Slobodan Kalajdziski

Comparative Analysis of Three Efficient Approaches for Retrieving Protein 3D Structures

In this paper, comparative analysis is presented of our three 3D structure-based approaches for the efficient retrieval of protein molecules. All approaches rely on the 3D structure of the proteins. In the first approach, the Spherical Trace Transform is applied to protein 3D structures in order to produce geometry based descriptors. Additionally, some biological properties of the protein are taken, thus forming better integrated descriptor. In the second approach, some modification of the ray based descriptor is applied on the backbone of the protein molecule. In the third approach, wavelet transformation is applied on the distance matrix of the Calpha atoms which form the backbone of the protein. The SCOP database was used to evaluate the retrieval accuracy. We provide some experimental results of the retrieval accuracy of our three approaches. The results show that the ray based approach gives the best retrieval accuracy (97,5%), while it is simpler and faster than the other two approaches.

Protein Classification by Matching 3D Structures

In this paper, a 3D structure-based approach is presented for the efficient classification of protein molecules. The method relies on the geometric 3D structure of the proteins. After proper positioning of the 3D structures, the spherical trace transform is applied to them to produce geometry - based descriptors, which are completely rotation invariant. Additionally, some biological properties of the protein are taken, and added to the geometry-based descriptor, thus forming better integrated descriptor. We have used nearest neighbour classification on the previously extracted descriptors. A part of the FSSP/DALI database, was used to evaluate the classification accuracy of this system. The results show that this method achieves more than 92 percent classification accuracy while it is simpler and faster than the DALI method. We provide some experimental results of the implemented system.

Protein Function Prediction Based on Neighborhood Profiles

The recent advent of high throughput methods has generated large amounts of protein interaction network (PIN) data. A significant number of proteins in such networks remain uncharacterized and predicting their function remains a major challenge. A number of existing techniques assume that proteins with similar functions are topologically close in the network. Our hypothesis is that the simultaneous activity of sometimes functionally diverse functional agents comprises higher level processes in different regions of the PIN. We propose a two-phase approach. First we extract the neighborhood profile of a protein using Random Walks with Restarts. We then employ a “chi-square method”, which assigns k functions to an uncharacterized protein, with the k largest chi-square scores. We applied our method on protein physical interaction data and protein complex data, which showed the later perform better. We performed leave-one-out validation to measure the accuracy of the predictions, revealing significant improvements over previous techniques.

Protein Secondary Structure Prediction Method Based on Neural Networks

Protein secondary structure prediction remains an open and important problem in life sciences as a first step towards the crucial tertiary structure prediction. In [3], a protein secondary structure prediction algorithm called PSIPRED presents an innovative approach - feeding the neural network (NN) with a position specific scoring matrix as input data. Starting from this idea, in this paper we propose a method based on breaking down the single first level NN classifier, into three separate ones, for each of the secondary structure elements (SSE) types, in order to achieve greater generalization qualities of the first level classifying algorithm. We also introduce the use of sparsely connected feed-forward NNs, instead of the classic fully interconnected one. This network architecture gains considerable speed improvements (for both the training and the testing part of the algorithm) by omitting the most of the remote units that have the poorest influence on the selected amino acid. The prediction results are encouraging - the predictions are similar to the target PDB data and we achieve better accuracy, compared to the predictions obtained from the original PSIPRED algorithm.

Spherical Mapping Based Descriptors for 3D Object Matching

Matching 3D objects by their similarity is a fundamental problem in computer vision, multimedia databases, molecular biology, computer graphics and a variety of other fields. A challenging aspect of this problem is to find a suitable shape signature/descriptor that can be constructed and compared quickly, while still discriminating between similar and dissimilar shapes. We find that the major problems in comparing 3D mesh objects lie in the non-uniform vertex sampling and level of detail distribution, in the non-uniform polygon topology and in mesh-representation anomalies, so the primary motivation behind the work presented in this paper is the introduction of mesh-parameterization which brings meshes into a form having uniform vertex sampling, uniform polygon topology and filtered anomalies, by spherically mapping the mesh surface. Further, we present two approaches in inferring shape-descriptors from the spherically mapped objects and the results from the conducted experiments.

