Gregory P. Way

Opportunities and obstacles for deep learning in biology and medicine

Deep learning describes a class of machine learning algorithms that are capable of combining raw inputs into layers of intermediate features. These algorithms have recently shown impressive results across a variety of domains. Biology and medicine are data-rich disciplines, but the data are complex and often ill-understood. Hence, deep learning techniques may be particularly well suited to solve problems of these fields. We examine applications of deep learning to a variety of biomedical problems—patient classification, fundamental biological processes and treatment of patients—and discuss whether deep learning will be able to transform these tasks or if the biomedical sphere poses unique challenges. Following from an extensive literature review, we find that deep learning has yet to revolutionize biomedicine or definitively resolve any of the most pressing challenges in the field, but promising advances have been made on the prior state of the art. Even though improvements over previous baselines have been modest in general, the recent progress indicates that deep learning methods will provide valuable means for speeding up or aiding human investigation. Though progress has been made linking a specific neural networks prediction to input features, understanding how users should interpret these models to make testable hypotheses about the system under study remains an open challenge. Furthermore, the limited amount of labelled data for training presents problems in some domains, as do legal and privacy constraints on work with sensitive health records. Nonetheless, we foresee deep learning enabling changes at both bench and bedside with the potential to transform several areas of biology and medicine.

Sex differences in a shoaling-boldness behavioral syndrome, but no link with aggression.

A behavioral syndrome is observed in a population when specific behaviors overlap at the individual level in different contexts. Here, we explore boldness and aggression personality spectra, the repeatability of shoaling, and possible associated correlations between the behaviors in a population of lab-reared zebrafish (Danio rerio). Our findings describe a sex-specific boldness-shoaling behavioral syndrome, as a link between boldness and shoaling behaviors is detected. The results indicate that bold males are likely to have a stronger shoaling propensity than shy males for unfamiliar conspecifics. Conversely, bold females are more likely to shoal than shy females, but only when presented with heterospecific individuals. Additionally, aggression does not correlate with boldness or shoaling propensity for either sex. A positive relationship between boldness and shoaling that differs by sex is contrary to most of the present literature, but could help to explain population dynamics and may also have evolutionary implications.

Comprehensive Cross-Population Analysis of High-Grade Serous Ovarian Cancer Supports No More Than Three Subtypes

Four gene expression subtypes of high-grade serous ovarian cancer (HGSC) have been previously described. In these early studies, a fraction of samples that did not fit well into the four subtype classifications were excluded. Therefore, we sought to systematically determine the concordance of transcriptomic HGSC subtypes across populations without removing any samples. We created a bioinformatics pipeline to independently cluster the five largest mRNA expression datasets using k-means and nonnegative matrix factorization (NMF). We summarized differential expression patterns to compare clusters across studies. While previous studies reported four subtypes, our cross-population comparison does not support four. Because these results contrast with previous reports, we attempted to reproduce analyses performed in those studies. Our results suggest that early results favoring four subtypes may have been driven by the inclusion of serous borderline tumors. In summary, our analysis suggests that either two or three, but not four, gene expression subtypes are most consistent across datasets.

A machine learning classifier trained on cancer transcriptomes detects NF1 inactivation signal in glioblastoma

