Deep Learning-Based Ligand Design Using Shared Latent Implicit Fingerprints from Collaborative Filtering.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Implicit fingerprints map ligands and proteins into a shared latent space for virtual screening.
• Existing methods often rely on fingerprint similarity searches, introducing potential empirical bias.

Purpose of the Study:

• To extend implicit fingerprints using deep learning for de novo molecule design.
• To translate latent descriptors into discrete molecular representations (SMILES) encoding druglike properties.
• To design new compounds based on protein latent representations, predicting binding affinities.

Main Methods:

• Utilized deep learning to translate latent descriptors into SMILES representations.
• Generated novel compounds by leveraging the latent space of proteins.
• Evaluated generated compounds for chemical feasibility, druglike properties, and binding affinity using predictive models.
• Assessed compound diversity using Tanimoto distance.

Main Results:

• Generated compounds exhibit chemical feasibility and druglike properties.
• Predicted excellent binding affinities to known proteins for the designed molecules.
• Demonstrated wide diversity among the generated compounds using Tanimoto distance.

Conclusions:

• The deep learning approach successfully generates novel, chemically feasible drug candidates.
• The method encodes druglike properties and predicts high binding affinities without fingerprint bias.
• The generated compounds show significant diversity, supporting broad drug discovery applications.