This manuscript presents an empirical evaluation framework for assessing global structural coherence in protein fold predictions generated by AlphaFold. The work introduces a predicted alignment error (PAE) based coherence metric designed to summarize whether predicted residue residue relationships form a single, globally consistent fold or fragment into independently uncertain structural regions. The analysis relies exclusively on standard AlphaFold outputs per-residue confidence (pLDDT) and predicted aligned error (PAE) and is conducted strictly at the post prediction evaluation level. No model retraining, architectural modification, or privileged access is employed. The coherence metric is evaluated across repeated prediction depth using publicly accessible AlphaFold predictions, with all statistics computed from exported confidence and uncertainty data. Version 2B refines the original manuscript by strengthening methodological clarity, expanding empirical context, and improving figure hygiene, while preserving the original conceptual contribution. The paper is explicitly scoped as an evaluation level study and avoids disclosure of any author specific folding workflows, orchestration strategies, or implementation details beyond what is required for scientific validity. Empirical results for a representative protein family demonstrate stable coherence behavior consistent with AlphaFold’s internal confidence signals, illustrating how PAE connectivity can complement existing confidence measures. The framework is intended as a lightweight diagnostic tool for post-prediction assessment rather than a modification or optimization of AlphaFold itself. Version 2B is a stabilized, submission ready revision that maintains full intellectual ownership of the coherence metric and its interpretation while improving reproducibility, depth, and external readability.
Kearon Allen (Sun,) studied this question.