BackgroundWe have identified molecules that exhibit synthetic lethality in cells with loss of the neurofibromin 1 (NF1) tumor suppressor gene. However, recognizing tumors that have inactivation of the NF1 tumor suppressor function is challenging because the loss may occur via mechanisms that do not involve mutation of the genomic locus. Degradation of the NF1 protein, independent of NF1 mutation status, phenocopies inactivating mutations to drive tumors in human glioma cell lines. NF1 inactivation may alter the transcriptional landscape of a tumor and allow a machine learning classifier to detect which tumors will benefit from synthetic lethal molecules.ResultsWe developed a strategy to predict tumors with low NF1 activity and hence tumors that may respond to treatments that target cells lacking NF1. Using RNAseq data from The Cancer Genome Atlas (TCGA), we trained an ensemble of 500 logistic regression classifiers that integrates mutation status with whole transcriptomes to predict NF1 inactivation in glioblastoma (GBM). On TCGA data, the classifier detected NF1 mutated tumors (test set area under the receiver operating characteristic curve (AUROC) mean = 0.77, 95% quantile = 0.53 – 0.95) over 50 random initializations. On RNA-Seq data transformed into the space of gene expression microarrays, this method produced a classifier with similar performance (test set AUROC mean = 0.77, 95% quantile = 0.53 – 0.96). We applied our ensemble classifier trained on the transformed TCGA data to a microarray validation set of 12 samples with matched RNA and NF1 protein-level measurements. The classifier’s NF1 score was associated with NF1 protein concentration in these samples.ConclusionsWe demonstrate that TCGA can be used to train accurate predictors of NF1 inactivation in GBM. The ensemble classifier performed well for samples with very high or very low NF1 protein concentrations but had mixed performance in samples with intermediate NF1 concentrations. Nevertheless, high-performing and validated predictors have the potential to be paired with targeted therapies and personalized medicine.

Challenges and Opportunities in Studying the Epidemiology of Ovarian Cancer Subtypes

Purpose of ReviewOnly recently has it become clear that epithelial ovarian cancer (EOC) is comprised of such distinct histotypes—with different cells of origin, morphology, molecular features, epidemiologic factors, clinical features, and survival patterns—that they can be thought of as different diseases sharing an anatomical location. Herein, we review opportunities and challenges in studying EOC heterogeneity,Recent FindingsThe 2014 World Health Organization diagnostic guidelines incorporate accumulated evidence that high- and low-grade serous tumors have different underlying pathogenesis, and that, on the basis of shared molecular features, most high-grade tumors, including some previously classified as endometrioid, are now considered to be high-grade serous. At the same time, several studies have reported that high-grade serous EOC, which is the most common histotype, is itself made up of reproducible subtypes discernable by gene expression patterns.SummaryThese major advances in understanding set the stage for a new era of research on EOC risk and clinical outcomes with the potential to reduce morbidity and mortality. We highlight the need for multidisciplinary studies with pathology review using the current guidelines, further molecular characterization of the histotypes and subtypes, inclusion of women of diverse racial/ethnic and socioeconomic backgrounds, and updated epidemiologic and clinical data relevant to current generations of women at risk of EOC.

Boldness, Aggression, and Shoaling Assays for Zebrafish Behavioral Syndromes.

A behavioral syndrome exists when specific behaviors interact under different contexts. Zebrafish have been test subjects in recent studies and it is important to standardize protocols to ensure proper analyses and interpretations. In our previous studies, we have measured boldness by monitoring a series of behaviors (time near surface, latency in transitions, number of transitions, and darts) in a 1.5 L trapezoidal tank. Likewise, we quantified aggression by observing bites, lateral displays, darts, and time near an inclined mirror in a rectangular 19 L tank. By dividing a 76 L tank into thirds, we also examined shoaling preferences. The shoaling assay is a highly customizable assay and can be tailored for specific hypotheses. However, protocols for this assay also must be standardized, yet flexible enough for customization. In previous studies, end chambers were either empty, contained 5 or 10 zebrafish, or 5 pearl danios (D. albolineatus). In the following manuscript, we present a detailed protocol and representative data that accompany successful applications of the protocol, which will allow for replication of behavioral syndrome experiments.

Cross-population analysis of high-grade serous ovarian cancer reveals only two robust subtypes

