Mapping the Space of Chemical Reactions using Attention-Based Neural Networks

Abstract

<div><div><div><p>Organic reactions are usually assigned to classes grouping reactions with similar reagents and mechanisms. Reaction classes facilitate communication of complex concepts and efficient navigation through chemical reaction space. However, the classification process is a tedious task, requiring the identification of the corresponding reaction class template via annotation of the number of molecules in the reactions, the reaction center and the distinction between reactants and reagents. In this work, we show that transformer-based models can infer reaction classes from non-annotated, simple text-based representations of chemical reactions. Our best model reaches a classification accuracy of 98.2%. We also show that the learned representations can be used as reaction fingerprints which capture fine-grained differences between reaction classes better than traditional reaction fingerprints. The unprecedented insights into chemical reaction space enabled by our learned fingerprints is illustrated by an interactive reaction atlas providing visual clustering and similarity searching. </p><p><br></p><p>Code: https://github.com/rxn4chemistry/rxnfp</p><p>Tutorials: https://rxn4chemistry.github.io/rxnfp/</p><p>Interactive reaction atlas: https://rxn4chemistry.github.io/rxnfp//tmaps/tmap_ft_10k.html</p></div></div></div>

DRACON: Disconnected Graph Neural Network for Atom Mapping in Chemical Reactions

“…Model interpretation is a significant component in any ML study. 46 In this section, we demonstrate how DRACON learns and memorizes useful information from disconnected reaction graphs. First, we investigate the learned latent vector representations of reactions.…”

Molecular Transformer-aided Biocatalysed Synthesis Planning

“…We selected the target molecules from the RetroBioCat's curated set of biocatalyzed pathways 30 based on the intersection between chemistry coverage in our data set ECREACT and the data set of RetroBioCat. In fact, the encoding of ECREACT and the RetroBioCat test set using rxnfp 31 shows that the RetroBioCat test set reactions are forming distinct clusters in the TMAP-embedded reaction space (Figure 13a), in which the fraction of nearest neighbors from the set itself is consistently higher compared to reactions from ECREACT (13b). This analysis highlights the different chemistry captured by the datasets and anticipates a poor performance for those Retrobiocat examples poorly covered in the ECREACT data set (more details can be found in the Method Section, see Figure S14 for a depiction of the reaction classes' statistics from Finnigan 30 ).…”

“…We annotated reaction SMILES for each step of the biocatalytic cascades considered in the test set from Finnigan et al 18 , excluding solvents information. For each reaction SMILES we extracted fingerprints using rxnfp 31 and we computed among the knearest neighbors (k=10), the fraction of neighbors belonging to RetroBioCat test set. The visualization of the embedded reactions was generated using TMAP.…”

Completion of Partial Reaction Equations

“…Deep-learning models applied to chemical reactions have received much attention in recent years: from the design of algorithms for forward reaction prediction [1][2][3] and retrosynthetic analysis [1,4,5] that help chemists plan the design and execution of chemical syntheses, to the generation of reaction fingerprints [6] and prediction of reaction classes [7,6], yields [8], or activation energies [9].…”