MCMFPP: A Multifunctional Peptides Prediction Method Based on Class Feature Enhancement and Classifier Fusion

With the increasing discovery of peptide sequences and the growing demand for peptide-targeted drugs, traditional wet-lab experiment methods have become inadequate for peptide function prediction due to their high cost and that they are time consuming. This has made the development of computational tools for the accurate identification of peptide functions particularly urgent. However, existing computational methods are limited by challenges such as data sparsity, long-tailed distribution imbalance, label dependency modeling and inadequate class feature representation in multifunctional therapeutic peptides (MFTP) prediction tasks, resulting in suboptimal performance. To address these limitations, we first introduce two subclassifiers: SLFE and CFEC. SLFE leverages the large language model ESMC to complement sequence representation and alleviate feature insufficiency in tail-class data, while CFEC improves class feature representation by enhancing the learning on single-function peptide samples combined with contrastive learning. Based on these subclassifiers, we propose MCMFPP, a deep learning method that integrates the predictions of SLFE and CFEC through weighted fusion. This method overcomes the constraints of single-classifier approaches, enabling more accurate prediction of challenging samples. MCMFPP outperforms state-of-the-art methods in multifunctional peptide prediction, achieving improvements of 3.1% in precision, 2.7% in coverage, 2.8% in accuracy, and 2.7% in absolute true while reducing the absolute false rate by 0.3%. We anticipate that MCMFPP will serve as a valuable tool for multifunctional peptide prediction, enabling more efficient and accurate identification of candidate peptides.