AAV-CapNet: Multi-Omics Deep Learning with Fourier Enhancement and Cross-Attention for Predicting Infectivity of Known AAV Capsid Variants

This work builds AAV-CapNet, a deep learning model integrating Fourier frequency enhancement and cross-attention fusion, to predict infectivity of 20 AAV capsid variants across 52 cancer cell lines with a self-built multi-omics dataset of 392 validated measurements. Validated via stratified splitting and repeated five-fold cross-validation, it extracts dual-domain capsid features fused with host transcriptomics for joint classification and regression. Surpassing mainstream ML/DL baselines in key metrics, it performs far better at low FPR for high-throughput screening. Attention mapping and ablation tests validate core modules’ necessity. Despite poor generalization to unseen serotypes, this interpretable tool facilitates intraserotype capsid screening and rational engineering.