Predicting protein network topology clusters from chemical structure using deep learning.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Inferring protein targets and functions from chemical structures is crucial but often limited by relying solely on chemical similarity.
• Existing methods may misinterpret functional relationships based on molecular structure alone.

Purpose of the Study:

• To develop and evaluate a novel deep neural network methodology for predicting target protein clusters.
• To improve the accuracy of inferring protein targets and functions by integrating network topology with interaction data.

Main Methods:

• A deep learning model was trained on compound clusters derived from combined compound-protein and protein-protein interaction data.
• Network topology analysis was employed to calculate similarities for model training.
• Several deep learning architectures, including convolutional and recurrent neural networks, were compared.

Main Results:

• The recurrent neural network architecture, MolPMoFiT, demonstrated superior performance, achieving an F1 score near 0.9 on a test set of 8907 compounds.
• In-depth analysis confirmed the method's predictive accuracy for most well-studied compounds.
• The approach successfully distinguished between compounds with similar structures but different functions, highlighting its advantage over simple chemical similarity.

Conclusions:

• Deep neural networks, particularly the MolPMoFiT architecture, offer a robust methodology for predicting target protein clusters.
• This approach enhances the accuracy of inferring protein function and targets beyond traditional chemical similarity metrics.
• The findings have significant implications for drug discovery and understanding chemical biology.