Hierarchical Reaction Logic Enables Computational Design of Complex Peptide Syntheses

The prevalent assumption in computer-assisted synthesis
planning
has been to rely on the wealth of reaction data and on the consideration
of this vast knowledge base at every stage of route planning. Yet
even if equipped with all requisite knowledge of individual reaction
transforms and state-of-the-art search algorithms, the existing programs
struggle when confronted with advanced targets, such as the complex
peptides this work considers. By contrast, when the searches are constrained
by hierarchical logic, dictating which subsets of reactions to apply
at different stages of synthesis planning, these algorithms are able
to plan, within minutes, complete routes to clinically relevant targets
as complex as vancomycin and as large as semaglutide. Despite not
being trained on any literature precedents, the routes designed by
the algorithm mimic the strategies used by human experts. The hierarchical
planning we describe incorporates protecting-group strategies and
realistic pathway pricing and can be performed in solid-state or solution
modes, in the latter case using either C-to-N or N-to-C peptide extension
strategies.