Efficient Customizable Synthesis Planning with Graph Grammars, with an application to Elastase B inhibitor candidates
Year: 2026
Type: preprint
Abstract: Computer-assisted synthesis planning (CASP) software allows experimentalists to streamline setups and reduce unsuccessful experiments. CASP has a long history, yet remains an active research area. Although various tools have been developed and published, few have remained publicly available and continue to be supported. Those available at no-cost typically operate on generalized global databases, ignoring domain-or laboratory-specific constraints. Here, we present LAGER, a graph-grammar-based retrosynthesis engine built on the Double Pushout (DPO) formalism of the MØD framework. Retrosynthetic pathways are resolved by a state-based A∗ search with a weighted evaluation function. The path cost, g(n), penalizes pathway length and chemoselectivity risks from competing side reactions, whereas the heuristic h(n) estimates chemical distance to available reactants using fragment-based synthetic accessibility score (SAscore) to prioritize structurally simpler precursors and limit search-space expansion. LAGER is designed to accelerate synthesis planning for subsequent generations of a specific compound series by leveraging synthetic knowledge from previous generations. Its reaction library therefore comprises well-established reactions routinely used within a particular laboratory alongside readily available starting materials. We evaluate LAGER on a benchmark of 109 investigated inhibitors of elastase B (LasB), a key virulence factor of the Gram-negative bacterium Pseudomonas aeruginosa. With appropriate heuristic weighting, LAGER recovers the experimentally used pathway among Top-3 proposed pathways for up to 81% of targets. Importantly, deviation from the reported route did not imply chemical infeasibility: at least one chemically valid pathway was identified within the Top-3 candidates for 98% of targets. However, our results indicate limitations of SAscore as a chemical distance metric for targeted retrosynthesis, highlighting the need for more robust, computationally efficient alternatives. Comparison with AiZynthFinder using identical starting materials further demonstrated the importance of laboratory-specific reaction knowledge, as general reactiontemplate sets rarely reproduced the complete literature route for the investigated compounds and often included only a subset of the key transformations in the proposed pathways. Scientific Contribution. This work demonstrates the value of incorporating laboratory-specific reaction knowledge into automated retrosynthetic planning. Unlike general-purpose approaches, LAGER explicitly encodes domain-specific transformations within an interpretable graph-grammar-based search framework, providing a customizable approach for synthesis planning within a defined chemical space.
Source: ChemRxiv
Institutions: Leipzig University, Center for Scalable Data Analytics and Artificial Intelligence, German Center for Infection Research, Helmholtz Centre for Infection Research, Helmholtz Institute for Pharmaceutical Research Saarland, Saarland University, University of Vienna, Santa Fe Institute, Universidad Nacional de Colombia, Max Planck Institute for Mathematics in the Sciences
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7220904213