REM3DI: Learning smooth, chiral 3D molecular representations from equivariant atomistic foundation models

Accurate molecular descriptors are central to property prediction and similarity screening in computational drug discovery, but prevailing approaches rely on two-dimensional graphs that underuse 3D geometry. We introduce REM3DI, a framework that builds 3D molecular representations by leveraging latent atomic features from equivariant machine-learned interatomic potentials (MLIPs). The local atom centred MLIP features are pooled with permutation-invariant attention aggregation to obtain molecule-level embeddings. To add chiral sensitivity, REM3DI constructs pseudo-scalar channels, which change sign under mirror reflection. Given the scarcity of experimental training data, we pretrain the model via self-supervised denoising. Across small-molecule benchmarks and dedicated chirality tests, REM3DI achieves competitive performance with strong 2D baselines and consistently distinguishes enantiomers on tasks where stereochemistry matters. Importantly, REM3DI sidesteps limitations of 2D cheminformatics descriptors which encode the presence of molecular substructures. REM3DI provides a unified route from physics-based atomic embeddings to versatile, chirality-aware molecular representations for property prediction and virtual screening.