A generative model cannot tell you why it proposed that molecule.

It can tell you the molecule scored well. For a filing, a patent, or a synthesis decision, the provenance of a candidate matters as much as its score, and a learned model does not have one to give.

No provenance

a learned generator produces a structure without a derivation. There is nothing to audit, reproduce, or defend later. Design rationale, geometric pipeline · our position

Not reproducible

stochastic sampling gives a different answer on a different seed, so the same request does not yield the same candidate set. Design rationale · our position

Validity is separate

from plausibility. A structure that scores well can still be chemically invalid, and the score does not check. Chemical validity checker · our position

Detailed balance

over an explicit energy gives a sampler whose distribution is stated rather than learned, so the acceptance criterion is inspectable. Property-targeted graph sampler · our position

This is the same argument as everywhere else on this platform. The answer matters, and so does whether you can show how it was reached. A geometric pipeline is auditable by construction; a learned one is not, whatever its benchmark says.

5 tools, and the failure each one prevents

Six that close the biggest gaps. 5 in the full molecular and drug discovery library.

Property-Targeted Graph Sampler HPC

The gap it closes. Explicit-energy detailed-balance MCMC, so the sampling distribution is a stated property rather than an emergent one.

See it run in your browser →

Chemical Validity Checker HPC

The gap it closes. Valence, charge and connectivity checked deterministically. A generator that never sees this can propose structures that cannot exist.

See it run in your browser →

Geometric Molecule Resolver HPC

The gap it closes. Resolves structure geometrically rather than sampling toward it, so the same input yields the same output every time.

See it run in your browser →

Molecular Embedding and Distance HPC

The gap it closes. A distance you can define is a distance you can defend in a similarity argument. A learned embedding gives a number without a metric you chose.

See it run in your browser →

Active-Learning Ligand Screening STD

The gap it closes. Screening under an explicit acquisition rule, so the reason a candidate was selected next is recorded rather than inferred.

See it run in your browser →

These are computational chemistry tools, not a discovery program and not a substitute for synthesis, assay or a medicinal chemist. Validity checking establishes that a structure is chemically well-formed, which is a much weaker claim than that it is synthesisable, stable, or active.

Run one on your own numbers

Every tool above runs entirely in this browser tab. Nothing is uploaded, nothing is sent to a model, and each result cites the rule it applied.

Free tier, no card, no install. An account saves your work and lets you export a citable record of a run. See all 5 molecular and drug discovery tools.