Evaluating Ligand-Binding Site Predictions: A Comparison of Geometry-based, Machine Learning, and Deep Learning Approaches

View at publisher Summarise with VIRAT
Authors: S. Gopalakrishnan S., C. Sureshkumar
Cited by: 0
References: 23

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Binding site prediction between proteins and ligands is essential in understanding functional mechanisms, thereby enabling advancements in pharmacological studies. For this study, three prediction models: fpocket, P2Rank, and DeepBindPoc, were considered to search for ligand pockets in proteins in various configurations. FPocket relies on geometric concepts to search for pockets employing alpha spheres, whereas P2Rank applies the use of machine learning to evaluate the physicochemical properties of the pockets. Binding sites prediction in DeepBindPoc applies the concept of deep learning, thereby delivering precise results. The dataset collected included protein structure models obtained from the experimental results of the Protein Data Bank PDB. Full accuracy was demonstrated in the near-native sites prediction by DeepBindPoc, recognising its efficiency to be greater in comparison to P2Rank and fpocket. None of these models were able to identify specific sites in pockets that are either flexible or cryptic. The limitations of binding site prediction by static structure-based models can be well understood, thereby justifying the need to incorporate dynamic models. The models generated significantly different results for the pocket priority position as well as confidence levels of assessment, thereby justifying the limitations of their performance as a result of the difference generated by their predictive capability and computing power. The study highlights the proper application of pocket prediction algorithms to aid structure-based drug design by understanding their limitations. Improvements in understanding the mobility of proteins along with different models of predictions will provide potential applications in improving the accuracy of pocket prediction algorithms to identify binding sites in complex proteins.

Fields of Study & Subjects

Computational Drug Discovery Methods Protein Structure and Dynamics Protein Interaction Studies and Fluorescence Analysis

Publication details

Journal / source
International Conference on Image Processing and Capsule Networks (ICIPCN)
Pages
168–174
Publication date
27 January 2026
Language
EN