PhysChem Forum 2025

A. Sazonovas ^1,2, K. Lanevskij ^1,2, R. Didziapetris ^1,2

¹ VšĮ „Aukštieji algoritmai“, A.Mickevičiaus 29, LT-08117 Vilnius, Lithuania

² ACD/Labs, Inc., 8 King Street East, Toronto, Ontario, M5C 1B5, Canada

Lipophilicity, which is often expressed in terms of 1-octanol/water partitioning coefficient logP, or the corresponding pH-dependent distribution coefficient logD, is one of the key physicochemical characteristics of any new drug candidates, as it has a major influence on a variety of the compounds’ properties constituting their ADME, pharmacokinetic, and drug safety profiles. Widely available in silico tools for predicting these properties are mostly based on experimental data for simple organic chemicals and marketed drugs. Consequently, as drug discovery projects are moving to increasingly novel regions of chemical space, utility of existing methods becomes more and more questionable. In several previously published evaluation studies, the mean logP prediction error for in house compound libraries of pharmaceutical companies was shown to exceed 1 log unit by almost all methods. Prediction of logD is even more challenging, as it requires accurate knowledge of both logP of neutral form and distribution of ionic forms of the compound in the relevant pH range.

The main objectives of the current study were:
(1) Collecting a data set of experimental logD values from recent publications dealing with novel congeneric compound series from drug discovery projects;
(2) Evaluating the performance of ACD/LogD predictor for the newly collected molecules using different combinations of available logP and pKa calculation algorithms;
(3) Investigating the potential for improving prediction accuracy for unknown compound classes by application of automated model training.

The compiled data set consisted of >1000 logD values measured at physiological pH conditions. According to the initial validation results, the highest accuracy of predictions based on the models employing only built-in compound libraries can be achieved using a combination of ACD/LogP Consensus and ACD/pKa Classic algorithms, yielding RMSE of about 1 log unit. However, utilizing the automatic training feature of ACD/LogP GALAS algorithm by the means of stepwise addition of collected data to the model self-training library allowed decreasing the RMSE of predictions for the reserved validation set to as low as 0.75 log units. These results demonstrate that performing experimental measurements for some selected representatives of a novel chemical series can be used to gradually adapt ACD/LogP and ACD/LogD predictors to provide reliable property estimates for the entire new class of compounds.

K. Lanevskij ^1,2, R. Didziapetris ^1,2, A. Sazonovas ^1,2

¹ VšĮ „Aukštieji algoritmai“, A.Mickevičiaus 29, LT-08117 Vilnius, Lithuania

² ACD/Labs, Inc., 8 King Street East, Toronto, Ontario, M5C 1B5, Canada

In silico studies of P-glycoprotein (P-gp) mediated efflux of pharmaceuticals usually treat it as a binary endpoint and only attempt to classify molecules as P-gp substrates or non-substrates. However, recently we have adopted a more advanced statistical approach that can circumvent the lack of accurate quantitative data by employing censored regression-based machine learning technique and can make use of experimental measurements recorded as open-ended intervals, or so called censored data points. Such models, parameterized using a minimal set of relevant physicochemical descriptors (lipophilicity, ionization, molecular size and topology), are capable of producing predictions in the form of numerical Efflux Ratio (ER) values, i.e., the ratios of bidirectional permeation rates observed in polarized transport assays.

The proposed approach can also be extended by applying an estimate of passive permeability in Caco-2 cells to split measured ER values into the contributions of passive and active transport routes, and subsequently fitting the model to represent pure P-gp efflux effect. Both model types achieve similar predictive power on the qualitative classification task (> 75% overall accuracy at a threshold of ER > 2 for substrates), while providing a clear basis for mechanistic interpretation. Practical utility of quantitative predictions is demonstrated by using calculated ER values to augment predictions of passive intestinal permeability for a set of high molecular weight ligand-directed degrader molecules. The resulting approach was able to classify the compounds into high and low bioavailability categories with AUC = 0.87, illustrating that the employed mechanistic calculations are applicable even to complex molecules belonging to beyond rule-of-five chemical space.