Next-Generation Molecular Design - Structure-Based Generative AI with groupGen and Ligand AIDE

AI Ligand Decoration using GroupGen

groupGen is a structure-aware generative neural network designed specifically for R-group optimization and ligand decoration during lead optimization.

Starting with a ligand containing a defined attachment point, groupGen analyzes the surrounding protein binding pocket and generates chemically reasonable substituents that are tailored to the available 3D space.

Structure-Aware Generation

Unlike approaches that generate substituents primarily from chemical patterns or learned molecular statistics, groupGen explicitly considers the 3D environment surrounding the attachment point. The model learns how different chemical groups can interact with the local binding site and uses this information to propose new R-groups. The neural network captures important properties of the local pocket, including:

• Pocket geometry and available 3D space
• Electrostatic environment
• Hydrophobic and hydrophilic regions
• Shape complementarity
• Interactions with nearby protein residues

An encoder based on a 3D graph convolutional neural network converts this local structural information into a latent representation of the binding environment. A generative decoder then uses this representation to construct chemically valid substituent structures. groupGen is particularly useful when a promising ligand has already been identified and the goal is to explore modifications around one or more positions on the molecule.

For example, a chemist can retain the core of a known ligand while asking groupGen to explore alternative substituents at a selected position. The resulting compounds can then be docked and scored to identify modifications that may improve binding affinity, interactions, or other desired properties.

This provides a practical way to explore chemical space while maintaining the important binding interactions of an existing lead.

Ligand AIDE - AI Design Evolution

Ligand AIDE is a de novo ligand generation workflow based on Artificial Intelligent Design Evolution. It uses an iterative evolutionary strategy to grow and optimize ligands directly in the context of a protein binding site.

The process begins with a population of docked small fragments. From this starting point, new designs are generated by adding or replacing R-groups using the groupGen neural network, as well as by performing atom-level substitutions using the Atom Predictor. Each newly generated cohort of compounds is then re-docked into the binding site.

Designs are refined through multiple evolutionary cycles, typically three to five iterations. At each step, a Darwinian selection process removes poor designs based on several criteria, including RTCNN scores, physics-based binding scores with penalty terms, drug-like property filters, and synthesizability.

Through successive generations, fragments evolve into more complete and optimized ligands. This iterative grow-and-select strategy enables efficient exploration of chemical space while maintaining binding quality, drug-like behavior, and practical synthetic feasibility.

• Iterative Growth: Starting from small docked fragments or crystal structures, the tool adds or replaces chemical groups using groupGen and an Atom Predictor over 3 to 5 generations.
• Darwinian Selection: At each step, a selection process removes poor designs based on binding quality, drug-like properties, and synthesizability.
• Synthesis Constraints: Mandatory synthesizability cutoffs ensure that the resulting molecules are practical for laboratory synthesis.