Michael F. Ochs

Outlier analysis and top scoring pair for integrated data analysis and biomarker discovery

Pathway deregulation has been identified as a key driver of carcinogenesis, with proteins in signaling pathways serving as primary targets for drug development. Deregulation can be driven by a number of molecular events, including gene mutation, epigenetic changes in gene promoters, overexpression, and gene amplifications or deletions. We demonstrate a novel approach that identifies pathways of interest by integrating outlier analysis within and across molecular data types with gene set analysis. We use the results to seed the top-scoring pair algorithm to identify robust biomarkers associated with pathway deregulation. We demonstrate this methodology on pediatric acute myeloid leukemia (AML) data. We develop a biomarker in primary AML tumors, demonstrate robustness with an independent primary tumor data set, and show that the identified biomarkers also function well in relapsed pediatric AML tumors.

Evaluating optimal therapy robustness by virtual expansion of a sample population, with a case study in cancer immunotherapy

Significance A successful cancer therapy induces a strong antitumor response while causing minimal side effects. The heterogeneous nature of cancer observed across different regions of the primary tumor, across metastatic sites, across time, and across patients makes designing such a successful therapy challenging. Both standard of care and finely tailored treatment protocols run the risk of not exhibiting a robust antitumor response in the face of these uncertainties. Here we introduce a platform for exploring this robustness question using treatment response data from a sample population. Our method integrates these experimental data with statistical and mathematical techniques, allowing us to quantify therapeutic robustness. Using this approach, we identified a robust therapeutic protocol that combines oncolytic viruses with an immunotherapeutic vaccine. Cancer is a highly heterogeneous disease, exhibiting spatial and temporal variations that pose challenges for designing robust therapies. Here, we propose the VEPART (Virtual Expansion of Populations for Analyzing Robustness of Therapies) technique as a platform that integrates experimental data, mathematical modeling, and statistical analyses for identifying robust optimal treatment protocols. VEPART begins with time course experimental data for a sample population, and a mathematical model fit to aggregate data from that sample population. Using nonparametric statistics, the sample population is amplified and used to create a large number of virtual populations. At the final step of VEPART, robustness is assessed by identifying and analyzing the optimal therapy (perhaps restricted to a set of clinically realizable protocols) across each virtual population. As proof of concept, we have applied the VEPART method to study the robustness of treatment response in a mouse model of melanoma subject to treatment with immunostimulatory oncolytic viruses and dendritic cell vaccines. Our analysis (i) showed that every scheduling variant of the experimentally used treatment protocol is fragile (nonrobust) and (ii) discovered an alternative region of dosing space (lower oncolytic virus dose, higher dendritic cell dose) for which a robust optimal protocol exists.

An argument for mechanism-based statistical inference in cancer

Cancer is perhaps the prototypical systems disease, and as such has been the focus of extensive study in quantitative systems biology. However, translating these programs into personalized clinical care remains elusive and incomplete. In this perspective, we argue that realizing this agenda—in particular, predicting disease phenotypes, progression and treatment response for individuals—requires going well beyond standard computational and bioinformatics tools and algorithms. It entails designing global mathematical models over network-scale configurations of genomic states and molecular concentrations, and learning the model parameters from limited available samples of high-dimensional and integrative omics data. As such, any plausible design should accommodate: biological mechanism, necessary for both feasible learning and interpretable decision making; stochasticity, to deal with uncertainty and observed variation at many scales; and a capacity for statistical inference at the patient level. This program, which requires a close, sustained collaboration between mathematicians and biologists, is illustrated in several contexts, including learning biomarkers, metabolism, cell signaling, network inference and tumorigenesis.

PatternMarkers & GWCoGAPS for novel data-driven biomarkers via whole transcriptome NMF

Summary: Non‐negative Matrix Factorization (NMF) algorithms associate gene expression with biological processes (e.g. time‐course dynamics or disease subtypes). Compared with univariate associations, the relative weights of NMF solutions can obscure biomarkers. Therefore, we developed a novel patternMarkers statistic to extract genes for biological validation and enhanced visualization of NMF results. Finding novel and unbiased gene markers with patternMarkers requires whole‐genome data. Therefore, we also developed Genome‐Wide CoGAPS Analysis in Parallel Sets (GWCoGAPS), the first robust whole genome Bayesian NMF using the sparse, MCMC algorithm, CoGAPS. Additionally, a manual version of the GWCoGAPS algorithm contains analytic and visualization tools including patternMatcher, a Shiny web application. The decomposition in the manual pipeline can be replaced with any NMF algorithm, for further generalization of the software. Using these tools, we find granular brain‐region and cell‐type specific signatures with corresponding biomarkers in GTEx data, illustrating GWCoGAPS and patternMarkers ascertainment of data‐driven biomarkers from whole‐genome data. Availability and Implementation: PatternMarkers & GWCoGAPS are in the CoGAPS Bioconductor package (3.5) under the GPL license. Contact: gsteinobrien@jhmi.edu or ccolantu@jhmi.edu or ejfertig@jhmi.edu Supplementary information: Supplementary data are available at Bioinformatics online.

