Abstract
Custom tokenization dictionary (CUSTODI) is introduced as a novel way for tackling the problem of molecular representations, and especially the challenge of molecular property prediction. Herein, the motivational theory and the actual representation and model are presented and shown to have performance that is in line with benchmark methodologies. The uniqueness of CUSTODI is its applicability on small training sets and the developed theory suggests its possible use for a-priori estimation of future fit quality on any given dataset, regardless of the method used for fitting.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have