Main navigation | Main content

HOME » SCIENTIFIC RESOURCES » Volumes

Abstracts and Talk Materials

November 17, 2011

In current genetic and clinical research, identification of disease specific variations particularly from non-coding RNA and cis-elements is a major bottleneck. Massively parallel sequencing of exome and transcriptome is widely being used to effectively interrogate the key protein-coding and non-coding RNA regions. In such scenarios, the deep sequencing data of exome and transcriptome could be used for estimating levels of allele-specific expression in diseased vs. control samples (case-control cohorts) and hence the identification of disease specific signatures. This provides a functional basis to identify the differentially expressed alleles, mono-allelic expression, imprinting of alleles and allele regulated alternative splicing. All such data and approaches together make a stronger strategy to predict the disease susceptibility alleles and their functional role in disease mechanism.
Our approaches at BioCOS Life Sciences using the Next Generation Sequencing (NGS) data analysis for the precise detection of allele’s differential expression becomes important in identifying causal/susceptibility genes by mapping their variance in both coding and non-coding DNA/RNA regions.

I will present our current research work on developing methods and data processing approaches, which can be applied in identification of the susceptibility alleles using the combined approaches from RNA-Seq and Exome-Seq data as well as directly predicting them from RNA-Seq data. The talk will also discuss the existing bottlenecks in the area and approaches to obtain high quality results with a focus on calling genotypes from RNA-Seq data.

I will present our current research work on developing methods and data processing approaches, which can be applied in identification of the susceptibility alleles using the combined approaches from RNA-Seq and Exome-Seq data as well as directly predicting them from RNA-Seq data. The talk will also discuss the existing bottlenecks in the area and approaches to obtain high quality results with a focus on calling genotypes from RNA-Seq data.

Read More...

Quantification from medical images involves three levels of developments:
- Modeling of the organs
- Extraction of the visual features
- Formulation of the quantification task
Regarding organ modeling, geometric encoding of the shape is designed as a tradeoff between flexibility and robustness. Encoding of the variability within a population is a complex task that can have drawbacks when handling pathological cases. On the other hand, generic anatomical knowledge, especially regarding the context of the organ, can provide rich and more robust information, with spatial relations for example.
Regarding the visual features, images are richer than they appear in terms of tissue signature, embedding multiscale information. The field of image processing has evolved slowly in the design of sophisticated organ-specific visual features, the majority of them remaining very basic. Future challenges remain open regarding the need to correlate multi-modal tissue signatures with physiological characteristics.
Formulation of the quantification task such as segmentation, tracking or detection of longitudinal changes can be formulated either with a deterministic or stochastic formalism. Algorithms remain poorly robust to image quality, lack of image calibration, parameter tuning and presence of pathologies. Finer interactions between algorithmic tuning and image content and better calibration of image content is currently under investigation to address this lack of robustness and reproducibility.

These three components of the pipeline will be discussed, with illustrations on brain, cardiac liver and obstetric data. Emphasis will be paid to the constraints of being fast and robust, in the context of handling large data sets with great variability and pathologies.

These three components of the pipeline will be discussed, with illustrations on brain, cardiac liver and obstetric data. Emphasis will be paid to the constraints of being fast and robust, in the context of handling large data sets with great variability and pathologies.

December 31, 1969

Deciphering the structure of gene regulatory networks is crucial to
understanding the functionality of genes as well as the behavior of
cells. To this end, a network topology estimator is developed in this
work based on the structural equation model (SEM) approach, which
capitalizes on naturally occurring genetic variations viewed as
statistical perturbations that enable inference of the causal
relationships between genes. The SEM offers a suitable framework for
the estimation of cyclic directed networks, but typically requires
searching over a huge parameter space, incurring prohibitively high
computational complexity. As gene networks are sparse, meaning that
the number of edges is relatively small when compared to the number of
all possible edges, the present work contributes a SEM-based
sparsity-aware inference methodology. Simulated tests demonstrate that
the novel method can markedly improve inference accuracy.

*A joint work with G. B. Giannakis *

December 31, 1969

We formulate the optimal intervention problem in genetic regulatory networks
as a minimal-perturbation of the network in order to force it to converge to
a desired steady-state distribution of gene regulation. We cast optimal
intervention in gene regulation as a convex optimization problem, thus
providing a globally optimal solution which can be efficiently computed
using standard techniques for convex optimization. The criteria adopted for
optimality is chosen to minimize potential adverse effects as a consequence
of the intervention strategy. We consider a perturbation that minimizes (i)
the overall energy of change between the original and controlled networks
and (ii) the time needed to reach the desired steady-state of gene
regulation. Moreover, we show that there is an inherent tradeoff between
minimizing the energy of the perturbation and the convergence rate to the
desired distribution. We further show that the optimal inverse perturbation
control is robust to estimation errors in the original network. The proposed
control is applied to the Human melanoma gene regulatory network.

