Research Project
10303852
There are significant advantages from translating genome sequences into proteins, where there is a large body of accumulated knowledge regarding their relationships among sequence, structure and function. Advances in genome sequencing are producing a deluge of data that can be used to train and test prediction methods to identify the characteristics of various mutants by building atop the large functional protein data. Clinicians need to know the functional behavior of mutants - whether they are neutral or deleterious - whether they affect protein structure ? whether they affect protein dynamics - whether they affect protein binding specificity. Protein structures have local environments for each amino acid in the sequence, and usually amino acids at each position are compatible with their local environment. This leads to strongly correlated amino acids as manifested in the multiple sequence alignments. This project will combine protein sequence and structure data together with amino acid properties and their correlations to characterize each site in the protein structure to investigate the hypothesis that outliers in the distributions over the important amino acid properties for each position will negatively impact functionality, i.e. they will be deleterious mutants. The project will drill down deeply to learn what is the nature of the impaired mechanism. Two diverse approaches will be taken in the two aims: Aim 1 will investigate the amino acid property distributions to identify the properties that best characterize each position in the sequence and structure, and determine how the outliers negatively impact the functional structures, dynamics and binding characteristics. Preliminary results show that the deleterious mutants usually have a significantly broader range of single amino acid properties for the deleterious mutants. Data from these analyses will be fed into Aim 2 where two type of machine learning approaches ? Extreme Learning Machines and Random Forests will be jointly applied. Preliminary results show that incorporating just one amino acid property yields significant gains over existing methods. One of the major strengths of this project is that results from the two Aims will be exchanged frequently to achieve improved predictions for both approaches. The project builds on the long experience of the PIs in datamining from protein structures and sequences, as well as previous machine learning applications. Important potential outcomes include a more reliable, more informed understanding of how mutants affect function. In addition, the project aims to predict connections of mutants to specific diseases. The results of the project will be important for drug development, because the specific part of the protein where function is impaired will be identified, to allow drug developers to narrow their focus onto more limited parts of a protein that is targeted for drug design. The predictors established by this project will also have the potential to screen for large numbers of previously unknown mutations that could be used to identify specific regions of a protein structure susceptible to further disease-related mutations.
