Q-BAFNet: A Hybrid Quantum Classical Approach for Drug-Target Binding Affinity Prediction.

To accelerate drug discovery, especially during high-throughput screening, accurate estimation of drug-target binding affinity (DTA) is essential. Existing deep learning models often fail to capture the complex, context-dependent relationships between ligands and proteins. To address this, we present Q-BAFNet, a hybrid quantum-classical deep learning architecture that integrates semantic, structural, and sequential molecular representations. Q-BAFNet leverages ChemBERTa for SMILES-based ligand embeddings, ProtT5 for protein sequence modeling, and graph convolutional networks (GCNs) for topological molecular features. A key innovation is the cross-modal attention fusion mechanism, which dynamically aligns ligand and protein substructures, allowing for more efficient capture of biologically meaningful interactions than through traditional averaging or convolution. To further improve expressiveness, Q-BAFNet includes a variable quantum circuit (VQC), which projects fused embeddings into quantum Hilbert space. This quantum layer captures nonlinear and entangled dependencies beyond the capabilities of classical models. We evaluate Q-BAFNet on three benchmark datasets, Davis, KIBA, and Metz, under four evaluation protocols: random pairing, drug cold-start, target cold-start, and drug-target cold-start. This model outperforms existing methods in terms of mean squared error (MSE), Pearson correlation coefficient (PCC), concordance index (CI), and ${\mathrm{R}}^{2}$, especially in zero-shot scenarios. Q-BAFNet demonstrates the promise of quantum-enhanced deep learning for robust and generalizable DTA prediction in data-poor and biologically diverse settings.