MetaMBP: Few-Shot Multilabel Prediction of Bioactive Peptides Based on Deep Metric Meta-Learning

Bioactive peptides are highly specific and have low toxicity, making them a promising treatment option. There are many different types of bioactive peptides, while some types have limited samples (under 500). Methods that can handle limited types of bioactive peptides are needed to enhance the predictive ability of multilabel tasks with few sample categories. In this work, we proposed a novel multilabel model MetaMBP, based on deep metric meta-learning to predict the function of bioactive peptides. The model used the meta-knowledge obtained in the meta-learning stage to help improve the performance of limited sample categories in the fine-tuning stage. Our proposed model, MetaMBP, outperformed existing methods on benchmark data sets, particularly in predicting limited sample categories. Experiments in few-shot scenarios confirmed the adaptability of MetaMBP. Moreover, we analyzed the relationships between different categories by visualizing the features learned by MetaMBP and the attention scores in the attention module. All of these results have demonstrated that MetaMBP can offer an accurate, low-sample-adaptive approach for screening multilabel bioactive peptides.