biologically meaningful groupings of genes that, in turn, point to health and disease relevant hypotheses.

The major goal of precision medicine is to improve human health. A feature that unites much research in the field is the use of large datasets such as genomic data and electronic health records. Research in this field includes examination of variation in the core bases of DNA and their methylation status, through variations in metabolic and signaling molecules, all the way up to broader systems level changes in physiology and disease presentation. Intermediate goals include understanding the individual drivers of disease that differentiate the cause of disease in each individual. To match this development of approaches to physical and activitybased measurements, computational approaches to using these new streams of data to better understand improve human health are being rapidly developed by the thriving biomedical informatics research community. This session of the 2017 Pacific Symposium of Biocomputing presents some of the latest advances in the capture, analysis and use of diverse biomedical data in precision medicine.

The past decade has seen exponential growth in the numbers of sequenced and genotyped individuals and a corresponding increase in our ability of collect and catalogue phenotypic data for use in the clinic. We now face the challenge of integrating these diverse data in new ways new that can provide useful diagnostics and precise medical interventions for individual patients. One of the first steps in this process is to accurately map the phenotypic consequences of the genetic variation in human populations. The most common approach for this is the genome wide association study (GWAS). While this technique is relatively simple to implement for a given phenotype, the choice of how to define a phenotype is critical. It is becoming increasingly common for each individual in a GWAS cohort to have a large profile of quantitative measures. The standard approach is to test for associations with one measure at a time; however, there are many justifiable ways to define a set of phenotypes, and the genetic associations that are revealed will vary based on these definitions. Some phenotypes may only show a significant genetic association signal when considered together, such as through principle components analysis (PCA). Combining correlated measures may increase the power to detect association by reducing the noise present in individual variables and reduce the multiple hypothesis testing burden. Here we show that PCA and k-means clustering are two complimentary methods for identifying novel genotype-phenotype relationships within a set of quantitative human traits derived from the Geisinger Health System electronic health record (EHR). Using a diverse set of approaches for defining phenotype may yield more insights into the genetic architecture of complex traits and the findings presented here highlight a clear need for further investigation into other methods for defining the most relevant phenotypes in a set of variables. As the data of EHR continue to grow, addressing these issues will become increasingly important in our efforts to use genomic data effectively in medicine.

The use of posterior probabilities to summarize genotype uncertainty is pervasive across genotype, sequencing and imputation platforms. Prior work in many contexts has shown the utility of incorporating genotype uncertainty (posterior probabilities) in downstream statistical tests. Typical approaches to incorporating genotype uncertainty when testing Hardy-Weinberg equilibrium tend to lack calibration in the type I error rate, especially as genotype uncertainty increases. We propose a new approach in the spirit of genomic control that properly calibrates the type I error rate, while yielding improved power to detect deviations from Hardy-Weinberg Equilibrium. We demonstrate the improved performance of our method on both simulated and real genotypes.

Most studies of disease etiologies focus on one disease only and not the full spectrum of multimorbidities that many patients have. Some disease pairs have shared causal origins, others represent common follow-on diseases, while yet other co-occurring diseases may manifest themselves in random order of appearance. We discuss these different types of disease co-occurrences, and use the two diseases "sleep apnea" and "diabetes" to showcase the approach which otherwise can be applied to any disease pair. We benefit from seven million electronic medical records covering the entire population of Denmark for more than 20 years. Sleep apnea is the most common sleep-related breathing disorder and it has previously been shown to be bidirectionally linked to diabetes, meaning that each disease increases the risk of acquiring the other. We confirm that there is no significant temporal relationship, as approximately half of patients with both diseases are diagnosed with diabetes first. However, we also show that patients diagnosed with diabetes before sleep apnea have a higher disease burden compared to patients diagnosed with sleep apnea before diabetes. The study clearly demonstrates that it is not only the diagnoses in the patient’s disease history that are important, but also the specific order in which these diagnosis are given that matters in terms of outcome. We suggest that this should be considered for patient stratification.

