The problem
The problem with generated molecules
Generative models now propose molecules by the million. Most would not survive a medicinal chemist’s first look. Some are chemically invalid or unstable. Some are known compounds, or trivial variants of them. Some score well only because the one model that scored them has a blind spot, which is the cheapest of all mistakes to make and the most expensive to discover later, at the bench.
A candidate list is only as good as the scrutiny behind it. That scrutiny is what we sell.
How it works
From objective to dossier in four steps
Evolve
A search that learns which strategy works for you
Evolve is an evolutionary search in the spirit of recent AI discovery systems. It keeps an archive of the best molecule found in each chemical series and scaffold family, so the search stays broad instead of collapsing onto one answer, and it breeds the next generation from that archive.
What makes it unusual is that the proposers compete. Every strategy is scored on the same terms. The budget follows whatever is actually producing confirmed results on your problem, whether or not it is the fashionable choice.
A generative model
Conditioned on your target and fine-tuned toward the properties you need.
Medicinal-chemistry edits
Fragment growing, scaffold hops and bioisosteric swaps applied to the strongest candidates so far.
A language model writing molecules
Shown its own past results so it can learn from them.
A language model writing search programs
Run in a locked sandbox with no network and no access to anything else.
Independent checks, one verdict
When a molecule counts as confirmed
A molecule counts as confirmed only when all six of these conditions hold.
01
Valid
A chemically valid structure, with sensible valences, charges and stereochemistry, that survives standardisation.
02
Binds
Its predicted binding to your target clears the threshold agreed at scoping, with the binding pose recorded.
03
A second opinion agrees
An independently built scoring method reaches the same verdict on binding.
04
Fit for purpose
Its predicted properties sit inside your window, and it trips no structural alerts for reactivity or toxicity.
05
New
No match among known compounds in public reference libraries, and no repeat within your run.
06
Realistic to make
A plausible synthesis route exists, and its synthetic accessibility sits inside the limit agreed at scoping.
When a shortlist deserves it, the final candidates can go one step further, to more rigorous physics-based binding calculations before anything is made.
We also watch how often the independent checks disagree as a search goes on. A rising disagreement rate is the signature of a search that has learned to fool its own judge, and we report it rather than bury it.
Who it is for
Built for teams that need a shortlist they can defend
01
Pharma and biotech R&D teams
Who want a ranked, defensible shortlist before committing chemistry and assay time.
02
Start-ups
With a target and no computational chemistry group of their own.
03
University groups and consortia
Who need a discovery work package with documented, reproducible methods.
Engagements
Start small, then scale with your requirements
Honest limits
What computational screening cannot do
Screening, not synthesis
This is computational screening. Every candidate still has to be made and tested.
Known weak spots
Predicted binding is an approximation, least reliable for flexible binding sites, metal-containing sites and covalent binders. We flag those cases in your report and recommend more rigorous checks before relying on them.
Synthesisability
A plausible route on paper is not a validated synthesis. Your chemists make the final call, with the evidence to make it.
Search time
Search time depends on the objective and on our queue. We quote it at scoping.
Questions
Asked and answered
- Is this the same as Google DeepMind’s AlphaEvolve?
- No. AlphaEvolve is a general system for evolving computer programs. Evolve applies the same broad idea, an evolutionary loop with automatic evaluation, to molecules, and adds a verification layer designed for chemistry. We are not affiliated with Google DeepMind.
- Which generative models do you use?
- Open-source generative chemistry models, fine-tuned toward your target, alongside the other strategies above. Every model and version used in your run is named in your dossier.
- Why should I trust “confirmed”?
- Because it is a precise term, not a mood. It means the six conditions above, every one recorded in your dossier. Where it falls short of experiment, we say so.
- Can you go beyond fast scoring on the shortlist?
- Yes. Shortlisted candidates can go to more rigorous physics-based binding calculations, as part of a programme or on request.
- What if nothing good turns up?
- Then the report says so, with the evidence, and explains why: an objective too tight, a chemical space already well explored, or the checks disagreeing. A negative result honestly reported is worth more than a list padded with candidates that will not survive the bench.
- How is this different from Materials Discovery AI?
- Materials Discovery AI designs inorganic crystals. Targeted Molecule Discovery designs small molecules. Both work on the same principle: nothing reaches your shortlist without independent evidence.
