让实验以计算速度重放：Leash Bio 如何发现模型在学化学，还是在走捷径

Leash treats scalable measurement as an engineering problem in which reproducibility, controls, and noise diagnosis must improve alongside the number of observations.

Encoding the same molecule with multiple DNA barcodes helps distinguish genuine chemical binding from proteins that preferentially bind the barcode itself.

A reliable measurement of a DNA-tagged molecule can train a conditional model, but advancing that molecule as a drug still requires off-DNA synthesis and separate functional validation.

The Kaggle competition showed that success on familiar building blocks did not transfer to a wholly different chemical library, where participating teams performed no better than chance.

Predicting a molecule's author from its structure revealed a shortcut through which activity models could exploit chemist identity and selection habits instead of learning the intended physical relationship.

Leash chose deeper characterization of smaller libraries across more proteins rather than assuming that a larger compound count alone would produce better generalization.

After Hermes appeared to memorize protein nicknames, Leash engineered pocket mutations and compared binding changes to make that shortcut less useful during training.

The agent sandbox replays previously measured experiments through database lookups, allowing rapid hypothesis revision without claiming that new wet-lab experiments happen instantaneously.

A promising binder and a six-month mouse PKPD result are early milestones, not evidence that a candidate is safe or effective in humans.

Ian argues that Leash's defensibility comes from accumulated process knowledge and frictionless metadata capture, while broader chemical reasoning remains an ambition beyond today's screened space.