3D Object Matching Using Spherical Mapping

Matching 3D objects by their similarity is a fundamental problem in computer vision, multimedia databases, molecular biology, computer graphics and a variety of other fields. A challenging aspect of this problem is to find a suitable shape signature/descriptor that can be constructed and compared quickly, while still discriminating between similar and dissimilar shapes. We find that the major problems in comparing 3D mesh objects lie in the non-uniform vertex sampling and level of detail distribution, in the non-uniform polygon topology and in mesh-representation anomalies, so the primary motivation behind the work presented in this paper is the introduction of mesh-parameterization which brings meshes into a form having uniform vertex sampling, uniform polygon topology and filtered anomalies, by spherically mapping the mesh surface. Further, we present two approaches in inferring shape-descriptors from the spherically mapped objects and the results from the conducted experiments

Novel Gene Ontology Based Distance Metric for Function Prediction via Clustering in Protein Interaction Networks

The increased availability of large-scale protein-protein interaction (PPI) data has made it possible to have a network level understanding of the basic components and organization of the cell machinery. A significant number of proteins in protein interaction networks (PIN) remain uncharacterized and predicting their function remains a major challenge. We propose a novel distance metric for PIN clustering. First we augment the graph representing the PIN with weights derived from Gene Ontology (GO) semantic similarity and we use this augmented representation in a random walk with restarts (RWR) process. The distance between a pair of proteins is calculated from the steady state distribution of the RWR. We validate our approach by function prediction via clustering in a purified and reliable Saccharomyces cerevisiae PIN. We show that the rise of function prediction performance when using the novel distance metric is significant, as compared to traditional approaches.

Model of Cloud-Based Services for Data Mining Analysis

New cloud-based services are being developed constantly in order to meet the need for faster, reliable and scalable methods for knowledge discovery. The major benefit of the cloud-based services is the efficient execution of heavy computation algorithms in the cloud simply by using Big Data storage and processing platforms. Therefore, we have proposed a model that provides data mining techniques as cloud-based services that are available to users on their demand. The widely known data mining algorithms have been implemented as Map/Reduce jobs that are been executed as services in cloud architecture. The user simply chooses or uploads the dataset to the cloud, makes appropriate settings for the data mining algorithm, executes the job request to be processed and receives the results. The major benefit of this model of cloud-based services is the efficient execution of heavy computation data mining algorithm in the cloud simply by using the Ankus - Open Source Big Data Mining Tool and StarfishHadoop Log Analyzer. The expected outcome of this research is to offer the integration of the cloud-based services for data mining analysis in order to provide researchers with reliable collaborative data mining analysis model.

SHOPIN: Semantic Homogeneity Optimization in Protein Interaction Networks

Protein interaction networks (PINs) are argued to be the richest source of hidden knowledge of the intrinsic physical and/or functional meanings of the involved proteins. We propose a novel method for computational protein function prediction based on semantic homogeneity optimization in PIN (SHOPIN). The SHOPIN method creates graph representations of the PIN augmented by inclusion of the semantics of the proteins and their interacting contexts. Network wide semantic relationships, modeled using random walks, are used to map the augmented PIN graphs in a new semantic metric space. The method produces a hierarchical partitioning of the PIN optimal in terms of semantic homogeneity by iterative optimization of the ratio of between clusters dissimilarities and within clusters similarities in the new semantic metric space. Function prediction is done using cluster wide-hierarchy high function enrichment. Results validate the rationale of the SHOPIN method placing it right next to state-of-the-art approaches performance wise.

Protein Classification Based on 3D Structures and Fractal Features

To understand the structure-to-function relationship, life sciences researchers and biologists need to retrieve similar structures and classify them into the same protein fold. In this paper, we propose a 3D structure-based approach for efficient classification of protein molecules. Classification is performed in three phases. In the first phase, we apply fractal descriptor matching as a filter. Then, protein structures which satisfy the fractal and radius tolerance are classified in the second phase. In this phase, 3D Fourier Transform is applied in order to produce rotation invariant descriptors. Additionally, some properties of primary and secondary structure are taken. In the third phase we use k nearest neighbor classifier. Our approach achieves 86% classification accuracy with applying fractal filter, and 92% without fractal filter. It is shown that fractal filter significantly shorten the classification time. Our system is faster (seconds) than DALI system (minutes, hours, days), and we still get satisfactory results.