Reformulating Reactivity Design for Data-Efficient Machine Learning

Toby Lewis‐Atwell, Daniel Beechey, Özgür Şimşek, Matthew N. GraysonACS 2023 · DOI

TL;DR

This paper introduces a novel machine learning approach that reformulates the problem of predicting reaction activation barriers. Instead of aiming for general prediction across vast chemical spaces, it focuses on efficiently finding a specific reaction within a predefined set that has a desired target activation barrier, requiring significantly fewer data points. This data-efficient method is crucial for rational design in areas like catalyst optimization and drug discovery.

What Problem Does This Paper Address?

Predicting the activation barriers of chemical reactions is vital for understanding reaction kinetics, designing new materials, and developing pharmaceuticals. Traditional methods, like quantum mechanical (QM) calculations or experiments, are computationally expensive and time-consuming. Machine learning (ML) offers a faster alternative, but current ML models typically require massive datasets (thousands to tens of thousands of reaction barriers) to achieve reasonable accuracy. This data requirement is a major bottleneck, as generating such datasets is costly. Furthermore, ML models often struggle to generalize to reaction types or chemical spaces not represented in their training data, necessitating bespoke datasets for each new problem. This paper addresses the challenge of achieving accurate barrier predictions with significantly less data, specifically by reframing the problem from general prediction to targeted search within a defined chemical space.

Methods

The core innovation lies in reformulating the ML problem. Instead of predicting an arbitrary barrier, the goal becomes identifying a reaction from a prespecified set that possesses a barrier close to a desired "target" value. The authors propose an ML-based search algorithm that iteratively:

  1. Samples an initial set of reactions and measures their accurate activation barriers.
  2. Trains an ML model on this small set of measured data.
  3. Predicts the barriers for the remaining unmeasured reactions.
  4. Selects the reaction predicted to be closest to the target barrier.
  5. Measures the accurate barrier of the selected reaction and adds it to the training set.
  6. Retrains the ML model and repeats the process until a reaction within 1 kcal/mol of the target is found.

A key aspect of their approach is the use of low-level, approximate activation barriers (e.g., from semiempirical methods or cheaper DFT functionals) as an input feature to the ML model, alongside a one-hot encoding of functional groups. This combination provides a compact and informative representation that helps the ML model learn efficiently with limited data. They also compared their ML search strategy against other methods like random search, local search, Bayesian optimization, and genetic algorithms, finding their ML approach to be the most successful.

Datasets

The authors evaluated their method on several diverse datasets:

  • Aza-Michael Addition: A new dataset generated for this study, focusing on reactions relevant to toxicology and covalent drug discovery. It includes 1000 Michael acceptors reacting with methylamine.
  • Dihydrogen Activation on Vaska's Complex: A dataset of 1947 dihydrogen activation barriers on derivatives of Vaska's complex, previously reported by Friederich et al.
  • E2 and SN2 Reactions: Datasets of E2 and SN2 reaction activation barriers, originally generated by von Rudorff et al., comprising 1000 E2 and 1720 SN2 reactions after accounting for symmetry.

These datasets vary in their "completeness" (percentage of reactions for which accurate barriers are available), with Aza-Michael Addition and Dihydrogen Activation being nearly complete, while E2 and SN2 have significant numbers of missing barriers (74% and 56% respectively).

Key Insights and Ideas

  • Reframing the problem is key: Shifting from general prediction to targeted search within a defined chemical space dramatically reduces data requirements. This circumvents the need for models to generalize across vastly different chemical environments.
  • Low-level barriers as features are powerful: Incorporating computationally cheaper, approximate barrier calculations as input features significantly aids the ML model in learning and guiding the search, even when the accurate barrier data is sparse.
  • Data completeness matters: The performance of the ML search is strongly influenced by the availability of accurate barrier data. Datasets with high "completeness" (like Aza-Michael Addition and Dihydrogen Activation) allow the ML model to learn more effectively and converge faster.
  • Targeted search is more efficient than blind sampling: Even for challenging target barriers at the "tail ends" of the distribution, their ML approach requires orders of magnitude fewer data points than traditional ML methods.

Key Results

  • The proposed ML search algorithm consistently outperformed other tested methods (random search, local search, Bayesian optimization, genetic algorithm) across all datasets.
  • For datasets with high completion rates (Aza-Michael Addition, Dihydrogen Activation), the ML approach typically found reactions within 1 kcal/mol of the target barrier using less than 20 accurately measured barriers.
  • Even for datasets with significant missing data (E2 and SN2), the ML method still required double-digit numbers of samples, which is substantially fewer than the hundreds or thousands typically needed for conventional ML barrier prediction.
  • In a case study optimizing a dihydrogen activation catalyst, the algorithm identified a catalyst structure with a barrier within 1 kcal/mol of the minimum target barrier by performing only 12 DFT reaction barrier calculations. This reaction turned out to have the second-lowest barrier in the dataset.

Limitations and Future Work

  • Defined Search Space: The primary limitation is that the method is restricted to finding a reaction within a predefined set. It does not predict barriers for entirely novel reaction classes outside this set.
  • Data Completeness Impact: The performance degrades for datasets with a high proportion of missing accurate barrier data, as the ML model cannot incorporate these samples into its training.
  • Future Work: The authors suggest that for datasets with low completion rates, it might be prudent to invest in more accurate, albeit still lower-level, barrier calculations for the input features. They also propose an adaptive strategy where the low-level barrier feature is only calculated if the ML model's performance is insufficient without it. The broader implication is that such data-efficient search methods can guide, rather than replace, experimental and computational efforts in chemical discovery.

Summary updated

Abstract

Machine learning (ML) can deliver rapid and accurate reaction barrier predictions for use in rational reactivity design. However, model training requires large data sets of typically thousands or tens of thousands of barriers that are very expensive to obtain computationally or experimentally. Furthermore, bespoke data sets are required for each region of interest in reaction space as models typically struggle to generalize. We have therefore reformulated the ML barrier prediction problem toward a much more data-efficient process: finding a reaction from a prespecified set with a desired target value. Our reformulation enables the rapid selection of reactions with purpose-specific activation barriers, for example, in the design of reactivity and selectivity in synthesis, catalyst design, toxicology, and covalent drug discovery, requiring just tens of accurately measured barriers. Importantly, our reformulation does not require generalization beyond the domain of the data set at hand, and we show excellent results for the highly toxicologically and synthetically relevant data sets of aza-Michael addition and transition-metal-catalyzed dihydrogen activation, typically requiring less than 20 accurately measured density functional theory (DFT) barriers. Even for incomplete data sets of E2 and S<sub>N</sub>2 reactions, with high numbers of missing barriers (74% and 56% respectively), our chosen ML search method still requires significantly fewer data points than the hundreds or thousands needed for more conventional uses of ML to predict activation barriers. Finally, we include a case study in which we use our process to guide the optimization of the dihydrogen activation catalyst. Our approach was able to identify a reaction within 1 kcal mol<sup>-1</sup> of the target barrier by only having to run 12 DFT reaction barrier calculations, which illustrates the usage and real-world applicability of this reformulation for systems of high synthetic importance.
BibTeX
@inproceedings{lewisatwell2023reformulating,
  author = {Toby Lewis‐Atwell and Daniel Beechey and Özgür Şimşek and Matthew N. Grayson},
  title = {{Reformulating Reactivity Design for Data-Efficient Machine Learning}},
  booktitle = {ACS},
  year = {2023},
  doi = {10.1021/acscatal.3c02513},
}
Figure