Unsupervised Representation Learning for Proteochemometric Modeling.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• In silico protein-ligand binding prediction is crucial for efficient drug discovery.
• Proteochemometric modeling (PCM) uses protein and ligand descriptors but often relies on handcrafted features.
• Existing methods can be limited by alignment requirements or protein-specific descriptors.

Purpose of the Study:

• To evaluate the utility of unsupervised representation learning for protein-ligand binding prediction.
• To compare the performance of various unsupervised protein and compound embeddings against handcrafted features.
• To demonstrate the superiority of learned representations in modeling protein-ligand interactions.

Main Methods:

• Utilized three different unsupervised protein embedding methods.
• Employed two distinct unsupervised compound embedding methods.
• Evaluated performance using benchmark and internal datasets with various data splits.

Main Results:

• Unsupervised-learned representations significantly outperformed handcrafted features.
• Demonstrated the effectiveness of language-modeling-based embeddings in capturing protein-ligand interactions.
• Showcased improved accuracy in predicting binding activities.

Conclusions:

• Unsupervised representation learning offers a powerful approach for enhancing protein-ligand binding prediction.
• Learned embeddings provide more effective and generalizable representations compared to traditional methods.
• This advancement can accelerate computational drug discovery by reducing time and resources.