MicroRNAs play important roles in the development of many complex diseases. Because of their importance, the analysis of signaling pathways including miRNA interactions holds the potential for unveiling the mechanisms underlying such diseases. However, current signaling pathway databases are limited to interactions between genes and ignore miRNAs. Here, we use the information on miRNA targets to build a database of miRNA-augmented pathways (mirAP), and we show its application in the contexts of integrative pathway analysis and disease subtyping. Our miRNA-mRNA integrative pathway analysis pipeline incorporates a topology-aware approach that we previously implemented. Our integrative disease subtyping pipeline takes into account survival data, gene and miRNA expression, and knowledge of the interactions among genes. We demonstrate the advantages of our approach by analyzing nine sample-matched datasets that provide both miRNA and mRNA expression. We show that integrating miRNAs into pathway analysis results in greater statistical power, and provides a more comprehensive view of the underlying phenomena. We also compare our disease subtyping method with the state-of-the-art integrative analysis by analyzing a colorectal cancer database from TCGA. The colorectal cancer subtypes identified by our approach are significantly different in terms of their survival expectation. These miRNA-augmented pathways offer a more comprehensive view and a deeper understanding of biological pathways. A better understanding of the molecular processes associated with patients' survival can help to a better prognosis and an appropriate treatment for each subtype.

Motivation: Large scale genomics studies have generated comprehensive molecular characterization of numerous cancer types. Subtypes for many tumor types have been established; however, these classifications are based on molecular characteristics of a small gene sets with limited power to detect dysregulation at the patient level. We hypothesize that frequent graph mining of pathways to gather pathways functionally relevant to tumors can characterize tumor types and provide opportunities for personalized therapies.

Results: In this study we present an integrative omics approach to group patients based on their altered pathway characteristics and show prognostic differences within breast cancer (p < 9:57E - 10) and glioblastoma multiforme (p < 0:05) patients. We were able validate this approach in secondary RNA-Seq datasets with p < 0:05 and p < 0:01 respectively. We also performed pathway enrichment analysis to further investigate the biological relevance of dysregulated pathways. We compared our approach with network-based classifier algorithms and showed that our unsupervised approach generates more robust and biologically relevant clustering whereas previous approaches failed to report specific functions for similar patient groups or classify patients into prognostic groups.

Conclusions: These results could serve as a means to improve prognosis for future cancer patients, and to provide opportunities for improved treatment options and personalized interventions. The proposed novel graph mining approach is able to integrate PPI networks with gene expression in a biologically sound approach and cluster patients in to clinically distinct groups. We have utilized breast cancer and glioblastoma multiforme datasets from microarray and RNA-Seq platforms and identified disease mechanisms differentiating samples. Supplementary information: Supplementary methods, figures, tables and code are available at https://github.com/bebeklab/dysprog.

The discovery of driver genes is a major pursuit of cancer genomics, usually based on observing the same mutation in different patients. But the heterogeneity of cancer pathways plus the high background mutational frequency of tumor cells often cloud the distinction between less frequent drivers and innocent passenger mutations. Here, to overcome these disadvantages, we grouped together mutations from close kinase paralogs under the hypothesis that cognate mutations may functionally favor cancer cells in similar ways. Indeed, we find that kinase paralogs often bear mutations to the same substituted amino acid at the same aligned positions and with a large predicted Evolutionary Action. Functionally, these high Evolutionary Action, non-random mutations affect known kinase motifs, but strikingly, they do so differently among different kinase types and cancers, consistent with differences in selective pressures. Taken together, these results suggest that cancer pathways may flexibly distribute a dependence on a given functional mutation among multiple close kinase paralogs. The recognition of this “mutational delocalization” of cancer drivers among groups of paralogs is a new phenomena that may help better identify relevant mechanisms and therefore eventually guide personalized therapy.

