A Generative Model For Electron Paths

Abstract

Chemical reactions can be described as the stepwise redistribution of electrons in molecules. As such, reactions are often depicted using "arrow-pushing" diagrams which show this movement as a sequence of arrows. We propose an electron path prediction model (ELECTRO) to learn these sequences directly from raw reaction data. Instead of predicting product molecules directly from reactant molecules in one shot, learning a model of electron movement has the benefits of (a) being easy for chemists to interpret, (b) incorporating constraints of chemistry, such as balanced atom counts before and after the reaction, and (c) naturally encoding the sparsity of chemical reactions, which usually involve changes in only a small number of atoms in the reactants. We design a method to extract approximate reaction paths from any dataset of atom-mapped reaction SMILES strings. Our model achieves excellent performance on an important subset of the USPTO reaction dataset, comparing favorably to the strongest baselines. Furthermore, we show that our model recovers a basic knowledge of chemistry without being explicitly trained to do so.

Dataset Bias in the Natural Sciences: A Case Study in Chemical Reaction Prediction and Synthesis Design

“…As such, although the training set can be split into reactants and reagents, splitting the input of the test set into reactants and reagents makes the reaction prediction problem circular because the split can only be done with the answer known a priori. This dataset split is, however, routinely performed in current approaches [11,5,4,12], where experiments are reported in which reagent labels have been provided to the model at test time. The distinction between providing and not providing reagent labels is illustrated in Figure 2.…”

“…[11] explicitly label reagents using separate tokens. [4] input reagent information as a context vector to their model and [12] report results where improved performance is obtained with labelled reagents. Improved performance is not surprising in this case given that the space of possible products is narrowed through the exclusion of side reactions with the reagent.…”

“…For example, "arrow pushing"mapping out electron paths -is a sanity check tool in organic chemistry. The innovative preprocessing step of [4] models arrow pushing and prunes the data to consider only reactions that can be explained by linear electron topology. We argue that more effort should go into identifying and removing "impossible" chemical reactions from the dataset.…”

“…The synthesis design problem as well as the associated problems of reaction prediction and reaction planning are illustrated in Figure 1. The application of Machine learning to this problem has a history at NIPS [2] and has recently been demonstrated to be the state-of-the-art approach both in reaction prediction [3,4,5] and in synthesis design for reactants [6,7,8].…”

Autonomous discovery in the chemical sciences part I: Progress

“…With a similar goal, Jacob and Lapkin build a stochastic block model (SBM) for the classification of reactions into true or false using reactions in Reaxys (true) and ones randomly generated from known chemicals (false) [199]. Other machine learning-based methods include ones that rank enumerated mechanistic [200][201][202] or pseudo-mechanistic [203] steps, score/rank reaction templates [153,204], score/rank candidate products generated from reaction templates [205], propose reaction products as resulting from sets of graph edits [206,207], and translate reactant SMILES strings to product SMILES strings using models built for natural language processing tasks [208][209][210]. These all formulate reaction prediction differently; for example, the model in ref.…”