Instance Based Learning in Protein Interaction Networks

Josifoski, Martin and Trivodaliev, Kire (2017) Instance Based Learning in Protein Interaction Networks. In: PROCEEDINGS of the 14th Conference on Informatics and Information Technology. Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University in Skopje, Macedonia, Skopje, Macedonia, pp. 200-205. ISBN 978-608-4699-07-1

[img]
Preview
Text
978-608-4699-07-1_pp200-205.pdf

Download (490kB) | Preview
Official URL: http://ciit.finki.ukim.mk

Abstract

One of the essential challenges in proteomics is the computational function prediction. Protein interaction networks (PINs), as the richest source of information, can be utilized to solve this problem. PINs are modeled via graphs where proteins correspond to nodes, interactions to edges, and protein function to labels associated to nodes. The problem of computational function prediction now becomes one for proper labeling of its corresponding node. In this paper the PIN graph is used to derive a continuous vector representation of its nodes using semisupervised learning. The approach used employs a biased random walk procedure with a flexible notion of a nodes neighborhood, which efficiently explores diverse neighborhoods. The produced vectors maximize the likelihood of preservation of the graph topology locally and globally. Once the vector representations of the nodes are produced learning is modeled as a set of binary classifications where a single classification corresponds to a single label (from a set of possible labels). In this single label classification the objective is to determine the existence (or non-existence) of a label to which aim the node vector representations are noted positive (if node has the label) or negative (node does not have label). Experiments are performed using a highly reliable human protein interaction network. Classification is done using two well known algorithms, namely, k-nearest neighbors and support vector machines. Results prove that this approach can be used to successfully tackle protein function prediction, but also provide insight to improvements that can make it comparable to state-of-the-art methods.

Item Type: Book Section
Subjects: International Conference on Informatics and Information Technologies > Students Session
Depositing User: Vangel Ajanovski
Date Deposited: 29 Nov 2017 18:27
Last Modified: 29 Nov 2017 18:27
URI: http://eprints.finki.ukim.mk/id/eprint/11404

Actions (login required)

View Item View Item