CoGAPS matrix factorization algorithm identifies transcriptional changes in AP-2alpha target genes in feedback from therapeutic inhibition of the EGFR network

Patients with oncogene driven tumors are treated with targeted therapeutics including EGFR inhibitors. Genomic data from The Cancer Genome Atlas (TCGA) demonstrates molecular alterations to EGFR, MAPK, and PI3K pathways in previously untreated tumors. Therefore, this study uses bioinformatics algorithms to delineate interactions resulting from EGFR inhibitor use in cancer cells with these genetic alterations. We modify the HaCaT keratinocyte cell line model to simulate cancer cells with constitutive activation of EGFR, HRAS, and PI3K in a controlled genetic background. We then measure gene expression after treating modified HaCaT cells with gefitinib, afatinib, and cetuximab. The CoGAPS algorithm distinguishes a gene expression signature associated with the anticipated silencing of the EGFR network. It also infers a feedback signature with EGFR gene expression itself increasing in cells that are responsive to EGFR inhibitors. This feedback signature has increased expression of several growth factor receptors regulated by the AP-2 family of transcription factors. The gene expression signatures for AP-2alpha are further correlated with sensitivity to cetuximab treatment in HNSCC cell lines and changes in EGFR expression in HNSCC tumors with low CDKN2A gene expression. In addition, the AP-2alpha gene expression signatures are also associated with inhibition of MEK, PI3K, and mTOR pathways in the Library of Integrated Network-Based Cellular Signatures (LINCS) data. These results suggest that AP-2 transcription factors are activated as feedback from EGFR network inhibition and may mediate EGFR inhibitor resistance.

Correcting Transcription Factor Gene Sets for Copy Number and Promoter Methylation Variations

Preclinical Research

Enter the matrix: Interpreting unsupervised feature learning with matrix decomposition to discover hidden knowledge in high-throughput omics data

Omics data contains signal from the molecular, physical, and kinetic inter- and intra-cellular interactions that control biological systems. Matrix factorization techniques can reveal low-dimensional structure from high-dimensional data that reflect these interactions. These techniques can uncover new biological knowledge from diverse high-throughput omics data in topics ranging from pathway discovery to time course analysis. We review exemplary applications of matrix factorization for systems-level analyses. We discuss appropriate application of these methods, their limitations, and focus on analysis of results to facilitate optimal biological interpretation. The inference of biologically relevant features with matrix factorization enables discovery from high-throughput data beyond the limits of current biological knowledge—answering questions from high-dimensional data that we have not yet thought to ask.High-dimensional data is currently standard for biological inquiry. Biological systems are comprised of interrelated gene regulatory mechanisms, gene-gene interactions, and cellular interactions. These interactions induce low-dimensional structure within the high-dimensional data. Matrix factorization, also known as compressed sensing, learns low-dimensional mathematical representations from high-dimensional data. These factorization techniques can embed assumptions about pleiotropy, epistasis, inter-relationships between complex traits, and context-dependent interactions. They have been applied to uncover new biological knowledge in a breadth of topics ranging from pathway discovery to time course analysis. These techniques have been applied to data from diverse high-throughput omics technologies, including bulk and single-cell data. There are numerous computational techniques within the class of matrix factorization, each of which provides a unique interpretation of the processes in high-dimensional data. We review the visualization and applications of matrix factorization to systems-level analyses, which are diverse and require standardization to enable biological interpretation. Codifying the techniques to decipher biologically relevant features with matrix factorization enables their broad application to discovery beyond the limits of current biological knowledge — answering questions from high-dimensional data that we have not yet thought to ask.

