Solving reactions as movements of electrons — making the mechanism itself the object of learning
Machine learning for chemical reactions has either generated product molecules afresh or applied heuristic edits over molecular bonds. This work models a reaction as discrete flow matching over electron occupation vectors.
Paper overview (our summary)
- Field (arXiv category)cs.AI
- AuthorsNguyen Xuan-Vu, Octavian Susanu, Daniel Armstrong, et al. (4)
- Submitted2026-08-27
- arXiv ID2608.27429v1
Key points
- MAELLE models chemical reactions as discrete flow matching over electron occupation vectors.
- Existing methods either generated product molecules afresh or applied heuristic graph edits over molecular topology.
- The reactant-to-product mapping is posed as a continuous-time Markov chain over an integer-valued electron occupation space across bonding, non-bonding and hydrogen sites.
- Generalizing the mixture path with optimal transport yields interpretable edit sequences without elementary step annotations.
- It is reported competitive on USPTO-480K and to hold performance in out-of-distribution settings of structural complexity and reaction type.
1What to put in the middle
Reaction prediction is usually posed as: put reactants in, get products out. What this work changes is what sits between the two. A reaction is expressed as the movement of electrons, and that movement becomes what the model learns.
2Three ways of doing it
| The aspect | Generating the product afresh | Heuristic edits over bonds | This work, MAELLE |
|---|---|---|---|
| What is output | Product molecules | Edits to the topology | Changes in electron occupation |
| The space it works in | Molecules | Molecular topology | An integer-valued electron occupation space over bonding, non-bonding and hydrogen sites |
| Where mechanism sits | Not in the output | Approximated as a sequence of edits | Emerges as interpretable edit moves |
| Elementary step annotations | — | — | Not required |
The premise doing the work is that chemical reactions are fundamentally transformations in electron space. Molecules and bonds are the result of an arrangement of electrons, not its cause; so, the reasoning runs, solve on the electron side.
3Mechanism without annotation
- 1FormulationPose the reactant-to-product mapping as a continuous-time Markov chain over the electron occupation space
- 2Building trajectoriesGeneralize the discrete flow matching mixture path to discrete electron rearrangements using optimal transport
- 3What resultsA sequence of edit moves that can be read as a mechanism
- 4What is not neededElementary step annotations, the intermediate labels a person would otherwise supply
Teaching a model reaction mechanism ordinarily requires someone to annotate the intermediate steps. This method asks for none. Because a path is built across the continuous process of electron rearrangement, something mechanism-like arrives from the side of the procedure.
4What happens outside the distribution
Evaluation ran on the USPTO-480K benchmark, where performance is reported as competitive with leading models. The weight of the paper falls after that. Robustness was measured in two out-of-distribution settings, structural complexity and reaction type, where performance is said to hold in territory that degrades existing methods.
Alongside that, the model is reported to recover mechanistic trajectories consistent with known chemistry and to predict side products.
Why it matters
Where the object of prediction is placed determines what a model learns. Solve a reaction as product molecules and mechanism is absent from the output; solve it as edits to topology and mechanism is approximated. Solve it as electron rearrangement and something mechanism-like arrives from the procedure. Carrying an understanding of what the subject fundamentally is into the formulation is reported to tell in out-of-distribution robustness as well.
FAQ
Why work on electrons?
What does needing no elementary step annotations mean?
What are the out-of-distribution settings?
Sources (primary)
Source: arXiv (descriptive metadata is CC0 public domain). Summaries are our own; see arXiv for the original text and PDF.
- arXiv abstract page (original, official)
- PDF (arXiv)
- arXiv ID: 2608.27429