Accurate prediction of HLA-peptide binding is vital for immunotherapy, vaccine design, and antigen discovery. We present a unified, pan-allele machine learning framework for classifying peptides as binders or non-binders to both HLA Class I and Class II molecules. Our method introduces an FFT-based cross spectral encoding that captures peptide-HLA interactions without requiring sequence alignment or structural modeling. Peptide sequences and structurally conserved regions of the HLA binding groove, including the two alpha helices and the groove base, are numerically embedded using the BLOSUM 100 matrix, then transformed into the frequency domain. The amplitudes and phases of the resulting cross-spectra represent alignment-free interaction features, capturing latent biochemical and spatial patterns. To ensure consistency across highly polymorphic alleles, we extract and standardize binding groove sequences using structural analysis and gene-group-specific alignments. This approach enables robust, scalable prediction across the full diversity of HLA Class I and II molecules, offering a novel and interpretable embedding strategy for immunoinformatics.
