Factorized Search and Cartography of Synthon-Based Chemical Spaces
Miroslav Lžičař
Abstract
Make-on-demand molecular design spaces are increasingly defined by reactions and synthons rather than by enumerated molecules, creating a grounding problem for synthesis-aware discovery workflows: they must search, audit, and visualize spaces too large to materialize. We present a factorized framework with two complementary operations over a shared synthon product surrogate. SynthonOR approximates each candidate product fingerprint as the bitwise OR of its participating synthon fingerprints and performs query-conditioned retrieval by Tanimoto ranking. SynthonMap uses the same surrogate for route-aware sampling, importance weighting, nonlinear embedding, and product-mass cartography. On an enumerable 29,200-product RDKit synthon-space slice, raw-connector ECFP4{} OR similarities preserve exact assembled-product pairwise similarity rankings with Spearman $\rho=0.928$; count Morgan fingerprints with sum composition reach $\rho=0.982$. On a deterministic 990,829-product slice with 6,273 synthons and 42 reactions, a 42-query all-reaction ECFP4{} check recovers 63.4\% mean exact top-100 overlap (median 67.0, std. 17.1), while a faster global-synthon policy recovers 59.2\% at 0.80s/query. We further instantiate the search benchmark on Freedom Space 4.0, where SynthonOR returns complete top-100 lists for all 100 Freedom-space queries and provides direct ECFP4 hit-similarity comparisons with FTrees and SpaceLight. The result is a transparent search-and-audit substrate for synthesis-aware molecular generators and library-inspection workflows.
Chat is not available.
Successful Page Load