FSM-DDTR: End-to-end feedback strategy for multi-objective De Novo drug design using transformers

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Drug discovery faces challenges due to polypharmacology and the need for selective compounds.
• Existing deep learning methods for de novo drug design often neglect crucial properties like validity and target selectivity.
• The multi-objective nature of pharmacological space requires sophisticated generative approaches.

Purpose of the Study:

• To develop a multi-objective Transformer-based architecture for generating drug candidates with enhanced molecular properties and target selectivity.
• To address limitations in current in silico drug design methods by incorporating multi-objective optimization.
• To generate novel, valid, and selective compounds for specific biological targets.

Main Methods:

• A Transformer-Decoder Generator was used to create novel compounds in SMILES format.
• A Transformer-Encoder Predictor estimated binding affinity to the target.
• A feedback loop and multi-objective optimization strategy ranked molecules and guided generation.
• The Adenosine A2A Receptor (AA2AR) was used as a relevant biological target for validation.

Main Results:

• The Transformer-based Generator achieved a 97.38% novelty rate, outperforming baselines.
• Generated molecules demonstrated high binding affinity to AA2AR and adhered to Lipinski's rule of five (99.36%).
• Multi-objective optimization successfully shifted molecular properties towards desired drug-like characteristics without prior specialized training sets.

Conclusions:

• The proposed Transformer-based architecture is effective for de novo drug design, generating novel molecules with improved pharmacological properties and target selectivity.
• This approach enables efficient exploration of chemical space for identifying potential drug leads.
• The study validates the use of AI for designing selective and synthesizable drug candidates.