Adaptation of a retrosynthetic tool to project-specific chemistry: oligomer synthesis case study
Year: 2026
Type: preprint
Abstract: Retrosynthetic planning for discrete oligomers presents challenges that differ from those encountered for conventional small molecules because their synthesis often relies on repeated transformations applied at multiple related sites within a multistep route. In conventional template-based search, each application of a single-center reaction rule is represented as an independent action, causing repeated chemistry to consume additional search depth and potentially generating order-equivalent trajectories. Here, we showcase the adaptation of the SynPlanner retrosynthetic platform to oligomeric chemistry through the introduction of project-specific priority rules and a newly implemented Rule Multi-Application (RMA) mechanism. RMA enables repeated applications of the same single-center rule to be represented within a single search expansion, with the multiplicity determined dynamically from the molecular structure. Composite rules were additionally used to provide sequence-level guidance for selected ordered transformations. The approach was evaluated on two literature-derived targets: a linear oligoacetylacetone-derived oligomer and a macrocyclic calix[3]pyrrole. For the linear target, baseline planning generated 161 complete routes across 20 strategic-bond-pattern (SBP) clusters but did not reproduce the literature strategy. Combining priority-rule guidance, RMA, and sequence-level abstraction enabled exact recovery of the reported pathway. For the macrocyclic target, baseline planning failed to identify a complete route even after increasing the search budget to 10,000 iterations, whereas the RMA-enabled guided configuration generated 65 complete routes across three SBP clusters, including routes reproducing the strategic-bond pattern of the literature synthesis. These results demonstrate that adapting a retrosynthetic planner to oligomeric chemistry requires not only access to relevant reaction knowledge but also an appropriate representation of repeated transformations during multistep search.
Source: ChemRxiv
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220908868