Clustering Diffusion Model for Improving Protein Structure Prediction Performance

Since protein structure prediction is very important for biochemical studies, researchers have developed many deep learning-based AI models to predict protein structure using sequence information only. In order to further enhance the performance of deep learning-based AI system for protein structure prediction, diverse and realistic new protein samples need be generated. These protein samples can provide additional patterns to enhance training of the computational system. Traditionally, researchers use the single generative model to explore the complex data distribution from the whole dataset, which is very challenging task. In order to overcome this problem, we propose Clustering Diffusion Model (CDM) which organizes multiple Diffusion Models (DMs) in the tree. Each DM in the tree can learn the unique local data distribution for one protein subfamily instead of attempting to understand the data distribution of the entire dataset. Consequently, CDM has great potentials to generate more diverse and reasonable protein samples than the single generative model. Experimental results demonstrate that CDM can further enhance the performance of protein structure prediction as compared to the single generative model while the construction time of CDM is reduced due to the parallel strategy.