Generative discovery of chemical reaction networks

arXiv:2610.38236 · q-bio.MN, physics.chem-ph · Submitted 2026-10-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Generative discovery of chemical reaction networks".

Paper summary: Ines: Thinking about the title, Generative discovery of chemical reaction networks, it really captures what they did: generating these networks from scratch using kinetic guidance rather than starting with a pre-existing set of rules. Marcus And I think the authors are pointing towards a future where we can explore chemical space much more broadly without needing exhaustive prior knowledge of what reactions exist. Ines They show that you can use generative models to propose transition states and predict full reaction trajectories, then verify those ideas with machine learned interatomic potentials for connectivity and electronic barriers.

Yuki: It seems like this approach could be valuable in areas where the chemical landscape is incredibly complex, like exploring new material synthesis or understanding biological pathways where the reaction mechanisms are not well documented. Marcus The paper shows that you can find lower barrier cooperative pathways and competition between different reaction branches, which gives us concrete mechanistic insights into how things actually happen.

Ines: So the implication here is that we gain a way to uncover those hidden connections in chemical systems where traditional methods would struggle because they rely too heavily on known reaction templates. Yuki It really shifts the focus from just finding known reactions to actively discovering new, unseen ways molecules can transform into each other.

Conclusion: Ines: So, to wrap up this discussion on "Generative discovery of chemical reaction networks," we're looking at how they built these entire sets of reactions just from graphs that show you the starting molecules.

Marcus: Yeah, it’s about using generative models to propose transition states and then simulating the whole journey from reactants all the way to products.

Ines: And they checked those proposed paths using machine learned interatomic potentials to make sure the connections actually hold up and the energy barriers look right.

Yuki: This means you don't need a list of known reactions or templates to start exploring chemical space; you just need the ingredients.

Marcus: The paper shows they found these networks uncover cooperative pathways and competition between different reaction branches, which gives us real mechanistic insight into how chemistry actually works.

Ines: It suggests we can use this generative approach to discover new reaction mechanisms without needing to predefine every possible product.

Yuki: That opens up the door for figuring out complex biological processes or new material synthesis where we don't have a map yet.

Marcus: The implication is that mass spectral data, which is often just a fingerprint, can be used to help interpret these complex reaction networks by suggesting candidate structures.

Ines: So it’s moving us toward an approach where we generate hypotheses about chemistry and then test them rigorously with computation before we even touch a lab.

Yuki: That’s the core shift: moving from pattern matching known reactions to actively generating new ones based on energetic guidance.

Deep Principle Inc. · Peking University Shenzhen Graduate School, Shenzhen, Guangdong, China. · Key Laboratory of Theoretical and Computational Photochemistry of Ministry of Education, College of Chemistry, Beijing Normal University

q-bio.MN, physics.chem-ph

Submitted: 2026-10-28

Updated: 2026-11-18

Code: https://github.com/ZimmermanGroup/pyGSM

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 91/100

The gist: ReactNexus is a generative framework designed for kinetic-guided network exploration, capable of constructing unknown reaction networks directly from input reactant graphs without requiring…

Key concepts

ReactNexus
A generative framework that explores chemical reaction networks starting only from reactant graphs. It uses generative models to propose transition states and predict full trajectories connecting reactants to products, guided by energy calculations.
Machine-Learned Interatomic Potentials (MLIPs)
Calculations using MLIPs, like Carina, are used to verify the connectivity and electronic barriers of proposed reaction pathways. These potentials are trained on data to accurately estimate the energy landscape of molecular interactions.
Navigation Heuristic
A mathematical rule used by ReactNexus to decide which new chemical step or reactant set to explore next. It prioritizes routes with lower cumulative energy barriers, ensuring the search focuses on the most energetically favorable reaction pathways.
Intrinsic Reaction Coordinate (IRC) Verification
A calculation method used to verify that a proposed transition state is indeed a true reaction pathway connecting reactants and products. This ensures that the generated network steps are chemically valid and physically meaningful.

