EquiBoost: An Equivariant Boosting Approach to Molecular Conformation Generation

TL;DR

EquiBoost introduces a novel SE(3)-equivariant boosting framework that stacks multiple equivariant graph transformers to iteratively refine molecular conformations, offering a compelling alternative to diffusion-based generation. By sharing parameters across weak learners, injecting the step index as input, and using constrained random initialization, it achieves high accuracy with only five inference steps while maintaining diversity. The method combines permutation-invariant RMSD (piRMSD) with internal-coordinate losses to enforce both global and local geometric plausibility, and employs optimal conformation mapping to align training targets with reference ensembles. On GEOM-QM9 and GEOM-DRUGS, EquiBoost surpasses prior methods in AMR and precision, demonstrating notable efficiency gains and restoring boosting as a viable approach for fast, accurate 3D molecular conformation generation with practical impact in drug design and docking workflows.

Abstract

Molecular conformation generation plays key roles in computational drug design. Recently developed deep learning methods, particularly diffusion models have reached competitive performance over traditional cheminformatical approaches. However, these methods are often time-consuming or require extra support from traditional methods. We propose EquiBoost, a boosting model that stacks several equivariant graph transformers as weak learners, to iteratively refine 3D conformations of molecules. Without relying on diffusion techniques, EquiBoost balances accuracy and efficiency more effectively than diffusion-based methods. Notably, compared to the previous state-of-the-art diffusion method, EquiBoost improves generation quality and preserves diversity, achieving considerably better precision of Average Minimum RMSD (AMR) on the GEOM datasets. This work rejuvenates boosting and sheds light on its potential to be a robust alternative to diffusion models in certain scenarios.

Figures (6)

• Figure 1: EquiBoost framework (a) The molecular features include three components: the atom and bond types, graph topology, and random atomic coordinates. (b) The graph embedding is constructed from these three molecular features. (c) The equivariant graph transformer contains $L$ blocks. Each block contains equivariant graph attention, layer normalization, and feed-forward network. This transformer model serves as the weak learner. (d) In the boosting paradigm, the generated conformation from the current weak learner is passed to the next, and this process is repeated for $M$ iterations. (e) The final conformation is generated after $M$ boosting iterations. The orange arrow indicates the output of the weak learner, which represents the displacement of atomic coordinates.
• Figure 2: Comparison between diffusion models and EquiBoost (a) A diffusion model contains a forward and reverse process. The forward process, depicted from right to left, involves the gradual addition of noise to the true conformation $\mathcal{C}^0$, resulting in the chaotic conformation $\mathcal{C}^T$. The reverse process, shown from left to right, involves the iterative removal of noise to recover the true conformation $\mathcal{C}^0$. (b) EquiBoost aims to directly generate a precise conformation $\mathcal{C}^M$ from the chaotic conformation $\mathcal{C}^0$.
• Figure 3: An example of substructure symmetries in molecular conformation generation Atom types are distinguished by different colors, while bond types are depicted uniformly without differentiation. (a) represents the ground truth conformation, (b) is the initial noise, (c) is generated from (b) via our model, (d) and (e) show the results after alignment, with (e) highlighting alignment issues caused by substructure symmetries.
• Figure 4: Example of Higher-order neighbors.
• Figure 5: Example of Node Symmetry in Molecular Conformations (a) A self-symmetric ring structure. (b) A symmetric substructure connected to the same node.