Molecular profiles of tumors and tumor-associated cells hold great promise as biomarkers of clinical outcomes. However, existing data sets are fragmented and difficult to analyze systematically. Here we present a pan-cancer resource and meta-analysis of expression signatures from ∼18,000 human tumors with overall survival outcomes across 39 malignancies. By using this resource, we identified a forkhead box M1 (FOXM1) regulatory network as a major predictor of adverse outcomes, and we found that expression of favorably prognostic genes, including KLRB1 (encoding CD161), largely reflect tumor-associated leukocytes. By applying CIBERSORT, a computational approach for inferring leukocyte representation in bulk tumor transcriptomes, we identified complex associations between 22 distinct leukocyte subsets and cancer survival. For example, tumor-associated neutrophil and plasma cell signatures emerged as significant but opposite predictors of survival for diverse solid tumors, including breast and lung adenocarcinomas. This resource and associated analytical tools (http://precog.stanford.edu) may help delineate prognostic genes and leukocyte subsets within and across cancers, shed light on the impact of tumor heterogeneity on cancer outcomes, and facilitate the discovery of biomarkers and therapeutic targets

We present LAPSI (Lineage-Aided Maximum-Parsimony Structural Variation Inference), a novel method for detecting and genotyping somatic structural variations (SVs) in multiple whole-genome sequencing tumor samples, taken from a cancer patient. In contrast to standard SV discovery approaches in cancer genomes, which do not leverage phylogenetic information, LASPI makes use of the multi-sample lineage tree structure reconstructed from ultra-deep sequencing somatic SNV datasets using a specialized method such as LICHeE. We demonstrate that leveraging lineage trees reduces false negatives in detecting and genotyping of SVs. Our method effectively pools samples that share a common ancestor in the tree and finds clusters of discordant paired-end reads that suggest the same SV breakpoint across these samples. Placement of SVs onto specific branches of the lineage tree results in a more comprehensive roadmap of the tumor's genome evolution that begins at the zygote.

Massively multitask neural architectures provide a learning framework for drug discovery that synthesizes information from many distinct biological sources. To train these architectures at scale, we gather large amounts of data from public sources to create a dataset of nearly 40 million measurements across more than 200 biological targets. We investigate several aspects of the multitask framework by performing a series of empirical studies and obtain some interesting results: (1) massively multitask networks obtain predictive accuracies significantly better than single-task methods, (2) the predictive power of multitask networks improves as additional tasks and data are added, (3) the total amount of data and the total number of tasks both contribute significantly to multitask improvement, and (4) multitask networks afford limited transferability to tasks not in the training set. Our results underscore the need for greater data sharing and further algorithmic innovation to accelerate the drug discovery process.

While the genomic and genetic basis of many cancers has been partly elucidated, less is known about the identity of specific cancer genes that are associated with clinical parameters such as clinical stage and metastasis.  To address this question, we compiled multiple genomic data types (mutations, copy number alterations, gene expression and methylation status) as well as clinical meta-data for twelve different cancers in the Cancer Genome Atlas Project.  We used an elastic-net regularized regression method on the combined genomic data to identify genetic aberrations and their associated cancer genes that are indicators of clinical parameters in each cancer type.  To facilitate the identification of clinically actionable genetic aberrations from this cancer genomic data, we utilized genomic and drug response data from the Cancer Cell Line Encyclopedia (CCLE).  These expansive data sets of hundreds of cancer cell lines include three genomic data types (mutations, copy number alterations, gene expression) and drug response data.

We identify the top ranked genes that delineate key clinical features such as clinical stages for 17 cancers. A fit of the elastic-net regularized regression to more than 200 samples/cancer type and integrative analysis of integrative genomic platforms identified the set of top gene predictors of clinical parameters (TCGA data) and drug response (CCLE data). There are several genes ranked top in multiple cancer types while most of them are unique to a cancer type. These results provide us with new insights into the underlying genetic similarities and differences among 17 cancers; namely, common genetic characteristics that have implications for clinical phenotype. Finally, we are identifying the core set of gene aberrations that are both clinically relevant and drug responsible.

While it is common practice to immunohistochemically stain tumor sections post-surgery in early stage breast cancer, the rich information in these slides is typically summarized in terms of a simple univariate score such as the proportion of positively staining tumor nuclei. We show here, using fully automated image processing, that there are morphological, relational and co-expression features which are significantly associated with survival and could therefore be used to improve prognosis and guide follow up treatment.

Feature selection is an important machine learning problem, and it is widely used for various types of applications such as gene selection from microarray data, document categorization, and prosthesis control, to name a few. The feature selection problem is a traditional and popular machine learning problem, and thus there exist many methods including the least absolute shrinkage and selection operator (Lasso). However, there are a few methods that can select features from large and ultra high-dimensional data (more than million features) in nonlinear way. To the best of our knowledge, feature screening approaches, which first compute d input-output relevance scores and then selects m features with the largest relevance to the output, are the only method that can select nonlinearly related features from large and ultra high-dimensional data. Feature screening methods are simple yet effective and theoretically well justified (Fan and Lv, 2008; Balasubramanian et al., 2013). However, since feature screening approaches do not use input-input relevance, they tend to select redundant features. In this talk, we first introduce a Hilbert-Schmidt Independence Criterion Lasso (HSIC Lasso) that can efficiently select non-redundant features from a small and high-dimensional data in nonlinear way. A key advantage of HSIC Lasso is that it is a convex method and can find a globally optimal solution.  Then we further extend the proposed method to handle ultra high-dimensional data by incorporating with distributed computing framework. Moreover, we show the proposed method is closely related to feature screening, minimum redundancy maximum relevance (mRMR), and two-step multiple kernel learning (MKL). Finally, we experimentally show that the proposed method can select features from a large and ultra high-dimensional data (more than 1 million features and more than 10K samples) in a few hours.

