Nucleotide Level

1. A Novel Approach for Integrating and Assessing Computational Tools to Detect Transcription Factor Binding Sites Kathryn Dempsey*, Daniel Quest*°, Mohammad Shafiullah*, Dhundy Bastola* and Hesham Ali*° *College of Information Science & Technology, University of Nebraska at Omaha, Omaha, NE 68182 °Department of Pathology & Microbiology, University of Nebraska Medical Center, Omaha, NE 68198 Abstract Transcription factor binding sites (TFBS) regulate the expression of genes in the cell. Their discovery remains one of the most challenging problems in molecular biology. To solve this problem, traditional techniques have been supplemented by the development of computational prediction methods. Today over 100 algorithms exist for detecting TFBS in various problem domains, yet it remains unclear how to assess their performance. Comparing these tools across datasets is nearly impossible due to a lack of standards for running programs and reporting results. This work proposes a novel method for standardizing runtime procedures and assessing tool performance. To appropriately compare each method, a standard reporting format was developed for each tool. We developed a pipeline framework to integrate, score, and evaluate TFBS identification for each method. We evaluated 9 computational methods for detecting TFBS and obtained statistics that describe their performance. These results allowed us to rank each method by performance on a range of datasets. Computational detection of TFBS remains a challenging problem. Sensitivity and specificity remain low for the algorithms regardless of dataset. Our method for comparing tools exposes a unique view into the behavior of TFBS detection and enables the improvement of methods in the future. Motivation One of the most complex problems facing scientists today is the prediction of TFBS patterns in silico. Over 100 computational prediction tools have been developed to address the problem of predicting TFBS. As a result, it can be unclear to scientists which available method is right for their specific domain. It is possible for a method that performs poorly on one dataset to perform better on another, but until recently there had been no way to determine this information quickly. Comparing even a small set of TFBS prediction tools manually requires extensive knowledge of each program, hours of manual computation, and the ability to streamline a number of different input/output formats into a format that is easily scored. Our objective is to make the process of comparing and evaluating tool performance automated and accurate. Our work allows many popular, publicly-available tools to be analyzed on a number of different datasets and parameters over the period of a few days. The Motif Tool Assessment Platform, or MTAP, allows scientists to learn more about which TFBS detection tool is right for their area of expertise. We are able to show that when tools are run through an automated pipeline, they perform just as well (if not better) than when they are executed manually. MTAP converts results for each tool we examine into one standard format. This format can be used for evaluation, visualization, and integrative prediction. On a representative set of data, we are able give insight into which programs perform best in varying domains. Finally, we can suggest the overall areas where TFBS prediction is lacking, and where TFBS prediction algorithms perform well. Methods The Motif Tool Assessment Platform, or MTAP is a platform for the integration of motif discovery tools. To accommodate the wide diversity of libraries and dependencies, the MTAP architecture was implemented using Java, Perl, Python, and C++. In order to handle computational demands, MTAP evaluates prediction tools simultaneously on a clustered computer. To ensure that our automation process reported results that were similar to results from manual investigation, we included 5 of 13 tools from Tompa et al.’s evaluation of TFBS detection programs [1]. The remaining 8 methods were not included because they were unobtainable, obsolete, or existed only as web versions. To expand our methods beyond what had already been evaluated, we included additional tools based on popularity, availability, and the ability to be executed on the command-line. Input / Output Input datasets were created using peer-reviewed public databases. Data was collected from RegulonDB, DBTBS, and Tompa et al.’s assessment on fly, human, yeast, and mouse. Outputs from each tool vary, and so a custom parsing class was designed for each individual tool. The results are stored in the pipeline scaffold framework for scoring, and stored in a standard output format for later retrieval. MTAP gives insight into program performance over different problem features.MTAP was run over the same 400bp sequences generated with RegulonDB as in Figure 5. Figure 6 was produced with MTAP by plotting the ratio of information content found in the motif relative to background sequence size versus sSn performance. Increasing the information content of the motif and decreasing the background sequence noise is expected to improve tool prediction performance. However, not all tools are able