Seminars

Today’s healthcare system requires clinicians to master an ever-expanding knowledge base while the time to master this knowledge and apply it to specific tasks is steadily shrinking. The convergence of an expanding knowledge base, expanding patient databases and escalating time constraints poses a serious healthcare problem that inevitably leads to physician’s errors. Solution to this problem requires the development of new tools to assist physicians to timely apply comprehensive, up-to-date knowledge and the available patient data to specific clinical problems. The long-term objective of our project is to improve the care of cardiac patients by developing an imaging expert system that is automatically updated with the latest scientific/clinical knowledge extracted from journal articles using natural language understanding techniques. The goal of this constantly updated decision support system is to assist physicians to appropriately perform and interpret ECG-gated myocardial perfusion SPECT studies for the diagnosis and prognosis of coronary artery disease.

Image Registration is one of today's most challenging tasks in medical
imaging. It helps to improve existing or even allows for new
diagnostics and therapies in medicine.

Roughly speaking, registration is computing a spatial
alignment of two images. More formally, given a reference image $R$ and a
template image $T$ the problem is to find a deformation fields $\phi$ such that
$R$ and $T(\phi)$ are similar.

The problem arises whenever one wants to compare or fuse images of objects
that for example have been taken at different times, with different imaging
devices, or form different perspectives. The spectrum of applications is
large and not only restricted to medicine. The demand for registration methods
comes basically in all disciplines where image data is used, e.g., in
geophysics and material science.

However, in my talk I will give an introduction to image registration and
outline existing methods and their building blocks. Thereby, the focus is on
medical applications. Finally, I will present some projects I have been
working on.

Viruses have been among the favorite targets of physicists interested in biology since right after World War II, when Max Delbrück and other physicists founded the so-called “Phage School”. (Phage are viruses that infect bacteria.) Viruses exist at the boundary between living and non-living objects. Some are as small as 50 nm in diameter and have quasi-icosahedral symmetry, but their apparent simplicity hides an amazingly rich set of mechanisms for infecting their hosts and reproducing themselves. This colloquium will examine the mechanisms of assembling icosahedral viruses, with a description of recent experimental advances (electron microscopy; single-molecule experiments) and a discussion of theoretical and computational approaches to investigating viral assembly.

Beyond "Bag of Words": A Framework for Conceptual Retrieval and an Application in Clinical Question Answering"

Although the field of information retrieval has made enormous progress
in the last half century, virtually all systems are still built on the
remarkably simple concept of "counting words", under assumptions of
term independence. Although these methods have been empirically
validated (e.g., in TREC evaluations), it is a simple fact that words
alone cannot capture the semantic content of documents and information
needs.

In this talk, I will discuss a framework for "conceptual retrieval"
that articulates the types of knowledge that are important for
information seeking. This general framework is instantiated in a
clinical question answering system that operationalizes the principles
of evidence-based medicine (EBM). Experiments show that an EBM-based
scoring algorithm outperforms a state-of-the-art baseline that employs
only term statistics. Ablation studies further yield a better
understanding of the performance contributions of different
components.

I will conclude by discussing how other domains can benefit from
knowledge-based approaches and the general applicability of this
proposed framework.


About the Speaker

Jimmy Lin is an assistant professor in the iSchool at the University
of Maryland. In addition, he's affiliated with both the
Human-Computer Interaction Laboratory (HCIL) and the Computational
Linguistics and Information Processing (CLIP) laboratory in UMD's
Institute for Advanced Computer Studies (UMIACS). He graduated with a
Ph.D. in computer science from MIT in 2004. Jimmy's research lies at
the intersection between information retrieval and natural language
processing. He leads the IBM/Google "Cloud Computing" initiative at
Maryland.

High-throughput data production technologies are revolutionizing modern biology. Translating this experimental data into discoveries of relevance to human health relies on sophisticated computational tools that can handle large-scale data.

One area of rapid ongoing data generation is whole genome sequencing. Comparisons between sequenced genomes can be a powerful tool to understand functional genomic regions, by going beyond the primary sequence to capture patterns in how functional regions evolve. Using data generated by the ENCODE project we will demonstrate the power of genome comparisons to distinguish cis-regulatory elements (critical for the control of gene expression). We will then describe a machine learning approach that goes beyond sequence conservation to capture broader and more informative sequence and evolutionary patterns that better distinguish different classes of elements. This approach has proven successful for a variety of classification problems. In particular, the "Regulatory Potential Score" has been used to identify putative regulatory elements with high rates of experimental validation.

