A geometry-aware generative framework integrating GPS-VAE and Transformer-SELFIES for structure-based de novo drug...

Designing novel ligands tailored to specific protein binding pockets remains a core objective in structure-based de novo drug design (SBDD). However, deep generative approaches encounter key challenges: standard graph neural networks fail to capture global pocket topology, SMILES-based models generate chemically invalid structures, and reinforcement learning remains unstable in multi-objective optimization. We evaluated our proposed framework on Janus Kinase 2 (JAK2) and Dopamine D2 Receptor (DRD2) targets. Compared to baseline models, our GPS-VAE successfully captured complex geometric dependencies, achieving robust active site representation and topological reconstruction. For molecular generation, we identified fragment-like and lead-like scaffolds demonstrating high predicted ligand efficiency (LE > 0.5) under the AutoDock Vina scoring function, discovering interesting macrocyclic adaptations targeting JAK2. We demonstrate that high-quality data representation combined with evolutionary search significantly enhances the efficiency of de novo drug design.

We developed a new generative framework combining geometric deep learning and evolutionary search. First, we built a graph interaction transformer variational autoencoder (GPS-VAE) utilizing local graph attention networks and global transformer self-attention to extract physicochemical and geometric features. Second, we employed a Transformer-SELFIES autoencoder to replace the RNN-SMILES architecture, guaranteeing 100% chemical validity. Finally, we designed a variational projection network to anchor protein features into the chemical latent space, followed by structural refinement using the STONED evolutionary algorithm. Molecule preprocessing, docking, and fitness evaluations were performed using OpenBabel and AutoDock Vina.