Quantitative genetic trait prediction based on high-density genotyping arrays plays an important role for plant and animal breeding, as well as genetic epidemiology such as complex diseases. The prediction can be very helpful to develop breeding strategies and is crucial to translate the findings in genetics to precision medicine. Epistasis, the phenomena where the SNPs interact with each other, has been studied extensively in Genome Wide Association Studies (GWAS) but received relatively less attention for quantitative genetic trait prediction. As the number of possible interactions is generally extremely large, even pairwise interactions is very challenging. To our knowledge, there is no solid solution yet to utilize epistasis to improve genetic trait prediction. In this work, we studied the multi-locus epistasis problem where the interactions with more than two SNPs are considered. We developed an effcient algorithm MUSE to improve the genetic trait prediction with the help of multi-locus epistasis. MUSE is sampling-based and we proposed a few different sampling strategies. Our experiments on real data showed that MUSE is not only effcient but also effective to improve the genetic trait prediction. MUSE also achieved very significant improvements on a real plant data set as well as a real human data set.

Given the diverse molecular pathways involved in tumorigenesis, identifying subgroups among cancer patients is crucial in precision medicine. While most targeted therapies rely on DNA mutation status in tumors, responses to such therapies vary due to the many molecular processes involved in propagating DNA changes to proteins (which constitute the usual drug targets). Though RNA expressions have been extensively used to categorize tumors, identifying clinically important subgroups remains challenging given the difficulty of discerning subgroups within all possible RNA-RNA networks. It is thus essential to incorporate multiple types of data. Recently, RNA was found to regulate other RNA through a common microRNA (miR). These regulating and regulated RNAs are referred to as competing endogenous RNAs (ceRNAs). However, global correlations between mRNA and miR expressions across all samples have not reliably yielded ceRNAs. In this study, we developed a ceRNA-based method to identify subgroups of cancer patients combining DNA copy number variation, mRNA expression, and microRNA (miR) expression data with biological knowledge. Clinical data is used to validate identified subgroups and ceRNAs. Since ceRNAs are causal, ceRNA-based subgroups may present clinical relevance. Using lung adenocarcinoma data from The Cancer Genome Atlas (TCGA) as an example, we focused on EGFR amplification status, since a targeted therapy for EGFR exists. We hypothesized that global correlations between mRNA and miR expressions across all patients would not reveal important subgroups and that clustering of potential ceRNAs might define molecular pathway-relevant subgroups. Using experimentally validated miR-target pairs, we identified EGFR and MET as potential ceRNAs for miR-133b in lung adenocarcinoma. The EGFR-MET up and miR-133b down subgroup showed a higher death rate than the EGFR-MET down and miR-133b up subgroup. Although transactivation between MET and EGFR has been identified previously, our result is the first to propose ceRNA as one of its underlying mechanisms. Furthermore, since MET amplification was seen in the case of resistance to EGFR-targeted therapy, the EGFR-MET up and miR-133b down subgroup may fall into the drug non-response group and thus preclude EGFR target therapy.

Gene set analysis methods continue to be a popular and powerful method of evaluating genome-wide transcriptomics data. These approach require a priori grouping of genes into biologically meaningful sets, and then conducting downstream analyses at the set (instead of gene) level of analysis. Gene set analysis methods have been shown to yield more powerful statistical conclusions than single-gene analyses due to both reduced multiple testing penalties and potentially larger observed effects due to the aggregation of effects across multiple genes in the set. Traditionally, gene set analysis methods have been applied directly to normalized, log-transformed, transcriptomics data. Recently, efforts have been made to transform transcriptomics data to scales yielding more biologically interpretable results. For example, recently proposed models transform log-transformed transcriptomics data to a confidence metric (ranging between 0 and 100%) that a gene is active (roughly speaking, that the gene product is part of an active cellular mechanism). In this manuscript, we demonstrate, on both real and simulated transcriptomics data, that tests for differential expression between sets of genes using are typically more powerful when using gene activity state estimates as opposed to log-transformed gene expression data. Our analysis suggests further exploration of techniques to transform transcriptomics data to meaningful quantities for improved downstream inference.

