Prediction Models for Identifying Ion Channel-Modulating Peptides via Knowledge Transfer Approaches

• Published in: IEEE journal of biomedical and health informatics 2022-12, Vol.26 (12), p.6150-6160
• Main Authors: Lee, Byungjo, Shin, Min Kyoung, Kim, Taegun, Shim, Yu Jeong, Joo, Jong Wha J., Sung, Jung-Suk, Jang, Wonhee
• Format: Article
• Language: English
• Subjects: Acetylcholine receptors; Acetylcholine receptors (nicotinic); Antiinfectives and antibacterials; Antimicrobial peptides; Calcium; Calcium channels; Calcium ions; Convolutional neural networks; Datasets; Deep learning; Humans; Ion channel-modulating peptides; Ion Channels; Knowledge management; Knowledge transfer; Machine Learning; multi-task learning; Neurological diseases; Peptides; Potassium; Potassium channels; Prediction models; Predictive models; Sodium; Sodium channels; Therapeutic targets; Transfer learning
• ISSN: 2168-2194; 2168-2208
• DOI: 10.1109/JBHI.2022.3204776

Ion channels, which can be modulated by peptides, are promising drug targets for neurological, metabolic, and cardiovascular disorders. Because it is expensive and labor-intensive to experimentally screen ion channel-modulating peptides (IMPs), in-silico approaches can serve as excellent alternatives. In this study, we present PrIMP, prediction models for screening IMPs that can target sodium, potassium, and calcium ion channels, as well as nicotine acetylcholine receptors (nAChRs). To overcome the data insufficiency of the IMPs, we utilized two types of knowledge transfer approaches: multi-task learning (MTL) and transfer learning (TL). MTL enabled model training for four target tasks simultaneously with hard parameter sharing, thereby increasing model generalization. TL transferred knowledge of pre-trained model weights from antimicrobial peptide data, which was a much larger, naturally-occurring functional peptide dataset that could potentially improve the model performance. MTL and TL successfully improved the prediction performance of prediction models. In addition, a hybrid approach by implementing deep learning along with traditional machine learning was utilized, with additional performance improvements. PrIMP achieved F1 scores of 0.924 (sodium ion channel), 0.937 (potassium ion channel), 0.898 (calcium ion channel), and 0.931 (nAChRs). The pre-processed dataset and proposed model are available at https://github.com/bzlee-bio/PrIMP .