SeedFold and SeedProteo (seedfold.github.io)

A team has introduced two pioneering protein folding models, SeedFold and SeedProteo, which leverage advanced architecture modifications and extensive datasets to set new benchmarks in biomolecular structure prediction and protein design. SeedFold achieves remarkable performance by scaling the Pairformer width and implementing an innovative attention mechanism that reduces computational complexity, allowing it to outperform AlphaFold3 in various protein-related tasks. Notably, it utilizes a 26.5 million sample dataset distilled from AlphaFold2, enhancing training data size by 147 times and significantly improving model capabilities. SeedProteo repurposes the foundational architecture of SeedFold into a generative framework focused on protein design. It excels in generating all-atom structures and binder designs, surpassing current open-source models by demonstrating superior novelty and performance across multiple benchmark targets. The model incorporates advanced techniques such as Markov Random Field for energy minimization and a unique Design View that utilizes self-conditioning features, enabling the generation of physically accurate protein structures of substantial size. Together, these innovations signify a substantial leap forward in both predictive accuracy and efficiency within the AI/ML community, paving the way for more effective drug design and biomolecular research.