🤖 machine learning

ReactionAtlas: Ab origine exploration of chemical reaction networks with machine learning

The paper introduces ReactionAtlas, a machine learning framework that autonomously constructs large-scale chemical reaction networks from a few seed molecules by combining generative models with a DFT-trained force field, successfully mapping approximately 47,000 reactions among 12,000 compounds in prebiotic carbohydrate chemistry with unprecedented accuracy and scale.

Original authors: Stefan Gugler, Max Eissler, Khaled Kahouli, Klaus-Robert Müller

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stefan Gugler, Max Eissler, Khaled Kahouli, Klaus-Robert Müller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to map every possible road, shortcut, and dead-end in a massive, ever-expanding city. In chemistry, this city is the "reaction network," where molecules are the buildings and chemical reactions are the roads connecting them. For a long time, building this map was like trying to draw every street in a metropolis by hand, one by one, using a very slow, expensive, and clumsy tool. You had to know the starting point and the destination before you could even try to find the road between them. If you didn't know the destination, you were stuck.

ReactionAtlas is a new, super-smart robot explorer that changes the game. Instead of needing to know the destination, it starts with just a few "seed" molecules (like a handful of basic building blocks: water, formaldehyde, carbon dioxide, etc.) and asks, "What can this turn into?" It then builds the map from scratch, discovering new roads and cities as it goes.

Here is how it works, using some everyday analogies:

1. The Two-Engine System

The robot uses two different "engines" working together to explore this chemical city:

The "Dreamer" (Generative Loop): Imagine a creative artist who looks at a single building (a molecule) and says, "What if I twist this wall or add a window?" It uses a type of AI called a diffusion model. Think of this like taking a clear photo of a building, adding a little bit of static noise to it (making it blurry), and then asking the AI to "clean up" the noise to reveal a new structure that is chemically possible. This new structure is a "Transition State"—the exact moment a reaction happens, like the split second a door is halfway open. The Dreamer proposes these new roads without being told where they lead.
The "Inspector" (Critic): Once the Dreamer proposes a new road, the Inspector checks if it's real. It uses a Machine Learning Force Field (a super-fast physics simulator) to ask: "Is this road stable? Does it actually connect two valid buildings? Is the energy cost reasonable?" If the road is a dead end or physically impossible, the Inspector throws it away. If it's valid, it gets added to the map.

2. The "Kinetic Compass"

The robot doesn't just wander randomly. It has a Kinetic Compass. It runs a simulation to see which molecules are the most "popular" or likely to exist in the real world based on the roads it has already found. It then focuses its energy on exploring the neighborhoods where these popular molecules live. This is like a tour guide who knows which streets are busy and decides to map those first, rather than wasting time on empty, abandoned alleys.

3. The Big Discovery: The Formose Cycle

The researchers tested this robot on a famous chemical puzzle called the Formose Reaction. This is a process where simple formaldehyde turns into complex sugars (like the ones in your body). It's a famous example of how life's building blocks might have formed naturally on early Earth.

The Old Way: Scientists knew some parts of the map, but the full picture was a tangled mess of millions of possibilities. They had to guess the rules and manually check them.
The ReactionAtlas Way: The robot started with just eight simple seeds and explored on its own. It discovered 47,000 reactions connecting 12,000 different compounds.
The Result: It didn't just find the known paths; it found the entire cycle, including a "shortcut" through a specific type of molecule (an enediol) that scientists hadn't fully mapped before. It also figured out the 3D shapes (stereochemistry) and electrical charges of these molecules, which previous maps missed.

4. Why It's a Big Deal

Speed: Traditional methods take days or weeks to find one reaction path. ReactionAtlas finds thousands of valid paths in a fraction of the time.
No Rules Needed: Old methods needed a list of "rules" (e.g., "carbon always bonds with oxygen this way"). ReactionAtlas doesn't need these rules. It learns the physics and chemistry from data and figures out the rules itself.
Accuracy: The robot's guesses are so good that when scientists checked them with ultra-precise (but slow) computer calculations, 85% of the robot's guesses were already perfect, and the rest were very close.

Summary

Think of ReactionAtlas as an autonomous drone that flies over a foggy landscape. Instead of needing a map or a destination, it starts at a few known landmarks, uses its AI to guess where the terrain might go, checks if the path is safe, and then flies there to discover new landmarks. It does this so fast and accurately that it has redrawn the map of sugar chemistry, revealing hidden shortcuts and proving that simple molecules can naturally build the complex ingredients needed for life.

The paper claims this method works for carbohydrates (sugars) and can be applied to other areas like combustion or catalysis, but it specifically highlights its success in mapping the Formose cycle and the origin of life chemistry without any human-imposed rules.

Technical Summary: ReactionAtlas

Problem Statement
Mapping chemical reaction networks (CRNs)—the graphs of minima, transition states (TS), and elementary reactions connecting them—is fundamental to understanding complex chemical phenomena, from catalysis to the origin of life. However, constructing these networks has historically been impractical due to the combinatorial explosion of possible intermediates and the computational cost of characterizing tens of thousands of transition states. Traditional methods, such as Density Functional Theory (DFT), are prohibitively slow for open-ended exploration. Furthermore, existing autonomous frameworks often rely on double-ended TS search methods that require pre-supplied reactant-product pairs, necessitating heuristic rules or prior knowledge of reaction mechanisms. This "cart-before-the-horse" approach limits generalization and prevents truly ab origine exploration starting solely from seed molecules.