DNA methylation has emerged as promising epigenetic markers for disease diagnosis. Both the differential mean (DM) and differential variability (DV) in methylation have been shown to contribute to transcriptional aberration and disease pathogenesis. The presence of confounding factors in large scale EWAS may affect the methylation values and hamper accurate marker discovery. In this paper, we propose a exible framework called methylDMV which allows for confounding factors adjustment and enables simultaneous characterization and identification of CpGs exhibiting DM only, DV only and both DM and DV. The proposed framework also allows for prioritization and selection of candidate features to be included in the prediction algorithm. We illustrate the utility of methylDMV in several TCGA datasets. An R package methylDMV implementing our proposed method is available at http://www.ams.sunysb.edu/~pfkuan/softwares.html#methylDMV.

Genomic sequencing studies in the past several years have yielded a large number of cancer somatic mutations. There remains a major challenge in delineating a small fraction of somatic mutations that are oncogenic drivers from a background of predominantly passenger mutations. Although computational tools have been developed to predict the functional impact of mutations, their utility is limited. In this study, we applied an alternative approach to identify potentially novel cancer drivers as those somatic mutations that overlap with known pathogenic mutations in Mendelian diseases. We hypothesize that those shared mutations are more likely to be cancer drivers because they have the established molecular mechanisms to impact protein functions. We first show that the overlap between somatic mutations in COSMIC and pathogenic genetic variants in HGMD is associated with high mutation frequency in cancers and is enriched for known cancer genes. We then attempted to identify putative tumor suppressors based on the number of distinct HGMD/COSMIC overlapping mutations in a given gene, and our results suggest that ion channels, collagens and Marfan syndrome associated genes may represent new classes of tumor suppressors. To elucidate potentially novel oncogenes, we identified those HGMD/COSMIC overlapping mutations that are not only highly recurrent but also mutually exclusive from previously characterized oncogenic mutations in each specific cancer type. Taken together, our study represents a novel approach to discover new cancer genes from the vast amount of cancer genome sequencing data.

Cancer metabolism differs remarkably from the metabolism of healthy surrounding tissues, and it is extremely heterogeneous across cancer types. While these metabolic differences provide promising avenues for cancer treatments, much work remains to be done in understanding how metabolism is rewired in malignant tissues. To that end, constraint-based models provide a powerful computational tool for the study of metabolism at the genome scale. To generate meaningful predictions, however, these generalized human models must first be tailored for specific cell or tissue sub-types. Here we first present two improved algorithms for (1) the generation of these context-specific metabolic models based on omics data, and (2) Monte-Carlo sampling of the metabolic model ux space. By applying these methods to generate and analyze context-specific metabolic models of diverse solid cancer cell line data, and primary leukemia pediatric patient biopsies, we demonstrate how the methodology presented in this study can generate insights into the rewiring differences across solid tumors and blood cancers.

The effort to personalize treatment plans for cancer patients involves the identification of drug treatments that can effectively target the disease while minimizing the likelihood of adverse reactions. In this study, the gene-expression profile of 810 cancer cell lines and their response data to 368 small molecules from the Cancer Therapeutics Research Portal (CTRP) are analyzed to identify pathways with significant rewiring between genes, or differential gene dependency, between sensitive and non-sensitive cell lines. Identified pathways and their corresponding differential dependency networks are further analyzed to discover essentiality and specificity mediators of cell line response to drugs/compounds. For analysis we use the previously published method EDDY (Evaluation of Differential DependencY). EDDY first constructs likelihood distributions of gene-dependency networks, aided by known genegene interaction, for two given conditions, for example, sensitive cell lines vs. non-sensitive cell lines. These sets of networks yield a divergence value between two distributions of network likelihoods that can be assessed for significance using permutation tests. Resulting differential dependency networks are then further analyzed to identify genes, termed mediators, which may play important roles in biological signaling in certain cell lines that are sensitive or non-sensitive to the drugs. Establishing statistical correspondence between compounds and mediators can improve understanding of known gene dependencies associated with drug response while also discovering new dependencies. Millions of compute hours resulted in thousands of these statistical discoveries. EDDY identified 8,811 statistically significant pathways leading to 26,822 compound-pathway-mediator triplets. By incorporating STITCH and STRING databases, we could construct evidence networks for 14,415 compound-pathway-mediator triplets for support. The results of this analysis are presented in a searchable website to aid researchers in studying potential molecular mechanisms underlying cells’ drug response as well as in designing experiments for the purpose of personalized treatment regimens.

