Accurate transition state generation with an object-aware equivariant elementary reaction diffusion model

Introduction

• Quantum-chemistry reaction-network construction requires many calculations, and most DFT energy evaluations are spent locating accurate transition-state structures from reactant and product information.Transition-state structures are needed to estimate reaction rates and identify dominant pathways.Source: Introduction · p. 1
• Earlier machine-learning approaches lacked full physical symmetry handling, required reconstruction and optimization of final 3D structures, and fell short of the stated barrier-height accuracy target.The cited target is 3 kcal/mol, corresponding to a one-order-of-magnitude change in reaction rate at 300 °C.Source: Introduction · p. 2
• OA-ReactDiff generates transition-state structures in seconds and reports an average RMSD of 0.18 Å, reduced to 0.13 Å with a confidence-ranking recommender.The model uses reactant and product geometries and avoids atom-order mapping and fragment alignment.Source: Introduction · p. 2
• OA-ReactDiff models reactant, transition state, and product as multiple objects, using object-level equivariance and inpainting to condition generation on known reactant and product structures.Figure 1 contrasts single-molecule diffusion with the multi-molecule reaction model and its conditional sampling scheme.Source: Introduction · p. 3, Figure 1 · p. 3
• For the most challenging systems for the model, OA-ReactDiff achieves a barrier-height MAE of 2.6 kcal/mol using only 14% of DFT-based optimizations, approaching the 3 kcal/mol accuracy required to explore reaction networks with unknown mechanisms.Source: Introduction · p. 3

Results

• Object-aware equivariance models reactions as multiple objects, preserving required object- and fragment-level symmetries that vanilla SE(3) GNNs do not capture.The model uses equivariant updates within objects and scalar message passing between them; it avoids atom-order matching and fragment alignment and directly outputs Cartesian coordinates.Source: Results · pp. 4–6
• On three unseen reaction examples, 128 generated candidates per reaction included multiple distinct TS structures, some corresponding to different elementary reactions than the intended one.The confidence model ranks candidates and favors low-RMSD structures, but is not guaranteed to select the lowest-RMSD candidate.Source: Results · pp. 8–9
• 0.058 Å is the median RMSD for recommended TS structures, versus 0.076 Å for randomly selected samples on 1,073 unseen reactions.The recommended structures take an average of 6 seconds to generate with batching on a V100 GPU, compared with 12 hours for DFT-based climbing-image NEB.Source: Results · p. 9
• 4.4 kcal/mol mean and 1.6 kcal/mol median absolute energy differences were achieved by OA-ReactDiff + recommender, with 71% of barrier errors below 3 kcal/mol.The reported MAE would improve only marginally to 4.0 kcal/mol if the lowest-RMSD sample were selected instead.Source: Results · p. 11
• OA-ReactDiff + recommender systematically outperforms prior approaches in both mean and median RMSD and barrier-height errors, though Table 1 comparisons use different data partitions and may be affected by their random seeds.The PSI-based model also uses a predecessor dataset, and some methods have inconclusive predictions from uncertainty filtering or convergence issues.Source: Results · p. 12, Table 1 · p. 13
• With confidence threshold 0.5, DFT-based NEB is used for 14% of reactions while 86% are accepted directly, yielding an overall accuracy of 2.6 kcal/mol.The authors report a monotonic relationship between confidence threshold and MAE, enabling a trade-off between accuracy and DFT calculations.Source: Results · p. 12, Figure 5 · p. 13