We formulate the optimal intervention problem in genetic regulatory networks as a minimal-perturbation of the network in order to force it to converge to a desired steady-state distribution of gene regulation. We cast optimal intervention in gene regulation as a convex optimization problem, thus providing a globally optimal solution which can be efficiently computed using standard techniques for convex optimization. The criteria adopted for optimality is chosen to minimize potential adverse effects as a consequence of the intervention strategy. We consider a perturbation that minimizes (i) the overall energy of change between the original and controlled networks and (ii) the time needed to reach the desired steady-state of gene regulation. Moreover, we show that there is an inherent tradeoff between minimizing the energy of the perturbation and the convergence rate to the desired distribution. We further show that the optimal inverse perturbation control is robust to estimation errors in the original network. The proposed control is applied to the Human melanoma gene regulatory network.

Solving modern biomedical problems, especially, those involving genome data, requires advanced computational and analytical methods. The huge quantities of data and escalating demands of modern biomedical research increasingly require the sophistication and power of computational techniques for their pattern discovery. Key techniques include relational data management, pattern recognition, data mining, modelling and visualization of biomedical data. In this talk, I will demonstrate recent methodologies and data structures for gathering high-quality approximations and modelling of genomic information, and will use these innovations as the basis for developing methods to cluster and visualize biomedical data in pattern discovery.

December 31, 1969

In recent years, biomedicine has been faced with difficult high-throughput
small-sample classification problems, which are typically validated with
re-sampling error estimation methods such as cross-validation. While
heuristically designed error estimation techniques may be acceptable in
problems where large amounts of data are available, the small-sample setting
is different because asymptotic results are not meaningful and validation
becomes a critical issue. A recently proposed classifier error estimator
places the problem in a signal estimation framework in the presence of
uncertainty, thereby permitting a rigorous optimal solution in a
minimum-mean-square error (MMSE) sense. The uncertainty in this model is
relative to the parameters of the feature-label distributions, resulting in
a Bayesian approach to error estimation. The same Bayesian framework also
produces the theoretical MSE for both Bayesian error estimators and
arbitrary error estimators, where uncertainty is again relative to the
unknown model parameters and conditioned on the observed sample. Thus, the
Bayesian error estimator has a unique advantage over classical error
estimators in that its mathematical framework naturally gives rise to a
practical expected measure of performance given a fixed sample.

November 14, 2011

Within only a decade since the first draft of the human genome, we’ve witness astonishing pace of development of technologies for high throughput molecular profiling that probe various aspects of genome biology and its relationship to tumorigenesis and cancer treatment. There have been giant steps towards cataloging massive amounts of data and providing fairly good annotation information. However, computational methods that tried to tease out the relationship between the genotype and its functional readout - in normal and cancer states – revealed a semantic gap that is yet to be bridged. Narrowing this gap is essential in order to develop meaningful clinical decision support technologies.

In addition to imaging modalities which give the gross tissue level properties, crucial decisions in the context of oncology therapy selection require molecular level information that are increasingly captured by the emerging sequencing modalities. We undertook part in multiple studies aiming to understand the tumor heterogeneity and response to chemotherapy. Our efforts span several complementary modalities. In this talk I will provide several examples from our recent high throughput genomic studies:

1. DNA Sequencing: Assembly and downstream analysis of genomic data from normal individuals to understand and establish variation within normal individuals at the single nucleotide and structural level as well as the functional impact of these variations.

2. RNA Sequencing, CNV and DNA methylation: analysis in the context of chemo- and biological therapy response in breast and ovarian cancer.

3. Integration into a computational framework that combines genome-wide DNA methylation, gene expression and copy number variation data in a comprehensive fashion with the aim of finding mechanistic associations as well as signatures indicative of therapy resistance.

Our goal is to include these modalities in a Comprehensive Clinical Decision Support system where we need to integrate sequencing with imaging, pathology and other clinical data.

