Advancing De Novo Drug Design with Geometric Transformer Models

Researchers have developed a framework that maps protein pocket geometry to valid chemical structures, overcoming the instability of previous reinforcement learning models. Historically, this has been difficult because models either ignore the global shape of the binding pocket or produce strings that represent chemically impossible molecules.

Refining De Novo Drug Design via Geometric Precision

Current structure-based de novo drug design often relies on SMILES strings, which are prone to syntax errors, or graph networks that fail to capture the macro-scale topology of an active site. This new method, GPS-VAE, integrates local graph attention with global transformer self-attention to extract physicochemical features. By replacing RNN-based SMILES with a Transformer-SELFIES architecture, the researchers ensured 100% chemical validity in their generated outputs. The study tested this framework on Janus Kinase 2 (JAK2) and Dopamine D2 Receptor (DRD2) targets. The system identified macrocyclic adaptations and lead-like scaffolds with predicted ligand efficiency scores exceeding 0.5 under AutoDock Vina evaluations. These results suggest that anchoring protein features directly into a chemical latent space allows for more efficient exploration of viable drug candidates than traditional random search. The integration of the STONED evolutionary algorithm allowed for structural refinement within the latent space, further optimising the candidates for specific binding affinities. This multi-step pipeline represents a shift from simple generative pattern matching to a more rigorous, physics-aware methodology. However, the study does not solve the inherent inaccuracy of docking scores, which often fail to mirror actual in vitro binding affinity.