Researchers led by Prof. LIU Qi from the University of Science and Technology of China (USTC) of the Chinese Academy of Sciences, collaborating with Marinka Zitnik's lab from Harvard Medical School, developed a novel deep generative model, PocketGen. This model, based on graph representation learning and protein language models, efficiently generates protein pocket sequences and spatial structures for binding small molecules. The study was published in Nature Machine Intelligence.

Functional protein design, particularly for proteins binding to small molecules such as enzymes and biosensors, is crucial for drug discovery and biomedical applications. Traditional methods based on energy optimization and template matching are time-consuming and yield low success rates. Meanwhile, deep learning models face challenges in modeling complex molecular-protein interactions and capturing sequence-structure dependencies. PocketGen addresses these issues, offering a high-efficiency and high-accuracy solution that adheres to physicochemical principles.

PocketGen builds on previous works FAIR and PocketFlow and consists of two core components. First is a dual-layer graph transformer encoder inspired by proteins’ hierarchical structures. This module is designed to learn different fine-grained interaction information and to update the representations and spatial coordinates of amino acids and atoms accordingly. The other part is a pre-trained protein language model where PocketGen efficiently fine-tunes the ESM2 model to assist in amino acid sequence prediction. By selectively adapting certain parameters, PocketGen enhances sequence-structure consistency through cross-attention mechanisms.

Experimental results demonstrated that PocketGen significantly outperforms traditional methods in affinity, structural plausibility, and computational efficiency, achieving over a 10-fold improvement in speed. Furthermore, its effectiveness in tasks such as protein pocket design for small molecules like fentanyl and ibuprofen was confirmed through comparisons with state-of-the-art generative models including RFDiffusion and RFDiffusionAA developed by Nobel Laureate David Baker’s lab. In addition, the attention matrices generated by PocketGen were compared with results from first-principle-based force field simulations, demonstrating the good interpretability of this deep learning-based model.

This study advances the application of deep generative models in functional protein design, laying a foundation for further biological experimentation and offering valuable insights into protein design principles. It also highlights the potential of AI to address challenges in drug discovery and bioengineering.