Russian version English version
Volume 18   Issue 1   Year 2023
Exploiting ensemble learning and negative sample space for predicting extracellular matrix receptor interactions

Abhigyan Nath1, Sudama Rathore1, Pangambam Sendash Singh2

1Department of Biochemistry, Pt. Jawahar Lal Nehru Memorial Medical College, Raipur, India
2Department of Computer Science, Banaras Hindu University, Varanasi, India

Abstract.  The extracellular matrix (ECM) is best described as a dynamic three-dimensional mesh of various macromolecules. These include proteoglycans (e.g., perlecan andagrin), non-proteoglycan polysaccharides (e.g., hyaluronan), and fibrous proteins (e.g., collagen, elastin, fibronectin, and laminin). ECM proteins are involved in various biological functions and their functionality is largely governed by interaction with other ECM proteins as well as trans-membrane receptors including integrins, proteoglycans such assyndecan, other glycoproteins, and members of the immunoglobulin superfamily. In the present work, a machine learning approach is developed using sequence and evolutionary features for predicting ECM protein-receptor interactions. Two different feature vector representations, namely fusion of feature vectors and average of feature vectors are used within corporation of the best representation employing feature selection. The current results show that the feature vector representation is an important aspect of ECM protein interaction prediction, and that the average of feature vectors performed better than the fusion of feature vectors. The best prediction model with boosted random forest resulted in 72.6 % overall accuracy, 74.4 % sensitivity and 70.7 % specificity with the 200 best features obtained using the ReliefF feature selection algorithm. Further, a comparative analysis was performed for negative sample subset selection using three sampling methods, namely random sampling, k-Means sampling, and Uniform sampling. k-Means based representative sampling resulted in enhanced accuracy (75.5 % accuracy with 80.8 % sensitivity, 68.1 % specificity and 0.801 AUC) for the prediction of ECM protein-receptor interactions in comparison to the other sampling methods. On comparison with other three state of the art protein-protein interaction predictors, it is observed that the latter displayed low sensitivity but higher specificity. The current work presents the first machine learning based prediction model specifically developed for ECM protein-receptor interactions.

Key words: ECM, receptor interaction, Boosting, Boosted Random Forest; ReliefF; Random Sampling; k-Means Sampling; Uniform Sampling.

Table of Contents Original Article
Math. Biol. Bioinf.
doi: 10.17537/2023.18.113
published in English

Abstract (eng.)
Abstract (rus.)
Full text (eng., pdf)
Supplementary data


  Copyright IMPB RAS © 2005-2023