Multi-scale temporal convolutional networks and continual learning based in silico discovery of alternative antibiotics to combat multi-drug resistance. (1st April 2023)

• Record Type: Journal Article
• Authors: Singh, Vishakha; Shrivastava, Sameer; Singh, Sanjay Kumar; Kumar, Abhinav; Saxena, Sonal
• Is Part Of: Expert systems with applications. Volume 215(2023)
• Publication Date: 2023-04-01
• Subjects: Temporal convolutional networks -- Deep learning -- Artificial intelligence -- Continual learning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.119295
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: The high incidence of diseases caused by multi-drug resistant (MDR) pathogens combined with the shortage of effective antibiotics has necessitated the development of in-silico machine and deep learning tools to facilitate rapid drug discovery. The construction of computational models to discover antibacterial peptides (ABPs) in proteins of various organisms to develop a new line of antibiotics has emerged as a possible recourse. To this end, we used multi-scale temporal convolutional networks (MSTCN) to develop a robust deep learning-based model called MSTCN-ABPpred (BL) that classifies ABPs with an accuracy of 98% (which is better than various state-of-the-art models). The main contribution of this proposed work is that we have incorporated a continual learning module in this model so that it keeps adapting itself dynamically by re-training on new data points. This re-trainable version of the baseline model (MSTCN-ABPpred (BL)) was termed as MSTCN-ABPpred (CL). We re-trained this model on the ABPs and non-ABPs predicted by it in some antibacterial proteins. It has been demonstrated that the proposed model does not exhibit any statistically significant deterioration in performance after extensive re-training, and it gains additional skills compared to the MSTCN-ABPpred (BL). We have also deployed a freely accessible web application based on our final model, available at https://mstcn-abppred.anvil.app/, which can identify and discover ABPs in a protein using which the model gets re-trained on its own. Highlights: Proposed an MSTCN-based model for classifying antibacterial peptides. Showed the model's better performance than other state-of-the-art models. Proposed and compared four continual learning methods for the model's retraining. Proved that the retraining process helps the model perform better than before. Performed ANOVA test to prove good learning stability of the proposed model.