Chen Lihua, Hu Zhenkang, Rong Yuzhi, Lou Bao
School of Perfume and Aroma Technology, Shanghai Institute of Technology, Shanghai 201418, China.
School of Perfume and Aroma Technology, Shanghai Institute of Technology, Shanghai 201418, China.
Comput Biol Chem. 2024 Apr;109:108021. doi: 10.1016/j.compbiolchem.2024.108021. Epub 2024 Jan 22.
Functional peptides are easy to absorb and have low side effects, which has attracted increasing interest from pharmaceutical scientists. However, due to the limitations in the laboratory funding and human resources, it is difficult to screen the functional peptides from a large number of peptides with unknown functions. With the development of machine learning and Deep learning, the combination of computational methods and biological information provides an effective method for identifying peptide functions. To explore the value of multi-functional active peptides, a new deep learning method named Deep2Pep (Deep learning to Peptides) was constructed, which was based on sequence encoding, embedding, and language tokenizer. It can achieve predictions of peptides on antimicrobial, antihypertensive, antioxidant and antihyperglycemic by converting sequence information into digital vectors, combined BiLSTM, attention-residual algorithm, and BERT Encoder. The results showed that Deep2Pep had a Hamming Loss of 0.095, subset Accuracy of 0.737, and Macro F1-Score of 0.734. which outperformed other models. BiLSTM played a primary role in Deep2Pep, which BERT encoder was in an auxiliary position. Deep learning algorithms was used in this study to accurately predict the four active functions of peptides, and it was expected to provide effective references for predicting multi-functional peptides.
功能性肽易于吸收且副作用小，这引起了制药科学家越来越浓厚的兴趣。然而，由于实验室资金和人力资源的限制，很难从大量功能未知的肽中筛选出功能性肽。随着机器学习和深度学习的发展，计算方法与生物信息的结合为识别肽的功能提供了一种有效方法。为了探索多功能活性肽的价值，构建了一种名为Deep2Pep（深度学习预测肽）的新型深度学习方法，该方法基于序列编码、嵌入和语言分词器。通过将序列信息转换为数字向量，并结合双向长短期记忆网络（BiLSTM）、注意力残差算法和BERT编码器，它可以实现对肽的抗菌、降压、抗氧化和降血糖功能的预测。结果表明，Deep2Pep的汉明损失为0.095，子集准确率为0.737，宏F1分数为0.734，优于其他模型。在Deep2Pep中，BiLSTM起主要作用，而BERT编码器处于辅助地位。本研究使用深度学习算法准确预测了肽的四种活性功能，有望为预测多功能肽提供有效参考。
