NameRxn was developed to recognise and categorise reactions from their connection tables. NameRxn usages include analysing corporate Electronic Laboratory Notebooks (ELN), journal or patent data. NameRxn is a rule-based expert system that recognises reactions by matching and applying transforms. Over 300 additional reactions are now recognised, taking the total beyond 1500. The recently added reactions together with improvements to existing reaction rules and data extraction increased the classification rate from ~ 66 to ~ 73 % on reactions extracted from patents (Pistachio). The coverage of NameRxn on cleaner data, such as a corporate ELN, is typically higher ( &gt;80 %). NameRxn produces atom maps as a result of applying its rules. This is in contrast to maximum common subgraph (MCS) methods where the atom mapping is the basis for the classification. Our approach allows for classification when a reagent is missing. One of the challenges when working with raw data is how to assess the quality of the data. Pistachio, an automatically extracted reaction database that utilises NameRxn, has become popular for training machine-learning algorithms and retro-synthetic analysis tools. We will discuss sources and examples of errors from automated reaction extraction and approaches we are taking to clean this up. Examples include: &lt; 1 % yield, did the reaction really happen; or the presence of a 31P NMR spectra when there is no phosphorous atom in the product indicates that something is not right. We will also look at some challenges for reaction classification schemes, such as unbalanced reactions, the long tail of less common reactions, tautomers, and matching alternative kekule forms.

NameRxn: More than just a reaction classifier

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.