A groundbreaking study introduces Ambient Protein Diffusion, a novel framework that leverages low-confidence AlphaFold structures as noisy data to train protein diffusion models. This approach significantly enhances the diversity and quality of generated protein structures, outperforming previous state-of-the-art models.
The key innovation lies in treating low-confidence AlphaFold predictions as corrupted data. Instead of filtering them out, the method adjusts the diffusion objective based on the corruption level of each structure, allowing the model to learn from both high and low-quality data. This results in major improvements in diversity and designability for long proteins