Intelligent Data Analysis: An International Journal

We have developed the Transcription Regulatory Regions Database (TRRD, http://www.bionet.nsc.ru/trrd/) designed for storage of data on the structural-functional organization of transcriptional regulatory regions of the eukaryotic genes and their expression patterns. TRRD contains experimentally supported data only. Knowledge extraction from scientific literature, storage and further application of the results are all stepwise, conforming to the Data Mining technology: i) knowledge extraction from scientific publications; ii) preprocessing (data cleaning, syntactic and semantic analysis; iii) data transformation; iv) application for prediction; v) interpretation of the obtained knowledge to resolve the timely issues in bioinformatics. TRRD contains a compilation of data on 2 344 genes, their 14 407 expression patterns, 3 490 regulatory units, and 10 135 transcription factor binding sites. TRRD is filled in by manual annotation of scientific publications. The information incorporated into TRRD is the result of annotations of 7 609 scientific papers. Sequence Retrieval System (SRS) is the main tool for search and navigation in TRRD. A large number of indexed fields in its SRS version allow the user to generate queries both within and between libraries. TRRD has thesauruses and search systems that provide additional options for data access. TRRD is currently linked to 20 worldwide information resources, including EMBL/GeneBank, Ensembl, EPD, SWISS-PROT, TRANSFAC, GDB, GeneCards, MGD, RGD, GO. The links serve as a framework for integration in a distributed network environment.

The network modelling technique is one of the most effective ways to represent the biological system models of almost any types, and also their further analysing and simulating. The work is aimed to develop the unified software toolkit for these purposes, capable to operate with different knowledge domains. Particularly, we are taking into account ecological and molecular-genetic systems of high complexity. The theoretical part of the work is concerned with adaptation of the visual modelling technique for this area, development of principles and methods to represent network models and related knowledge domains in formal way, and approaches to handle the complex networks of thousands and even more elements. The other part of the work includes the requirements analysis, developing the software components prototypes. The main components are: the database with capabilities mentioned, and visual modelling tool for biological networks with a high grade of adaptability and flexibility.

We developed Relational Data Mining approach which allows to overcome essential limitations of the Data Mining and Knowledge Discovery techniques. In the paper the approach was implemented to adapt the original `Discovery' system to the computational biology needs. The objects under consideration, eukaryotic transcription regulatory regions, are characterized by the great variety of context physicochemical and conformational DNA features. The currently available tools aimed at the regulatory regions analysis are sensitive to specific DNA features; therefore they produce poor results on complex heterogeneous data. Development of a method integrating the results of different recognition programs is a challenging task. We have developed the `ExpertDiscovery' system, which discovers the hierarchically complicating set of complex signals based on different elementary signals. It provides a powerful tool to construct a model of regulatory region generalizing the results of different programs. Besides, the system is an independent tool for analysis. In the paper we demonstrate that `ExpertDiscovery' outperforms the position weight matrix in the case when the elementary signals introduced to the system are nucleotides at specific positions. The system is able to discover biologically significant, simple to complex models of potential transcription factor binding sites for regulatory regions of interferon-inducible genes.

A principally new approach to the classifications of nucleotide sequences based on the “natural” classification concept is proposed. As a result of “natural” classification of the nucleotide sequences, we obtain regularity matrices, where nucleotides are interconnected by regularities. Method, algorithm and software system DNANatClass for performing the “natural” classification have been developed. Experimental results comparing weight matrices with regularity matrices are presented. In this experiment, site recognition by regularity matrices appears to be more accurate than by weight matrixes.

Development of reliable transcription factor binding site (TFBS) recognition methods is an important step in the large-scale genome analysis. The most of currently applied methods to predict functional TFBSs are hampered by the high false-positive rates that occur when too few functionally characterised sequences are available and only sequence conservation within a site core is considered. We propose two methods to search for binding sites (BSs) of peroxisome proliferator-activated receptor (PPAR) (peroxisome proliferator response elements, PPREs). The first method is the optimized dinucleotide position weight matrix (PWM) model, the second method represented by SiteGA model that used genetic algorithm with a discriminant function of locally positioned dinucleotides to infer the most important positions and dinucleotides. We used in our analysis two PPRE datasets, consisting of 37 and 98 BSs, correspondingly. We showed that dataset extension improved the accuracy of SiteGA, but not PWM model. Finally we combined both models (PWM and SiteGA) to the dataset of annotated human promoters (EPD). We demonstrated that the larger dataset and the longer window length supported notable growth of accuracies for PWM and SiteGA models. Consequently, a combined PWM and SiteGA application may better restrict the number of potential targets in the EPD promoter dataset.