Generative machine learning, which aims to find the distribution in the training data set to generate new samples, can take the role of a chemist not only in the formulation of testable hypotheses, but also in the creative aspect, in the assembly of innovative molecules. We have implemented and challenged 'chemistry-savvy' deep learning models in prospective molecular design projects that aimed to obtain synthetically easily accessible new chemical entities. In the first study, recurrent networks were trained with structures of known synthetically accessible, druglike compounds. By applying transfer learning, the learned feature distributions were biased towards certain pharmacologically desired enpoints. The computationally generated de novo designs were subsequently prioritized, chemically synthezised and biochemically tested for the predicted activities with high success rates. At this point, human expert knowledge turned out to be beneficial for selecting synthetically accessible chemical designs. In the second study, we developed a novel virtual synthetic assembly method that combines a rule-based approach with a neural network trained on successful synthetic routes described in chemical patent literature. This unique combination enabled a balance between ligand-similarity based generation of innovative compounds by scaffold hopping and forward-synthetic feasibility of the designs. In a prospective proof-of-concept application, the software successfully produced sets of de novo designs for four approved drugs that were in agreement with the desired structural and physicochemical properties. Target prediction indicated more than 50% of the computer-generated molecules as biologically active. Selected computer-generated compounds were successfully synthesized in accordance with the synthetic route proposed by this method. If successful in the long run, these concepts will combine a continuously learning machine intelligence with the synthesis and testing of pharmacologically relevant chemical matter. Such an envisaged automated drug design engine may not only imitate but exceed human decision making as a core aspect of the drug discovery process.

Designing synthesizable bioactive compounds with chemistry-savvy machine intelligence

2019 Fall National Meeting

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