SeedFold
paperA next-generation biomolecular folding model that scales up model capacity and performance beyond AlphaFold3. SeedFold introduces a width-scaling strategy for the Pairformer and a Linear Triangular Attention mechanism that reduces computational complexity from $O(n^3)$ to $O(n^2)$. Trained on a massive distillation dataset of 26.5 million samples, SeedFold outperforms AlphaFold3 on protein monomers, antibody-antigen interfaces, and protein-RNA interface predictions.
Paper
arXiv: 2512.24354