| dc.description.abstract | Understanding Protein-Protein and Protein-DNA interaction is of fundamental importance in deciphering gene regulation and other biological processes in living cells. Traditionally, new interaction knowledge is discovered through biochemical experiments that are often labor intensive, expensive and time-consuming. Thus, computational approaches are preferred. Due to the abundance of sequence data available today, sequence-based interaction analysis becomes one of the most readily applicable and cost-effective methods.
One important problem in sequence-based analysis is to identify the functional regions from a set of sequences within the same family or demonstrating similar biological functions in experiments. The rationale is that throughout evolution the functional regions normally remain conserved (intact), allowing them to be identified as patterns from a set of sequences. However, there are also mutations such as substitution, insertion, deletion in these functional regions. Existing methods, such as those based on position weight matrices, assume that the functional regions have a fixed width and thus cannot not identify functional regions with mutations, particularly those with insertion or deletion mutations. Recently, Aligned Pattern Clustering (APCn) was introduced to identify functional regions as Aligned Pattern Clusters (APCs) by grouping and aligning patterns with variable width. Nevertheless, APCn cannot discover functional regions with substitution, insertion and/or deletion mutations, since their frequencies of occurrences are too low to be considered as patterns.
To overcome such an impasse, this thesis proposes a new APC discovery algorithm known as Pattern-Directed Aligned Pattern Clustering (PD-APCn). By first discovering seed patterns from the input sequence data, with their sequence positions located and recorded on an address table, PD-APCn can use the seed patterns to direct the incremental extension of functional regions with minor mutations. By grouping the aligned extended patterns, PD-APCn can recruit patterns adaptively and efficiently with variable width without relying on exhaustive optimal search. Experiments on synthetic datasets with different sizes and noise levels showed that PD-APCn can identify the implanted pattern with mutations, outperforming the popular existing motif-finding software MEME with much higher recall and Fmeasure over a computational speed-up of up to 665 times. When applying PD-APCn on datasets from Cytochrome C and Ubiquitin protein families, all key binding sites conserved in the families were captured in the APC outputs.
In sequence-based interaction analysis, there is also a lack of a model for co-occurring functional regions with mutations, where co-occurring functional regions between interaction sequences are indicative of binding sites. This thesis proposes a new representation model Co-Occurrence APCs to capture co-occurring functional regions with mutations from interaction sequences in database transaction format. Applications on Protein-DNA and Protein-Protein interaction validated the capability of Co-Occurrence APCs. In Protein-DNA interaction, a new representation model, Protein-DNA Co-Occurrence APC, was developed for modeling Protein-DNA binding cores. The new model is more compact than the traditional one-to-one pattern associations, as it packs many-to-many associations in one model, yet it is detailed enough to allow site-specific variants. An algorithm, based on Co-Support Score, was also developed to discover Protein-DNA Co-Occurrence APCs from Protein-DNA interaction sequences. This algorithm is 1600x faster in run-time than its contemporaries. New Protein-DNA binding cores indicated by Protein-DNA Co-Occurrence APCs were also discovered via homology modeling as a proof-of-concept. In Protein-Protein interaction, a new representation model, Protein-Protein Co-Occurrence APC, was developed for modeling the co-occurring sequence patterns in Protein-Protein Interaction between two protein sequences. A new algorithm, WeMine-P2P, was developed for sequence-based Protein-Protein Interaction machine learning prediction by constructing feature vectors leveraging Protein-Protein Co-Occurrence APCs, based on novel scores such as Match Score, MaxMatch Score and APC-PPI score. Through 40 independent experiments, it outperformed the well-known algorithm, PIPE2, which also uses co-occurring functional regions while not allowing variable widths and mutations. Both applications on Protein-Protein and Protein-DNA interaction have indicated the potential use of Co-Occurrence APC for exploring other types of biosequence interaction in the future. | en |
