An empirical analysis of instance-based transfer learning approach on protease substrate cleavage site prediction

Abstract

Classical machine learning algorithms presume the supervised data emerged from the same domain. Transfer learning on the contrary to classical machine learning methods; utilize the knowledge acquired from the auxiliary domains to aid predictive capability of diverse data distribution in the current domain. In the last few decades, there is a significant amount of work done on the domain adaptation and knowledge transfer across the domains in the field of bioinformatics. The computational method for the classification of protease cleavage sites is significantly important in the inhibitors and drug design techniques. Matrix metalloproteases (MMP) are one such protease that has a crucial role in the disease process. However, the challenge in the computational prediction of MMPs substrate cleavage persists due to the availability of very few experimentally verified sites. The objective of this paper is to explore the cross-domain learning in the classification of protease substrate cleavage sites, such that the lack of availability of one-domain cleavage sites can be furnished by the other available domain knowledge. To achieve this objective, we employed the TrAdaBoost algorithm and its two variants: dynamic TrAdaBoost and multisource TrAdaBoost on the MMPs dataset available at PROSPER. The robustness and acceptability of the TrAdaBoost algorithms in the substrate site identification have been validated by rigorous experiments. The aim of these experiments is to compare the performances among learner. The experimental results demonstrate the potential of dynamic TrAdaBoost algorithms on the protease dataset by outperforming the fundamental and other variants of TrAdaBoost algorithms.