Summary

ReactNexus is a generative framework designed for kinetic-guided network exploration, capable of constructing unknown reaction networks directly from input reactant graphs without requiring predefined products or reaction templates. p. 1

Where the paper says it

44,45 Here we introduce ReactNexus, a framework for generating chemical reaction networks directly from reactants, without predefined products, reaction templates or prior enumeration of candidate reactions. Single-step benchmarks establish the basis for iterative, kinetic-guided exploration (Fig.

The core methodology involves coupling pathway proposals to endpoint identification and energetic evaluation. p. 1
Where the paper says it

1 Generative network construction and single-step validation Method overview. ReactNexus grows a network by coupling pathway proposals to endpoint identification and energetic evaluation (Fig.

Generative models propose transition states (TS) geometries from an input set of reactants, and a simulator predicts full reaction trajectories connecting these TS geometries to identify distinct products. p. 0
Where the paper says it

Here we introduce ReactNexus, a generative framework for kinetic-guided network exploration starting solely from reactant graphs. Generative models propose transition states and predict the full reaction trajectories connecting reactants and products, while calculations using a machine-learned interatomic potential verify connectivity and evaluate electronic barriers.

Connectivity and electronic barriers are then verified using calculations employing machine-learned interatomic potentials (MLIPs), such as Carina, which is based on the Equiformer-V3 architecture. p. 0
Where the paper says it

Here we introduce ReactNexus, a generative framework for kinetic-guided network exploration starting solely from reactant graphs. Generative models propose transition states and predict the full reaction trajectories connecting reactants and products, while calculations using a machine-learned interatomic potential verify connectivity and evaluate electronic barriers.

The framework operates through several sequential stages: candidate screening, stationary-point and endpoint verification, energy referencing, mapped identity assessment, and global candidate pool expansion. p. 2

Where the paper says it

The remaining candidates are grouped by endpoint connectivity and mapped bond changes. Within each group, Carina single-point energy (SPE)-based screening and farthest-point sampling select low-energy, geometrically diverse representatives, allowing multiple TS arrangements and pathways connecting the same reactants and products to be retained.

Candidate screening involves generating a batch of TS candidates per input and using heavy-atom RMSDs after Kabsch alignment to perform initial clustering (e.g., complete-linkage clustering at 0. p. 2
Where the paper says it

Selection uses a persistent global candidate pool and a kinetic-guided policy adapted from previous work. 12 Candidates are ranked using local electronic barriers along their best available complete routes from the starting reactants, with the largest barrier contributing most.

15 Å) to reduce the number of Simulator evaluations by selecting representative medoids for each cluster. p. 2
Where the paper says it

When further verification is required, Carina TS optimization is performed, followed by Simulator prediction or IRC calculations on the Carina potential energy surface and endpoint verification. To reduce the computational cost of these calculations, TS candidates are first clustered by geometrical similarity, and up to k representatives per cluster are sent to Simulator, where k is configurable.

Stationary points are verified by obtaining forces and Hessians from Carina energy differentiation, followed by a test for a single significant imaginary frequency below-10 cm-1. p. 3
Where the paper says it

Records retain structures and trajectories together with energies (E), forces (F) and Hessians (H). (c) Outcomes of DFT TS optimization, frequency analysis and IRC calculations for 10,000 Sampler-generated TS structures.

Energy references are established as local Carina-computed electronic-energy differences (delta E double dagger e = E Carina(X double dagger e) - E Carina(X ref R,j)), which serve as relative electronic heights rather than absolute barriers. p. 7

Where the paper says it

Solid and dashed arrows indicate reactions with and without water, respectively. Arrow values are local Carina-computed electronic barriers (kcal/mol), each relative to the corresponding optimized reactants.

These local values are used for navigation scoring. p. 7
Where the paper says it

Solid and dashed arrows indicate reactions with and without water, respectively. Arrow values are local Carina-computed electronic barriers (kcal/mol), each relative to the corresponding optimized reactants.

Network expansion is guided by a persistent global pool of discovered, eligible reactant sets. p. 2

Where the paper says it

After each round of calculations, further network exploration requires selecting which accepted product sets to expand. Selection uses a persistent global candidate pool and a kinetic-guided policy adapted from previous work.

