Pearl outperforms AlphaFold 3 cofolding

Pearl matters because better cofolding turns structure prediction from a nice visualization tool into a practical filter for which molecules a drug team should actually make. Genesis is claiming that Pearl is not just better at drawing a plausible pose, it is better on external tests like OpenBind and on broad protein ligand benchmarks, which suggests the model can rank real design options with less human cleanup and fewer wasted synthesis cycles.

• The core technical move is more training data. Public protein ligand structures are limited, so Genesis adds a proprietary set of physics generated synthetic complexes. That is important because cofolding models learn from examples, and more varied examples can improve zero shot performance on unseen targets.
  1 sacra 2 arxiv
• The benchmark mix matters as much as the headline result. OpenBind is a dense single target dataset with structure and affinity data for hundreds of compounds, while Runs N' Poses and PoseBusters test broader pose prediction quality across many systems. Doing well across both types suggests Pearl is learning transferable binding geometry, not just memorizing one assay setup.
  3 github 4 nature 5 nature
• This is becoming a competitive frontier in AI drug discovery. Isomorphic has also emphasized beating AlphaFold 3 on hard protein ligand tasks, which shows that model quality on binding poses is now a product differentiator. The company with the best pose engine can screen larger libraries with higher confidence before spending lab dollars.
  6 sacra 7 genesistherapeutics

The next step is turning benchmark wins into faster program progression in the lab. If Pearl keeps improving while inference costs fall through distillation, Genesis can evaluate many more molecule protein pairs per cycle, feed more experimental results back into GEMS, and compound an advantage in proprietary training data over time.