Property-targeted molecular-graph generation by EXPLICIT-ENERGY, detailed-balance MCMC - the geometric, VERIFIABLE alternative to a learned energy-based or discrete-diffusion generator.
Property-targeted molecular-graph generation by EXPLICIT-ENERGY, detailed-balance MCMC - the geometric, VERIFIABLE alternative to a learned energy-based or discrete-diffusion generator. There is NO neural net, no training, no tuning to the target. A molecule is a benzene-ring graph with a fixed set of editable substituent-group and bond-order slots; proposals are a symmetric single-slot re-draw so Metropolis-Hastings needs no proposal-ratio correction and provably targets the Gibbs measure exp(-beta V). The energy V is CITED, not fit: a Hansch-Leo 1979 logP pi-additivity term toward your target logP plus a defined valence-validity penalty (and an optional defined geometric term). Two modes: a gem-analog two-phase transport-plus-mixing sampler, and a diffusion-analog corrupt-to-noise then reverse-transport-down-the-energy sampler. Because the working state space is small enough to ENUMERATE, the plugin computes the EXACT Gibbs measure and the EXACT detailed-balance residual of the transition kernel, and measures the sampler mixing histogram against ground truth (total-variation distance, reported). HONEST + LOAD-BEARING: the detailed-balance band is on the SAMPLER (detailed-balance residual), NOT on whether a molecule is good, novel, or synthesizable; EXACT/PRECISE describes the computation, never chemical truth. The geometric embedding is a defined embedding - the alternative geometric embedding was benchmarked as information-equivalent to standard methods (no chemical advantage) - and geometric distance is NOT claimed to equal chemical similarity (now tested: does not track composition; 3D shape resolves stereochemistry; embedding matches standard methods). No competitor benchmark is claimed; a "beats discrete diffusion" statement would need a matched-frame benchmark that has NOT been run here.
Leave your email and we will tell you when we add or update Scientific Computing & Physics tools like this one, at most once a month.
We email you once to confirm. After that, unsubscribe from any issue in one click. Your address goes to VaultSync only (privacy).
Open it inside GDBS to save runs to Sandbox, attach results to a Worklog case, or share through a Gate client portal - all in the browser, nothing uploaded to anyone’s cloud.