Machine Learning for Organic Cage Property Prediction

We use machine learning to predict
shape persistence and cavity
size in porous organic cages. The majority of hypothetical organic
cages suffer from a lack of shape persistence and as a result lack
intrinsic porosity, rendering them unsuitable for many applications.
We have created the largest computational database of these molecules
to date, numbering 63,472 cages, formed through a range of reaction
chemistries and in multiple topologies. We study our database and
identify features which lead to the formation of shape persistent
cages. We find that the imine condensation of trialdehydes and diamines
in a [4 + 6] reaction is the most likely to result in shape persistent
cages, whereas thiol reactions are most likely to give collapsed cages.
Using this database, we develop machine learning models capable of
predicting shape persistence with an accuracy of up to 93%, reducing
the time taken to predict this property to milliseconds, and removing
the need for specialist software. In addition, we develop machine
learning models for two other key properties of these molecules, cavity
size and symmetry. We provide open-source implementations of our models,
together with the accompanying data sets, and an online tool giving
users access to our models to easily obtain predictions for a hypothetical
cage prior to a synthesis attempt.