Sophisticated methods for the analysis of biological data are of little value if they are not accessible. Powerful analysis tools, data warehouses, and browsers exist, but for the average experimental biologists with limited computer expertise, making effective use of these resources is still out of reach. We have developed "Galaxy", which solves this problem by providing an integrated web-based workspace that bridges the gap between different tools and data sources. For computational tool developers, Galaxy eliminates the repetitive effort involved in creating high-quality user interfaces, while giving them the benefit of being able to provide their tools in an integrated environment. For experimental biologists, Galaxy allows running complex analysis on huge datasets with nothing more than a web browser, and without needing to worry about details like resource allocation, format conversions, etc. Galaxy makes high-end computational biology more accessible, efficient, and reproducible.

Hepatitis B virus is a small DNA virus with an icosahedral (spherical) core. No matter how simple it may be, an infectious virus is a complicated machine. Unlike other machines, viruses self-assemble. By interfering with self assembly, we may open new therapeutic avenues, but first we must understand this multi-step reaction. In this talk I will discuss our understanding of the assembly reaction and where it is vulnerable to interference. Assembly is treated as a nucleated cascade of low order reactions. The energy surface of assembly indicates that only a small subset of reactants are important and that inhibiting assembly by any "traditional" mechanism will be easily evaded by the virus. These computational results will be compared with solution experiments that test our conclusions. The normal hepatitis B virus core includes a capsid constructed from 240 copies of the core protein (Cp). Using Cp expressed in E coli, we have examined normal assembly, the ability of defective protein to interfere with assembly, the effects of small molecules that actually enhance assembly. These enhancers alter association energy, kinetics, and can misdirect assembly.

The ability to measure genome-wide expression holds great promise for characterizing cells and distinguishing diseased from normal tissues. Thus far, microarray technology has only been useful for measuring relative expression between two or more samples, which has handicapped its ability to classify tissue types. By determining absolute expression, the gene expression barcode is the first method that can successfully predict tissue type based on data from a single hybridization. We have shown that this ability is useful for predicting clinical outcomes in breast cancer patients similar to the Oncotype DX and MammaPrint tests, but the barcode can easily be extended to other cancers, making the barcode a powerful diagnostic and prognostic tool for oncology.

Advances in the theory of the elastic rod model for DNA enhanced our
understanding of a wide range of phenomena associated with biological
manipulation and storage of DNA. Most recently, the development of a
base-pair level model for DNA made it possible to incorporate the effects
of nucleotide sequence and the negative charge of DNA in new mesoscale
models of complex DNA-protein assemblies, which yielded insights into the
role of DNA deformability in gene regulation. This talk will give an
overview of resent research of the speaker and his collaborators on the
above topics, in particular, methods for solving equilibrium equations of
the elastic rod model for DNA in cases when self-contact is present,
determining the likely configurational states and their free energy, and
the importance of sequence dependent DNA elasticity in modeling of looped
DNA-protein assemblies. Emphasis will be on models of Lac
repressor-mediated DNA loop, which are supported by available data and
yield experimentally verifiable conclusions about the influence of DNA
deformability on the mechanism of gene regulation. Students and
non-specialists are welcome.

**Refreshments will be served at 3:30pm in the Physics Commons

My talk will consist of two parts, image-driven simulation of
brain-tumor brain interactions and simulation of deformable biological
vesicles.

In the first part, I will discuss image-driven simulation of
mechanical effects of tumor growth. In particular, I will present the
application of an optimization framework to the deformable
co-registration of images of brains with tumors to images of normal
brains. I will discuss a simplified reaction-diffusion model for the
glioma growth, a linearly elastic model for the brain tumor, their
nonlinear coupling, and a landmark-based registration problem that is
constrained by the physics. The registration problem is used to
recover unknown model parameters, like the growth rate of the
tumor. Using this framework, we can use only four parameters to
recover mass effects caused by the tumor.

In the second part, I will discuss the simulation of fluid vesicles
suspended in Stokesian fluids. Fluid membranes are area-preserving
interfaces that resist bending. They are common in biophysics as
they model cell membranes, vesicles, and viral particles. I will
discuss time-stepping and stability restrictions, a new semi-implicit
time marching scheme for such problems, and present numerical results
that demonstrate the effectiveness of the overall algorithm.


Short bio: George Biros is an assistant professor in Mechanical
Engineering and Applied Mechanics, and Computer and Information
Science at the University of Pennsylvania. He received his BS in
Mechanical Engineering from Aristotle University Greece (1995), his MS
in Biomedical Engineering from Carnegie Mellon (1996), and his PhD in
Computational Science and Engineering also from Carnegie Mellon
(2000). He joined Penn in 2003 after a postdoctoral appointment at the
Courant Institute of Mathematical Sciences. He is affiliated with the
Computer Science Research Institute (CSRI) at Sandia National
Laboratories.