Table 1. Average phenotype consistency across all test organisms for models gap filled using the four evaluated algorithms.

	Biolog data		Essentiality data
	Sensitivity	Specificity	Sensitivity	Specificity
Targeted parsimony-based	56%	69%	86%	67%
Targeted parsimony PP	56%	69%	84%	68%
Targeted likelihood-based	56%	69%	84%	67%
Iterative parsimony-based	66%	59%	86%	64%
Iterative parsimony PP	66%	59%	85%	64%
Iterative likelihood-based	67%	56%	85%	65%

Iterative gap filling greatly increased the sensitivity (more correct positive growth conditions) and reduced the specificity (more incorrect positive growth conditions) of Biolog simulations. The use of likelihoods did not have a significant effect on the specificity or sensitivity of Biolog simulations. The overall model accuracy for essentiality data was similar for all four algorithms because genes added due to likelihood-based gap filling represented only at most about 7% of the genes in the model. See Figure 6 for the results of knockout simulations using only genes added to gap filling solutions. “PP” means post-processed to add genes to gap filled reactions.