Pierre Baldi, UCI University, California, USA

• Carbon-Based Computing Vs Silicon-Based Computing: A New Theory of Circadian Rhythms Abstract

  Carbon-based and silicon-based computing systems are very different. One key difference is the pervasive presence of circadian rhythms in living systems at multiple levels. At the molecular level, circadian rhythms are regulated by a central clock consisting of a key negative transcription-translation feedback loop involving a dozen of genes. However, integrative systems biology analyses of high-throughput transcriptomic and metabolomic data reveal that roughly 10% of genes or metabolites oscillate in a circadian manner in any given cell or tissue. Furthermore, when data is aggregated across different tissues and genetic or environmental conditions, the overlap in circadian species beyond the core clock is very small. Thus a large fraction of molecular species in the cell is capable of oscillating in a circadian manner under some set of conditions. We will present a novel theory of circadian rhythms to explain these puzzling findings. In this theory, molecular networks are viewed as networks of coupled-oscillators sculpted by 3.5 billion years of evolution. Under a given set of genetic and environmental conditions, a cell can reprogram itself and select its own subset of oscillatory species out of a vast repertoire. The oscillating species provide a physiological signature of the state of the cell.

  • V. R. Patel, K. Eckel-Mahan, P. Sassone-Corsi, and P. Baldi. How Pervasive Are Circadian Oscillations? Trends in Cell Biology, in press, DOI:10.1016/j.tcb.2014.04.005, (2014).
  • K. L. Eckel-Mahan1, V. R. Patel, S. de Mateo, N. J. Ceglia, S. Sahar, S. Dilag, Kenneth A. Dyar, R. Orozco-Solis, P. Baldi, and P. Sassone-Corsi. Reprogramming of the Circadian Clock by Nutritional Challenge. Cell, 155, 7, 1464-1478, (2013).
  • V. Patel, K. Eckel Mahan, P. Sassone-Corsi, and P. Baldi. CircadiOmics: Integrating Circadian Genomics, Transcriptomics, Proteomics, and Metabolomics. Nature Methods, 9, 8, 772-773, (2012).
  • K. L. Eckel-Mahan, V. R. Patel, K. S.Vignola, R. P. Mohney, P. Baldi, and P. Sassone-Corsi. Coordination of Metabolome and Transcriptome by the Circadian Clock. PNAS, 109 (14) 5541-5546, (2012)

Anne-Laure Boulesteix, Department of Computational Molecular Medecine, IBE, Ludwig-Maximilians Universität München, Germany

• Statistical testing and variability in real-data-based benchmark experiments for supervised learning methods Abstract

  Resampling-based methods such as, e.g., k-fold cross-validation or repeated splitting into training and test sets are routinely used in the context of supervised statistical learning to assess the prediction performance of prediction methods. In this talk, I discuss two important issues related to the use of such methods: the design of resampling-based benchmark experiments from the perspective of statistical testing and the variability of resampling-based procedures for the choice of tuning parameters.

  The first part of the talk deals with benchmark experiments that aim at comparing the performance of different algorithms and presents a statistical framework for hypothesis testing in real data comparison studies. In computational literature, most abstracts of articles presenting new supervised learning methods end with a sentence like "our method performed better than existing methods on real data sets", e.g. in terms of error rate. However, these claims are often not based on proper statistical inference and, if statistical hypothesis tests are performed, the tested hypothesis is not clearly defined and poor attention is paid to the type I and type II error. We propose a proper statistical framework for hypothesis tests comparing the performance of supervised learning methods based on several real data sets with unknown underlying distributions. After giving a statistical interpretation of ad-hoc paired t-tests commonly performed in practice, we devote special attention to power issues and outline a simple method to determine the number of data sets to be included in a comparison study to reach an adequate power. As an extension of this testing framework, we will also consider regression models considering the relative performance of prediction methods as a dependent variable and data sets’ characteristics as independent variables, with the long-term aim to guide the choice of the prediction method depending on the type of data set at hand. The methods are illustrated through an application of PLS-based classification to 50 microarray gene expression data sets.

  The second part of the talk is devoted to the problem of the choice of tuning parameters - for instance the number of PLS components - based on resampling methods. I will again show results of a PLS-based classification method applied to microarray data sets. These results empirically demonstrate the instability induced by the random character of resampling-based procedures. As suggested by this study and substantiated by theoretical results, we recommend to perform a large number of resampling iterations whenever computationally possible to yield a better stability.

Karsten Borgwardt, Department of Biosystems Science and Engineering (D-BSSE), ETH Zürich, Switzerland

• Machine Learning for Personalized Medicine Abstract

  Over the last decade, enormous progress has been made on recording the health state of an individual patient down to the molecular level of gene activity and genomic information – even sequencing a patient’s genome for less than 1000 dollars is within reach. However, the ultimate hope to use all this information for personalized medicine, that is to tailor medical treatment to the needs of an individual, remains largely unfulfilled. To turn the vision of personalized medicine into reality, many methodological problems remain to be solved: there is a lack of methods that allow us to gain a causal understanding of the underlying disease mechanisms, including gene-gene and gene-environment interactions. Similarly, there is an urgent need for integration of the heterogeneous patient data currently available, for improved and robust biomarker discovery for disease diagnosis, prognosis and therapy outcome prediction. The field of machine learning, which tries to detect patterns, rules and statistical dependencies in large datasets, has also witnessed dramatic progress over the last decade and has had a profound impact on the Internet. Amongst others, advanced methods for high-dimensional feature selection, causality inference, and data integration have been developed or are topics of current research. These techniques address many of the key methodological challenges that personalized medicine faces today and keep it from rising to the next level. In this talk, we will describe the challenges and opportunities for machine learning in personalized medicine and we will present our recent research results in this direction.