Many researchers now have available multiple high-dimensional molecular and clinical datasets when studying a disease. As we enter this multi-omic era of data analysis, new approaches that combine different levels of data (e.g. at the genomic and epigenomic levels) are required to fully capitalize on this opportunity. In this work, we outline a new approach to multi-omic data integration, which combines molecular and clinical predictors as part of a single analysis to create a prognostic risk score for clear cell renal cell carcinoma. The approach integrates data in multiple ways and yet creates models that are relatively straightforward to interpret and with a high level of performance. Furthermore, the proposed process of data integration captures relationships in the data that represent highly disease-relevant functions.

Autism has been shown to have a major genetic risk component; the architecture of documented autism in families has been over and again shown to be passed down for generations. While inherited risk plays an important role in the autistic nature of children, de novo (germline) mutations have also been implicated in autism risk. Here we find that autism de novo variants verified and published in the literature are Bonferroni-significantly enriched in a gene set implicated in synaptic elimination. Additionally, several of the genes in this synaptic elimination set that were enriched in protein-protein interactions (CACNA1C, SHANK2, SYNGAP1, NLGN3, NRXN1, and PTEN) have been previously confirmed as genes that confer risk for the disorder. The results demonstrate that autism-associated de novos are linked to proper synaptic pruning and density, hinting at the etiology of autism and suggesting pathophysiology for downstream correction and treatment.

A wide range of patient health data is recorded in Electronic Health Records (EHR). This data includes diagnosis, surgical procedures, clinical laboratory measurements, and medication information. Together this information reflects the patient’s medical history. Many studies have efficiently used this data from the EHR to find associations that are clinically relevant, either by utilizing International Classification of Diseases, version 9 (ICD-9) codes or laboratory measurements, or by designing phenotype algorithms to extract case and control status with accuracy from the EHR. Here we developed a strategy to utilize longitudinal quantitative trait data from the EHR at Geisinger Health System focusing on outpatient metabolic and complete blood panel data as a starting point. Comprehensive Metabolic Panel (CMP) as well as Complete Blood Counts (CBC) are parts of routine care and provide a comprehensive picture from high level screening of patients’ overall health and disease. We randomly split our data into two datasets to allow for discovery and replication. We first conducted a genome-wide association study (GWAS) with median values of 25 different clinical laboratory measurements to identify variants from Human Omni Express Exome beadchip data that are associated with these measurements. We identified 687 variants that associated and replicated with the tested clinical measurements at p<5×10^-08. Since longitudinal data from the EHR provides a record of a patient’s medical history, we utilized this information to further investigate the ICD-9 codes that might be associated with differences in variability of the measurements in the longitudinal dataset. We identified low and high variance patients by looking at changes within their individual longitudinal EHR laboratory results for each of the 25 clinical lab values (thus creating 50 groups – a high variance and a low variance for each lab variable). We then performed a PheWAS analysis with ICD-9 diagnosis codes, separately in the high variance group and the low variance group for each lab variable. We found 717 PheWAS associations that replicated at a p-value less than 0.001. Next, we evaluated the results of this study by comparing the association results between the high and low variance groups. For example, we found 39 SNPs (in multiple genes) associated with ICD-9 250.01 (Type-I diabetes) in patients with high variance of plasma glucose levels, but not in patients with low variance in plasma glucose levels. Another example is the association of 4 SNPs in UMOD with chronic kidney disease in patients with high variance for aspartate aminotransferase (discovery p-value: 8.71×10^-09 and replication p-value: 2.03×10^-06). In general, we see a pattern of many more statistically significant associations from patients with high variance in the quantitative lab variables, in comparison with the low variance group across all of the 25 laboratory measurements. This study is one of the first of its kind to utilize quantitative trait variance from longitudinal laboratory data to find associations among genetic variants and clinical phenotypes obtained from an EHR, integrating laboratory values and diagnosis codes to understand the genetic complexities of common diseases.

