Unveiling optimal molecular features for hERG insights with automatic machine learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Accurate prediction of compound properties is crucial for efficient drug discovery.
• Existing methods often struggle with novel chemical scaffolds and generalizability.
• Integrating diverse molecular representations can potentially improve predictive model performance.

Purpose of the Study:

• To develop a universal framework (MaxQsaring) for accurate compound property prediction.
• To evaluate MaxQsaring's performance on hERG blockage prediction and general benchmarks.
• To assess the contribution of deep learning representations and model interpretability.

Main Methods:

• Developed MaxQsaring, a framework integrating molecular descriptors, fingerprints, and deep learning representations.
• Applied MaxQsaring to human ether-à-go-go-related gene (hERG) blockage prediction on external datasets.
• Utilized automatic optimal feature combinations and identified interpretable features.
• Benchmarked performance on Therapeutics Data Commons (TDC) tasks.

Main Results:

• MaxQsaring achieved state-of-the-art performance in hERG blockage prediction.
• Identified 10 important interpretable features for decision tree modeling.
• Models demonstrated interpretability aligning with empirical optimization strategies.
• Ranked first in 19 out of 22 tasks in the TDC benchmarks.
• Deep learning representations showed moderate performance enhancement but limited generalizability improvement.

Conclusions:

• MaxQsaring offers a universal and accurate approach for compound property prediction.
• The framework facilitates early drug discovery by improving prediction accuracy and success rates.
• Model interpretability aids in understanding structure-activity relationships for practical applications.
• While deep learning representations offer benefits, further research is needed for novel scaffold generalizability.