A DEEP LEARNING FRAMEWORK FOR ACCURATE IDENTIFICATION OF ANTIFUNGAL PEPTIDES USING HYBRID FEATURE EMBEDDINGS

Abstract

Antifungal peptides (AFPs) represent a vital class of therapeutic agents capable of combating resistant fungal infections through multi-mechanistic action. However, their experimental identification remains costly and time-consuming. We propose a novel deep learning framework that integrates both handcrafted and automatically learned features to accurately identify AFPs. Specifically, our model extracts and fuses sequence descriptors such as amino acid composition (AAC), dipeptide composition (DPC), and pseudo amino acid composition (PseAAC) with deep contextual Embeddings from a pertained ProtT5 transformer, evolutionary profiles from BLOSUM62, and physicochemical descriptors via composite property profiles (CPP). A co-attention-based fusion mechanism is employed to enhance the representation power by dynamically weighing inter-feature relationships. The fused features are evaluated using classifiers including Extra Trees (ET), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN).

The model was trained and validated on benchmark datasets (Antifp_Main and Deep-AFPpred) using five-fold cross-validation and independent test evaluation. The proposed hybrid model, particularly the MLP classifier, outperformed all baselines, achieving a top accuracy of 96.2%, AUC of 0.99, and MCC of 0.93. This demonstrates the utility of hybrid multi-view learning and co-attention fusion in improving AFP identification. The framework is generalizable, scalable, and can be readily adopted in computational antifungal drug discovery pipelines.

Index Terms— AAC (Amino Acid Composition), AFP (Antifungal Peptides), BLOSUM62 (Blocks Substitution Matrix), CNN (Convolutional Neural Network), CPP (Composite Physicochemical Properties), DPC (Dipeptide Composition), ET (Extra Trees Classifier), MLP (Multilayer Perceptron), ProtT5 (Protein Transformer Embedding), PseAAC (Pseudo Amino Acid Composition).