• OA-ReactDiff + recommender achieves 0.129 Å RMSD and 4.4 kcal/mol MAE against climbing-image NEB TS; NEB on the 14% most uncertain reactions reduces MAE to 2.6 kcal/mol.Source: Discussion · pp. 13–14
• The 3N-atom representation makes fully connected scalar message passing a single-GPU bottleneck above 100 atoms; stochastic sampling also leaves sample quality uncertain and can accumulate runtime.Source: Discussion · p. 14
• OA-ReactDiff can also model single-ended or zero-ended scenarios, and barrier prediction could use a swapped model output layer.Source: Discussion · p. 14
• The 17-times-larger RGD1 dataset may benefit OA-ReactDiff because diffusion-model quality depends strongly on training-data size, potentially aiding reaction-network construction with fewer DFT calculations.Source: Discussion · p. 14
• The object-aware SE(3) model may apply to other chemical problems involving multiple 3D objects whose interactions do not depend on their Euclidean locations.Source: Discussion · p. 15

Diffusion models

• The diffusion process adds noise toward a Gaussian prior, then trains a denoiser to predict the added noise and reverse the process.For conditional generation, inpainting includes the known input in the joint distribution during training and combines it with the generated region during inference.Source: Methods · pp. 15–16

Equivariant diffusion models

• SE(3)-equivariant diffusion requires an invariant prior and equivariant transition kernel; the model breaks reflection symmetry to represent molecular chirality.Translation symmetry is handled by limiting the distribution to the linear subspace where the center of mass is at the origin.Source: Methods · pp. 15–16

LEFTNet and geometric processing

• LEFTNet uses local frames to scalarize geometric quantities, while tensorization converts scalars back into geometric tensors.The methods also describe MPNNs as message, update, and readout operations, with equivariant messages incorporating scalarized coordinates.Source: Methods · pp. 16–18

Object-aware SE(3) implementation

• Scalar-only message passing connects objects after separate SE(3)-equivariant updates, incorporating cross-object interactions without introducing their relative positions.The object-aware block adapts a vanilla SE(3) update layer with scalarization, concatenation, and scalar message passing.Source: Methods · p. 18

Training data and model

• Transition1x contains 10,073 reactions; 9,000 were used for training and validation, and 1,073 were held out for testing.OA-ReactDiff used LEFTNet and was trained for 2,000 epochs with 5,000 diffusion steps.Source: Methods · p. 19

Confidence model

• The confidence model is trained as a binary classifier on generated reactions, labeling a TS as good when its RMSD from the true TS is below 0.2 Å.Its predicted probability serves as the confidence score for estimated TS quality.Source: Methods · p. 19

• The title fragment describes an object-aware equivariant elementary-reaction diffusion model.Source: Supplementary Information for… · p. 25
• The title identifies the model's application as elementary-reaction diffusion.Source: Supplementary Information for… · p. 25
• The listed authors are Chenru Duan, Yuanqi Du, Haojun Jia, and Heather J. Kulik.Source: Supplementary Information for… · p. 25

• OA-ReactDiff means an object-aware SE(3) GNN for generating 3D molecule sets in elementary reactions under diffusion.Source: Abbreviation · p. 25
• RMSD means root mean square deviation, SE(3) means the Special Euclidean group in 3D space, and TS means transition state.Source: Abbreviation · p. 25
• MAE means mean absolute error.Source: Abbreviation · p. 25

• The reaction representation includes reactant fragments, a transition state, and product fragments, with symmetry under atom reorderings, fragment reorderings, and each fragment’s rotation and translation.Source: S1 Physical symmetries and constraints… · p. 25
• Table S1 compares RMSD performance across models to assess object-wise symmetry in vanilla SE(3) LEFTNet and vanilla SE(3) in EGNN.Source: Table S1 · p. 25
• 0.03 Å RMSD separates the generated TS from the DFT-optimized true TS in an example with randomized atom mapping and fragment alignment.Source: Figure S1 · p. 26
• Figure S2 shows reactions generated from a C, H, N, and O composition absent from the Transition1x training dataset.Source: Figure S2 · p. 26
• Supplementary evaluations cover runtime and GPU memory by batch size, performance by training-data amount and constraints, and barrier-height error across methods.Source: Table S2 · p. 26, Table S3 · p. 27, Table S4 · p. 27
• Figure S4 presents RMSD distributions grouped by system size for the 1,073 test reactions.Source: Figure S4 · p. 28