Correction: we stress-tested our own docking, and one of our headline claims didn't hold up

At launch we said our docking results were "in the range of approved drugs", and we pointed out a curious pattern: all three candidates had the same best-scoring target. Before moving to the lab we decided to stress-test that with control experiments.

What we did: we docked 23 unrelated molecules (steroids, triterpenoids, polyphenols, sugars, polyamines) and 7 known inhibitors into the same six pockets with identical settings, repeated the candidates with 4 random seeds, re-scored with a second scoring function, and redocked a known inhibitor into its own crystal structure.

What we found: (1) the "same best target" is an artifact of one large cavity: 22 of 26 test molecules scored best there. (2) The Vina score tracks molecule size (|r| about 0.7 to 0.94); after correcting for size none of our candidates stands out, and random steroids scored as well as our best "hit". (3) Known inhibitors are not distinguishable from size-matched random molecules in that box, and the best-known one failed to reproduce its experimental pose there (about 7.5-8 angstroms off), although it reproduced it to 0.96 angstroms in its own crystal structure. (4) Under a second scoring function the pattern mostly disappears.

So for us docking is a way to prioritize hypotheses, not evidence of binding. We have removed the "approved-drug range" wording from this launch page, our website and our chatbot. This does not mean the compounds don't work; it means docking cannot show that by itself. Independent lab assays are the decisive step, and we are arranging them now.

Full write-up with the numbers: https://rasayana.andrii-it.de/blog/docking-controls.html

If you work with docking and spot a flaw in these controls, I would genuinely like to hear it.