(Fig.1) AI's success rate of binder molecules was too bad = only 16%
The 2nd, 4th, 5th, 10th paragraph of this hyped news (8/29/2025)
say
"Researchers have used neural networks to help develop new proteins called binders that are designed to attach to therapeutically relevant targets (= proteins )."
"BindCraft grew out of a desire to develop a more accessible, user-friendly tool that would only need to test a handful of proteins to get a binder."
"used structures fed into Google DeepMind's AlphaFold2 system to generate sequences for new binders based on a set of desired functional properties—like binding to a specific target."
"Overall, experiments showed that the team's binders attached to their intended targets with an average success rate of 46% (= hype ), offering the possibility (= still useless ) of greater therapeutic control."
↑ This research paper ↓
p.2-Fig.1b says "Values in the blue box indicate the number of successful designs, where binding was observed on SPR measurement versus the total number of designs tested"
↑ So the success rate of finding binder molecules attached to the target CbAgo protein (= Fig.1b-right-lower-blue box ) is only 2/12 = 16%, which is too low success rate to predict side effect (= cannot predict the designed binders may attach to non-target undesired proteins ).
↑ This research used SPR which can vaguely detect the refractive index of target proteins slightly changed after attaching to binders by light, which can Not directly confirm binder molecules really attach to the designated sites of target proteins.
The same paper's p.9-right-2nd-paragraph says
"Despite the successes outlined here, there are limitations to the BindCraft design approach... We assessed the possibility of using the recently released
AlphaFold3 model for filtering, but still found a large proportion of false positive predictions" ← This AI Bindcraft + Alphafold cannot predict proteins (= drugs ), after all.
The same paper's p.24-Extended-Fig.5b shows local resolution of proteins by cryo-electron microscope is too bad (= ~ 7Å ) to see individual atoms that need 1Å resolution.
The same paper's p.27-Extended-Table1-Resolution shows X-ray crystallography (= bad resolution ~45Å ) also cannot see individual atoms.
↑ So AI, Alphafold trained on these current useless cryo-electron microscopes and X-ray-crystallography with bad resolution are useless, unable to predict molecules or drugs effective against target proteins forever.
↑ This above research tried to find binder molecules attaching to exogenous (bacterial = CbAgo ) proteins which are completely different from humans' normal proteins (= easier to distingush ) by AI.
↑ This current AI with too low success rate cannot distinguish normal proteins and slightly-different mutated proteins (= Alphafold cannot predict mutated proteins ) of cancers (or Alzheimer ), so AI is useless for finding drugs against cancers and Alzheimer
Feel free to link to this site.