Table 1. Performance of nine models for four different vehicle pairs where the modelled dataset, N, was 50 compounds or more and the balanced accuracy was at or above the 99^th percentile of the distribution of balanced accuracies for 300 y-randomised models of the same data. Interpolation/extrapolation is how the area under the dose-toxicity curve has been treated: interpolation only means no extrapolation has taken place; high and high-low indicates that the data has been extrapolated at the high-dose or both the high- and low-dose ranges. AUC is the difference in areas under the dose-survival curve needed before it is considered significant. Equivocal is the number of compounds, not included in the modelling, which showed lower toxicity in one vehicle in some experiments but no difference in others. Contradictory is the number of compounds which showed lower toxicity for one vehicle in some experiments and for the other vehicle in other experiments. Percentile is the percentile of the balanced accuracy of the real model in the distribution of balanced accuracies for the y-randomised models. HPC is hydroxypropylcellulose (Klucel).

Vehicle pair	Interpolation/extrapolation	AUC/%	N (less toxic in first vehicle: second vehicle)	Equivocal	Contradictory	Modelling method	Balanced accuracy	Percentile
Saline vs. CMC	Extrapolate high-low	40	123 (69 : 54)	100	14	PLS	76%	99.7
Saline vs. CMC	Extrapolate high-low	40	123 (69 : 54)	100	14	DT	69%	100
Saline vs. CMC	Interpolation only	60	66 (43 : 23)	55	2	PLS	83%	99.0
Saline vs. CMC	Extrapolate high-low	60	79 (49 : 30)	76	4	PLS	85%	99.7
Saline vs. HPC	Extrapolate high-low	30	116 (69 : 47)	42	3	PLS	80%	99.3
Saline vs. HPC	Extrapolate high	30	131 (72 : 59)	44	3	DT	69%	99.7
Saline vs. saline with Tween-80	Extrapolate high-low	30	165 (92 : 73)	81	9	PLS	75%	99.0
HPC vs. saline with Tween-80	Extrapolate high-low	30	107 (56 : 51)	45	7	PLS	86%	99.7
HPC vs. saline with Tween-80	Extrapolate high	30	116 (59 : 57)	45	8	PLS	86%	100