Application of ALOGPS 2.1 to predict log D distribution coefficient for Pfizer proprietary compounds

Area of Science:

• Computational chemistry
• Drug discovery and development

Background:

• Accurate prediction of the distribution coefficient (log D) is crucial for drug discovery.
• Existing computational tools like ALOGPS, ACD Labs LogD, and PALLAS PrologD showed limitations in log D prediction accuracy.
• High root-mean-squared error (RMSE) of 1.0-1.5 log was observed for these tools on large in-house datasets.

Purpose of the Study:

• To evaluate the accuracy of ALOGPS, ACD Labs LogD, and PALLAS PrologD for log D coefficient prediction.
• To identify the factors limiting the accuracy of log D estimation by these algorithms.
• To assess the impact of the ALOGPS self-learning feature (LIBRARY mode) on log D prediction accuracy.

Main Methods:

• Evaluation of three log D prediction suites: ALOGPS, ACD Labs LogD, and PALLAS PrologD.
• Utilized two large in-house Pfizer datasets comprising 17,861 and 640 compounds.
• Investigated the role of log P prediction accuracy in overall log D estimation.
• Assessed the performance enhancement offered by the ALOGPS LIBRARY mode.

Main Results:

• Standard implementations of ALOGPS, ACD Labs LogD, and PALLAS PrologD exhibited high RMSE (1.0-1.5 log) for log D prediction.
• Inaccuracy in predicted octanol-water partition coefficient (log P) was identified as a key limitation.
• The ALOGPS LIBRARY mode demonstrated a significant improvement in log D prediction accuracy.
• The enhanced ALOGPS LIBRARY mode achieved an RMSE of 0.64-0.65 for both datasets.

Conclusions:

• Commercial log D prediction software requires improvement, particularly in log P estimation.
• The ALOGPS LIBRARY mode offers a more accurate approach to log D prediction.
• Self-learning algorithms show promise for enhancing computational drug discovery tools.