DeepTargetClass: a web-based platform for predicting protein target classes of small molecules.

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacology and drug discovery
• Machine learning in bioinformatics

Background:

• Identifying protein target classes is crucial for efficient drug discovery and repurposing.
• Existing computational methods require robust and interpretable models for accurate classification.

Purpose of the Study:

• To develop a deep-learning pipeline for predicting pharmacological protein target classes.
• To create an accessible tool for classifying novel compounds based on their protein targets.

Main Methods:

• A multilayer perceptron (MLP) model was trained on 15,804 compounds using extended connectivity fingerprints (ECFP4).
• Model performance was evaluated using internal cross-validation and an external test set.
• SHAP values were employed for model interpretability, highlighting key substructures.

Main Results:

• The MLP model achieved 96% accuracy in internal cross-validation and 87% on an external test set.
• Predictions were robust and balanced across four major target classes: GPCRs, kinases, nuclear receptors, and transporters.
• Model interpretability revealed pharmacophore-like substructures, consistent with known ligand-target interactions.

Conclusions:

• The developed deep-learning framework provides a reliable and interpretable method for protein target class prediction.
• The MLP model demonstrates performance comparable to ensemble methods and is validated by its application to reference drugs.
• A user-friendly web application facilitates accessible protein class prediction, supporting drug discovery and repositioning.