Identification of defensins using transformer-derived protein embeddings and discrete cosine transformation-enhanced...

Defensins, cationic antimicrobial peptides (AMPs) with broad-spectrum activity and a lower propensity for resistance development, represent promising candidates for combating multidrug-resistant bacteria and contributing to host-directed therapies. However, current computational models for defensin prediction often suffer from limited accuracy due to inadequate feature representation and model design. To address these limitations, we introduce GAC-BTCNN-Pred, a novel predictor for defensin identification. Our approach integrates advanced feature extraction with a hybrid model architecture. Initially, we derive evolutionary information using a segmented Position-Specific Scoring Matrix (seg-PSSM), which is subsequently enhanced through Discrete Cosine Transformation (DCT) to reduce noise and emphasize local sequence motifs, creating the Seg-PSSM-DCT descriptor. To capture complementary information, we incorporate ProtGen-LLM, a transformer-based protein language model, which effectively encodes contextual, physicochemical, and long-range dependencies within protein sequences. The fusion of Seg-PSSM-DCT and ProtGen-LLM yields a comprehensive, multi-faceted feature set termed PGL-PD. This rich feature representation is then fed into our proposed Generative Adversarial Capsule Bidirectional Temporal Convolutional Neural Network (GAC-BTCNN). Extensive comparative evaluations against traditional machine learning and existing deep learning methods demonstrate the superior performance of GAC-BTCNN-Pred across various evaluation metrics, establishing it as a scalable tool for accurate defensin identification within the context of novel antibacterial agent discovery.