Sneak peek: Raptors search for prey using stochastic head turns

ABSTRACT The strategies by which foraging predators decide when to redirect their gaze influence both prey detection rates and the preys ability to detect and avoid predators. We applied statistical analyses that have been used to study neural decision-making for gaze redirection in primates to 3 species of predatory birds with different sizes, visual systems, habitats, and hunting behaviors: the Northern Goshawk (Accipiter gentilis), Coopers Hawk (A. cooperii), and Red-tailed Hawk (Buteo jamaicensis). The timing of head saccades was measured during visual searches using field video recordings of foraging raptors, and during a variety of behaviors using a miniature camera mounted on the head of a Northern Goshawk. The resulting statistical distribution of latencies (time between successive head saccades) was compared to predictions from various models proposed to describe visual search strategies. Our results did not support models that assume a constant probability of gaze redirection per unit time, a constant time for “giving up” on the visual search, or an initial setup time before visual search initiation. Instead, our data were fit best by a log-normal distribution, consistent with the raptors stochastically changing their gaze direction on the basis of accumulated environmental information. Specifically, this suggests that saccade initiation arises from a neural computation based on detection of a threshold level of a dynamically updated decision signal that encodes noisy sensory data, similar to the processes inferred from previous studies of visual search strategies in primates. The only significant between-species difference we found was a slower mean gaze-redirection rate for 2 larger species compared to the Coopers Hawk, even though the latter has hunting behavior and maneuverability similar to that of the Northern Goshawk. Head-saccade latencies measured for a Northern Goshawk during different behaviors showed that the bird changed gaze direction significantly less frequently, on average, while perched than while in motion.

Toward Signaling-Driven Biomarkers Immune to Normal Tissue Contamination

The goal of this study was to discover a minimally invasive pathway-specific biomarker that is immune to normal cell mRNA contamination for diagnosing head and neck squamous cell carcinoma (HNSCC). Using Elseviers MedScan natural language processing component of the Pathway Studio software and the TRANSFAC database, we produced a curated set of genes regulated by the signaling networks driving the development of HNSCC. The network and its gene targets provided prior probabilities for gene expression, which guided our CoGAPS matrix factorization algorithm to isolate patterns related to HNSCC signaling activity from a microarray-based study. Using patterns that distinguished normal from tumor samples, we identified a reduced set of genes to analyze with Top Scoring Pair in order to produce a potential biomarker for HNSCC. Our proposed biomarker comprises targets of the transcription factor (TF) HIF1A and the FOXO family of TFs coupled with genes that show remarkable stability across all normal tissues. Based on validation with novel data from The Cancer Genome Atlas (TCGA), measured by RNAseq, and bootstrap sampling, the biomarker for normal vs. tumor has an accuracy of 0.77, a Matthews correlation coefficient of 0.54, and an area under the curve (AUC) of 0.82.

PatternMarkers and Genome-Wide CoGAPS Analysis in Parallel Sets (GWCoGAPS) for data-driven detection of novel biomarkers via whole transcriptome Non-negative matrix factorization (NMF)

NMF algorithms associate gene expression changes with biological processes (e.g., time-course dynamics or disease subtypes). Compared with univariate associations, the relative weights of NMF solutions can obscure biomarkers identification. Therefore, we developed a novel PatternMarkers statistic to extract unique genes for biological validation and enhanced visualization of NMF results. Finding novel and unbiased gene markers with PatternMarkers requires whole-genome data. However, NMF algorithms typically do not converge for the tens of thousands of genes in genome-wide profiling. Therefore, we also developed GWCoGAPS, the first robust Bayesian NMF technique for whole genome transcriptomics using the sparse, MCMC algorithm, CoGAPS. This software contains additional analytic and visualization tools including a Shiny web application, patternMatcher, which are generalized for any NMF. Using these tools, we find granular brain-region and cell-type specific signatures with corresponding biomarkers in GTex data, illustrating GWCoGAPS and patternMarkers unique ability to detect data-driven biomarkers from whole genome data.Summary Non-negative Matrix Factorization (NMF) algorithms associate gene expression with biological processes (e.g., time-course dynamics or disease subtypes). Compared with univariate associations, the relative weights of NMF solutions can obscure biomarkers. Therefore, we developed a novel PatternMarkers statistic to extract genes for biological validation and enhanced visualization of NMF results. Finding novel and unbiased gene markers with PatternMarkers requires whole-genome data. However, NMF algorithms typically do not converge for the tens of thousands of genes in genome-wide profiling. Therefore, we also developed Genome-Wide CoGAPS Analysis in Parallel Sets (GWCoGAPS), the first robust whole genome Bayesian NMF using the sparse, MCMC algorithm, CoGAPS. This software contains analytic and visualization tools including a Shiny web application, patternMatcher, which are generalized for any NMF. Using these tools, we find granular brain-region and cell-type specific signatures with corresponding biomarkers in GTex data, illustrating GWCoGAPS and patternMarkers ascertainment of data-driven biomarkers from whole-genome data. Availability PatternMarkers & GWCoGAPS are in the CoGAPS Bioconductor package (3.5) under the GPL license. Contact gsteinobrien@jhmi.edu; ccolantu@jhmi.edu; ejfertig@jhmi.edu