Seminars

Today, a vast number of associations (hundreds of thousands to millions) between small molecule structures and their respective target affinities are known. Those associations are exploited here for the prediction of on-target as well as off-target effects of novel compounds, and to analyze factors contributing to promiscuity of new structures. Namely, we describe the development of in silico models to predict targets of compounds, such as those contained in common safety profiling panels1, to prioritize targets to be screened in safety evaluations. For the analysis of compound promiscuity, a usually undesired feature of active ingredients, we present computational models which firstly provide estimates for compound promiscuity, and secondly give insight into features frequently associated with promiscuous compounds.2 Finally, we will discuss ideas how phenotypic profiling of compounds can be integrated with in silico techniques to derive a comprehensive assessment of the biomodulatory capabilities of a compound.3 Applications will in particular be presented on the prediction of compound activity against particular mutants of HIV Reverse Transcriptase, using so-called proteochemometric modeling approaches.

Segmenting tissues in histological images is challenging due to their textural appearance, size, and the need to generalize results for different tissues and stains. Recent results show that in some
cases an unsupervised segmentation can be achieved by using spatial distributions of cellular and subcellular components as segmentation features. The spatial distribution features have a theoretical basis that enables fast and deterministic calculation using convolutions, and their peculiar representation in feature space permits robust clustering. This talk will discuss recent segmentation results and developments in feature clustering and computation on GPU and parallel systems.

Computer modeling, coupled with knowledge discovery and data-mining (KDD), is ushering in a new era in biological investigation, mandated by the massive flow of data emerging from new experimental tools. Systems Biology, deriving from the demands made by the data streams of genomes and proteomes, must now be extended to neurobiology in response to the growing streams from electrome (signal flow within the neuron), connectome (neural connectivity) and neurophysiome (multiunit recording). This change in perspective can be usefully compared to the change that accompanied the shifts in painting from Academicism to Impressionism and Cubism. To illustrate the approach, I will discuss results from our simulation of neocortical connectivity and dynamics, and from our data-mining of multiunit recordings to examine neural coding in normals and in animal models of schizophrenia.

The execution of data intensive workflows in Grid environments involves the movement of several gigabytes of data. Strategies used to stage data in and out of compute resources may have a significant impact on the overall execution of workflows. My research focuses on identifying the execution time data management strategies that allow for the most efficient (in terms of reducing overall execution time) execution of a given scientific workflow.

In this talk, I will describe my work in understanding the data processing requirements of various scientific workflows and describe a framework that allows for the evaluation of the impact of different data staging strategies on the execution of these workflows. I will then focus on the execution of data intensive workflows on storage constrained resources and present the results of heuristics and genetic-algorithm based approaches that try to minimize workflow execution times while meeting storage limitations.

Large, strongly coupled neural networks tend to produce chaotic spontaneous activity. This might appear to make them unsuitable for generating reliable sensory responses or repeatable motor patterns. However, this is not the case. Inputs can induce a phase transition, leading to responses uncontaminated by chaotic "noise". Likewise, appropriately trained feedback units can control the chaos, resulting in a wide variety of repeatable output patterns. These issues will be discussed accompanied by examples, comparisons with experimental data and demonstrations.

Neurons encode and process sensory information in the form of spatiotemporal activity patterns. The implementation of these temporal codes in the brain requires neurons to be capable of following a stimulus with millisecond precision, as well as of responding consistently to repetitions of the same stimulus. I will summarize recent experimental, computational and mathematical work revealing the biophysical properties of neurons that allow them to encode sensory information in a reliable, reproducible fashion. Time permitting, I will also talk about mechanisms underlying reliable neural dynamics at the network level.

Whereas protein function is widely acknowledged to follow from form, actually predicting ligand recognition from structure remains challenging. Conversely, there is a long history in pharmacology and medicinal chemistry of relating ligands to understand the properties of their receptors, even though ligand structures encode much less information than do those of proteins. More confounding still, related drugs and ligands can bind to targets that appear biologically unrelated. To quantify this, we compared drug targets based on the chemical similarity of their sets of ligands. The similarity score between each ligand set was calculated, using a BLAST-based statistical model to rank significance. Expressed as a minimum spanning tree, the sets may be mapped together. Although these maps are connected solely by chemical similarity, biologically sensible clusters emerge. Links among unexpected targets also emerged, several of which were experimentally testable. Returning to protein structure, we have also attempted to predict new ligands for some of the same targets by docking to their structures. Experimental testing of ligands proposed based on complementarity to protein structure, and based on the ligand polypharmacology hypothesis, will be discussed.

Eukaryotic chromatin is the fundamental protein/nucleic acid unit that stores the genetic material. Understanding how chromatin fibers fold and unfold in physiological conditions is important for interpreting fundamental template-directed biological processes like DNA replication and transcription regulation. Using a mesoscopic model of oligonucleosome chains and tailored sampling protocols, we elucidate the energetics of oligonucleosome folding/unfolding and the role of each histone tail, linker histones, and divalent ions in regulating chromatin structure. The resulting compact topologies reconcile features of the zigzag model with straight linker DNAs with the solenoid model with bent linker DNAs for optimal fiber organization and reveal dynamic and energetic aspects involved.

The CardioVascular Research Grid (CVRG) Project is an NHLBI-funded R24 resource that is developing a grid infrastructure for sharing and analyzing a broad range of cardiac data. In order to avoid "re-inventing the wheel", the CVRG is re-using and extending software components from other major biogrid projects such as caBIG and BIRN, while focusing its resources on developing tools targeted to meeting the specific needs of the cardiovascular research community. In this talk, I will present an overview of the project, describing both data and analytic services that are currently available on the CVRG. I will also describe mechanisms by which the CVRG can stimulate and support collaborative projects.

In this talk we illustrate some real case studies in vascular medicine and surgery that we have recently investigated with the support of mathematical models and numerical simulations. We present some examples where our investigation has different purposes, ranging from a better understanding of phenomena of clinical interest to the optimization of surgical procedures. Each example provides the conceptual framework to introduce new mathematical models and numerical methods whose applicability, however, goes beyond the specific case that is addressed. The perspective underlying these research is that a strong integration of medical imaging tools, statistical analysis and numerical modeling can strongly increase informations available to medical doctors for supporting their decisions.