MRI Prediction Beyond Tissue Probabilities in a Human Mitochondrial Brain Atlas

**Background:** MRI-derived mitochondrial maps motivate a direct comparison with simpler tissue-probability predictors. Correlation with observed biochemistry does not quantify the accuracy added by other MRI features. **Objective:** Quantify incremental held-out accuracy of 22 MRI features over the same grey-matter and white-matter probabilities for six source-processed biochemical outcomes, and assess anatomical-label support, matched training counts and a quadratic tissue comparator. **Methods:** An exploratory reanalysis used 539 linked voxel rows from one post-mortem biochemical donor and template MRI from other cohorts. Fixed ordinary least-squares/ridge fits used training-only predictor standardization and penalties 0, 1, 10 and 100. Paired squared errors were evaluated under fivefold random holdout and, in a selected 336-row pool, exclusion of each of 16 exact labels and ten count-matched random partitions per label. All outcomes and penalties were retained. Predictive R2 and its paired increment are finite descriptive quantities; no population intervals were estimated. **Results:** At penalty 1, full-atlas random-holdout incremental R2 over linear tissue probabilities ranged from 0.0230 to 0.0626 across the six outcomes (539 voxel rows; one donor). Within the selected pool, exact-label increments were negative for all six outcomes, ranging from −0.1508 to −0.0572; count-matched random increments ranged from −0.0204 to 0.0491. The five-term quadratic comparator retained all six negative label-held increments at penalty 1. At penalty 100, label-held increments over the linear baseline became positive for CI, TRC and MRC, but remained negative for CII, CIV and MitoD. Excluding the single row absent from the released original split preserved that six-outcome sign pattern at penalties 1 and 100. **Conclusions:** The additional channels provide conditional accuracy within this fixed atlas. Comparator, evaluated voxel pool and penalty materially change that accuracy. The reusable loss tables support explicit baseline and label-support reporting; they establish neither a causal anatomical effect nor donor-level, individual-map or clinical validation.