Networks are a fundamental tool for understanding the intricate interconnections that govern biological systems. This talk will describe two ways in which networks, in combination with mathematical models and algorithmic techniques, can yield valuable biological insights.
Causal regulatory networks help reveal the hidden regulators of gene expression patterns. To facilitate their analysis we established an efficient method for evaluating the significance of the overlap of ternary signals, which generalizes Fisher's exact test. We used this method to analyze a large-scale causal regulatory network and uncovered new regulators of cardiac hypertrophy.
Metabolic networks help identify novel drug targets. We uncovered structural features of these networks that had been missed by previous researchers, and developed a theoretical framework based on duality for analyzing them in a consistent fashion. We used this theoretical framework to create a new metabolic network for Mycobacterium tuberculosis by algorithmically merging two existing networks, and identified several putative drug targets.

High-dimensional data representation is an important problem in many different areas of science. Nowadays, it is becoming crucial to interpret the data of varying dimensionality correctly. Dimensionality reduction methods process the data in order to help visualize the data, reduce its complexity, or find latent representation of the original problem.
In particular, in biological domain, flow cytometry (FC) is an important versatile tool that produces a high-dimensional quantitative measurements of intracellular and surface variables for a large number of cells. However, analyzing such high-dimensional data can be a challenge. Conventional methods, such as manual gating using a series of biaxial plots is very subjective, labour-intensive task and often provide incomplete and very localized information. Even when using automated clustering algorithms (such as FLOCK or flowClust) important properties of the data, such as phenotype continuity, can be lost. Thus, it becomes essential to use a good visualization technique to assess the structure of the data and observe patterns and natural clusters of different cells.
In this talk I will review several techniques for dimensionality reduction, in particular aiming for methods that provide good, adaptive visualization of complex, highly-structured datasets. I will cover both linear (e.g. PCA, MDS) and non-linear (e.g. ISOMAP, t-SNE) methods, discuss their similarities and differences, and provide an intuition into how those algorithms can be used efficiently. Finally, I will extend those methods to be applicable to large-scale (up to several millions points) high-dimensional (up to a million of dimensions) datasets.

This tutorial will survey several topics related to the use of the \ell_1 norm to obtain a sparse or structured solution to a given problem. The first topic is on the \ell_1 norm and \ell_1 regularization itself: several interpretations, some basic theoretical results, and some practical heuristics that are often useful. The second topic is algorithms for solving problems
involving \ell_1 norms and related functions. These are called proximal methods and rely on the concept of the proximal operator of a convex function. They generalize familiar algorithms: the projected gradient method is a special case. The final topic is examples and variations on the basic theme of promoting sparsity via the \ell_1 norm. The examples include topics like basis pursuit, sparse coding, the elastic net, the group lasso, the structured or overlapping group lasso, multi-task learning, the fused lasso, the graphical lasso, low rank matrix completion, matrix decomposition, and robust and sparse PCA, among others. The point is that it is possible to cover many examples because they just reuse the same building blocks in various ways. By the end, it should be clear how to use these tools in other problems of interest.

A central challenge of biology is to understand how complex phenotypes are controlled by individual molecules and their interactions. We report the first computational model which explains the life cycle of an entire organism, /Mycoplasma genitalium/, including metabolism, macromolecule synthesis, and cytokinesis, from the level of individual molecules and their chemical interactions. The hybrid computational model consists of submodels of 28 cellular processes integrated through 16 cellular states, accounts for the specific function of every annotated gene product, and predicts the dynamics of every molecule. Using the model we
identified the molecular determinates of cellular growth and replication. We found the /M. genitalium/ cell cycle is 9.0 ± 0.5 h, and that most of its variance is due that of metabolism and of thymidylate and acetate kinase expression. Additionally, we found that replication initiation and DnaA binding dynamics are a significant source of cell cycle length variation among fast growing cells. We examined the genetic requirements of single cell growth and found four distinct classes of single-gene deletion strains: strains indistinguishable from wild-type, strains with early growth cessation, strains with slowly decaying growth, and non-dividing strains. The model correctly predicts the experimentally observed essentially of > 80% of genes. We believe that gene-complete models will accelerate biomedical discovery and bioengineering by enabling rapid, low cost /in silico /experimentation, facilitating experimental design and interpretation, and guiding rational engineering of biological systems and medical therapies.

In the era of "omics" biology, microarray and next-generation sequencing technologies are widely used to study complex biological systems. Without the aid of appropriate analytical tools, however, the huge outputs of these methods are of limited utility. Unsupervised clustering represents one major class of data-mining approaches for partitioning large datasets into coherent subsets more suitable for analysis. Since common clustering methods are generally limited by dataset size, cluster shape, detection of outliers, and in particular, cluster number detection, we designed and implemented a new method based on a serial application of well-established techniques from machine learning, cartography, and graph theory. Without prior knowledge of cluster number or structure, our strategy, called AutoSOME, benchmarks favorably against state-of-the-art methods, and effectively identifies both discrete and fuzzy relationships among high dimensional data points, such as genome-scale gene expression data. In this talk, I will describe the AutoSOME method and illustrate its utility for genome biology using a number of anecdotal examples, including results obtained from analyzing stem cell and cancer cell microarray data. Our findings indicate that AutoSOME provides a valuable approach for cluster analysis of genome-wide data.