FROM THE RELEASED TABLES
CASCADE
Two results in the data Anthropic released, neither of which is a re-sort of their ranking score. The lab result still depends on which program drew the protein. And for two tight binders, that ranking names a toxin the protein does not bind.
FINDING
The score does not erase the tool
10 of the 17 binders came from 44 designs, the ones drawn by RFdiffusion3 or Genie 3. A random 44 of the 211 would hold about 3.5. Inside the designs that scored 0.90 or better, PXDesign is 0 for 13 and Genie 3 is 4 for 9. Same score band. Different tools.
Cn2-B1 and AaHII-B1, the two tightest hits they reported, were both drawn by RFdiffusion3. That tool was 17 of the 211 designs ordered.
FINDING
The average names the wrong toxin
Each design was scored against all five toxins. For 14 of the 17 binders, the highest average is the toxin that bound. For these, it is not. AaHII-B2 and AaHII-B3 bind AaH II at 3.8 and 4.0 nM. The average's top toxin is LqhIII, a no-hit. ESMFold2 ipSAE, one term inside that average, names AaH II.
- AaHII-B23.8 nM
Bound AaH II. The average says LqhIII, which did not bind. ipSAE says AaH II.
- AaHII-B34.0 nM
Bound AaH II. The average says LqhIII, which did not bind. ipSAE says AaH II.
- Cn2-B550 nM
Bound Cn2. The average says Ts1, which did not bind. ipSAE says Ts1, which did not bind.
- AaHI-B359 nM
Bound AaH I. The average says AaH I. ipSAE says LqhIII, which did not bind.
The seventeen
Tool is the program that drew the backbone, from the origin table.
AaHII-B2 binds AaH II at 3.8 nM. Backbone from RFdiffusion3. The highest six-score average is LqhIII. The highest ESMFold2 ipSAE is AaH II.
SAEEQAKKVKEVLEAAKKDPKILEDALETSKKNKASEESIKTIEEMLEAVKAKDYAKYEELFLSNLEEPVREAL