Study findings

Binder0154 contains 64 residues and Binder0026 contains 67. Each was modeled with NCALD/P61601 residues 1–193 using seeds 7, 42 and 101 and five samples per seed. The dataset contains 30 original coordinate files and 30 confidence summaries. Representative structures duplicate individual samples and are not additional observations.

The median binder RMSD at positions near each design’s own RF reference is 3.477 Å for Binder0154 and 4.852 Å for Binder0026. At the fixed 5 Å reference-contact definition, median retained pair counts are 6/8 and 11/19. Different design-specific reference masks prevent a direct ranking between binders.

QF192–193 is within 8 Å of the binder in 13/15 Binder0154 models and 7/15 Binder0026 models. The associated relative placement remains uncertain by PAE. One model per binder also contains an interchain minimum distance below 1.5 Å; these geometrically problematic examples remain included.

The older 191-residue NCALD series is a separate comparison and is not included here. Changes in inference features, sequence alignments, templates and token representation prevent interpreting the 191/193 comparison as an isolated causal effect of the terminal QF residues.

Interpretation and reproducibility

The results describe model geometry. They establish neither direct binding, affinity, NCALD reduction, a physiological target conformation nor SMA-related function.

The 30 models are dependent samples, not 30 independent experiments. Broad sample and seed variation limits claims of construct-independent stability.

Compact inputs contain exact sequences, constructs and seeds, but omit full inference features. The older RF reference coordinates and fixed masks are not part of this bundle, so it cannot by itself reproduce every inference or comparison metric.

Dataset download

All 30 models