Enabling Starting Material-Oriented Strategies in Computer-Aided Synthesis Planning With a Bidirectional Search Algorithm

Abstract

Retrosynthesis, in which one proposes a reaction pathway towards a target molecule from simpler starting materials, is a fundamental task in synthetic chemistry. Current computational search methods assume the sufficiency of reaching arbitrary building blocks but fail to address the common real-world constraint where the use of specific starting materials is desirable. To this end, this thesis reformulates computer-aided retrosynthesis as a starting material-constrained problem, in which one or more starting materials are given as input in addition to the target structure. Under this formulation, we are able to apply novel strategies to more efficiently navigate the combinatorial explosion of reactions to consider during synthesis planning. First, we demonstrate how training on multi-step synthesis routes inferred from a reaction base allows a neural network to predict the number of steps needed to synthesize targets from other specified building blocks. Using this learned value function in combination with recent advances in bottom-up synthesis planning, this thesis proposes a novel bidirectional CASP algorithm, DESP (Double-Ended Synthesis Planning). We demonstrate the utility of DESP through a number of empirical benchmarks and case studies.