Dynamic Fourier wavelet positional encoding for anticancer peptide prediction via transfer learning

Computational identification of anti-tumor peptides (ACPs) is challenged by limited training data and the multi-scale nature of biological sequences. Standard transformer positional encodings fail to capture the periodicities and local motifs that govern peptide function. We propose dynamic Fourier wavelet positional encoding (DFWPE), which replaces fixed sinusoidal encodings with parallel adaptive Fourier and multi-scale wavelet branches, fused by a context-aware gating mechanism. A dual-stage transfer learning strategy pre-trains on 5,785 curated peptide sequences before fine-tuning on ACP740. DFWPE achieves 85.4% accuracy on ACP740 and 83.8% on the independent ACP240 test set, outperforming recent transformer-based predictors on overall metrics. Ablation studies confirm the necessity of both encoding branches and transfer learning. Interpretability analyses reveal length-dependent gating behavior consistent with biological structure, demonstrating that spectral-spatial decomposition provides robust and interpretable peptide representation for drug discovery.