A comparative analysis of machine learning classifiers for predicting protein-binding nucleotides in RNA sequences

• Published in: Computational and structural biotechnology journal 2022-01, Vol.20, p.3195-3207
• Main Authors: Agarwal, Ankita, Singh, Kunal, Kant, Shri, Bahadur, Ranjit Prasad
• Format: Article
• Language: English
• Subjects: Machine learning; Protein-binding nucleotides; Random forest classifier; RNA-protein interactions; Stratified cross validation
• ISSN: 2001-0370; 2001-0370
• DOI: 10.1016/j.csbj.2022.06.036

[Display omitted] •RNA are master players in various cellular and biological processes and RNA-protein interactions are vital for proper functioning of cellular machineries.•Knowledge of binding sites is crucial to decipher their functional implications.•RNA NC-triplet and NC-quartet features could give reasonably high performance.•RF model outperformed other machine learning classifiers with 85% accuracy and 0.93 AUC and performed better than few existing methods.•An online webserver “Nucpred” is developed with trained model and freely accessible for scientific community. RNA-protein interactions play vital roles in driving the cellular machineries. Despite significant involvement in several biological processes, the underlying molecular mechanism of RNA-protein interactions is still elusive. This may be due to the experimental difficulties in solving co-crystallized RNA-protein complexes. Inherent flexibility of RNA molecules to adopt different conformations makes them functionally diverse. Their interactions with protein have implications in RNA disease biology. Thus, study of binding interfaces can provide a mechanistic insight of the molecular functioning and aberrations caused due to altered interactions. Moreover, high-throughput sequencing technologies have generated huge sequence data compared to available structural data of RNA-protein complexes. In such a scenario, efficient computational algorithms are required for identification of protein-binding interfaces of RNA in the absence of known structures. We have investigated several machine learning classifiers and various features derived from nucleotide sequences to identify protein-binding nucleotides in RNA. We achieve best performance with nucleotide-triplet and nucleotide-quartet feature-based random forest models. An overall accuracy of 84.8%, sensitivity of 83.2%, specificity of 86.1%, MCC of 0.70 and AUC of 0.93 is achieved. We have further implemented the developed models in a user-friendly webserver “Nucpred”, which is freely accessible at “http://www.csb.iitkgp.ac.in/applications/Nucpred/index”.