QSAR with experimental and predictive distributions: an information theoretic approach for assessing model quality

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for predicting drug properties.
• Current QSAR predictions often lack explicit uncertainty quantification.
• Accurate uncertainty estimation is vital for reliable drug development decisions.

Purpose of the Study:

• To represent QSAR predictions as probability distributions for robust uncertainty quantification.
• To evaluate machine learning algorithms and error estimation methods for generating predictive distributions.
• To assess the utility of predictive distributions in estimating compound property profiles.

Main Methods:

• Utilized Kullback-Leibler (KL) divergence to measure the quality of predictive distributions against experimental data.
• Assessed various machine learning algorithms and error estimation techniques on AstraZeneca's DMPK datasets.
• Developed reliability indices to associate predictive distribution tightness with prediction reliability.

Main Results:

• Identified specific algorithm and error estimation combinations that yield accurate and valid compound-specific predictive distributions.
• Demonstrated that KL divergence effectively quantifies the quality of QSAR predictive distributions.
• Showcased the use of reliability indices to differentiate between high- and low-confidence predictions.

Conclusions:

• Representing QSAR predictions as probability distributions enhances model interpretability and reliability.
• The KL-divergence framework provides a robust method for evaluating predictive distribution quality.
• Validated predictive distributions can inform the probability of a compound meeting specific target profiles, aiding in compound selection.