The blood thinner warfarin has a narrow therapeutic range and high inter- and intra-patient variability in therapeutic doses. Several studies have shown that pharmacogenomic variants help predict stable warfarin dosing. However, retrospective and randomized controlled trials that employ dosing algorithms incorporating pharmacogenomic variants under perform in African Americans. This study sought to determine if: 1) including additional variants associated with warfarin dose in African Americans, 2) predicting within single ancestry groups rather than a combined population, or 3) using percentage African ancestry rather than observed race, would improve warfarin dosing algorithms in African Americans. Using BioVU, the Vanderbilt University Medical Center biobank linked to electronic medical records, we compared 25 modeling strategies to existing algorithms using a cohort of 2,181 warfarin users (1,928 whites, 253 blacks). We found that approaches incorporating additional variants increased model accuracy, but not in clinically significant ways. Race stratification increased model fidelity for African Americans, but the improvement was small and not likely to be clinically significant. Use of percent African ancestry improved model fit in the context of race misclassification.

Recent technological developments allow gathering single-cell measurements across different domains (genomic, transcriptomics, proteomics, imaging etc). Sophisticated computational algorithms are required in order to harness the power of single-cell data. This session is dedicated to computational methods for single-cell analysis in various biological domains, modelling of population heterogeneity, as well as translational applications of single cell data.

Next generation sequencing of the RNA content of single cells or single nuclei (sc/nRNA-seq) has become a powerful approach to understand the cellular complexity and diversity of multicellular organisms and environmental ecosystems. However, the fact that the procedure begins with a relatively small amount of starting material, thereby pushing the limits of the laboratory procedures required, dictates that careful approaches for sample quality control (QC) are essential to reduce the impact of technical noise and sample bias in downstream analysis applications. Here we present a preliminary framework for sample level quality control that is based on the collection of a series of quantitative laboratory and data metrics that are used as features for the construction of QC classification models using random forest machine learning approaches. We’ve applied this initial framework to a dataset comprised of 2272 single nuclei RNA-seq results and determined that ~79% of samples were of high quality. Removal of the poor quality samples from downstream analysis was found to improve the cell type clustering results. In addition, this approach identified quantitative features related to the proportion of unique or duplicate reads and the proportion of reads remaining after quality trimming as useful features for pass/fail classification. The construction and use of classification models for the identification of poor quality samples provides for an objective and scalable approach to sc/nRNA-seq quality control.

