Enhancing latent cross-attention-based diffusion models using graph neural networks for generating pocket-aware and target protein-specific novel therapeutic peptide designs

Abstract

While existing therapies ultimately face challenges, therapeutic peptides are designed to produce targeted biological effects, minimising off-target side effects and providing more tailored and cost-effective treatment options for a wide range of pathologies. However, designing them can be challenging due to a multitude of reasons, like the scarcity of peptide data. Several Deep Learning methods have been applied to reduce development costs and time. Nevertheless, these methods either require large datasets or struggle to incorporate the functional information necessary to generate therapeutic peptides that are structurally and functionally complementary to the target.

In our research project, we introduced a new method for generating novel, pocket-aware, target protein-specific peptide sequences by building upon the work of Sayuti et al. [36]. Our approach limits reliance on prior knowledge by integrating a latent diffusion model-centric architecture with an E(3)-invariant structure representation of the binding pocket extracted using a Graph Attention Network (GAT). The binding-site representation allows us to embed critical functional and structural information of the target protein’s pocket. Whereas, by introducing binding-site embeddings as constraints in the Cross-Attention block of the architecture, this approach allows us to indirectly capture multi-modal information and incorporate target specificity. The validity of our approach is demonstrated by its superior performance on evaluation metrics, outperforming those reported by Sayuti et al. [36], as well as models based on Variational Autoencoders (VAE) and Wasserstein Autoencoders (WAE). The results of the study underscore the importance of incorporating binding-site information in generative models to design peptides that are not only structurally and functionally relevant but also more likely to interact effectively with target proteins.