Improving Computational Predictions of Cis-regulatory Binding Sites in Genomic Data

Cis-regulatory elements are the short regions of DNA to which specific regulatory proteins bind and these interactions subsequently influence the level of transcription for associated genes, by inhibiting or enhancing the transcription process. It is known that much of the genetic change underlying morphological evolution takes place in these regions, rather than in the coding regions of genes. Identifying these sites in a genome is a non-trivial problem. Experimental (wet-lab) methods for finding binding sites exist, but all have some limitations regarding their applicability, accuracy, availability or cost. On the other hand computational methods for predicting the position of binding sites are less expensive and faster. Unfortunately, however, these algorithms perform rather poorly, some missing most binding sites and others over-predicting their presence. The aim of this thesis is to develop and improve computational approaches for the prediction of transcription factor binding sites (TFBSs) by integrating the results of computational algorithms and other sources of complementary biological evidence. Previous related work involved the use of machine learning algorithms for integrating predictions of TFBSs, with particular emphasis on the use of the Support Vector Machine (SVM). This thesis has built upon, extended and considerably improved this earlier work. Data from two organisms was used here. Firstly the relatively simple genome of yeast was used. In yeast, the binding sites are fairly well characterised and they are normally located near the genes that they regulate. The techniques used on the yeast genome were also tested on the more complex genome of the mouse. It is known that the regulatory mechanisms of the eukaryotic species, mouse, is considerably more complex and it was therefore interesting to investigate the techniques described here on such an organism. The initial results were however not particularly encouraging: although a small improvement on the base algorithms could be obtained, the predictions were still of low quality. This was the case for both the yeast and mouse genomes. However, when the negatively labeled vectors in the training set were changed, a substantial improvement in performance was observed. The first change was to choose regions in the mouse genome a long way (distal) from a gene over 4000 base pairs away - as regions not containing binding sites. This produced a major improvement in performance. The second change was simply to use randomised training vectors, which contained no meaningful biological information, as the negative class. This gave some improvement over the yeast genome, but had a very substantial benefit for the mouse data, considerably improving on the aforementioned distal negative training data. In fact the resulting classifier was finding over 80% of the binding sites in the test set and moreover 80% of the predictions were correct. The final experiment used an updated version of the yeast dataset, using more state of the art algorithms and more recent TFBSs annotation data. Here it was found that using randomised or distal negative examples once again gave very good results, comparable to the results obtained on the mouse genome. Another source of negative data was tried for this yeast data, namely using vectors taken from intronic regions. Interestingly this gave the best results