The availability of gene expression data at the single cell level makes it possible to probe the molecular underpinnings of complex biological processes such as differentiation and oncogenesis. Promising new methods have emerged for reconstructing a progression ’trajectory’ from static single-cell transcriptome measurements. However, it remains unclear how to adequately model the appreciable level of noise in these data to elucidate gene regulatory network rewiring. Here, we present a framework called Single Cell Inference of MorphIng Trajectories and their Associated Regulation (SCIMITAR) that infers progressions from static single-cell transcriptomes by employing a continuous parametrization of Gaussian mixtures in high-dimensional curves. SCIMITAR yields rich models from the data that highlight genes with expression and co-expression patterns that are associated with the inferred progression. Further, SCIMITAR extracts regulatory states from the implicated trajectory-evolvingco-expression networks. We benchmark the method on simulated data to show that it yields accurate cell ordering and gene network inferences. Applied to the interpretation of a single-cell human fetal neuron dataset, SCIMITAR finds progression-associated genes in cornerstone neural differentiation pathways missed by standard differential expression tests. Finally, by leveraging the rewiring of gene-gene co-expression relations across the progression, the method reveals the rise and fall of co-regulatory states and trajectory-dependent gene modules. These analyses implicate new transcription factors in neural differentiation including putative co-factors for the multi-functional NFAT pathway.

Pooled sample analysis by mass cytometry barcoding carries many advantages: reduced antibody consumption, increased sample throughput, removal of cell doublets, reduction of cross-contamination by sample carryover, and the elimination of tube-to-tube-variability in antibody staining. A single-cell debarcoding algorithm was previously developed to improve the accuracy and yield of sample deconvolution, but this method was limited to using fixed parameters for debarcoding stringency filtering, which could introduce cell-specific or sample-specific bias to cell yield in scenarios where barcode staining intensity and variance are not uniform across the pooled samples. To address this issue, we have updated the algorithm to output debarcoding parameters for every cell in the sample-assigned FCS files, which allows for visualization and analysis of these parameters via flow cytometry analysis software. This strategy can be used to detect cell type-specific and sample-specific effects on the underlying cell data that arise during the debarcoding process. An additional benefit to this strategy is the decoupling of barcode stringency filtering from the debarcoding and sample assignment process. This is accomplished by removing the stringency filters during sample assignment, and then filtering after the fact with 1- and 2-dimensional gating on the debarcoding parameters which are output with the FCS files. These data exploration strategies serve as an important quality check for barcoded mass cytometry datasets, and allow cell type and sample-specific stringency adjustment that can remove bias in cell yield introduced during the debarcoding process.

Mouse brain transcriptomic studies are important in the understanding of the structural heterogeneity in the brain. However, it is not well understood how cell types in the mouse brain relate to human brain cell types on a cellular level. We propose that it is possible with single cell granularity to find concordant genes between mouse and human and that these genes can be used to separate cell types across species. We show that a set of concordant genes can be algorithmically derived from a combination of human and mouse single cell sequencing data. Using this gene set, we show that similar cell types shared between mouse and human cluster together. Furthermore we find that previously unclassified human cells can be mapped to the glial/vascular cell type by integrating mouse cell type expression profiles.

Tumors are composed of heterogeneous populations of cells. Somatic genetic aberrations are one form of heterogeneity that allows clonal cells to adapt to chemotherapeutic stress, thus providing a path for resistance to arise. In silico modeling of tumors provides a platform for rapid, quantitative experiments to inexpensively study how compositional heterogeneity contributes to drug resistance. Accordingly, we have built a spatiotemporal model of a lung metastasis originating from a primary bladder tumor, incorporating in vivo drug concentrations of first-line chemotherapy, resistance data from bladder cancer cell lines, vascular density of lung metastases, and gains in resistance in cells that survive chemotherapy. In metastatic bladder cancer, a first-line drug regimen includes six cycles of gemcitabine plus cisplatin (GC) delivered simultaneously on day 1, and gemcitabine on day 8 in each 21-day cycle. The interaction between gemcitabine and cisplatin has been shown to be synergistic in vitro, and results in better outcomes in patients. Our model shows that during simulated treatment with this regimen, GC synergy does begin to kill cells that are more resistant to cisplatin, but repopulation by resistant cells occurs. Post-regimen populations are mixtures of the original, seeded resistant clones, and/or new clones that have gained resistance to cisplatin, gemcitabine, or both drugs. The emergence of a tumor with increased resistance is qualitatively consistent with the five-year survival of 6.8% for patients with metastatic transitional cell carcinoma of the urinary bladder treated with a GC regimen. The model can be further used to explore the parameter space for clinically relevant variables, including the timing of drug delivery to optimize cell death, and patient-specific data such as vascular density, rates of resistance gain, disease progression, and molecular profiles, and can be expanded for data on toxicity. The model is specific to bladder cancer, which has not previously been modeled in this context, but can be adapted to represent other cancers.