In addition to imaging modalities which give the gross tissue level properties, crucial decisions in the context of oncology therapy selection require molecular level information that are increasingly captured by the emerging sequencing modalities. We undertook part in multiple studies aiming to understand the tumor heterogeneity and response to chemotherapy. Our efforts span several complementary modalities. In this talk I will provide several examples from our recent high throughput genomic studies:

1. DNA Sequencing: Assembly and downstream analysis of genomic data from normal individuals to understand and establish variation within normal individuals at the single nucleotide and structural level as well as the functional impact of these variations.

2. RNA Sequencing, CNV and DNA methylation: analysis in the context of chemo- and biological therapy response in breast and ovarian cancer.

3. Integration into a computational framework that combines genome-wide DNA methylation, gene expression and copy number variation data in a comprehensive fashion with the aim of finding mechanistic associations as well as signatures indicative of therapy resistance.

Our goal is to include these modalities in a Comprehensive Clinical Decision Support system where we need to integrate sequencing with imaging, pathology and other clinical data.

Single-particle cryo electron microscopy provides
images of biological macromolecular complexes with
spatial sampling on the order of 1-2 Angstrom.
Combining on the order of 100,000 such images can result
in 3-D reconstructions of the electron scattering
intensity of the complex with a spatial resolution as
fine as 4-5 Angstrom. Due to damage in the imaging
process, each complex is imaged only once and therefore
having a homogeneous ensemble of complexes is
important. Algorithms and results will be presented
for the case where the complexes are not homogeneous
and the reconstruction yields a statistical description
of the electron scattering intensity rather than a
single unique intensity. Related work on computed
electron tomography, where the electron scattering
intensity of individual complexes are determined but at
lower resolution will also be presented.

A perusal of the contemporary biological literature involving high-throughput data sets reveals the generation of a vast amount of data and an enormous number of models (classifiers, clusters, networks) derived from this data via a plethora of algorithms. There tends to be four interrelated characteristics common to these publications: (1) no experimental deign, (2) data sets where the number of measured variables greatly exceeds the number of replications, (3) algorithms whose performance is unknown for the populations to which they are applied – and often known to work poorly when applied to a small number of replicates, and (4) models that are epistemologically meaningless because they have not been validated. Hence, we find ourselves in a position somewhat akin to that confronted by Immanuel Kant in the Eighteenth Century when he famously asked, “How is metaphysics as a science possible?” Certainly there was a lot of “metaphysical” talk in the air, but to what sureties had it led? To address the problem, Kant had to tackle the meaning of science and then appreciate what constraints had to be placed on metaphysical statements to make them “scientific.” Fortunately for us, we do not have to take on the monumental task of characterizing scientific knowledge, an endeavor that stretched from Galileo to Einstein. But we do have to consider what constraints must be placed on biological statements to make them meaningful, that is, so that they constitute biological scientific knowledge. Moreover, we need to address a critical methodological scientific issue addressed by Kant: What differentiates productive observation of Nature from “groping in the dark,” to use his phrase?

December 31, 1969

Dynamic processes such as cell motility and deformation are key components of numerous scenarios including cell division & differentiation, morphogenesis, immune response strategies, but also parasite invasion, cancer development & proliferation and host-pathogen interactions. Continuous advances in microscopy imaging techniques have allowed scientists to shed light on many of these processes over extended periods of time, yielding huge amounts of time-lapse imaging data in multiple colors and various experimental conditions. In such context, visual interpretation and manual analysis have proved to be limited by the lack of reproducibility, user bias and fatigue. Scientists thus progressively turn to automatic quantification methods able to process spatiotemporal data in a robust and systematic manner. In this work we present a novel framework for automatic cell segmentation and tracking based on the theory of deformable models, and show how such a versatile mathematical tools can be used to extract various information related to cellular motility, shape analysis and morpho-dynamic studies.

One method used to generate a 3D image in medical imaging with Computerized Tomography (CT) is the helical cone beam scan. This method, obviously, generates a large amount of data. As in all methods of using X-ray scans, one is motivated to reduce the amount of radiation to which the patient is exposed. Combined with the desire to reduce the amounts of data being generated, one is motivated to develop efficient sampling methods. In the work presented here, the essential frequency support of a scanned body is estimated, bounded and then used to develop a sparse sampling method resulting in minimal loss of quality in the reconstructed image. Our initial results show that we can use this sparse sampling to reduce the data by at least factor of two.*coauthored by Tamir Ben-Dory.*

We present our work toward a statistical model of changes in gene
expression through four stages in the development of cervical cancer.
These stages are characterized in part by changes in the proportion of
cells of particular types. For example, normal tissue is organized in
layers with more well-differentiated cells at the surface and with
less differentiated, but more actively dividing cells further inside
the tissue. Neoplastic lesions shift the balance of types, at least
partly by having relatively more of the less differentiated types and
having fewer of the well-differentiated types. In our model, we make
use of this insight by postulating the existence of several distinct
(and unknown) types of cells which are present in all stages of the
progression, but whose relative proportions (also unknown) change
during the course of the progression. We then study differential
expression across the postulated cell types.

