Efficient virtual high-content screening using a distance-aware transformer model.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Molecular similarity search is crucial for drug discovery and virtual screening.
• Traditional 2D similarity metrics are fast, but complex 3D methods are computationally expensive for large databases.

Purpose of the Study:

• To develop a transformer model for efficient molecular similarity search.
• To reduce the search space in virtual screening while maintaining accuracy.

Main Methods:

• Trained a transformer model to autoencode tokenized SMILES strings.
• Utilized a custom loss function to conserve molecular similarities in latent space.
• Reduced molecular similarity to Euclidean distance in the generated latent space.

Main Results:

• The model reduced the search space by 5 orders of magnitude on a 1.5 billion molecule database.
• A custom loss function improved the prediction of highly similar molecules.
• The method demonstrated adequate generalization from a small dataset.

Conclusions:

• The developed model substantially reduces the relevant search space in virtual screening.
• This approach significantly increases the throughput of drug discovery processes.
• The method's efficiency is independent of the underlying similarity metric used.