Sometimes the world of informatics overlaps strongly with the universe of human comprehension, and reaction component alignment is one of these cases: when reactants and products are drawn with a common orientation, it can be made very easy and immediately apparent to anyone with basic chemistry experience what is going on in the reaction. An arbitrary molecule layout on the other hand can impose a fairly high cognitive burden.
Continue readingUsing models to propose recommendations for chemical reactions is appropriate for many of the steps needed to fill out the scheme, but sometimes it’s more effective to pick an existing reaction and use it as a template to build out the missing pieces. This approach can be applied to forward and backward syntheses (starting from a reactant or product respectively) and also to finding and proposing stoichiometric reagents.
Continue reading
Once the reaction scheme components are all present and marked up, the last step in this toolchain is to propose a coarse-grained set of conditions, namely the duration, temperature and catalyst concentration that is likely to provide the best yield.
Continue readingSometimes the best way to validate a prediction is to go check the literature and look for the most similar examples, to see if it looks reasonable. Fortunately the data that these reaction prediction tools are based only mostly comes with a DOI attached, so this feature can be added automagically.
Continue reading
The process of completing a reaction scheme includes four preliminary co-dependent steps: proposing formal byproducts, obtaining pairwise atom-to-atom mappings, balancing the stoichiometry, and assigning roles to each of the reaction components.
Continue readingPrediction models for proposing and ranking catalysts and solvents are all very well, but some predictions are more reliable than others. Coming up with some kind of metric for evaluating the difference is a major improvement to utility.
Continue reading
In this chapter we will explore models that can propose and rank solvents for a partially specified reaction. The methodology uses graph-based deep learning models trained on a moderate sized corpus of very well curated reactions with each of the solvents represented as a chemical structure.
Continue reading
This is the first article in a series about chemical reaction prediction, in particular a work-in-progress site that combines a number of original tools for designing reactions. The general idea is that you start with an incomplete reaction scheme, and the models and algorithms will guide you toward filling out the rest, so you end up with a useful starting point for an actual experiment.
Continue reading
Cheminformatics 2.0 