A Biological Property-Aware Quantum-Classical Fusion Model for Predicting AAV Capsid Production Fitness

Accurate prediction of capsid production fitness remains a major challenge in the rational design of adeno-associated virus (AAV) variants, owing to the highly nonlinear and heterogeneous relationship between amino acid sequence patterns and production outcomes. To address this problem, we propose a biological property-aware quantum-classical adaptive fusion model for predicting AAV capsid production fitness directly from amino acid sequences. The model represents sequences using biologically meaningful physicochemical properties and introduces a dual-signal adaptive controller that generates two task-dependent latent signals for regulating quantum encoding and entanglement operations. Based on these learned signals, the model dynamically adjusts quantum feature mapping and quantum-classical feature fusion, enabling sample-dependent regulation of quantum processing and heterogeneous feature integration. Experimental results over five independent runs show that the proposed model achieves a Pearson correlation coefficient of $0.919\pm 0.003$ and an MAE of $0.38\pm 0.02$, outperforming the traditional machine learning and general neural network baselines while using only 0.14M trainable parameters. Compared with the strongest task-relevant baseline, it achieves comparable Pearson correlation performance and a statistically supported reduction in MAE, demonstrating a favorable predictive-performance-parameter-efficiency trade-off. The results suggest that biologically informed quantum-classical hybrid learning is a promising direction for predictive modeling in AAV capsid engineering.