Finding. Machine-learning predictors of enzyme turnover number (kcat) look only modestly accurate on BRENDA and collapse on EnzyExtract. When those numbers are dropped into yeast enzyme-constrained genome-scale models, the error that actually changes the phenotype is an underpredicted mitochondrial ADP/ATP carrier: adenine nucleotide exchange throttles, and the model invents a cytosolic ATP shortage.
Why this paper matters
Kinetic parameters are the quiet prior inside every enzyme-constrained metabolic model. Measured kcat values are sparse, so the field has been buying machine-learning fill-ins and scoring them with global regression. Rimón Martínez, Lottermoser, Zanghellini and colleagues show that this is the wrong exam. A tool can win the benchmark and still teach a genome-scale model the wrong biology, because a few high-leverage reactions dominate the phenotype.
The high-leverage reaction they catch is mitochondrial. The ADP/ATP carrier (the adenine nucleotide translocator of yeast) is how matrix ATP becomes cytosolic ATP. Underpredict its turnover and the model does not politely increase error bars. It announces that cytosolic ATP supply is limiting. That is a mitochondrial misdiagnosis generated by a spreadsheet.
What they actually measured
Six current kcat predictors go onto a curated BRENDA-derived set. Five also face EnzyExtract. Accuracy is already only moderate on BRENDA. On EnzyExtract, every tool falls to an R-squared of 0.20 or lower. Training-set overlap shrinks with that drop (exact sequence matches 24 to 78 percent on BRENDA, 9 to 26 percent on EnzyExtract), but overlap is not a full explanation of who generalizes.
They then do the test the leaderboard skips. Predicted kcat values parameterize Saccharomyces cerevisiae enzyme-constrained genome-scale models. Growth is predicted across 19 conditions and compared with experiment. No tool-specific model consistently reproduces the observed growth variation.
The reversal on glucose minimal medium is the memorable result. The weakest benchmark performer gives the most accurate growth. Higher-ranked predictors miss by more. The authors trace that mismatch to localized errors at high-leverage nodes, specifically underpredicted mitochondrial ADP/ATP carrier turnover numbers that restrict adenine nucleotide exchange and impose an apparent limit on cytosolic ATP supply. Relax that constraint and predicted growth moves toward the experimental reference.
How to read the score
This is an 80-plus methods-and-systems paper because it changes how mitochondrial parameters should be trusted inside models, not because it discovers a new carrier gene. Confidence is high for the computational claim as stated: benchmark rank is not downstream rank, and the carrier is a demonstrated leverage point on glucose minimal medium. Confidence is lower that every wrong yeast phenotype in the wild is an AAC problem.
What to do with it
If you run enzyme-constrained models, add a mitochondrial-carrier audit before you believe an ATP-supply or growth hit. If you build kcat predictors, stop treating global R-squared as the product. The product is whether the downstream model still knows how mitochondria pay the cytosol. Mammalian reconstructions should get the same stress test. Do not cite this as experimental proof that yeast AAC is kinetically slow. Cite it as proof that a bad number at that node is enough to rewrite the phenotype.