SPiCE: Symmetry-Preserving Conformer Ensemble Networks for Molecular Representation Learning

SPiCE learns molecular properties from conformer ensembles while preserving joint equivariance to geometric transformations of individual conformers and permutations of the ensemble. Key components:

• Geometric Mixture-of-Experts (GMoE): Separate per-atom routers for scalar and vector features using Gumbel softmax
• Hierarchical Ensemble Encoding: Cross-attention between topology (GIN) and conformer representations
• Upcycling: Progressive expert specialization from a shared initialization

SPiCE GitHub

Publications

Zhu, Y.; Shi, Y.; Chen, Y.; Sun, F. Sun, Y.; Wang, W. Symmetry-Preserving Conformer Ensemble Networks for Molecular Representation Learning NeurIPS 2025, 39.