Automated design of multi-target ligands by generative deep learning

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Generative deep learning models, including chemical language models (CLM), are powerful tools for data-driven de novo molecular design.
• CLM trained on molecular string representations like SMILES have demonstrated success in designing novel chemical entities with confirmed activity against specific targets.

Purpose of the Study:

• To investigate the application of CLM for the de novo design of multi-target ligands, enabling designed polypharmacology.
• To leverage CLM's ability to learn from small fine-tuning datasets to bias molecular design towards drug-like properties and similarity to known ligands of target pairs.

Main Methods:

• Fine-tuning CLM on small datasets of known ligands for specific target pairs.
• Generating novel molecular designs predicted to be active on multiple protein targets.
• Synthesizing and experimentally testing computationally favored CLM designs.

Main Results:

• Generated molecules were predicted to be active on both target proteins and incorporated pharmacophore elements from ligands of both targets.
• Experimental validation of twelve CLM designs showed modulation of at least one intended protein, with potencies up to double-digit nanomolar.
• Seven compounds were confirmed as designed dual ligands, demonstrating successful multi-target ligand generation.

Conclusions:

• Chemical language models are effective for the multi-target de novo design of ligands.
• This approach serves as a significant source of innovation in drug discovery for developing polypharmacological agents.