Four gene expression subtypes of high-grade serous ovarian cancer (HGSC) have been previously described. In these studies, a fraction of samples that did not fit well into the four subtype classifications were excluded. Therefore, we sought to systematically determine the concordance of transcriptomic HGSC subtypes across populations without removing any samples. We created a bioinformatics pipeline to independently cluster the five largest mRNA expression datasets using k-means and non-negative matrix factorization (NMF). We summarized differential expression patterns to compare clusters across studies. While previous studies reported four subtypes, our cross-population comparison does not support four. Because these results contrast with previous reports, we attempted to reproduce analyses performed in those studies. Our results suggest that early results favoring four subtypes may have been driven by including serous borderline tumors. In summary, our analysis suggests that either two or three, but not four, gene expression subtypes are most consistent across datasets. CONFLICTS OF INTEREST The authors do not declare any conflicts of interest. OTHER PRESENTATIONS Aspects of this study were presented at the 2015 AACR Conference and the 2015 Rocky Mountain Bioinformatics Conference.Background Four gene expression-based subtypes of high-grade serous ovarian cancer (HGSC), variably associated with differential survival, have been previously described. However, in these studies, clustering heuristics were consistent with only three subtypes and reproducibility of the subtypes across populations and assay platforms has not been formally assessed. Therefore, we systematically determined the concordance of transcriptomic HGSC subtypes across populations. Methods We used a unified bioinformatics pipeline to independently cluster (k = 3 and k = 4) five mRNA expression datasets with >130 tumors using k-means and non-negative matrix factorization (NMF) without removing hard-to-classify samples. Within each population, we summarized differential expression patterns for each cluster as moderated t statistic vectors using Significance Analysis of Microarrays. We calculated Pearsons correlations of these vectors to determine similarities and differences in expression patterns between clusters. We identified sets of clusters that were most correlated across populations to define syn-clusters (SC), and we associated SC expression patterns with biological pathways using geneset overrepresentation analyses. Results Across populations, for k = 3, moderated t score correlations for clusters 1, 2 and 3 ranged between 0.77-0.85, 0.80-0.90, and 0.65-0.77, respectively. For k = 4, correlations for clusters 1-4 were 0.77-0.85, 0.83-0.89, 0.51-0.76, and 0.61-0.75, respectively. Within populations, comparing analogous clusters (k = 3 versus k = 4), correlations were high for clusters 1 and 2 (0.91-1.00), but lower for cluster 3 (0.22-0.80). Results were similar using NMF. SC1 corresponds to mesenchymal-like, SC2 to proliferative-like, SC3 to immunoreactive-like, and SC4 to differentiated-like subtypes reported previously. Conclusions While previous single-population studies reported four HGSC subtypes, our cross-population comparison finds strong evidence for only two subtypes and our re-analysis of previous data suggests that results favoring four subtypes may have been driven, at least in part, by the inclusion of samples with low malignant potential. Because the mesenchymal-like and proliferative-like subtypes are highly consistent across populations, they likely reflect intrinsic biological subtypes and are strong candidates for targeted therapies. The other two previously described subtypes (immunoreactive-like and differentiated-like) are considerably less consistent and may represent either a single subtype or signal that is not amenable to clustering.Background Three to four gene expression-based subtypes of high-grade serous ovarian cancer (HGSC) have been previously reported. We sought to systematically determine the similarity of HGSC subtypes between populations. Methods We performed k-means clustering (k = 3 and k = 4) in five publicly-available HGSC mRNA expression datasets with >130 tumors. Within each population, we summarized differential expression patterns for each cluster as moderated t statistic vectors using Significance Analysis of Microarrays. We calculated Pearsons correlations of these vectors to determine similarities and differences in expression patterns between clusters. We defined syn-clusters (SC) as sets of clusters that were strongly correlated across populations, and associated their expression patterns with biological pathways using geneset overrepresentation analyses. Results Across populations, for k = 3, moderated t score correlations for clusters 1, 2 and 3, respectively, ranged between 0.77-0.85, 0.80-0.90, and 0.65-0.77. For k = 4, correlations for clusters 1-4, respectively, ranged between 0.77-0.85, 0.83-0.89, 0.51-0.76, and 0.61-0.75. Within populations, comparing analogous clusters (k = 3 versus k = 4), correlations were high for clusters 1 and 2 (0.91-1.00), but were lower for cluster 3 (0.22-0.80). The results are similar using non-negative matrix factorization. SC1 corresponds to previously-reported mesenchymal-like, SC2 to proliferative-like, SC3 to immunoreactive-like, and SC4 to differentiated-like subtypes. Conclusions The ability to robustly identify correlated clusters across number of centroids, populations, and clustering methods provides strong evidence that at least three different HGSC subtypes exist. The mesenchymal-like and proliferative-like subtypes are remarkably consistent and could be uniquely targeted for treatment.

