we expected mutants that accumulated these
compounds to improve ethanol tolerance. However, when additional
metabolome data from several metabolic enzyme mutants
that increased these compounds were supplied to reconstruct the
model, the resultant model had low accuracy (RMSEP ¼ 0.381)
probably due to over-fitting and thus could not predict ethanol
tolerance. We found that the predictive model using metabolome
data from only the metabolic enzyme mutants had high accuracy
(RMSEP ¼ 0.0736), suggesting that mutants of genes from different
categories should be used for construction of the predictive model.
Therefore, construction of the predictive model using metabolome
data from mutants of different gene categories under ethanol stress
conditions was expected to get a highly accurate predictive model.
We also had tested Dtrp2 Dtrp3, a double KO strain of tryptophan
biosynthesis in preliminary experiment. However, we did not used
it in the present study because the strain caused the over-fit of
prediction model by its dramatically retarded growth.