Diagnostic cancellation and benchmark mismatch in SOD1 folding free energy calculations
Abstract
Alchemical free-energy calculations are increasingly used to predict how missense mutations change protein folding stability, and their reliability is usually judged by aggregate convergence diagnostics and agreement with experimental controls. This working paper reports a validation campaign for the apo, disulfide-reduced monomer of human Cu/Zn superoxide dismutase 1 (SOD1), the protein behind a subset of familial ALS, computed with GROMACS, pmx hybrid topologies and equilibrium λ-window sampling analyzed by MBAR. Gate thresholds (Pearson r ≥ 0.70, RMSE ≤ 1.5 kcal/mol, median net hysteresis ≤ 0.75 kcal/mol against eight experimental controls) were recorded before the first evaluation. The gate failed (r = 0.326, RMSE = 2.123 kcal/mol across seven usable variants), and prospective predictions were withheld. Post-verdict audits identified two limitations of the validation design itself. First, the standard net forward/reverse hysteresis metric sums signed discrepancies and can conceal large local disagreement: in one F64A ladder, net hysteresis was 0.023 kcal/mol while the summed absolute discrepancy was 1.678 kcal/mol (98.6% cancellation), and separately specified continuations showed the estimate moving by −2.08 kcal/mol between disjoint 3-ns blocks. Second, the experimental references were not state-matched to the simulated species in redox state, construct, oligomeric transition, or temperature. These are reported as limitations of one validation design, not as a validated diagnostic or an accuracy estimate for SOD1. This record contains the manuscript (PDF and Word) and the S1 Appendix (A4V sampling-extension and G93A disulfide sensitivity studies). Code, configuration, and analysis workflows are available at https://github.com/Meowsers4/sod1-free-energy-validation. This is a working paper and has not been peer reviewed. AI tools assisted with code development, auditing, manuscript drafting and formatting; the authors are responsible for the scientific content
