Fig. 3.

Model robustness to partial data. a The correlation coefficient r² was computed for the Fission yeast dataset using as training set different fractions of the data. b Using only the data for the first 250 generations of the Budding yeast dataset fitness was predicted for 500 generations. The coefficient of correlation r² as computed between the predicted and the observed fitness at 500 generations for each model (Power Law r² = 0.48, Stickbreaking r² = 0.03, Saturation r² = 0.18, Thermodynamics r² = −0.02, and FGM r² = 0.05). c Using the informations on the initial 10,000 generations the next 40,000 generations of the LTEE were predicted. The coefficient of correlation r² as computed between the predicted and the observed fitness for the last 40,000 generations for each model (linear r² = 0.82, Power Law r² = 0.78, Stickbreaking r² = −1.6, Saturation r² = 0.48, Thermodynamics r² = −1.7, and FGM r² = −1.8)