Diffusion Models for Protein Structure Prediction: Benchmarking Against AlphaFold2

Diffusion-based generative models have demonstrated remarkable capabilities in image synthesis and molecule generation, prompting investigation of their applicability to protein structure prediction. We present ProtDiff, a conditional diffusion model for end-to-end protein 3D structure prediction from amino acid sequences. ProtDiff employs equivariant neural networks as the score function and conditions generation on multiple sequence alignment features. Evaluated on CASP15 targets, ProtDiff achieves mean GDT-TS of 76.8, competitive with AlphaFold2 (78.4) on domains with fewer than 300 residues, while offering 5× faster inference and the ability to generate diverse structural ensembles capturing conformational flexibility absent in AlphaFold2 predictions.