DeepAC – conditional transformer-based chemical language model for the prediction of activity cliffs formed by bioactive compounds

Abstract

Activity cliffs (ACs) are formed by pairs of structurally similar or analogous active small molecules with large differences in potency. In medicinal chemistry, ACs are of high interest because they often reveal structure–activity relationship (SAR) determinants for compound optimization. In molecular machine learning, ACs provide test cases for predictive modeling of discontinuous (non-linear) SARs at the level of compound pairs. Recently, deep neural networks have been used to predict ACs from molecular images or graphs via representation learning. Herein, we report the development and evaluation of chemical language models for AC prediction. It is shown that chemical language models learn structural relationships and associated potency differences to reproduce ACs. A conditional transformer termed DeepAC is introduced that accurately predicts ACs on the basis of small amounts of training data compared to other machine learning methods. DeepAC bridges between predictive modeling and compound design and should thus be of interest for practical applications.