Genome-Wide Prediction of Intrinsic Disorder; Sequence Alignment of Intrinsically Disordered Proteins

Abstract

Intrinsic disorder (ID) is defined as a lack of stable tertiary and/or secondary structure under physiological conditions in vitro. Intrinsically disordered proteins (IDPs) are highly abundant in nature. IDPs possess a number of crucial biological functions, being involved in regulation, recognition, signaling and control, e.g. their functional repertoire complements the functions of ordered proteins. Intrinsically disordered regions (IDRs) of IDPs have a different amino-acid composition than structured regions and proteins. This fact has been exploited for development of predictors of ID; the best predictors currently achieve around 80% per-residue accuracy. Earlier studies revealed that some IDPs are associated with various human diseases, including cancer, cardiovascular disease, amyloidoses, neurodegenerative diseases, diabetes and others. We developed a methodology for prediction and analysis of abundance of intrinsic disorder on the genome scale, which combines data from various gene and protein databases, and utilizes several ID prediction tools. We used this methodology to perform a large-scale computational analysis of the abundance of (predicted) ID in transcripts of various classes of disease-related genes. We further analyzed the relationships between ID and the occurrence of alternative splicing and Molecular Recognition Features (MoRFs) in human disease classes. An important, never before addressed issue with such genome-wide applications of ID predictors is that - for less-studied organisms - in addition to the experimentally confirmed protein sequences, there is a large number of putative sequences, which have been predicted with automated annotation procedures and lack experimental confirmation. In the human genome, these predicted sequences have significantly higher predicted disorder content. I investigated a hypothesis that this discrepancy is not correct, and that it is due to incorrectly annotated parts of the putative protein sequences that exhibit some similarities to confirmed IDRs, which lead to high predicted ID content. I developed a procedure to create synthetic nonsense peptide sequences by translation of non-coding regions of genomic sequences and translation of coding regions with incorrect codon alignment. I further trained several classifiers to discriminate between confirmed sequences and synthetic nonsense sequences, and used these predictors to estimate the abundance of incorrectly annotated regions in putative sequences, as well as to explore the link between such regions and intrinsic disorder. Sequence alignment is an essential tool in modern bioinformatics. Substitution matrices - such as the BLOSUM family - contain 20x20 parameters which are related to the evolutionary rates of amino acid substitutions. I explored various strategies for extension of sequence alignment to utilize the (predicted) disorder/structure information about the sequences being aligned. These strategies employ an extended 40 symbol alphabet which contains 20 symbols for amino acids in ordered regions and 20 symbols for amino acids in IDRs, as well as expanded 40x40 and 40x20 matrices. The new matrices exhibit significant and substantial differences in the substitution scores for IDRs and structured regions. Tests on a reference dataset show that 40x40 matrices perform worse than the standard 20x20 matrices, while 40x20 matrices - used in a scenario where ID is predicted for a query sequence but not for the target sequences - have at least comparable performance. However, I also demonstrate that the variations in performance between 20x20 and 20x40 matrices are insignificant compared to the variation in obtained matrices that occurs when the underlying algorithm for calculation of substitution matrices is changed.