KAN-based symbolic QSPR modeling for molecular property prediction

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Molecular property prediction faces a trade-off between accuracy and interpretability.
• Graph neural networks (GNNs) offer high accuracy but lack interpretability.
• Traditional Quantitative Structure-Property Relationship (QSPR) models are interpretable but cannot capture complex nonlinear relationships.

Purpose of the Study:

• To propose a novel QSPR framework based on Kolmogorov-Arnold Networks (KANs).
• To bridge the gap between highly accurate but uninterpretable GNNs and interpretable but limited traditional QSPR models.
• To achieve both high predictive accuracy and model interpretability in molecular property prediction.

Main Methods:

• Developed a QSPR framework utilizing Kolmogorov-Arnold Networks (KANs).
• Employed explicit molecular descriptors as input features.
• Incorporated learnable B-spline activation functions to model nonlinear relationships.
• Extracted symbolic equations from the trained KAN model.

Main Results:

• KAN achieved competitive performance compared to GNNs on eight benchmark datasets.
• KAN models required an average of only six explicit molecular descriptors.
• Extracted symbolic equations demonstrated superior fitting ability over traditional linear QSPR formulas.
• Learned functional forms aligned with established chemical principles.

Conclusions:

• KAN presents a viable solution to the accuracy-interpretability trade-off in molecular property prediction.
• The KAN framework offers a powerful alternative to black-box deep learning models and simplistic linear QSPR approaches.
• This method enhances our ability to understand and predict molecular properties through interpretable symbolic equations.