Investigating Impact of Quantum Ansatz Complexity in Quantum Hybrid Fusion Models to Estimate Protein-Ligand Binding Energy

Calculating binding energy for proteins is essential for drug design. It is computationally expensive and difficult, especially for large molecules, due to the number of complex interactions that must be considered between proteins and ligands. Current research explores quantum machine learning (QML) for augmenting classical models. This paper presents analysis of a fusion model that extracts features from a 3DCNN and an SGCNN, then feeds them into a simulated quantum layer to predict binding energy. This paper quantifies the impact of replacing the unitary circuit, analyzes variability across circuit choices, compares fixed and parameterizable circuits, and quantifies effects of gate composition. From the scope of this paper, there is little difference and variability between fixed and parameterizable circuits at this scale, and furthermore, no statistically significant change in predictability of protein-ligand binding affinity through swapping 1-5 layers with unitary circuit layers; it remains possible that they are learning to fit around the noise. As such, it remains essential to have robust evaluations for QNNs.