to take advantage of this feature. Some methods, such as AnnSpec, AlignAce, MEME, and Weeder, (MEME and Weeder highlighted) exploit an increase in information. As a result, site sensitivity increases with number of sequences. Other programs do not appear to take advantage of the amount of information given. This highlights one area where TFBS prediction improvement can be focused. Discussion and Conclusions Computational detection of TFBS remains a challenging problem. Automation of computational TFBS prediction makes assessment a possibility where there had been no genome wide means of assessment before. Overall, sensitivity and specificity remain low for the algorithms regardless of dataset, but by examining patterns exposed by our work we can better suggest critical areas for improvement. We have shown that not only is the task of automating this process possible, it performs just as well as when runs are executed by hand. Automation cuts down the amount of time and labor required, it also eliminates opportunities for human/user error when running the tools, scoring the output, etc. We have also shown that phylogenetic programs do not always outperform non-phylogenetic programs. These results beg more questions about phylogenetic programs and the additional information used in their implementation, such as number and proximity of orthologs used, and size of the phylogenetic tree. All these things can now be investigated with the aide of the MTAP. Finally, we have presented representative data sets that give a glimpse into how many different ways we can rank and compare TFBS prediction tools. Our method for comparing tools exposes a unique view into the behavior of TFBS detection and enables the improvement of methods in the future. References [1] M. Tompa et al., Assessing Computational Tools for the Discovery of Transcription Factor Binding Sites. Nature Biotechnology, vol. 23, no. 1, January 2005, 137 - 144. [2] H. Salgado et al., Regulondb (version 4.0): transcriptional regulation, operon organization and growth conditions in escherichia coli k-12., Nucleic Acids Res 32 (2004), no. 1,: 65-67. [3] http://biobase.ist.unomaha.edu/ Acknowledgement This project was supported by the NIH grant number P20 RR016469 from the INBRE Program of the National Center for Research Resources. We would like to thank the developers of the motif detection tools for their help and for making their tools available. Nucleotide Level We define four statistics to describe how results are scored at the nucleotide level. If a base is part of an actual site and is predicted by the tool, it is a true positive. If it is only predicted, it is a false positive. If it is part of an actual site and not predicted, it is a false negative, and if it is neither predicted nor an actual site, it is a true negative Site Level We define three measures to describe how results are scored on the site level, including the use of a threshold which defines what percentage of a motif can be considered a correct prediction. If a site is an actual site and is predicted by the tool by at least the threshold percentage (6.7% for these tests), it is a true positive. If it is only predicted, it is a false positive. If it is part of an actual site and not predicted, it is a false negative. False negatives are not counted on the site level. Phylogenetic vs. non-Phylogenetic Some TFBS detection methods take prediction one step further. Similar to allowing students to use supplemental material on a test, these programs make use of orthologous sequence data and a phylogenetic tree (both generated by the user) in an attempt to raise sensitivity and specificity. We implement methods for automatically generating orthologous sequence data and phylogenetic trees based off ribosomal RNA to incorporate phylogenetic programs into MTAP. Results Program outputs and scoring revealed a wide array of observations on current TFBS detection methods. Some key observations include the following: Our method of automation returns sensitivity and specificity that are comparable to “hand-made” results.Figure 4 shows the nucleotide and site sensitivity, and nucleotide specificity for Tompa’s “hand-made” method versus our automated method. The graph suggests is that our automated method returns results that are comparable to Tompa, evidence that our method is reliable for producing accurate results. Nucleotide and site sensitivity are higher overall with our method, suggesting that automation is an acceptable route compared to manual assessment. Phylogenetic programs are comparable in performance to non-phylogenetic programs over RegulonDB. Figure 5 depicts statistics for phylogenetic vs. non-phylogenetic programs over sequences 400 bp upstream of the coding sequence from RegulonDB using genomes NC_000913, NC_007946, and AC_000091. Phylogenetic program PhyloGibbs compares favorably with MEME, Weeder, and AlignAce. PhyME returns low sensitivity, perhaps due to its resistance to make predictions. More evaluations are needed over different genome selections to explore the benefits of phylogenetic approaches. Regardless, this illustrates the capabilities of MTAP for evaluation of different classes of algorithms.