Multimodal data fusion for supervised learning-based identification of USP7 inhibitors: a systematic comparison

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Ubiquitin-specific-processing protease 7 (USP7) is a key target for cancer therapy.
• Developing USP7 inhibitors is challenging due to limitations in traditional virtual screening accuracy.

Purpose of the Study:

• To build accurate supervised learning models for identifying USP7 inhibitors.
• To evaluate different molecular representations and machine learning models for this task.
• To provide guidance on handling imbalanced datasets in drug discovery.

Main Methods:

• Calculated physicochemical descriptors, MACCS keys, ECFP4 fingerprints, and SMILES for compounds.
• Constructed two deep learning (DL) and nine classical machine learning (ML) models.
• Implemented 75 experiments across 15 groups using various molecular representations and activity cutoffs.

Main Results:

• Ensemble learning models performed optimally on balanced and severely imbalanced datasets.
• SMILES-based DL models excelled on slightly imbalanced datasets.
• Data fusion, SMOTE, unbiased decoy selection, and SMILES enumeration improved model performance, particularly for imbalanced data.

Conclusions:

• Highly accurate supervised learning models were established to accelerate USP7 inhibitor development.
• Guidance is provided for selecting models, molecular representations, and methods for imbalanced datasets in drug discovery research.