A(C)VPpred: Transfer Learning-Enhanced Prediction of Antiviral and Anticoronavirus Peptides from Sequence Data.

“A(C)VPpred: Transfer Learning-Enhanced Prediction of Antiviral and Anticoronavirus Peptides from Sequence Data. Wang Q(1), Chu Y(2), Mao X(1), Wei DQ(3)(4)(5). Author information: (1)School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai 200240, China. (2)Arc Institute, Palo Alto, California 94304, United States. (3)State Key Laboratory of Microbial Metabolism and School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai 200240, P.R. China. (4)Qihe Laboratory, Qishui Guang East, Qibin District, Hebi, Henan, 458030, China. (5)Zhongjing Research and Industrialization Institute of Chinese Medicine, Zhongguancun Scientific Park, Meixi, Nanyang, Henan, 473006, P.R. China. Viral infections pose an ongoing public health challenge and create a need for innovative therapies. Antiviral peptides (AVPs) offer a promising solution due to their high specificity, low toxicity, and reduced risk of inducing drug resistance. The COVID-19 pandemic has heightened the need for targeted anticoronavirus peptides (ACVPs). We introduce the term A(C)VPs (where "(C)" indicates the optional coronavirus specificity) to encompass both broad-spectrum AVPs and ACVPs. However, experimental screening of effective A(C)VPs remains slow and costly. To overcome these limitations, we propose A(C)VPpred, a computational pipeline integrating transfer learning via a fine-tuned BERT model, for the simultaneous prediction of AVPs and ACVPs. This framework first trains a general AVP predictor and then fine-tunes it for ACVP specificity, enabling the precise identification of peptides effective against diverse viruses, including emerging coronaviruses. Comparative analysis reveals that our A(C)VPpred model outperforms the state-of-the-art models on the test data set. Specifically, the AVP module attains peak performance at 98.6% accuracy and 97.2% Matthews correlation coefficient (MCC), whereas the ACVP module achieves 97% accuracy with 94.1% MCC. Additionally, analysis of BERT's attention mechanism revealed that A(C)VPpred captures structural features directly from sequence data. A(C)VPpred is expected to contribute to computer-aided screening and the design of AVP antiviral drugs. To further enhance accessibility for researchers, we have developed a user-friendly web interface, available at https://fond-defensive-day.anvil.app/.”