ArcMol Enables Task-Adaptive Spherical Representation Learning for Molecular Property Prediction

Molecular properties are measured under different experimental and representational settings, each capturing a distinct aspect of molecular behavior. Integrating such task-dependent molecular information remains challenging, especially when labeled data are scarce and structurally similar molecules exhibit markedly different properties. Here we introduce ArcMol, a task-adaptive framework for molecular property prediction that combines explicit physicochemical descriptors with learned molecular representations. ArcMol organizes these features in a hyperspherical latent space, where molecular similarity is defined by angular relationships. Across more than 150 molecular property prediction benchmarks, ArcMol achieves performance comparable to state-of-the-art methods, while showing improved robustness in low-data settings and for structurally similar molecules. Analyses of the learned representations reveal task-relevant separation, ordered property gradients, and coherent local neighborhoods. We demonstrate ArcMol in a real-world application targeting the kinase IRAK4, where it prioritizes candidate molecules for drug repurposing and leads to the identification of nanomolar type II inhibitors. These results suggest ArcMol as a general framework for molecular property prediction in discovery settings.