The team had 4CFE exactly backwards, and deprioritised the one target where a win could actually be attributed to structure

What we expected

A target with “no learnable signal” (a ligand-only floor indistinguishable from zero) is a weak or unpromising target to keep investing GPU time in.

What happened

The notebook records that 4CFE/AMPK had been read that way and deprioritised on that basis — a reading later described as wrong, and wrong in the expensive direction. Scaffold-similarity leakage measured across the whole panel puts 4CFE far outside every other target:

8D0M 82%  ·  2YXJ 75%  ·  4HG7 74%  ·  4L7B 66%
4JSV 64%  ·  7AAD 61%  ·  4I5I 60%  ·  4CFE 16%

Why it happened

A low ligand-only floor was conflated with “nothing there,” when a low floor on a low-leakage target actually means the opposite — there is no memorisable 2D shortcut for a model to exploit instead of reading structure. The absence of a memorisable shortcut had been treated as the absence of a target, when it is the opposite: the one target where a structure-model win could actually be attributed to structure, because there is nothing there for a 2D model to memorise instead.

What changed

4CFE was redesignated as the recommended next GPU target, reconfirmed after a separate benchmark-builder bug was fixed and the floor re-measured (+0.191 [-0.11,+0.46] → +0.250 [-0.08,+0.53], overlapping confidence intervals — no floor moved measurably). By contrast, targets where GPU effort had concentrated — 8D0M, 2YXJ and 4I5I — were found to support no generalisation claim in either direction, and 2YXJ’s potency was reconfirmed as 90% recoverable from molecular weight alone (ρ +0.603 [+0.41,+0.73]).

“I had 4CFE exactly backwards… I treated the absence of a memorisable shortcut as the absence of a target.”