LABi: Laboratory for Advanced Biomedical Informatics @ LSUS

.:. Selected Research Projects & Publications .:.


.:. iNNfovis: Neural-Network-Augmented Information Visualization

Large amounts of high-dimensional biomedical data not only create the need for the analysis of data and interpretation of results, but also the need for the development of tools and methods that can handle such data. Many techniques commonly used to analyze such data are graphical in nature with low intrinsic dimensionality - ability to unambiguously represent the data dimensionality. In fact, they can represent only a small number of variables at a time, requiring users to discard potentially useful data. The algorithms we have developed are based on Kohonen’s self-organizing map algorithm where the (dis)placement of records in a neural-network augmented plot is directed by all (or a subset of) dimensional values, which significantly improves the treatment of dimensional values. Therefore, users no longer have to choose which dimensions to use to visualize the data in a classic low-dimensional visualization. Moreover, these techniques reveal non-trivial patterns and correlations in data as well as organize large data sets of multi-dimensional records into meaningful groups. Records exhibiting similar patterns across multiple dimensions map in proximity of each other, while preserving their original dimensional values and positions along the primary mapping coordinates in classic, well understood visualizations. Neural network-augmented visualization techniques enhance knowledge extraction, targeting complex data and providing for a very small, if any, loss of information. Our user studies, where users were tasked to find relationships in different datasets, showed increased insight with minimal or no user training. Additionally, it is interesting to note that most users indicated that they could clearly see patterns that were not intuitively obvious using original low-dimensional visualization techniques.

.:. Application of machine learning and visualization of heterogeneous data sets to uncover relationships between translation and developmental stage expression of C. elegans mRNAs

The relationships between genes in neighboring clusters in a self-organizing map (SOM) and properties attributed to them are sometimes difficult to discern, especially when heterogeneous datasets are used. We have developed a novel approach to identify correlations between heterogeneous data sets. One data set, derived from microarray analysis of polysomal distribution, contained changes in the translational efficiency of Caenorhabditis elegans mRNAs resulting from loss of specific eIF4E isoform. The second data set contained expression patterns of mRNAs across all developmental stages. Two algorithms were applied to these data sets: a classical scatter plot and an SOM and their outputs were linked using a two-dimensional color scale. This approach revealed that an mRNA’s eIF4E dependent translational efficiency is strongly dependent on its expression during development. This correlation was not detectable with traditional, widely utilized one-dimensional color scale.

.:. Identification, Tracking and Visualization of Platelets in Intravital Microscopy

Intravital microscopy permits observation of live events in intact tissues to study a variety of issues, including quantifying cell-vessel wall interactions. The analysis of the parameters is labor-intensive, subjective and limited to broad categories of blood cell-vessel wall interactions. We have developed an algorithmic approach that aids the analysis, automatically and objectively detects and tracks platelets and expands information derived from such videos. We integrate computer vision and break the identification, tracking and visualization into well defined steps. In addition, we enable computationally efficient means of eliminating movement within a video, based on positional shifting of an identified feature.

.:. Visual and analytical tools for biomedical informatics

We have been applying visualization, data mining and statistical techniques to large multivariate or high-dimensional data sets in life sciences, from gene expression arrays, protein structure models to miRNAs and motifs. Chemists and biologists are visually oriented and find it easier to analyze and memorize graphical presentations of structures and compounds and they greatly benefit from visualization tools that help them not only present, but also analyze and process the results using the appropriate computational methods coupled with effective and efficient visualizations. Under the NCRR/IDEA program, we investigated clustering and existing cluster analysis methods in order to devise new techniques for analysis of multiple clustering results. Many optimization problems in bioinformatics have been identified as computationally intractable, and that includes the number of clusters for a clustering algorithm which remains NP-hard to compute. The visual tools were developed as part of a suite called Castle and include new analytical and visual tools and techniques to provide insights into single and multiple clustering algorithm results. This work facilitates the analysis of high-dimensional data utilizing visual techniques for data presentation. The new approaches for measuring cluster differences provide information on proximity and provide a means for visual projection of the records including the largest number of common cluster memberships among the records. This work is to be published within the six months, and the techniques developed are used as part of SBIR proposals. This project alone provided research experience for 21 undergraduate and 9 graduate/master’s students in the past 5 years. We will utilize the suite as part of available tools in the NCRR/IDEA BioComputing/Bioinformatics Core (2010-2015).

.:. Identification of markers for bladder premalignancy

This work includes DNA microarray analysis of the UPII-S40Tag mouse model for invasive bladder transitional cell carcinoma (TCC) that we worked on as part of a multidisciplinary team with Dr. John Clifford (LSUHSC-S, Biochemistry) and Dr. Anita L. Sabichi (MD Anderson Cancer Center) and others. We combined a transgenic mouse model for invasive bladder cancer (UPII-SV40Tag mice) with DNA microarray technology to determine molecular mechanisms involved in early TCC development and to identify new biomarkers for detection, diagnosis, and prognosis of TCC. We identified genes that are differentially expressed between the bladders of UPII-SV40Tag mice and their age-matched wild-type littermates at 3, 6, 20, and 30 weeks of age. These are ages that correspond to premalignant, carcinoma in situ, and early-stage and later stage invasive TCC, respectively.

• The biomedical laboratory provided the animal models and created the molecular and cell biological studies based on a hypothesis, and its results were computationally processed using different algorithms for microarray data analysis and visualization, and combined with techniques to enrich the data with additional information from existing databases.

• The computation identified an initial set of differentially expressed genes, which we further filtered in order to obtain of a list of genes that were potentially interesting to explore further.

• These genes are currently in the clinical trials (Phase III Chemoprevention) and under investigation as biomarkers by MD Anderson Cancer Center.

• The results of step 3 will be used for generation of additional hypotheses and will feed back into the biomedical laboratory for further studies that will determine the mechanisms through which they operate.