Benchmarks for interpretation of QSAR models.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Interpreting Quantitative Structure-Activity Relationship (QSAR) models is crucial for understanding biological and physicochemical processes.
• Highly predictive models, especially neural networks, are complex and challenging to interpret.
• Evaluating and comparing emerging model interpretation methods is difficult due to a lack of standardized benchmarks.

Purpose of the Study:

• To develop benchmark datasets with pre-defined patterns for evaluating the performance of QSAR model interpretation approaches.
• To propose quantitative metrics for assessing the effectiveness of interpretation methods.
• To provide a framework for comparing and validating new techniques for interpreting complex machine learning models.

Main Methods:

• Creation of benchmark datasets with varying complexity, from simple additive properties to pharmacophore hypotheses.
• Development of quantitative metrics to measure the performance of interpretation approaches.
• Application of a universal ML-agnostic interpretation method to conventional models and graph convolutional neural networks.

Main Results:

• Demonstrated the utility of the developed benchmarks and metrics on diverse QSAR models.
• Successfully evaluated the ability of interpretation approaches to retrieve pre-defined patterns.
• Showcased the application of benchmarks to complex "black box" models like graph convolutional neural networks.

Conclusions:

• The developed benchmarks and metrics facilitate the rigorous evaluation of QSAR model interpretation methods.
• These resources will aid in understanding the decision-making processes of complex machine learning models.
• The study provides a foundation for advancing the field of interpretable machine learning in cheminformatics.