DMPE: A Dual-Branch Molecular Property Encapsulation Framework With Kolmogorov-Arnold Networks.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Accurate molecular representation is crucial for predicting drug properties.
• Existing methods may not fully capture complex molecular structures and relationships.
• Accelerating drug discovery requires efficient and reliable predictive models.

Purpose of the Study:

• To introduce the Dual-branch Molecular Property Encapsulation (DMPE) framework.
• To enhance molecular representation learning for property prediction.
• To identify potential hematopoietic progenitor kinase 1 (HPK1) inhibitors for hepatocellular carcinoma.

Main Methods:

• Utilizing a Refined Interactive Graph Attention Framework (RIGAF) to capture intramolecular and intermolecular features.
• Employing a Kolmogorov-Arnold Network (KAN)-based Embedding and Fusion (KAEF) module for integrating graph features and molecular fingerprints.
• Implementing Monte Carlo dropout for uncertainty estimation and ensemble strategies.

Main Results:

• Achieved superior ROC-AUC scores on BBBP (0.927) and SIDER (0.691) benchmarks.
• Demonstrated promising accuracy across nine of 14 breast cancer cell lines.
• Validated key architectural components through ablation studies.

Conclusions:

• The DMPE framework offers competitive performance in molecular property prediction.
• The KAEF module enhances interpretability and generalizability.
• DMPE shows potential for identifying novel drug candidates, exemplified by its application to HPK1 inhibitors.