Automated Discovery of Reactive Events via Hypergraph Mining of Ab Initio Atomistic Simulations

Alexandra Stan, Paolo Pellizzoni, Karsten Borgwardt, Christian OchsenfeldJournal of Chemical Theory and Computation 2026 · DOI

TL;DR

This paper introduces an automated workflow that combines ab initio molecular dynamics with hypergraph mining to discover and analyze chemical reaction networks. By treating reaction networks as directed hypergraphs, the method efficiently identifies frequently occurring reaction pathways and their statistical correlation with environmental conditions, significantly accelerating chemical discovery.

What Problem Does This Paper Address?

The exploration of chemical reaction space is crucial for discovering new molecular systems and understanding complex chemical processes. While computational methods like ab initio molecular dynamics (AIMD) can simulate these reactions, analyzing the vast amount of data generated to identify meaningful reaction pathways remains a significant bottleneck. Existing methods often rely on manual interpretation, which is time-consuming, prone to bias, and can miss statistically relevant events. Furthermore, understanding how environmental factors (like temperature or solvent presence) influence these reactions requires robust statistical analysis, which is often overlooked. This paper aims to bridge this gap by providing an automated, statistically rigorous framework for analyzing complex reaction networks derived from AIMD simulations.

Methods

The core of this work is an automated workflow that integrates several key computational techniques:

  • Ab Initio Hyperreactor Dynamics (AI-HRD): This method, previously developed by the authors, is used for efficient exploration of the potential energy surface. It combines molecular dynamics with hyperdynamics and spherical confinement to control temperature and enhance sampling without causing molecular fragmentation.
  • Chemical Reaction Networks (CRNs) as Directed Hypergraphs: CRNs are represented as directed hypergraphs, a generalization of graphs where edges (hyperedges) can connect multiple vertices. This representation accurately captures the multi-reactant and multi-product nature of chemical reactions, unlike traditional graphs.
  • Frequent Pattern Mining on Hypergraphs: The authors adapt frequent pattern mining algorithms, commonly used in computer science for finding recurring patterns in datasets, to identify frequently occurring reaction pathways within the hypergraph representation of CRNs. This is achieved by transforming the hypergraphs into itemsets, allowing the use of efficient algorithms like LCM.
  • Statistical Correlation Analysis (Fisher's Exact Test): To identify reactive events statistically correlated with environmental conditions (e.g., temperature, water concentration), Fisher's exact test is employed. The family-wise error rate (FWER) is controlled using Tarone's correction to ensure statistical significance and minimize false positives.
  • Quantum Mechanical Refinement (Double-Ended Growing String Method - DE-GSM): Once frequent and statistically significant reaction pathways are identified, their minimum energy paths (MEPs) and transition states are refined using the DE-GSM method at a higher level of theory (ωB97X-3c). This provides accurate kinetic and thermodynamic data.

Datasets

The primary "dataset" used for evaluation is the thermally controlled interstellar synthesis of carbamic acid. This system was chosen because it has been experimentally investigated, allowing for direct validation of the proposed computational workflow. The authors generated multiple simulation sets under varying conditions (temperature, water concentration) to test the statistical analysis capabilities. No new benchmark datasets were introduced; rather, an existing experimental system was used to generate computational data for analysis.

Key Insights and Ideas

  • Hypergraphs as a Natural Representation for Reaction Networks: The paper highlights that directed hypergraphs are a more accurate and unambiguous way to represent complex chemical reaction networks compared to traditional graphs, especially when dealing with multi-component reactions or catalysts.
  • Frequent Pattern Mining for Efficient Discovery: Applying frequent pattern mining to these hypergraphs allows for the automated identification of recurring reaction pathways across numerous simulations, overcoming the limitations of manual analysis. This significantly reduces the effort required to sift through large datasets.
  • Statistical Rigor for Environmental Effects: The integration of Fisher's exact test with FWER control provides a statistically sound method to determine if specific reaction pathways are significantly influenced by environmental factors, moving beyond simple observation.
  • Bridging Exploration and Refinement: The workflow seamlessly connects the initial, computationally cheaper exploration phase (using AI-HRD and pattern mining) with the more expensive, high-level quantum mechanical refinement, focusing computational resources on the most relevant and statistically significant findings.

Key Results

  • The workflow successfully reproduced experimental findings regarding the formation of carbamic acid and its dimer under interstellar conditions.
  • It identified the crucial role of water as a "solvent" in the astrochemical system, noting that while it reduces overall reactivity, it supports the formation of carbonic acid and its ammonium salt.
  • The statistical analysis revealed that at 39 K, patterns related to ammonium carbamate formation are significantly enriched, while at 62 K, the formation of carbamic acid is favored.
  • The presence of water was shown to indirectly aid in the formation of ammonium carbamate at 39 K and reduce the probability of carbamic acid formation at 62 K.
  • The method demonstrated its ability to handle a large number of simulations (hundreds of 50 ps runs) and extract meaningful insights.

Limitations and Future Work

  • Computational Cost of Exhaustive Sampling: While the method is efficient, achieving sufficient statistical power still requires a large number of simulations, especially for systems with subtle temperature differences.
  • Generalization of "Reactive Encounters": The authors acknowledge that defining a general metric for "reactive encounters" (to normalize pattern frequencies) is challenging and can be system-specific. They note that for a large number of samples, binary trajectory mapping suffices.
  • Future Work: The authors plan to extend the framework to include importance sampling for selected reaction pathways to obtain more accurate free energy estimates. They also aim to further develop the statistical analysis to include false discovery rate (FDR) control.

Summary updated

Abstract

The field of generative chemistry and automated exploration of chemical reaction space has gained much interest in recent years as it provides a feasible alternative to performing resource-intensive experiments by enabling important computational insights into new molecular systems. The results are often summarized in reaction networks, which reveal intricate relations between different key reactive events. Although various approaches to explore the available chemical space have been introduced, the information contained in the resulting reaction networks has not been fully exploited so far. We propose an automated workflow for the analysis of chemical reaction networks by applying frequent pattern mining on the corresponding directed hypergraphs to identify frequently occurring reactive patterns across a set of simulations. Furthermore, we identify reactive events that are statistically correlated with given environmental conditions by applying Fisher's exact test and controlling the family-wise error rate to ensure high statistical relevance. Minimum energy paths for frequent and statistically significant patterns are obtained with the molecular double-ended growing string method at ωB97X-3c level of theory. We showcase the pattern-mining-based analysis on the thermally controlled interstellar synthesis of carbamic acid, where we retrieve results in line with experimental data and further investigate the role of water as a protic solvent therein.
BibTeX
@article{stan2026automated,
  author = {Alexandra Stan and Paolo Pellizzoni and Karsten Borgwardt and Christian Ochsenfeld},
  title = {{Automated Discovery of Reactive Events via Hypergraph Mining of Ab Initio Atomistic Simulations}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2026},
  doi = {10.1021/acs.jctc.5c01682},
}
Figure