Predicting Binding from Screening Assays with Transformer Network Embeddings.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Cheminformatics utilizes molecular properties for chemistry applications.
• Deep learning and large chemical datasets accelerate computer-aided drug discovery.
• Virtual high-throughput screening (vHTS) faces limitations in binding affinity prediction due to scarce ground-truth data.

Purpose of the Study:

• To develop a method for single-target binding affinity prediction using text representations of molecules.
• To leverage transfer learning with pretrained transformer models for improved binding prediction.
• To analyze the organization of molecular properties within the embedding space.

Main Methods:

• Utilized text representations of 83 million molecules for binding affinity prediction.
• Employed an end-to-end transformer neural network trained on a text-based translation task.
• Repurposed the transformer's embedding via transfer learning for binding affinity classification.
• Quantified performance improvement using Area Under the Curve (AUC) on binding prediction tasks.

Main Results:

• Achieved successful single-target binding affinity prediction using molecular text representations.
• Demonstrated improved prediction accuracy by using pretrained transformer embeddings compared to untrained embeddings.
• Visualized the embedding space, revealing organization of structural and functional properties relevant to binding.

Conclusions:

• Transformer-based embeddings effectively capture molecular characteristics for binding affinity prediction.
• Transfer learning significantly enhances the performance of binding prediction models, especially with limited data.
• The developed method and resources offer a valuable tool for accelerating drug discovery processes.