image: Figure 1. AtomFlow generates reactant structures from product structures in three-dimensional space. The flow model preserves atom types and counts throughout generation. The trajectory illustrates the model’s generation process, not a measured reaction pathway. Credit: AtomFlow research team.
Credit: Enji Li, Siyu Hu, Xiao Tian, Wentao Zhang, Tao Liu, Linwang Wang and Weile Jia
Learning to generate reactant structures
AtomFlow represents generation as atomic transport through a learned velocity field, starting from a product structure and ending at a predicted reactant structure. Atom types and counts remain unchanged throughout this process. The learned path is a generative trajectory between structures, rather than a demonstrated physical reaction mechanism. Figure 1 summarizes the model and its architecture.
Accuracy and structure quality
On Transition1x, the 20-million-parameter MatRIS version, AtomFlow_P2R_MR, reached 79.30% Top-1 and 96.20% Top-10 accuracy. Transition1x contains roughly one hundredth as many reactions as commonly used USPTO benchmarks. However, differences in molecular size, representation and evaluation mean that cross-dataset results are not a controlled comparison or proof that the same task requires only 1% of the training data.
The study also assessed the geometry and energy of generated structures. The reported median root-mean-square deviation (RMSD) was 0.0147 Å, while the mean was 0.2007 Å, indicating that some predictions had appreciably larger errors. Figure 2 shows how these errors change with the number of function evaluations, or nfe. The authors selected nfe = 100 as a balance between generation quality and runtime.
Generating and ranking multiple candidates
Repeated generation starts from product coordinates perturbed with controlled noise. The resulting candidate structures are deduplicated and ranked by their frequency of occurrence. Figure 3 includes cases in which the matching reactant appears at different ranks, as well as failures. Sampling frequency provides a ranking criterion, not a calibrated probability that a reaction is correct.
Batch inference and next steps
In a separate efficiency test, the smaller GVP version, with 6 million parameters, completed 1,000 predictions in about 20 seconds and 10,000 in about 50 seconds on one H100-80G GPU. These timings describe the reported small-molecule batch setting and should not be attributed to the MatRIS model that achieved the highest Top-10 accuracy.
The current results are limited to small molecules in Transition1x. Atom conservation alone does not establish chemical feasibility or accurate energetics. Future work could test larger systems, metals and solvent effects, improve energy accuracy, and explore transition-state conditioning for forward prediction.
By generating explicit 3D structures while enforcing atom conservation, AtomFlow provides candidates that can undergo geometric and energetic checks. Its batch throughput suggests potential for virtual screening and efficient retrosynthesis prediction, subject to further validation for practical chemical systems. The source code, trained weights and reproduction scripts are publicly available.
The research has been recently published in the online edition of AI for Science, an international journal published by IOP Publishing.
Citation: Enji Li, Siyu Hu, Xiao Tian, Wentao Zhang, Tao Liu, Linwang Wang, Weile Jia. AtomFlow: Generative Prediction of Chemical Reactions in 3D Space[J]. AI for Science. DOI: 10.1088/3050-287X/ae9e53
Journal
AI for Science