Evaluation of log D and pKa predictions for the SAMPL8 blind challenge
4 affiliations
- 1Department of Pharmaceutical Sciences
- 2University of California
- 3Technische Universität Dortmund, Otto-Hahn-Str. 4a, Dortmund, Germany, 44227
- 4Department of Chemistry
Abstract
The goal of the Statistical Assessment of Modeling of Proteins and Ligands (SAMPL) challenge is to improve the accuracy of computational methods used in drug discovery and design by providing a platform to benchmark different methods and draw the community’s attention to specific areas of improvement. The SAMPL8 physical properties challenge focused on the prediction of pKa and log D for multi-protic compounds with two or more widely separated polar groups. Here, we present an overview of the results from the SAMPL8 challenge. We adopted two different workflows to compare experimental and predicted pKa data due to the presence of multiple macroscopic pKa values for individual compounds. These multiple pKa values meant that it was often unclear which of the microscopic transitions contributed to the observed macroscopic pKa. We find that the performance of different methods for pKa predictions is heavily dependent on how we pair predicted pKa values with measured ones–that is, the underlying mechanism. In contrast to some previous challenges, where analysis focused on pairing calculated pKa values with those which were experimentally closest, we here instead use alternate approaches, presenting two different methods of evaluating pKa predictions. Our first approach uses community consensus, pairing the most popular predicted transition(s) with the experimental value. The second involves selecting the correct transition based on additional experimental data–specifically, pH-dependent solubility curves. We used these two approaches to compare SAMPL submissions from participants using a wide variety of different methods. Results showed that while the numerical predictions for pKa values were often reasonable, different participating methods might disagree about the mechanism behind the pKa, indicating that pKa prediction methods may need to focus more on correct prediction of protonation states, not just pKa values. In the log D analysis we find that most methods performed better in prediction of log D for the octanol-water solvent combination, likely due to the greater abundance of training data for these solvents.
Related
- Assessing the state of the art in ADMET prediction: Lessons from the OpenADMET-ExpansionRx community blind challenge
- Crystal structure, Hirshfeld surface analysis, computational study, molecular docking predictions and biological evaluation of a new Schiff base (E)-2-((2,4-dichlorobenzylidene)amino)aniline
- Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
- Statistical predictions on the encapsulation of single molecule binding pairs into sized-dispersed nanocontainers
- Universal Model for Atoms (UMA) in Computational Asymmetric Catalysis: Comparing Enantioselectivity Predictions
BibTeX
@misc{nandkeolyar2026evaluation,
title = {{Evaluation of log D and pKa predictions for the SAMPL8 blind challenge}},
author = {Aakankschit Nandkeolyar and Nicolas Tielker and Stefan M. Kast and David L. Mobley},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009213/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009213/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Evaluation of log D and pKa predictions for the SAMPL8 blind challenge
AU - Aakankschit Nandkeolyar
AU - Nicolas Tielker
AU - Stefan M. Kast
AU - David L. Mobley
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009213/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009213/v1
AB - The goal of the Statistical Assessment of Modeling of Proteins and Ligands (SAMPL) challenge is to improve the accuracy of computational methods used in drug discovery and design by providing a platform to benchmark different methods and draw the community’s attention to specific areas of improvement. The SAMPL8 physical properties challenge focused on the prediction of pKa and log D for multi-protic compounds with two or more widely separated polar groups. Here, we present an overview of the results from the SAMPL8 challenge. We adopted two different workflows to compare experimental and predicted pKa data due to the presence of multiple macroscopic pKa values for individual compounds. These multiple pKa values meant that it was often unclear which of the microscopic transitions contributed to the observed macroscopic pKa. We find that the performance of different methods for pKa predictions is heavily dependent on how we pair predicted pKa values with measured ones–that is, the underlying mechanism. In contrast to some previous challenges, where analysis focused on pairing calculated pKa values with those which were experimentally closest, we here instead use alternate approaches, presenting two different methods of evaluating pKa predictions. Our first approach uses community consensus, pairing the most popular predicted transition(s) with the experimental value. The second involves selecting the correct transition based on additional experimental data–specifically, pH-dependent solubility curves. We used these two approaches to compare SAMPL submissions from participants using a wide variety of different methods. Results showed that while the numerical predictions for pKa values were often reasonable, different participating methods might disagree about the mechanism behind the pKa, indicating that pKa prediction methods may need to focus more on correct prediction of protonation states, not just pKa values. In the log D analysis we find that most methods performed better in prediction of log D for the octanol-water solvent combination, likely due to the greater abundance of training data for these solvents.
ER -