The problem of variable selection is useful for identifying the principal drivers of differential
gene-response under one or more treatments, phenotypes, or conditions. Once identified, such drivers can be targeted as potential
knockouts or enhancers in drug discovery or diagnostic testing. In high throughput data such as gene or
protein expression the large number of variables has made it impractical to implement all but the
simplest univariate methods for variable selection, e.g., detecting significant shifts in t-test or Wilcoxon test statistics.
We propose an alternative approach based on detecting significant shifts in patterns of
connectivity of genes in a correlation graph or concentration graph. Remarkably, it is precisely when the sample size is small
that the approach is scalable, e.g., to whole genome analysis. Furthermore a statistical performance analysis establishes phase transition behaviors and tight approximations to false discovery rate that can be used for error control. We will illustrate the approach on
several gene expression datasets.

Principal component analysis (PCA) is a widely applied method for extracting structure from samples of high dimensional biological data.
Often there exist misalignments between different samples and this can cause severe problems in PCA if not properly taken into account. For example, subject-dependent temporal differences in gene expression response to a treatment will create relative time shifts in the samples that decohere the PCA analysis. The sensitivity of PCA to such misalignments is severe, leading to phase transitions that can be studied using the spectral the theory of high dimensional matrices. With this as motivation, we propose a new method of PCA, called misPCA, that explicitly accounts for the effects of misalignments in the samples. We illustrate misPCA on clustering longitudinal temporal gene expression data.

* With Arnau Tibau-Puig, Ami Wiesel, and Raj Rao Nadakuditi*

December 31, 1969

This paper extends the problem of vaccine adverse reaction detection by incorporating historical medical conditions. We propose a novel measure called dual-lift for this task, and formulate this problem in the framework of constraint pattern mining. We present a pattern mining algorithm DLiftMiner which utilizes a novel approach to upper bound the dual-lift measure for reducing the search space. Experimental results on both synthetic and real world datasets show that our method is effective and promising.

In this work, we develop and evaluate a semiautomatic
algorithm for segmentation and morphological assessment of
the dimensions of the ciliary muscle in Visante Anterior
Segment Optical Coherence Tomography images.
Furthermore, we investigate the morphology of the ciliary
muscle during the act of accommodation in a population of
children. Increasing accommodative response was correlated
with increases in the thickness of CMTMAX (p=<0.001) and
CMT1 (p=<0.001), and decreases in the thickness of CMT3
(p=<0.001).

November 15, 2011

There has been an explosion of non-invasive biomedical sensing
modalities that have revolutionized our ability to probe the
biomedical world. Often decisions have to be made on the basis of
these increasingly high-dimensional observations. An example would be
the determination of cancer or stroke from indirect tomographic
projection measurements. The problem is frequently exacerbated by the
lack of labeled training samples from which to learn class models. In
many cases, however, there exists a latent low-dimensional sensing
structure that can potentially be exploited for inferencing aims.
This work investigates the impact of latent sensing structure on
supervised classification performance when the data dimension scales
to infinity faster than the number of samples. In contrast to some
existing studies, here the classification difficulty is held fixed and
finite as the data dimension scales. For a binary supervised
classification problem with Gaussian likelihood functions, it is shown
that the asymptotic error probability converges to that of pure
guessing if the sensing structure is totally ignored, whereas it
converges to the Bayes risk if the sensing structure is sufficiently
regular and the classification method is "sensing aware". It is also
shown, however, that without suitable regularity in the latent
low-dimensional sensing structure, it is impossible to attain
nontrivial asymptotic error probability. These findings are validated
through various simulations. Additional numerical results for support
vector machines and sensitivity to mismatch between true and assumed
structure are also provided.