Methodology
The authors introduce ReactionAtlas, a framework for ab origine CRN exploration that builds reaction networks from a handful of seed molecules without hand-crafted reaction rules. The system operates via two coupled loops that expand a shared reaction graph:

Generative Loop (Intermolecular Reactions):
- Sampling: Reactants are drawn from the current network based on steady-state concentrations calculated via a kinetic model (solving coupled ODEs). The sampler selects either single molecules or pairs (for bimolecular reactions).
- TS Proposal: A denoising diffusion model, MoreRed, proposes transition state geometries. Trained on the $\omega$ B97X-D3 Grambow dataset, MoreRed takes a reactant geometry (partially noised) and denoises it to propose a TS candidate. This is a strictly single-ended process, requiring no product information.
- Validation: A machine-learned force field (MLFF) critic, MD-ET (a transformer-based model trained on QCML data), acts as a zero-shot predictor for forces and Hessians. It filters candidates by checking for:
  - A Hessian with exactly one imaginary eigenvalue.
  - Two distinct minima at the Intrinsic Reaction Coordinate (IRC) endpoints.
  - Physically bounded forward and backward barrier heights.
PES Loop (Intramolecular & Local Exploration):
- For newly discovered minima above a concentration threshold, a short Langevin molecular dynamics (MD) trajectory samples the local configuration space.
- Saddle Search: Frames from the MD trajectory serve as starting points for Partitioned Rational Function Optimization (P-RFO) to climb to nearby saddle points.
- Validation: Validated saddle points are checked via IRC. If the endpoints correspond to the same compound, they are added as conformational transitions; if they correspond to different compounds (or fragments), they are registered as new reactions ("escaped reactions").
Kinetic Guidance:
- The network is periodically integrated as a mass-action system using Eyring transition-state theory to compute steady-state concentrations. These concentrations weight the sampling in the generative loop, directing computational resources toward kinetically plausible regions of the chemical space.

Key Contributions

Single-Ended Generative TS Search: The paper introduces a diffusion-based approach to propose TSs from a single reactant, eliminating the need for rule-based enumeration or pre-supplied product pairs.
MLFF-Critic Framework: It combines a generative proposer with a robust MLFF critic (MD-ET) to validate TSs and refine geometries, significantly reducing the reliance on expensive DFT calculations during the exploration phase.
Ab Origine Exploration: The framework successfully explores a reaction network starting from only eight prebiotic seed molecules ( $CH_2O, H_2O, OH^-, H_3O^+, CO_2, H_2CO_3, HCO_3^-, H$ ) without imposing specific chemical mechanisms.
Resolution of Charge and Stereochemistry: Unlike many prior datasets, ReactionAtlas resolves both charge states and stereochemistry (R/S, E/Z) across the network.

Results

Scale: Starting from eight seeds, ReactionAtlas discovered approximately 12,000 unique compounds (up to 10 heavy atoms), 31,000 conformers, and ~47,000 elementary reactions.
Relevance: The network recovered >95% of Wikipedia-named species at C2 and >80% at C3 and C4, despite the combinatorial space exploding at higher carbon counts. It also recovered 91% of closed-shell C/H/O species detected in interstellar space.
Quality: 86% of proposed rearrangement TSs and 41% of harder reactions were valid first-order saddle points according to MLFF criteria. After a single DFT (PBE0) refinement step, validity rates rose to 99% and 94%, respectively. The MLFF TSs matched PBE0 references within 0.5 Å RMSD in 85% of cases.
Speed: ReactionAtlas finds a valid TS approximately every 23 minutes (mean wall time on L4 GPU nodes). This is orders of magnitude faster than double-ended methods (e.g., NEB, DE-GSM) which require valid reactant-product pairs and take hundreds of minutes per reaction on CPU nodes.
Formose Cycle Discovery: The framework autonomously reconstructed the formose auto-catalytic cycle, including:
- The initiation step ( $2 \times \text{Formaldehyde} \to \text{Glycolaldehyde}$ ) with a barrier of ~33.4 kcal/mol.
- The Lobry de Bruyn–van Ekenstein (LdB–AvE) interconversion mechanism via enediol intermediates.
- Two distinct auto-catalytic cycles: the canonical Breslow route and a previously unreported "compressed" C3-enediol shortcut.
- Identification of a Type IV autocatalytic motif, a topological structure previously predicted but never observed in real chemical systems.

Significance
The paper claims that ReactionAtlas provides a viable, accurate, and efficient basis for exploring arbitrary CRNs across reaction chemistry. By removing the need for chemistry-specific rules and enabling open-ended discovery from seed molecules, it overcomes the scalability and generalization limits of previous methods. The resulting dataset offers unprecedented scale and detail for carbohydrate chemistry, resolving full stereochemistry and charge states, and provides novel mechanistic insights into the formose reaction, a central model for the chemical origins of life. The authors position this work as a step toward solving the open challenge of applying machine learning to the systematic construction and exploration of prebiotic reaction networks.