Combining a Chemical Language Model and the Structure-Activity Relationship Matrix Formalism for Generative Design of...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Optimizing drug compounds requires generating analogue series (AS) to understand structure-activity relationships (SARs).
• Identifying the most promising analogues for synthesis is a key challenge in compound optimization.
• Potency progression is a critical metric for advancing AS in drug development.

Purpose of the Study:

• To introduce a novel computational methodology for extending analogue series (AS) with potent compounds.
• To enable modifications at multiple sites, including core structure and substituents.
• To address the challenge of selecting which analogues to synthesize next in medicinal chemistry.

Main Methods:

• The approach integrates a transformer chemical language model (CLM) with an expanded SAR matrix (SARM) methodology for multisite AS.
• SARMs were utilized to identify and organize structurally related AS, with consensus series from potency gradients used for CLM training.
• Various model variants were developed and evaluated for their predictive capabilities.

Main Results:

• The developed models successfully predicted known potent analogues, ranking them highly in probability-based compound lists.
• The methodology achieved chemical diversification of AS through core structure modifications and multi-site substituent replacements.
• Both general and fine-tuned CLM variants demonstrated effectiveness in generating potent and diverse analogues.

Conclusions:

• The new computational approach effectively extends analogue series with potent compounds, aiding medicinal chemistry optimization.
• This method facilitates the exploration of structure-activity relationships by suggesting diverse and potent analogues.
• The integration of CLM and SARM offers a powerful tool for accelerating drug discovery and development.