Jean-Loup Faulon, IBSS, CNRS, Genopole, University of Evry, France

• Using Machine Learning in Synthetic Biology: The Design-Build-Test and Learn cycle Abstract

  Synthetic biology and metabolic engineering have succeeded in the biosynthesis of numerous commodity or high value compounds. Yet, the choice of pathways and enzymes used for such successful applications was many times made ad hoc, or required expert knowledge of the specific biochemical reactions. In order to rationalize this process we have developed the computer-aided design (CAD) tool RetroPath [1] that explores and enumerates metabolic pathways connecting the endogenous metabolites of a chassis cell to a target compound. Namely, our tool queries for target activities the list of enzymes found in metabolic databases based on their annotated and predicted activities based on a tensor product kernel [2]. Next, it ranks pathways based on the predicted efficiency of the available enzymes, the toxicity of the intermediate metabolites and the calculated maximum product flux. As an illustration of the power of rational design, RetroPath compiled the top-ranking pathways producing the flavonoid pinocembrin (a antibacterials targeting Staphylococcus aureus), narrowing down a list of nine million possible enzyme combinations to a number that could be easily assembled and tested. We next constructed the top-ranked enzyme combinations, four of which displayed significant yields. One round of metabolic network optimization based on RetroPath output further increased pinocembrin titers 17-fold [3]. In total, 12 out of the 13 enzymes tested in this work displayed a performance that was in accordance with its predicted score. These results validate the ranking function of our CAD tool, and open the way to its utilization in the biosynthesis of novel compounds.

  • 1. Carbonell P, Parutto P, Herisson J, Pandit S.B, Faulon J.L. XTMS: pathway design in an eXTended metabolic space. Nucleic Acids Res. in press 2014, [PMID: 24792156].
  • 2. Faulon J.L., Misra M., Martin S., Sale, K., Sapra R. Genome Scale Enzyme-metabolites and Drug-Target interaction predictions using the signature molecular descriptor. Bioinformatics, 24, 225-233, 2008.
  • 3. Fernandez-Castane A, Feher T, Carbonell P, Pauthenier C, Faulon J.L. Computer-aided design for metabolic engineering. J Biotechnol. in press 2014, [PMID: 24704607].

Nicola Segata, Centre for Integrative Biology, University of Trento, Italy

• Machine learning challenges in computational meta'omics Abstract

  The study of the microbial diversity with sequencing-based cultivation-free approaches (metagenomics) is currently revolutionizing our understanding of the biology associated with many natural systems including the human body. Outnumbering our own cells 10 to 1, the cloud of microbial organisms (microbiome) living in symbiosis with our body has, in fact, a profound effect on human health and is co-responsible of several complex diseases when the microbiome/host equilibrium is broken. Following the recent biotechnological revolution that enabled shotgun metagenomics, the research community focused primarily on developing computational tools to extract meaningful taxonomic [1], phylogenetic [2], and functional profiles [3] of the microbiome from the large amount of raw-data produced. These methods are now providing reliable quantitative snapshots of the microbiome that can thus be used for advanced learning tasks of clinical relevance including biomarker discovery [4] and subtype identification. However, specific characteristics of metagenomic data are affecting the effectiveness of such tasks and the availability of complementary meta’omic techniques (e.g. metatranscriptomics) is posing novel issues [5]. In this talk, I will discuss the machine learning challenges that the field is facing and will present some of the most relevant learning tasks in the study of the human-associated microbiome.

  • [1] Segata, N., Waldron, L., Ballarini, A., Narasimhan, V., Jousson, O., & Huttenhower, C. (2012). Metagenomic microbial community profiling using unique clade-specific marker genes. Nature Methods, 9, 811–814.
  • [2] Segata, N., Börnigen, D., Morgan, X., & Huttenhower, C. (2013). PhyloPhlAn is a new method for improved phylogenetic and taxonomic placement of microbes. Nature Communications, 4, 2304.
  • [3] Abubucker, S., Segata, N., Goll, J., Schubert, A., Izard, J., Cantarel, B. L., … Huttenhower, C. (2012). Metabolic Reconstruction for Metagenomic Data and Its Application to the Human Microbiome. PLoS Computational Biology, 8(6), e1002358.
  • [4] Segata, N., Izard, J., Waldron, L., Gevers, D., Miropolsky, L., Garrett, W. S., & Huttenhower, C. (2011). Metagenomic Biomarker Discovery and Explanation. Genome Biology, 12, R60.
  • [5] Segata, N., Boernigen, D., Tickle, T. L., Morgan, X., Garrett, W. S., & Huttenhower, C. (2013). Computational meta’omics for microbial community studies. Molecular Systems Biology, 9(666), 1–15.

Kathleen Marchal, Depts. Plant Biotechnology and Bioinformatics; Information Technology (INTEC, iMINDS), U.Ghent; Dept. of Microbial and Molecular Systems, K.U.Leuven, Belgium

• Network-based data-integration: applications to clonal systems Abstract

  In the last 10 years huge efforts have been made to infer interaction networks from large scale omics data. Networks have been inferred at different molecular levels (transcriptional, metabolic, protein interaction, signaling), allowing us to compile for a model organism ‘integrated networks’ in which nodes represent molecular entities (genes, proteins etc) and edges the interactions between those entities. Such networks that span different molecular levels, despite being noisy and static and thus overconnected, do comprehensively summarize all available (reliable and less reliable) molecular knowledge on an organism of interest. In this presentation we will show how the information contained within those networks can be exploited to guide experimentalists with the interpretation of their own in house generated omics experiments or how some data-integration efforts depend on network-based guidance (such as for the interpretation of genomic variations in clonal systems).

Program