December 31, 1969

DNA copy number variations (CNVs) are biological indicators that characterize cancer genomes. Identifying cancer-causing CNVs is a challenging problem due to the high dimensionality of CNV features and the heterogeneity of patients. Our objective is to build robust predictive models based on CNV data using machine learning techniques for accurate cancer phenotype prediction and identification of cancer-causing CNVs by integrating multiple datasets and modeling biological prior knowledge. We introduce an alignment-based kernel (probe-alignmetn kernel )and a graph-based learning model (HyperPrior) to analyze cancer DNA CNV data for identifying biomarkers and classifying tumor samples. Two computational problems are tackled 1) integration of multiple CNV datasets generated from arrayCGH platforms with unmatched probes; 2) introducing spatial constraints among arrayCGH probe features into predictive models. We applied the methods to predict tumor grade and subtypes, and discover biomarkers from several bladder cancer datasets. In the experiments, while achieving competitive classification results, the proposed methods also identified common cancer CNV regions that contain loci and genes with known association with bladder cancer. Further analysis shows that the identified marker genes enriched several biological functions and gene-gene interaction subnetworks with cancer relevance.*Joint work with Ze Tian*

November 16, 2011

Recent breakthroughs in optical microscopy have enabled in vivo imaging of the embryonic heart as it develops and gains function. Despite these advances, it remains difficult to simultaneously characterize heart morphology, heart function (the embryonic heart is beating before it is fully developed), and gene expression levels. We have developed computational tools to capture, process, and combine images acquired with different microscopy modalities, at different temporal and spatial scales, and over multiple samples, in an effort to build a multi-dimensional model of the beating and developing heart where morphology, function, and genetics can be simultaneously studied. Here, I will discuss image acquisition protocols and reconstruction strategies to overcome instrumentation and biological limitations that prevent simultaneous acquisition of these large, high-dimensional data sets. These tools will facilitate quantitative and systematic characterization of both morphology and function and study their relationship to genetic and epi-genetic factors that affect development in normal and diseased hearts.

Read More...

In this talk I will discuss the novel experimental designs for
large-scale multiple hypothesis testing problems. Testing to determine
which genes are differentially expressed in a certain disease is a
classic instance of multiple testing in medical informatics. Tremendous
progress has been made in high-dimensional inference and testing
problems by exploiting intrinsic low-dimensional structure. Sparsity is
perhaps the simplest model for low-dimensional structure. It is based on
the assumption that the signal of interest can be represented as a
combination of a small number of elementary components. Sparse recovery
is the problem of determining which components are needed in the
representation based on measurements of the signal. For example,
diseases are often characterized by a relatively small number of genes,
which can be identified using high-throughput experimental techniques.
This talk focuses on two issues related to this line of research.

1. Most theory and methods for sparse recovery are based on non-adaptive measurements. I will discuss the advantages of sequential measurement schemes that adaptively focus sensing using information gathered throughout the measurement process. In particular, I will show that sequential testing procedures can be signiﬁcantly more powerful than non-sequential methods in the high-dimensional setting.

2. The standard sparse recovery problem involves inferring sparse linear functions. I will discuss generalizations of the standard problem to the recovery of sparse multilinear functions. Such functions are characterized by multiplicative interactions between the input variables, with sparsity meaning that relatively few of all conceivable interactions are present. This problem is motivated by the study of interactions between processes in complex networked systems (e.g., among genes and proteins in living cells). Our results extend the notion of compressed sensing from the linear sparsity model to nonlinear forms of sparsity encountered in complex systems. In contrast to linear sparsity models, in the multilinear case the pattern of sparsity can significantly affect sensing requirements.

1. Most theory and methods for sparse recovery are based on non-adaptive measurements. I will discuss the advantages of sequential measurement schemes that adaptively focus sensing using information gathered throughout the measurement process. In particular, I will show that sequential testing procedures can be signiﬁcantly more powerful than non-sequential methods in the high-dimensional setting.

2. The standard sparse recovery problem involves inferring sparse linear functions. I will discuss generalizations of the standard problem to the recovery of sparse multilinear functions. Such functions are characterized by multiplicative interactions between the input variables, with sparsity meaning that relatively few of all conceivable interactions are present. This problem is motivated by the study of interactions between processes in complex networked systems (e.g., among genes and proteins in living cells). Our results extend the notion of compressed sensing from the linear sparsity model to nonlinear forms of sparsity encountered in complex systems. In contrast to linear sparsity models, in the multilinear case the pattern of sparsity can significantly affect sensing requirements.

The lecture will present biological imaging topics ranging from fundamentals in microscopy to specific methods and algorithms for the processing and quantification of 2- and 3-D+t images sequences in biological microscopy. We will demonstrate algorithms of PSF approximations for image deconvolution, image segmentation, multi-particle tracking and active contours models for cell shape and deformation analysis. We will illustrate the application of our methods in projects related to the study of the dynamics of genes in cell nuclei, the movement of parasites in cells and the detection and tracking of microbes in cells. One specific goal in biological imaging is indeed to automate the quantification of dynamics parameters or the characterization of phenotypic and morphological changes occurring as a consequence of cell/cell or pathogens/host cells interactions. The availability of this information and its thorough analysis is indeed of key importance to help deciphering underlying molecular mechanisms of e.g. infectious diseases.

