Search

Given that wet experimental methods were no longer adequate for the rapid identification of bitter peptides, this paper presented Bitter-Fus, a novel predictive deep learning method incorporating traditional manual features and pre-trained deep features. Firstly, the method automatically extracted deep learning features from peptide sequences using a pre-trained protein sequence language model, then fed the deep learning features into a long short-term memory (LSTM) network for dimensionality reduction to retain the most relevant features. Finally, the reduced-dimensional deep features were fused with the manual features composed of traditional amino acids composition (AAC) method and passed into the feedforward neural network to construct a prediction model. The validation experimental results showed that the prediction method Bitter-Fus obtained an accuracy precision value of 0. 902 and a Mathews correlation coefficient value of 0. 805 in a 10-fold cross-validation, and an accuracy precision value of 0. 930 and a Mathews correlation coefficient value of 0. 862 in the independent dataset test, which significantly outperformed the current state-of-the-art bitter peptide prediction methods BERT4Bitter and iBitter-SCM. [ABSTRACT FROM AUTHOR]