Nobody publishes the data of a HTE plate where most wells give zero yield. We kept ours.

Nobody publishes the data of a HTE plate where most wells give zero yield. We kept ours.

That’s the data behind ZeroShotAPI, the part of @ReactWise that gives you a warm start before you’ve run anything.

Most reaction datasets are built from published chemistry, which means they’re built almost entirely from successes. Conditions reported at their best, failures left out, a record of what worked rather than what was tried. A model trained on that has seen the middle of the space and none of the edges. It can tell you what tends to work. It has no idea where things stop working.

Our own HTE lab exists partly to fix that. We’ve run over 25,000 reactions, and a good amount of them didn’t give good yields - wrong ligand, wrong base, a temperature that decomposed the substrate. In most labs that result gets noted and forgotten. For us it’s half the signal.

ZeroShotAPI uses models pretrained on that data, successes and failures together, so a new campaign starts with a model that has some idea of the terrain before the first plate goes in. Not an answer, a prior. You’re starting somewhere informed rather than somewhere arbitrary.

The uncomfortable truth underneath this: many planned reactions don’t work, and the field has built its datasets as though it does. Closing that gap meant collecting the results nobody wants to publish.

So we did.

‍