Being able to determine the conformation of protein-bound DNA is extremely important for biological and medical research, including improvement of drug creation and administration methods. There are many protein-bound DNA complexes whose conformations have been cataloged in the Protein Data Base (PDB). However, most large complexes have not been cataloged and it would be difficult to do so. The topology of such large complexes can be solved through a process called tangle analysis, more specifically using a difference topology method. We will define and describe tangle analysis and how difference topology can be used to determine the topology of large protein-bound DNA complexes. Furthermore, we will discuss why topology alone is not enough and how we can add geometry to topological solutions in order to determine an accurate conformation for large protein-bound DNA complexes.

While the chemical structure of DNA is well understood, determining how genome-encoded components function in an integrated manner to perform cellular and organismal function is still an open challenge. The talk will motivate that imaging, more specifically the extraction of quantitative information, plays a critical role in this process. Such measurements will enable the automatic monitoring of cellular and intracellular events, and providing information about specific molecular mechanisms in individual cells.

By providing some specific examples it will be illustrated how specific computer vision algorithms enable the analysis of data sets and complex biological specimens that cannot be analyzed through manual inspection. The talk will highlight specific examples on how image analysis algorithms can be used to extract high content data. Specifically I will show how image segmentation methods are used to extract protein expression information in a novel sequential multiplexing process GE developed. In addition it will be discussed how statistical shape analysis methods can be applied to assess cellular morphology as well as the structure of entire organisms. Finally, it will be shown how the analysis of apparent motion can be used to monitor cardiomyocyte populations.

While imaging data potentially has much to add to models for systems biology, the usefulness of imaging information is dependent on the quantitative nature of the data and other aspects of its quality. Developing an awareness of the important long-term factors and challenge will help ensure acceptance of image analysis methods. Today image analysis methods are already used to study complex biological processes.

By providing some specific examples it will be illustrated how specific computer vision algorithms enable the analysis of data sets and complex biological specimens that cannot be analyzed through manual inspection. The talk will highlight specific examples on how image analysis algorithms can be used to extract high content data. Specifically I will show how image segmentation methods are used to extract protein expression information in a novel sequential multiplexing process GE developed. In addition it will be discussed how statistical shape analysis methods can be applied to assess cellular morphology as well as the structure of entire organisms. Finally, it will be shown how the analysis of apparent motion can be used to monitor cardiomyocyte populations.

While imaging data potentially has much to add to models for systems biology, the usefulness of imaging information is dependent on the quantitative nature of the data and other aspects of its quality. Developing an awareness of the important long-term factors and challenge will help ensure acceptance of image analysis methods. Today image analysis methods are already used to study complex biological processes.

November 14, 2011

Modern optical microscopy has grown into a multi-dimensional imaging
tool. It is now possible to record dynamic processes in living specimens in
their spatial context and temporal order, yielding information-rich 5-D
images
(3-D space, time, spectra).Of particular interest are complex and dynamic
tissuemicroenvironments that play critical roles in health and disease,
e.g., tumors,
stem-cell niches, brain tissue surrounding neuroprosthetic devices, retinal
tissue, cancer stem-cell niches, glands, and immune system tissues.

The task of analyzing these images exceeds human ability due to the sheer volume of the data (images routinely exceed 20GB in size), its structural complexity, and the dynamic behaviors of cells and organelles. First, there is a need for automated systems to assist the human analyst to map the tissue anatomy, quantify structural associations, identify critical events, map event locations and timing to the tissue anatomic context, identify and quantify spatial and temporal dependencies, produce meaningful summaries of multivariate measurement data, and compare perturbed and normal datasets for testing hypotheses, exploration, and systems modeling. Beyond automation, there is a need for ³computational sensing² of tissue patterns and cell behaviors that are too subtle for the human visual system to detect.

In this talk, I will describe large-scale application of image processing, active machine learning, multivariate clustering, and parallel computation methods that enable scalable analysis of multi-dimensional microscopy data. A particularly valuable application of these methods is to validate the large-scale automated analysis results. All of the software from this work is free and open source (www.farsight-toolkit,org).

