ROD-WGAN hybrid: A Generative Adversarial Network for Large-Scale Protein Tertiary Structures

The tertiary structures of proteins play a critical role in determining their functions, interactions, and bonding in molecular chemistry. Proteins are known to demonstrate natural dynamism under various physiological conditions, which enables them to adjust their tertiary structures and effectively interact with the surrounding molecules. The present study utilized the remarkable progress made in Generative Adversarial Networks (GANs) to generate tertiary structures that accurately mimic the inherent attributes of actual proteins, which includes the backbone conformation as well as the local and distal characteristics of proteins. The current study has introduced a robust model, ROD-WGAN hybrid, that is able to generate large-scale tertiary protein structures that greatly mimic those found in nature. We have made several noteworthy contributions in pursuit of this objective by integrating the ROD-WGAN model with a hybrid function that incorporates adversarial loss, perceptual loss, distance-by-distance loss, and structural similarity loss. Through this innovative approach, we achieved remarkable results, particularly in the generation of protein structures of considerable length. The ROD-WGAN hybrid model has transcended the limitations inherent in the ROD-WGAN model, demonstrating its capability in the successful synthesis of high-quality proteins, characterized by a length of up to 256 amino acids. This accomplishment serves as a demonstrating the effectiveness and potential of our proposed methodologies. Furthermore, these generated protein structures serve as a valuable resource for data augmentation in crucial applications such as molecular structure prediction, inpainting, dynamics, and drug design. Interested individuals can access the data, code, and trained models at https://github.conmena01/ROD-WGAN-and-ROD-WGAN-hybird-models.