Effect\nof Data Quality and Data Quantity on the Estimation\nof Intrinsic Solubility: Analysis Based on a Single-Source Data Set

Aqueous solubility is one of the most important physicochemical\nproperties of drug molecules and a major driving force for oral drug\nabsorption. To date, the performance of in silico models for the estimation\nof solubility for novel chemical space is limited. To investigate\npossible reasons and remedies for this, the Johnson and Johnson in-house\naqueous solubility data with over 40,000 compounds was leveraged.\nAll data were generated through the same high-throughput assay, providing\na unique opportunity to explore the relationship between data quality,\nquantity, and model estimations. Six intrinsic solubility data sets\nwith different sizes and noise levels were generated by making use\nof three different approaches: (i) inclusion or exclusion of amorphous\nsolid residue, (ii) measured or experimental log <i>D</i> to identify the intrinsic solubility, and (iii) adopting or omitting\na quality check process in the data processing workflow. A random\nforest regressor was trained on the data sets with three different\nsets of descriptors calculated from RDKit, ADMET predictor, or Mordred,\nand the performances were evaluated with nested cross-validation as\nwell as ten refined test sets. The models confirm, as expected, that\nwith the same data set size, high-quality data leads to better model\nperformance; however, also, models trained with larger data sets containing\nanalytical variability can give equally accurate estimations compared\nto models trained with small, clean, and diverse data sets. However,\nnoise introduced by including the presence of amorphous solid postsolubility\nmeasurement in the training data set cannot be overcome by increasing\ndata size, as they are introducing a biased systematic positive error\nin the data set, confirming the importance of critical data review.\nFinally, two top-performing models were tested on the first test set\nfrom the second solubility challenge, achieving RMSE values of 0.74\nand 0.72 and log <i>S</i> ± 0.5 of 46 and 48%,\nrespectively. These results demonstrated improved performance compared\nto those reported in the findings of the competition, highlighting\nthat a single-source curated data set can enhance the prediction of\nintrinsic solubility.