Decision benchmarking of cryo-EM re-refinement

Map fit is widely used in cryo-EM re-refinement. We benchmarked selection against retaining the deposited model (no operation), using packed-side-chain agreement with higher-resolution references. In a prespecified 3.0–3.6 Å cohort (167 entries, 131 sequence groups), highest-correlation selection increased χ 1 disagreement by 2.4° (95% interval 2.1–2.7°), with positive loss in 91% of groups. Group-weighted fractions with >20° increases and decreases were 7.2% and 3.7%; the fraction exceeding 1 Å local side-chain RMSD rose by 2.4 percentage points. Selection by a half map withheld from re-refinement increased disagreement (2.29°; 1.70–2.97°; 99 entries). In 53 prespecified within-study low/high-resolution pairs, loss was 3.43° (2.85–4.02°), positive in 50 pairs; half-map loss was 2.35° (1.60–3.21°; 44 pairs). Primum provides fixed cohorts, residue sets and scoring tools to benchmark refiners and selection rules against no operation.