Automatic kinetic mechanism generation, reagent screening in drug discovery, and automatic transition state search are a few examples of current trending applications where a large space of molecules is to be explored. Such large scale search requires predictive methods that allow fast and accurate estimation of molecules’ properties of interest, such as thermochemistry, solubility, toxicity and etc.<br/>Ideally predictive methods such as machine learning models should take 3D optimized geometry infor- mation as input. However, such information is not always available in practical settings (e.g., large scale search). Besides, most machine learning algorithms take 1D vector as input. Thus, experienced researchers tend to reduce 3D representation to 2D and craft a list of key descriptors which converts molecules to feature vectors (manual featurization). The success of this approach largely depends on how well the problem is understood by the experts.<br/>The authors of this study tried to construct a general architecture that can automatically learn fea- turization from 2D molecular data using convolutional neural networks. Many prediction tasks including solubility and melting points have been carried out. As an extension to molecular property prediction, the same architecture can be used for reaction kinetics by taking active sites as input. One of the most successful applications so far is thermochemistry prediction for polycyclic structures. It significantly reduces Hf (298 K) estimation error from over 60 kcal/mol (group additivity methods) to around 2.5 kcal/mol.<br/>In this talk, the author will discuss in detail how convolution neural network is applied, analyze mean- ingfulness of learned feature vectors and demonstrate its power in achieving high accuracy in certain tasks. In addition, the author will illustrate a continuous learning pipeline which is constructed recently to make such estimators evolve over time.

Adaptive molecular property estimator using convolutional neural networks

255th National Meeting (2018)

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