STRUCTURAL CLASSIFICATION OF PROTEINS USING IMAGE BASED MACHINE LEARNING

Classification of proteins is an important area of research that enables better grouping of proteins either by their function, evolutionary similarities or in their structural makeup. Structural classification is the area of research that this thesis focuses on. We use visualizations of proteins to build a machine learning class prediction model, that successfully classifies proteins using the Structural Classification of Proteins (SCOP) framework. SCOP is a well-researched classification with many approaches using a representation of a proteins secondary structure in a linear chain of structures. This thesis uses a novel approach of rendering a three dimensional visualization of the protein itself and then applying image based machine learning to determine a protein’s SCOP classification. The resulting convolutional neural network (CNN) method has achieved average accuracies in the range 78-87% on the 25PDB dataset, which is better than or equal to the existing methods.

STRUCTURAL CLASSIFICATION OF PROTEINS USING IMAGE BASED MACHINE LEARNING

Classification of proteins is an important area of research that enables better grouping of proteins either by their function, evolutionary similarities or in their structural makeup. Structural classification is the area of research that this thesis focuses on. We use visualizations of proteins to build a machine learning class prediction model, that successfully classifies proteins using the Structural Classification of Proteins (SCOP) framework. SCOP is a well-researched classification with many approaches using a representation of a proteins secondary structure in a linear chain of structures. This thesis uses a novel approach of rendering a three dimensional visualization of the protein itself and then applying image based machine learning to determine a protein’s SCOP classification. The resulting convolutional neural network (CNN) method has achieved average accuracies in the range 78-87% on the 25PDB dataset, which is better than or equal to the existing methods.

MoonProt 3.0: an update of the moonlighting proteins database

MoonProt 3.0 (http://moonlightingproteins.org) is an updated open-access database storing expert-curated annotations for moonlighting proteins. Moonlighting proteins have two or more physiologically relevant distinct biochemical or biophysical functions performed by a single polypeptide chain. Here, we describe an expansion in the database since our previous report in the Database Issue of Nucleic Acids Research in 2018. For this release, the number of proteins annotated has been expanded to over 500 proteins and dozens of protein annotations have been updated with additional information, including more structures in the Protein Data Bank, compared with version 2.0. The new entries include more examples from humans, plants and archaea, more proteins involved in disease and proteins with different combinations of functions. More kinds of information about the proteins and the species in which they have multiple functions has been added, including CATH and SCOP classification of structure, known and predicted disorder, predicted transmembrane helices, type of organism, relationship of the protein to disease, and relationship of organism to cause of disease.

DECISION TREE BASED INFORMATION INTEGRATION FOR AUTOMATED PROTEIN CLASSIFICATION

We propose a novel technique for automatically generating the SCOP classification of a protein structure with high accuracy. We achieve accurate classification by combining the decisions of multiple methods using the consensus of a committee (or an ensemble) classifier. Our technique, based on decision trees, is rooted in machine learning which shows that by judicially employing component classifiers, an ensemble classifier can be constructed to outperform its components. We use two sequence- and three structure-comparison tools as component classifiers. Given a protein structure and using the joint hypothesis, we first determine if the protein belongs to an existing category (family, superfamily, fold) in the SCOP hierarchy. For the proteins that are predicted as members of the existing categories, we compute their family-, superfamily-, andfold-level classifications using the consensus classifier. We show that we can significantly improve the classification accuracy compared to the individual component classifiers. In particular, we achieve error rates that are 3–12 times less than the individual classifiers' error rates at the family level, 1.5–4.5 times less at the superfamily level, and 1.1–2.4 times less at the fold level.

Comparison of proteins based on segments structural similarity.

We present here a simple method for fast and accurate comparison of proteins using their structures. The algorithm is based on structural alignment of segments of Calpha chains (with size of 99 or 199 residues). The method is optimized in terms of speed and accuracy. We test it on 97 representative proteins with the similarity measure based on the SCOP classification. We compare our algorithm with the LGscore2 automatic method. Our method has the same accuracy as the LGscore2 algorithm with much faster processing of the whole test set, which is promising. A second test is done using the ToolShop structure prediction evaluation program and shows that our tool is on average slightly less sensitive than the DALI server. Both algorithms give a similar number of correct models, however, the final alignment quality is better in the case of DALI. Our method was implemented under the name 3D-Hit as a web server at http://3dhit.bioinfo.pl/ free for academic use, with a weekly updated database containing a set of 5000 structures from the Protein Data Bank with non-homologous sequences.