Roughness of Molecular Property Landscapes and Its Impact on Modellability

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Molecular property landscapes are crucial for drug discovery and chemical space navigation.
• Landscape roughness is linked to activity cliffs and optimization difficulty.
• Existing methods for assessing roughness are limited.

Purpose of the Study:

• To introduce a general, quantitative measure for molecular property landscape roughness.
• To provide a tool for better understanding and navigating chemical space.
• To correlate landscape roughness with machine learning model performance.

Main Methods:

• Developed a novel roughness index (ROGI) inspired by fractal dimension.
• Applied ROGI to various molecular property landscapes.
• Correlated ROGI with out-of-sample errors of machine learning regression models.

Main Results:

• The proposed roughness index (ROGI) provides a quantitative assessment of landscape roughness.
• ROGI strongly correlates with the out-of-sample error of machine learning models across multiple regression tasks.
• Higher roughness, as measured by ROGI, indicates greater optimization challenges.

Conclusions:

• ROGI offers a valuable new metric for characterizing molecular property landscapes.
• This index can guide drug discovery efforts by predicting optimization challenges.
• ROGI's correlation with machine learning error suggests its utility in model development and validation.