The priority of a new node v is determined by a navigation heuristic S(v) = pi in PL(r,v) (sum from i=1 to pi 10-(i-1)b(i)), where b(i) are the local electronic heights along the route pi from the root r to v. p. 2
Where the paper says it

12 Candidates are ranked using local electronic barriers along their best available complete routes from the starting reactants, with the largest barrier contributing most. The configurable branching factor n sets the number of highest-priority eligible candidates selected from this pool per round.

This process allows for iterative exploration, prioritizing routes with lower cumulative energy barriers. p. 1
Where the paper says it

Single-step benchmarks establish the basis for iterative, kinetic-guided exploration (Fig. In multicomponent and structurally complex systems, the generated pathways uncover lower-barrier cooperative transformations and conformational alternatives (Fig.

The framework explores complex systems, including multicomponent reactions and flexible systems. p. 1

Where the paper says it

Single-step benchmarks establish the basis for iterative, kinetic-guided exploration (Fig. In multicomponent and structurally complex systems, the generated pathways uncover lower-barrier cooperative transformations and conformational alternatives (Fig.

Specific explorations include transition-state conformational sampling using CREST to refine TS geometries, single-point energy comparisons using methods like B3LYP/D3 with IEF-PCM n-hexane solvation, and finite network exploration which combines causal replay of archived calculations with new calculations for competing reactant sets. p. 6
Where the paper says it

Sampling settings, geometrical grouping and numerical diagnostics are detailed in Supplementary Section S7. For comparison with earlier work, B3LYP-D3 single-point calculations with continuum n-hexane solvation were performed on the Carina-optimized structures.

The results are used to uncover lower-barrier cooperative pathways and competition between different reaction branches, providing mechanistic insights that can guide experimental design by linking candidate product families to unexplained mass spectral signals. p. 0
Where the paper says it

The resulting networks uncover lower-barrier cooperative pathways and competition between catalyst release and catalyst-bound branches. In the examined carbocation cascades, coverage of known product precursors saturates after limited expansion and distinct entries converge onto shared intermediates, indicating a compact accessible reaction space.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Generative discovery of chemical reaction networks can be performed directly from reactants without predefined products, reaction templates or prior enumeration of candidate reactions. This allows for a generative approach to reaction-network discovery and mechanistic analysis without predefined products or templates.

  2. The system can propose transition states and predict the full reaction trajectories connecting reactants and products using Nexus-Sampler and Nexus-Simulator. This capability enables the exploration of unknown reaction networks by proposing pathways from reactants alone and identifying new species that can undergo further reactions.

  3. Kinetic-guided network growth can be implemented using a selection policy ranked by local electronic barriers along best available complete routes from starting reactants, where the largest barrier contributes most to the score. This ensures that explorations prioritize pathways with higher kinetic relevance.

  4. The framework can uncover lower-barrier cooperative transformations and conformational alternatives in complex systems by expanding five first-layer product nodes reached through lowest-barrier pathways (branching factor n = 5). This reveals mechanisms like cooperative proton transfer and alternative reaction channels that are difficult to specify in advance.

  5. The AI system can connect candidate product families to unexplained LC–MS signals by linking generated candidate species to their possible formation pathways. This provides structural hypotheses that can be examined by changing reaction conditions, guiding an iterative exchange between experimental observation and mechanistic hypothesis testing.

Abstract

Chemical kinetics determines molecular stability, reaction selectivity and the evolution of reacting mixtures, but discovering the underlying reaction networks remains challenging. Here we introduce ReactNexus, a generative framework for kinetic-guided network exploration starting solely from reactant graphs. Generative models propose transition states and predict the full reaction trajectories connecting reactants and products, while calculations using a machine-learned interatomic potential verify connectivity and evaluate electronic barriers. On an in-distribution benchmark of 10,000 generated candidates, 80.8% connect to the input reactant after density functional theory (DFT) optimization and intrinsic reaction coordinate verification. The resulting networks uncover lower-barrier cooperative pathways and competition between catalyst release and catalyst-bound branches. In the examined carbocation cascades, coverage of known product precursors saturates after limited expansion and distinct entries converge onto shared intermediates, indicating a compact accessible reaction space. Network-derived candidate structures further assist mass-spectral interpretation of a complex nucleophilic aromatic substitution reaction. Together, these results establish a generative approach to reaction-network discovery and mechanistic analysis without predefined products or reaction templates.

Sources

Related papers