How Gnosis Forge designs inside the constraints

There is a version of this post where we unveil a novel kinase scaffold discovered before lunch. We are not writing that version, because we do not have that result, and inventing one would be a strange way to introduce a system whose whole point is not bluffing.

What we do have is a design choice worth explaining. Most generative chemistry optimises a score, usually potency, and then sends the molecule downstream to discover its other problems. Gnosis Forge generates inside the combined constraints of the predictive modules from the start: target engagement, disease-cell selectivity, safety, downstream response, and ADMET. The generator is not asked "what binds?" It is asked "what could survive the whole argument?"

Why generation alone was never the bottleneck

Modern generators can fill a hard drive with plausible structures before the coffee cools. That solved the wrong scarcity. The scarce thing is a molecule that is still interesting after selectivity, exposure, and toxicity have had a vote. If those votes happen after synthesis, you pay for them in weeks and in compounds. If they happen inside the generator, you pay for them in compute, which is the cheaper currency.

What a design cycle looks like, minus the mythology

A team states an objective profile. Not a vibe, a profile: the engagement they want, the selectivity they will not compromise, the safety flags that are disqualifying, the exposure they need. CNS, for instance, is a different country from liver. Forge samples within those constraints. Lens then shows which features the models think are doing the work, so a chemist can reject a candidate for a reason rather than a feeling. The calibration layer marks anything that should not be trusted without an experiment.

The output is a shortlist, not a revelation. Hours, not a montage. And the shortlist is a hypothesis about what to make next, which is a much less cinematic object than a "novel scaffold" and a much more honest one.

The constraints that make this hard

Multi-property design is hard because the properties argue. A change that helps binding can wreck clearance. A linker that helps a ternary complex can wreck permeability. RNA-targeting and induced-proximity molecules make this worse, not better: more interactions, more geometry, more ways to be right about one thing and wrong about the rest. Forge exists because sequential optimisation is how those molecules fail late.

We will publish a real case study when a real series has been made and measured. Until then the method is the news, and the method is this: generate under the constraints, explain the score, and let the bench have the last word.