Structural confidence fails as a design filter in lanmodulin study
A function-guided model trained on measured metal binding strongly outperformed confidence-based selection when designing terbium-binding protein variants, according to a new arXiv preprint.
In plain words
Researchers tested two ways to pick the best computer-designed proteins. One relied on how structurally plausible each design looked; the other used a model trained on real lab measurements of metal binding. The measurement-based picks bound the target metal far more often, while the plausible-looking picks mostly failed.
What was announced
The authors trained a model on measured metal-binding affinity together with protein sequence, ligand identity and assay conditions, then used it to design variants of lanmodulin, a protein that naturally binds lanthanides, aimed at rare earth element capture. They screened 297 designs from three strategies against terbium and dysprosium, two adjacent lanthanides that are among the hardest rare earth pairs to separate. The panel chosen for high structural confidence actually carried the best confidence scores and expressed well, yet only 16.7 percent of its designs bound terbium, compared with 83.3 percent for the function-guided panel, whose confidence scores were the lowest. The researchers report the odds of binding differed 25-fold between the panels.
No structural-confidence metric separated binders from non-binders consistently across panels, while the function-trained model's own affinity predictions ranked best in every panel and tracked measured affinity closely among its binders. On terbium-versus-dysprosium selectivity, the function-guided designs also outranked a randomly drawn natural lanmodulin relative in 84 percent of comparisons across 149 orthologs. The authors conclude that structural confidence signals plausibility rather than function, and that direct supervision on measured function is what ranks designs.
What it does not show
- The source is an abstract only; full methods, data and peer review are not available in the supplied material.
- All results are the authors' own reported measurements and statistics, with no independent confirmation.
- The findings concern lanmodulin variants and lanthanide binding; generalization to other proteins or targets is not established in the excerpt.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.