Single-cell analysis can uncover the mysteries in the state of individual cells and enable us to construct new models about the analysis of heterogeneous tissues. State-of-the-art technologies for single-cell analysis have been developed to measure the properties of single-cells and detect hidden information. They are able to provide the measurements of dozens of features simultaneously in each cell. However, due to the high-dimensionality, heterogeneous complexity and sheer enormity of single-cell data, its interpretation is challenging. Thus, new methods to overcome high-dimensionality are necessary. Here, we present a computational tool that allows efficient visualization of high-dimensional single-cell data onto a low-dimensional (2D or 3D) space while preserving the similarity structure between single-cells. We first construct a network that can represent the similarity structure between the high-dimensional representations of single-cells, and then, embed this network into a low-dimensional space through an efficient online optimization method based on the idea of negative sampling. Using this approach, we can preserve the high-dimensional structure of single-cell data in an embedded low-dimensional space that facilitates visual analyses of the data.

Precision medicine is a health management approach that accounts for individual differences in genetic backgrounds and environmental exposures. With the recent advancements in high-throughput omics profiling technologies, collections of large study cohorts, and the developments of data mining algorithms, big data in biomedicine is expected to provide novel insights into health and disease states, which can be translated into personalized disease prevention and treatment plans. However, petabytes of biomedical data generated by multiple measurement modalities poses a significant challenge for data analysis, integration, storage, and result interpretation. In addition, patient privacy preservation, coordination between participating medical centers and data analysis working groups, as well as discrepancies in data sharing policies remain important topics of discussion. In this workshop, we invite experts in omics integration, biobank research, and data management to share their perspectives on leveraging big data to enable precision medicine.

Workshop website: http://tinyurl.com/PSB17BigData; HashTag: #PSB17BigData.

With the booming of new technologies, biomedical science has transformed into digitalized, data intensive science. Massive amount of data need to be analyzed and interpreted, demand a complete pipeline to train next generation data scientists. To meet this need, the transinstitutional Big Data to Knowledge (BD2K) Initiative has been implemented since 2014, complementing other NIH institutional efforts. In this report, we give an overview the BD2K K01 mentored scientist career awards, which have demonstrated early success. We address the specific trainings needed in representative data science areas, in order to make the next generation of data scientists in biomedicine.

The following sections are included:

• Bioinformatics is a Mature Discipline
• The Golden Era of Bioinformatics Has Begun
• No-Boundary Thinking in Bioinformatics
• References

The modern healthcare and life sciences ecosystem is moving towards an increasingly open and data-centric approach to discovery science. This evolving paradigm is predicated on a complex set of information needs related to our collective ability to share, discover, reuse, integrate, and analyze open biological, clinical, and population level data resources of varying composition, granularity, and syntactic or semantic consistency. Such an evolution is further impacted by a concomitant growth in the size of data sets that can and should be employed for both hypothesis discovery and testing. When such open data can be accessed and employed for discovery purposes, a broad spectrum of high impact end-points is made possible. These span the spectrum from identification of de novo biomarker complexes that can inform precision medicine, to the repositioning or repurposing of extant agents for new and cost-effective therapies, to the assessment of population level influences on disease and wellness. Of note, these types of uses of open data can be either primary, wherein open data is the substantive basis for inquiry, or secondary, wherein open data is used to augment or enrich project-specific or proprietary data that is not open in and of itself. This workshop is concerned with the key challenges, opportunities, and methodological best practices whereby open data can be used to drive the advancement of discovery science in all of the aforementioned capacities.