How we validate our AI docking pipeline before touching real molecules

Before we trust a docking score on any of our real candidates, we run the pipeline on molecules where we already know the right answer.

The method is called redocking validation: you take a protein structure from the PDB that was solved together with a bound ligand, strip the ligand out, then ask the pipeline to dock it back in blind. If the predicted pose lands close to where the crystal structure says it should, the pipeline is trustworthy for that target class.

We currently run three public, well-known pairs this way (none of these are our actual candidates - just neutral validation cases):

• Caffeine -> adenosine A2A receptor (PDB 3EML)

• - Ibuprofen -> cyclooxygenase-1 (PDB 1EQG)

• - Simvastatin acid -> HMG-CoA reductase (PDB 1HW9)

Same open-source stack for all three: RDKit + Meeko for ligand prep, AutoDock Vina for the actual docking, py3Dmol for the 3D visualization. You can watch a live run of all three examples on our site (link in the listing) - full logs, not a canned demo.

Once a target class checks out this way, we apply the same pipeline to our real longevity candidates. We're not naming those yet (there's a separate thread on why), but the validation methodology itself is fully open.

Happy to answer questions about the docking setup, the PDB selection process, or anything else.