The task of analyzing these images exceeds human ability due to the sheer volume of the data (images routinely exceed 20GB in size), its structural complexity, and the dynamic behaviors of cells and organelles. First, there is a need for automated systems to assist the human analyst to map the tissue anatomy, quantify structural associations, identify critical events, map event locations and timing to the tissue anatomic context, identify and quantify spatial and temporal dependencies, produce meaningful summaries of multivariate measurement data, and compare perturbed and normal datasets for testing hypotheses, exploration, and systems modeling. Beyond automation, there is a need for ³computational sensing² of tissue patterns and cell behaviors that are too subtle for the human visual system to detect.

In this talk, I will describe large-scale application of image processing, active machine learning, multivariate clustering, and parallel computation methods that enable scalable analysis of multi-dimensional microscopy data. A particularly valuable application of these methods is to validate the large-scale automated analysis results. All of the software from this work is free and open source (www.farsight-toolkit,org).

In this talk I will describe some of our efforts in the area of translational medical imaging,
and illustrate how mathematics and formalism play a fundamental role.
I will start with our work on brain imaging, where we have developed
entire analysis pipelines, going from
fixing basic mathematical errors in the classical formulas
of high resolution diffusion imaging (HARDI), all the way
to studying gender and kinship in brain connectivity networks and
to helping neuro-surgeons in deep brain stimulation procedures.
I will then present some of our work on the analysis of the
structure of HIV and other viruses with data
obtained from cryo-tomography, a critical step in vaccine
development. Additional applications for
helping surgeons in the operating room will be mentioned as well.

In this talk, we will describe a new interactive procedure for segmenting 3D data sets using a mixture of ideas from control and image processing. More precisely, using a Lyapunov control design, a balance is established
between the influence of a data-driven gradient flow
and the human’s input over time. Automatic segmentation
is thus smoothly coupled with interactivity. An application
of the mathematical methods to orthopedic segmentation
is shown, demonstrating the expected transient and steady
state behavior of the implicit segmentation function and
auxiliary observer.

Large Data Challenges in Medical Imaging and Bioinformati

November 18, 2011

Recent increases in detector coverage and trigger frequencies have opened up new clinical applications in modern Computed Tomography, but have also led to an explosion in the volume of CT raw datasets. This represents a particular challenge for accurate tomographic image reconstruction, particularly when using a model-based iterative framework based on Maximum A Posteriori estimation. Inclusion of detector physics, tube response, noise statistics, and image modeling involves certain complexity that drives up reconstruction time. However, recent results have started to demonstrate the significant potential of model-based iterative reconstruction for ultra-low-dose imaging aimed at improving patient safety, as well as high quality results in other targeted applications such as low contrast complex abdomen imaging and high-resolution medullar and cortical bone. This poses a particular challenge to come up with fast convergent algorithms that do not trade-off significant quality for speed, and are amenable to modern parallel computing hardware.

This talk will present the modeling framework for high quality model-based tomographic reconstruction and its advantages relative to alternative iterative approaches designed primarily with concern about reconstruction speed. In the proposed approach, speed and quality can be thought of as relatively orthogonal design elements, to the extent that convergence is reasonably achieved. First, the formulation of the optimization problem fully defines the target quality level as a function of the number and accuracy of the models designed to explicitly explain x-ray attenuation measurements based on realistic modeling of scanner behavior and non-idealities. Second, a globally convergent optimization algorithm chosen among a variety of potential alternatives is optimized to realize the performance targets for fast convergence, efficient implementation, and massive parallelization for practical applications. The development and continuous amelioration of such tools and models for tomographic reconstruction promise the establishment of a new platform for iterative reconstruction in modern CT that may someday replace standard analytical methods for routine high-quality low-dose imaging.

This talk will present the modeling framework for high quality model-based tomographic reconstruction and its advantages relative to alternative iterative approaches designed primarily with concern about reconstruction speed. In the proposed approach, speed and quality can be thought of as relatively orthogonal design elements, to the extent that convergence is reasonably achieved. First, the formulation of the optimization problem fully defines the target quality level as a function of the number and accuracy of the models designed to explicitly explain x-ray attenuation measurements based on realistic modeling of scanner behavior and non-idealities. Second, a globally convergent optimization algorithm chosen among a variety of potential alternatives is optimized to realize the performance targets for fast convergence, efficient implementation, and massive parallelization for practical applications. The development and continuous amelioration of such tools and models for tomographic reconstruction promise the establishment of a new platform for iterative reconstruction in modern CT that may someday replace standard analytical methods for routine high-quality low-dose imaging.

