Completion of Partial Reaction Equations

Abstract

We present a deep-learning model for inferring missing molecules in reaction equations. Such an algorithm features multiple interesting behaviors. First, it can infer the necessary reagents and solvents in chemical transformations specified only in terms of main compounds, as often resulting from retrosynthetic analyses. The completion with necessary reagents ensures that reaction equations are compatible with deep-learning models relying on a complete reaction specification. Second, it can cure existing datasets by detecting missing compounds, such as reagents that are essential for given classes of reactions. Finally, this model is a generalization of models for forward reaction prediction and retrosynthetic analysis, as both can be formulated in terms of incomplete reaction equations. We illustrate that a single trained model, based on the transformer architecture and acting on reaction SMILES strings, can address all three points.<br><br>Workshop paper at the Machine Learning for Molecules Workshop at NeurIPS 2020.<br>

Quantum Chemical Data Generation as Fill-In for Reliability Enhancement of Machine-Learning Reaction and Retrosynthesis Planning

“…Then, help may be provided by additional data-driven models, such as models providing the missing compounds in incomplete chemical equations. 50 Since SCINE Chemoton cannot steer an exploration into a desired direction based on the minimal input, the combinatorial explosion resulting from attempts to react every atom and atom pair in every molecule with every atom and atom pair in every other molecule will be computationally demanding and must be tamed. By providing Chemoton with additional information on a reactive system under consideration, the exploration can be made much more efficient.…”

“…With state of the art AI models, however, it is hard to determine byproducts automatically. In principle, the aforementioned chemical equation completion 50 is able to provide missing byproducts. In practice, however, data sets of chemical reactions do not provide this information, which is either unknown or of no practical interest to chemists.…”

SELFIES and the future of molecular string representations

“…Such sequence-2-sequence methods for forward prediction, retrosynthesis, and agent completion can be atom-mapping independent, as the reactant and product atoms do not have to be linked in the training reactions. 139 , 140 , 141

Figure 11 An example of a molecular transformer, which uses Smiles to represent and transform reactant and agent molecules into the product of the reaction, as used by Schwaller et al. 139 The tokenization of the Smiles is shown by the bold characters separated with spaces.

…”

Reagent prediction with a molecular transformer improves reaction data quality

“…Unlike the existing approaches designed specically for reagent prediction, our formulation is not conned to a predened set of possible reagents and allows the prediction of reagents for arbitrary reaction types. Whereas in principle our model is not the rst transformer suitable for reagent prediction, 31,32 it is the rst one to be trained specically for reagent prediction in a machine translation setting. We also demonstrate that the reagent prediction model can be used to improve a product prediction model trained on the USPTO dataset.…”