Prediction of Antifungal Peptides by Deep Learning with Character Embedding

Abstract

Conventional computerbased methods for identifying antifungal peptides (AFPs) are primarily based on large amounts of task-specific knowledge in feature extraction. This paper introduces a method, called AFPDeep, which only uses peptide sequences as input, without calculating any features by human intervention for the prediction of AFPs. First, an embedding layer is used to automatically code the sequence and to learn a dense vector representation for each kind of amino acid appearing in the training dataset. Second, a convolutional neural network (CNN) layer followed by a long short term memory (LSTM) layer is used to capture the local clues that are good indicators of AFPs, and to learn the long-term dependence and contextual information of the input data in a flexible way. Lastly, all the above layers are intertwined into hybrid neural network architecture for antifungal peptide prediction. Upon comparison with other methods on the same datasets, the AFPDeep yielded competitive results, achieving an AUC of on the Antifp_Main dataset, on the Antifp_DS1 dataset, and on the Antifp_DS2 dataset. These results suggest that the hybrid architecture of CNN-LSTM combined with character embedding is an effective model for improving the prediction of AFPs.