December 31, 1969

Cigarette smoking is a well-known risk factor of lung carcinogenesis. It is still unclear how smoking affects lung tumor progression, recurrence and metastasis, and whether smoking could be used as a predictor of lung cancer treatment outcome. In this study, we investigated the effect of smoking intensity on lung cancer treatment failure, and the interactions between smoking and clinicopathological factors in lung cancer progression.

Sparse decomposition methods are effective tools in a myriad biomedical inverse problems. However, in many settings reconstruction is only an intermediate goal preceding additional quantitative analysis. For instance, we may wish to classify tissue types in microscope images or identify tumors or lesions based on computed tomography data. This talk describes how sparse image decomposition methods can be used in conjunction with multiscale set estimation methods to improve subsequent quantitative analyses on large medical datasets. For instance, sparse decomposition for tissue differentiation breaks down in images with boundaries, but multiscale set estimation can be used to accurately identify regions where sparse decomposition can be effectively applied. Similarly, sparse image reconstruction methods alone can spend significant computing resources on estimating features irrelevant to the quantitative goals, but by incorporating multiscale set estimation metrics into the objective function we can perform accurate quantitative analysis much more quickly. This talk will cover both the theoretical underpinnings of these methods and their application to challenging large-scale problems in microendoscopy and tomography.

Recent advances in high-throughput technologies for measuring protein-protein interactions have yielded genome-scale protein interaction networks of various organisms. These large-scale protein-protein interaction data can provide invaluable resources for studying the complex machineries in these organisms. In this work, we present several computational methodologies that can be used to transform these data into novel biological insights. Furthermore, we discuss how we can utilize the existing protein interaction data for more effective and robust analysis of gene expression data.

November 17, 2011

Recent advancements in DNA sequencing technologies have led to wide dissemination of instrumentation, resulting data and excitement. As a result of declining costs and increasing in throughput, there is a rapid growth trajectory in the amount of sequence data production. It is predicted that DNA sequence data will soon become one of the largest data types requiring powerful infrastructure development and deployment in both software and hardware in order to enable routine and robust handling and analysis.

This tutorial will guide participants through multiple topics regarding the next generation sequencing (NGS) data production and processing. Emphasis will be placed on both didactic presentation and group discussion in the following areas: (1) What is happening; (2) The excitement; (3) Best practice-lessons from the 1000 Genomes Project; (4) Remaining bottlenecks in data handling; and (5) A view toward the future.

The HGSC has been pioneering the deployment of multiple NGS platforms (Roche 454, Illumina, SOLiD, PacBio, Ion Torrent), and spearheaded personal genomics (Waston Genome, Lupski Genome, and Beery Family), population genomics (1000 Genomes), cohort disease mapping (ARIC Studies), and Cancer Studies (TCGA, familial cancer). A great deal of experience in processing and handling NGS data and variant calling have been accumulated, which form a solid foundation to meet future challenges.

My group has been a major part of the 1000 Genomes Project for variant calling, imputation and integration for both low-coverage (~4X/genome) and exome data. We developed integrative variant analysis pipelines-Atlas2 and SNPTools (http://www.hgsc.bcm.tmc.edu/cascade-tech-software-ti.hgsc), which achieved high quality SNP and INDEL datasets in the 1000 Genomes Phase I project. I will share this experience as one example.

This tutorial will guide participants through multiple topics regarding the next generation sequencing (NGS) data production and processing. Emphasis will be placed on both didactic presentation and group discussion in the following areas: (1) What is happening; (2) The excitement; (3) Best practice-lessons from the 1000 Genomes Project; (4) Remaining bottlenecks in data handling; and (5) A view toward the future.

The HGSC has been pioneering the deployment of multiple NGS platforms (Roche 454, Illumina, SOLiD, PacBio, Ion Torrent), and spearheaded personal genomics (Waston Genome, Lupski Genome, and Beery Family), population genomics (1000 Genomes), cohort disease mapping (ARIC Studies), and Cancer Studies (TCGA, familial cancer). A great deal of experience in processing and handling NGS data and variant calling have been accumulated, which form a solid foundation to meet future challenges.

My group has been a major part of the 1000 Genomes Project for variant calling, imputation and integration for both low-coverage (~4X/genome) and exome data. We developed integrative variant analysis pipelines-Atlas2 and SNPTools (http://www.hgsc.bcm.tmc.edu/cascade-tech-software-ti.hgsc), which achieved high quality SNP and INDEL datasets in the 1000 Genomes Phase I project. I will share this experience as one example.