Interactions between aggression, boldness and shoaling within a brood of convict cichlids (Amatitlania nigrofasciatus)

A behavioral syndrome is considered present when individuals consistently express correlated behaviors across two or more axes of behavior. These axes of behavior are shy-bold, exploration-avoidance, activity, aggression, and sociability. In this study we examined aggression, boldness and sociability (shoaling) within a juvenile convict cichlid brood (Amatitlania nigrofasciatus). Because young convict cichlids are social, we used methodologies commonly used by ethologists studying social fishes. We did not detect an aggression-boldness behavioral syndrome, but we did find that the aggression, boldness, and possibly the exploration behavioral axes play significant roles in shaping the observed variation in individual convict cichlid behavior. While juvenile convict cichlids did express a shoaling preference, this social preference was likely convoluted by aggressive interactions, despite the small size and young age of the fish. There is a need for the development of behavioral assays that allow for more reliable measurement of behavioral axes in juvenile neo-tropical cichlids.

Deconvolution of DNA methylation identifies differentially methylated gene regions on 1p36 across breast cancer subtypes

Breast cancer is a complex disease consisting of four distinct molecular subtypes. DNA methylation-based (DNAm) studies in tumors are complicated further by disease heterogeneity. In the present study, we compared DNAm in breast tumors with normal-adjacent breast samples from The Cancer Genome Atlas (TCGA). We constructed models stratified by tumor stage and PAM50 molecular subtype and performed cell-type reference-free deconvolution to control for cellular heterogeneity. We identified nineteen differentially methylated gene regions (DMGRs) in early stage tumors across eleven genes (AGRN, C1orf170, FAM41C, FLJ39609, HES4, ISG15, KLHL17, NOC2L, PLEKHN1, SAMD11, WASH5P). These regions were consistently differentially methylated in every subtype and all implicated genes are localized to the chromosomal cytoband 1p36.3. Seventeen of these DMGRs were independently validated in a similar analysis of an external data set. The identification and validation of shared DNAm alterations across tumor subtypes in early stage tumors advances our understanding of common biology underlying breast carcinogenesis and may contribute to biomarker development. We also discuss evidence of the specific importance and potential function of 1p36 in cancer.

PathCORE-T: identifying and visualizing globally co-occurring pathways in large transcriptomic compendia

BackgroundInvestigators often interpret genome-wide data by analyzing the expression levels of genes within pathways. While this within-pathway analysis is routine, the products of any one pathway can affect the activity of other pathways. Past efforts to identify relationships between biological processes have evaluated overlap in knowledge bases or evaluated changes that occur after specific treatments. Individual experiments can highlight condition-specific pathway-pathway relationships; however, constructing a complete network of such relationships across many conditions requires analyzing results from many studies.ResultsWe developed PathCORE-T framework by implementing existing methods to identify pathway-pathway transcriptional relationships evident across a broad data compendium. PathCORE-T is applied to the output of feature construction algorithms; it identifies pairs of pathways observed in features more than expected by chance as functionally co-occurring. We demonstrate PathCORE-T by analyzing an existing eADAGE model of a microbial compendium and building and analyzing NMF features from the TCGA dataset of 33 cancer types. The PathCORE-T framework includes a demonstration web interface, with source code, that users can launch to (1) visualize the network and (2) review the expression levels of associated genes in the original data. PathCORE-T creates and displays the network of globally co-occurring pathways based on features observed in a machine learning analysis of gene expression data.ConclusionsThe PathCORE-T framework identifies transcriptionally co-occurring pathways from the results of unsupervised analysis of gene expression data and visualizes the relationships between pathways as a network. PathCORE-T recapitulated previously described pathway-pathway relationships and suggested experimentally testable additional hypotheses that remain to be explored.