Topology-Aligned Architectures Boost Molecular Property Prediction Efficiency.

In quantum chemistry, predicting molecular properties often faces challenges due to limited data and computational resources, making parameter-efficient learning crucial. A new approach introduces a "topology-aligned inductive bias," where the model's architecture directly reflects the molecular bond graph. This means atoms correspond to computational units, and bonds dictate interactions between them using shared learnable parameters. This principle was applied to two distinct architectures: Iso-QGNN, a variational quantum circuit, and Iso-CGNN, a classical message-passing model with matched parameters. Benchmarked on tasks like HOMO-LUMO gap and dipole moment prediction using the QM9 dataset, both models demonstrated impressive parameter efficiency. They achieved strong test AUCs with only 64 trainable parameters, reaching 90% of their peak performance with relatively few training molecules, suggesting the topology-aligned bias is key to their efficiency.