Solvaformer: Minimizing Geometric Redundancy for Scalable Solubility Prediction

Accurate prediction of small molecule solubility requires balancing physical fidelity with computational scalability. While geometric deep learning offers superior inductive biases, applying full SE(3)-equivariance to dynamic multi-component systems introduces geometric redundancy and high computational cost. We introduce Solvaformer, a graph transformer designed to achieve simplicity at scale by selectively grounding interactions in geometry. The architecture challenges the need for global equivariance: it applies strict SE(3)-equivariant attention only to rigid intramolecular structures, while modeling fluid intermolecular interactions via computationally efficient scalar attention. While Solvaformer demonstrates strong performance (approaching the DFT baseline), we also report a significant negative result: a simpler MPNN augmented with physics-informed partial charges (MLIPs) slightly outperforms the explicit Solvaformer architecture. This suggests that for scalar solubility prediction, high-quality electronic descriptors may render end-to-end equivariant architectures redundant. Our findings highlight that geometric redundancy can be minimized either architecturally (Solvaformer) or via decoupled feature generation (MPNN w/ MLIPs), offering two scalable paths for solution-phase modeling.