Gaolin Zheng

Distinct transcriptional profiles characterize oral epithelium‐microbiota interactions

Transcriptional profiling, bioinformatics, statistical and ontology tools were used to uncover and dissect genes and pathways of human gingival epithelial cells that are modulated upon interaction with the periodontal pathogens Actinobacillus actinomycetemcomitans and Porphyromonas gingivalis. Consistent with their biological and clinical differences, the common core transcriptional response of epithelial cells to both organisms was very limited, and organism‐specific responses predominated. A large number of differentially regulated genes linked to the P53 apoptotic network were found with both organisms, which was consistent with the pro‐apoptotic phenotype observed with A. actinomycetemcomitans and anti‐apoptotic phenotype of P. gingivalis. Furthermore, with A. actinomycetemcomitans, the induction of apoptosis did not appear to be Fas‐ or TNFα‐mediated. Linkage of specific bacterial components to host pathways and networks provided additional insight into the pathogenic process. Comparison of the transcriptional responses of epithelial cells challenged with parental P. gingivalis or with a mutant of P. gingivalis deficient in production of major fimbriae, which are required for optimal invasion, showed major expression differences that reverberated throughout the host cell transcriptome. In contrast, gene ORF859 in A. actinomycetemcomitans, which may play a role in intracellular homeostasis, had a more subtle effect on the transcriptome. These studies help unravel the complex and dynamic interactions between host epithelial cells and endogenous bacteria that can cause opportunistic infections.

SBLAST: Structural Basic Local Alignment Searching Tools using Geometric Hashing

While much research has been done on finding similarities between protein sequences, there has not been the same progress on finding similarities between protein structures. Here we report a new algorithm (SBLAST) which discovers the largest common substructures between two proteins using a triangle-based variant of the geometric hashing of protein structures algorithm. The algorithm selects triples (triangles) of selected Ca atoms from all proteins in a protein structure database and creates a hash table using a key based on the three inter-atomic distances. Hash table hits from the triangles of a query protein are extended recursively to determine the largest common substructures less than a threshold deviation level (rmsd). Comparisons between a query protein and a preprocessed protein database can be performed in parallel. Because SBLAST does not rely on protein sequence alignment, common substructures can be detected in the absence of sequence conservation. SBLAST has been tested using the ASTRAL subset of the PDB.

Microarray Data Analysis Using Neural Network Classifiers and Gene Selection Methods

Different research groups have conducted independent gene expression studies on tissue samples from human lung adenocarcinomas [Bhattacharjee et al. 2001; Beer et al. 2002]. In this paper we (a) investigate methods to integrate data obtained from independent studies, (b) experiment with different gene selection methods to find genes that have significantly differential expression among different tumor stages, (c) study the performance of neural network classifiers with correlated weights, and (d) compare the performance of classifiers based on neural networks and its many variants on gene expression data. Raw cell intensity data were preprocessed for our analyses. Affymetrix array comparison spreadsheets were used to extract the overlapping probe sets for the data integration study. We considered neural network classifiers with random weights selected from a univariate normal distribution and optimized using Bayesian methods. The performance of the neural network was further enhanced using ensemble techniques such as bagging and boosting. The performance of all the resulting classifiers was compared using the Michigan and Harvard data sets from the CAMDA website. Three gene selection methods were used to find significant genes that could discriminate between the various stages of lung cancer. Significant genes, which were mined from the Gene Ontology (GO) database using the GoMiner and AmiGO packages, were found to be involved in apoptosis, angiogenesis, and cell growth and differentiation. Neural networks enhanced with bagging exhibited the best performance among all the classifiers we tested.

AN APPLICATION OF ASSOCIATION RULE MINING TO HLA-A*0201 EPITOPE PREDICTION

This paper presents a novel approach to epitope prediction based on the clustering of known T-cell epitopes for a given MHC class I allele (HLA-A*0201). A combination of association rules (ARs) and sequence-structure patterns (SSPs) was used to do the clustering of training set epitopes from the Antijen database. A regression model was then built from each cluster and a peptide from the test set was declared to be an epitope only if one or more of the models gave a positive prediction. The sensitivity (TP/TP+FN) of the AR/SSP regression models approach was higher than that of a single regression model built on the entire training set, and was also higher than the sensitivity measures for SYFPEITHI, Rankpep, and ProPred1 on the same test set.

A branch-and-bound approach to knowledge-based protein structure assembly

With the unprecedented growth in the size of sequence and structure databases, knowledge-based methods have become increasingly feasible for protein structure prediction. We developed a branch-and-bound method for structlets-based protein structure assembly. We explore the effectiveness of this approach by examining its capability to reconstruct the 3D structure of some proteins with known 3D structures. Although our algorithm involves exhaustive search, our BestFirst implementation of a branch-and bound strategy is able to eliminate around 2/3 of the total search space in order to find the optimal 3D assembly for a protein of interest.

Discovering sequence-structure patterns in proteins with variable secondary structure

Proteins that share a similar function often exhibit conserved sequence patterns. Sequence patterns help to classify proteins into families where the exact function may or may not be known. Research has shown that these domain signatures often exhibit specific three-dimensional structures. We have previously shown that sequence patterns combined with structural information, in general, have superior discrimination ability than those derived without structural information. However in some cases, divergent backbone configurations and/or variable secondary structure in otherwise well-aligned proteins make identification of conserved regions of sequence and structure problematic. In this paper, we describe improvements in our method of designing biologically meaningful sequence-structure patterns (SSPs) starting from a seed sequence pattern from any of the existing sequence pattern databases. Improved pattern precision is achieved by including conserved residues from coil regions that are not readily apparent from examination of multiple sequence alignments alone. Pattern recall is improved by systematically comparing the structure of all known true family members and to include all the allowable variations in the pattern residues.

Pooling evidence to identify cell cycle–regulated genes

Most of the biological studies have embraced statistical approaches to make inferences. It is common to have several independent experiments to test the same null hypothesis. The goal of research on pooling evidence is to combine the results of these tests to ask if there is evidence from the collection of studies to reject the null hypothesis. In this study, we evaluated four different pooling techniques (Fisher, Logit, Stouffer and Liptak) to combine the evidence from independent microarray experiments in order to identify cell cycle-regulated genes. We were able to identify a better set of cell cycle-regulated genes using the pooling techniques based on our benchmark study on budding yeast (Saccharomyces cerevisiae). Our gene ontology study on time series data of both the budding yeast and the fission yeast (Schizosaccharomyces pombe) showed that the GO terms that are related to cell cycle are significantly enriched in the cell cycle-regulated genes identified using pooling techniques.