All predictions are ours. The measured values are BigSolDB’s. Each measurement is compared with the engine’s prediction for the same compound and solvent, interpolated to the temperature of that measurement.
What the comparison shows
Predictions track measurements well in the organic solvents: R² ranges from 0.47 to 0.68 across the 14 non-aqueous solvents, with mean absolute errors of about 0.4 to 0.8 logS. Water is the exception. There the engine’s R² is 0.12 and its RMSE is 1.74 logS, and only 38% of predictions fall within one log unit, so aqueous solubility predictions from this engine should be treated with much more caution. Accuracy also falls above about 343 K, but only 575 of the 39,026 measurements are at those temperatures.By solvent
Every solvent, with its measurement count under the name. Pick a class to focus on it.Predicted against measured
Pick a solvent, or all of them. Each point is one measurement; the dashed line is perfect agreement.How the errors are distributed
Does the engine know when it is wrong?
Each prediction comes with an uncertainty. Splitting the measurements into four equal groups by predicted uncertainty shows whether it is informative.Temperature and molecular size
How the comparison was made
Predictions and measurements
Predictions and measurements
The engine predicted logS for 1,448 compounds in 15 solvents (five polar protic, five polar aprotic and five non-polar) on a 10-point temperature grid from 298.15 to 373.15 K, 217,200 predictions in all. BigSolDB v2.0 holds about 104,000 measurements. We kept every measurement for a predicted compound and solvent between 298.15 and 373.15 K, which gives 39,026 measurements of 1,381 compounds. The other 67 compounds have no measurement in those solvents and that range.
How the numbers are calculated
How the numbers are calculated
For each measurement, the engine’s prediction is interpolated linearly in temperature between the two nearest grid points and compared with the measured logS. R² is the squared Pearson correlation over all points, and RMSE and MAE are in logS units. Every measurement counts once, so well-measured solvents such as ethanol weigh more in the overall figures than rarely measured ones.
Reading these results
A retrospective comparison
A retrospective comparison
These are predictions compared with an existing database, not a blind prospective test. We have not been able to check whether any of the BigSolDB measurements were part of the data the engine was developed on, so read these results as agreement with published data, not as proof of accuracy on compounds the engine has never seen.
No competing method is shown
No competing method is shown
This page reports the engine’s accuracy on its own. It does not compare it with other solubility predictors, so it makes no claim about being better or worse than them.
The result does not depend on how measurements are matched
The result does not depend on how measurements are matched
The main figures interpolate each prediction to the temperature of the measurement. Restricting to measurements within 2 K of a prediction grid temperature, so that little or no interpolation is needed, gives R² of 0.53 and a mean absolute error of 0.60 logS on 23,875 measurements. Within 0.5 K it gives 0.52 and 0.62, and using only measurements at exactly 298.15 K it gives 0.53 and 0.62 on 4,968. Averaging repeated measurements of the same compound, solvent and temperature gives 0.53 and 0.59. In every case 78% to 79% of predictions are within one log unit.
Measurements are heterogeneous
Measurements are heterogeneous
BigSolDB collects values from many laboratories and papers. Repeat measurements of the same compound can disagree by a good fraction of a log unit, which puts a floor under the error any predictor can reach.
Water is harder
Water is harder
One possible reason is that aqueous solubility depends strongly on ionisation and on the solid form of the compound, neither of which is corrected for in the measurements used here. The data on this page cannot say whether that explains the gap.
Sources
- Measured solubilities: BigSolDB v2.0, a public database of experimental solubilities of compounds in organic solvents and water over a range of temperatures.
- Predictions: Revilico solubility engine, 217,200 predictions, compared with BigSolDB as described above.