GenPTA: Enhancing De Novo and Scaffold-Based Molecular Generation with Attention Refinement Mechanism and Adaptive Positional Encoding

Drug design is conventionally constrained by the limited chemical diversity present in existing databases. Deep learning-based molecular generation models, such as Variational Autoencoders (VAEs) and Transformers, show great promise for both de novo and scaffold-based drug design. However, these widely adopted architectures face challenges in fully leveraging their encoding space due to the loss of key molecular features during pooling. Furthermore, modeling long-range dependencies is difficult because of inherent limitations in absolute positional encoding. Here, we present GenPTA, a model that enhances both de novo and scaffold-based molecular generation through an attention refinement mechanism, adaptive positional encoding, and an innovative Memory-cat module. The attention refinement mechanism dynamically extracts the most critical molecular features, improving latent variable expressiveness and ensuring generated molecules possess the desired properties. Meanwhile, the adaptive positional encoding captures complex spatial relationships and global features. Additionally, the Memory-cat module enables spatial sampling near the encoding space of a given molecule, further diversifying generated structures and optimizing design efficiency. Experimental results show that GenPTA achieves the lowest Fréchet ChemNet Distance (FCD) values of 0.0631 ± 0.0461 and 0.4799 ± 0.0023 on classic and scaffold-based reference sets, respectively, indicating strong alignment with reference distributions. It also attains the highest Nearest Neighbor Similarity (SNN) scores of 0.6278 ± 0.0323 and 0.5891 ± 0.0007, reflecting high chemical plausibility. Comparison of GenPTA-generated molecules with the native ligand of Cytochrome P450 1B1 (CYP1B1) reveals additional interactions that stabilize the molecular-protein binding, demonstrating the model's potential to enhance affinity. GenPTA’s performance in de novo and scaffold-based generation provides an efficient approach for hit discovery and lead optimization, accelerating drug development.