SeedFold Scaling Biomolecular Structure Prediction
A next-generation folding model that scales up model capacity through width scaling and large-scale data distillation. We also provide SeedFold-Linear, a more efficient variant with linear triangular attention. Both models achieve state-of-the-art results on FoldBench, outperforming AlphaFold3 on most protein-related tasks.
923M Parameters
26.5M Training Samples
#1 on FoldBench
Model Scaling
Scale the Pairformer width from 128 to 512, increasing model capacity. Training such wide networks poses significant challenges including memory constraints and training instability—we developed engineering solutions to overcome these obstacles.
- Width scaling > Depth scaling
- 128 → 256 → 384 → 512 dimensions
- Solved training stability challenges
Linear Triangular Attention
A novel attention mechanism that reduces computational complexity from O(n³) to O(n²), enabling efficient scaling while maintaining prediction quality.
- Additive & Gated variants
- 2-3× memory reduction
- Optimized Triton kernels
Large-Scale Distillation
Construct a 26.5M sample dataset through distillation from AlphaFold2, expanding training data by 147× compared to experimental structures.
- PDB + AFDB + Mgnify
- 147× data expansion
- Diverse sequence coverage
| Task | SeedFold | SeedFold-Linear | AlphaFold 3 | Boltz-1 | Chai-1 | Protenix-0.5 |
|---|---|---|---|---|---|---|
| Monomer (lDDT) |
0.8889 | 0.8861 | 0.88 | 0.87 | 0.87 | 0.8773 |
| Protein-Protein (SR%: DockQ ≥ 0.23) |
74.03% | 74.14% | 72.93% | 68.25% | 68.53% | 71.50% |
| Ab-Ag (SR%: DockQ ≥ 0.23) |
53.21% | 46.91% | 47.90% | 33.54% | 23.64% | 41.00% |
| Protein-Ligand (SR%: lRMSD < 2Å and lDDT-PLI > 0.8) |
63.12% | 66.48% | 64.90% | 55.04% | 51.23% | 62.30% |
| Protein-RNA (SR%: DockQ ≥ 0.23) |
65.31% | 61.80% | 62.32% | 56.90% | 50.91% | 50.70% |
| Protein-DNA (SR%: DockQ ≥ 0.23) |
72.60% | 76.00% | 79.18% | 70.97% | 69.97% | 71.38% |
↑
+5.31% on Ab-Ag vs AlphaFold3
↑
+1.58% on Prot-Lig vs AlphaFold3
↑
+2.99% on Prot-RNA vs AlphaFold3
Model Architecture
SeedFold inherits the AlphaFold3 architecture with key modifications for scaling:
- MSA Module: Extracts evolutionary features from multiple sequence alignments
- Pairformer: Updates pair representations via triangular operations (scaled to 512 width)
- Structure Module: Diffusion-based all-atom structure generation
Width vs Depth Scaling
Our experiments demonstrate that width scaling of the Pairformer is the most effective strategy. The pair representation dimension is the critical bottleneck—increasing it directly enhances the model's capacity to encode complex pairwise interactions.
Base (128-width)
432M params
Medium (256-width)
533M params
Large (512-width)
923M params
Linear Triangular Attention
We propose two variants of linear triangular attention to replace the computationally expensive vanilla triangular attention:
Additive Linear TriAtt
φ(Q)φ(K)ᵀ + ψ(B)
Inherits advantages of vanilla attention with well-established linear attention designs
Gated Linear TriAtt
φ(Q)φ(K)ᵀ ⊙ ψ(B)
Gating mechanism controls information flow; superior on DNA/RNA tasks
Project Leads
*Equal contribution. Listing order is random.
Contributors
Yiming Ma
Wei Qu
Fei Ye
Kexin Zhang
Lan Wang
Minrui Gui
Technical Lead
@misc{zhou2025seedfoldscalingbiomolecularstructure,
title={SeedFold: Scaling Biomolecular Structure Prediction},
author={Yi Zhou and Chan Lu and Yiming Ma and Wei Qu and Fei Ye and Kexin Zhang and Lan Wang and Minrui Gui and Quanquan Gu},
year={2025},
eprint={2512.24354},
archivePrefix={arXiv},
primaryClass={q-bio.BM},
url={https://arxiv.org/abs/2512.24354},
}SeedProteo Accurate De Novo All-Atom Design of Protein Binders
A diffusion-based model for de novo all-atom protein design. SeedProteo repurposes cutting-edge folding architecture into a powerful generative framework, achieving state-of-the-art performance in both unconditional generation and binder design.
All-Atom Design Level
Top In-silico Success Rate
Comparison of binder design success and diversity. SeedProteo-R (Robust) and SeedProteo-D (Diverse) modes vs. baselines.
| Method | TrkA | PD-L1 | Insulin | BHRF1 | IL-7RA | SC2RBD | VEGF-A | H1 | IL-17A | TNFA |
|---|---|---|---|---|---|---|---|---|---|---|
| Ours | ||||||||||
| SeedProteo-D | 0.829 | 0.832 | 0.837 | 0.822 | 0.840 | 0.819 | 0.836 | 0.823 | 0.806 | 0.870 |
| SeedProteo-R | 0.905 | 0.913 | 0.911 | 0.872 | 0.917 | 0.858 | 0.901 | 0.890 | 0.855 | -- |
| Baselines | ||||||||||
| BindCraft | 0.849 | 0.856 | 0.864 | 0.847 | 0.861 | 0.863 | 0.850 | 0.830 | 0.818 | -- |
| PXDesign | 0.914 | 0.929 | 0.928 | 0.924 | 0.928 | 0.917 | 0.913 | 0.888 | 0.906 | -- |
| BoltzGen | 0.908 | 0.924 | 0.929 | 0.928 | 0.885 | 0.915 | 0.902 | 0.885 | 0.863 | -- |
| RFDiffusion | 0.932 | 0.934 | 0.927 | 0.946 | 0.916 | 0.912 | 0.940 | -- | 0.938 | -- |
| RFDiffusion3 | 0.808 | 0.834 | 0.876 | 0.845 | 0.840 | -- | -- | 0.930 | 0.800 | -- |
All-Atom Modeling
Directly designs proteins at the all-atom level using the atom14 schema, ensuring physically accurate structures.
- atom14 representation
- Full atomic detail
Self-Conditioning
Reuses denoised structures and predicted secondary structures to stabilize the sampling process.
- Design View integration
- Structure feedback
- Stable sampling
MRF Sequence Decoding
Uses Markov Random Field for energy minimization in sequence space, ensuring global consistency.
- Energy minimization
- Global consistency
- Optimal sequences
Design View Framework
SeedProteo mimics the architecture of AlphaFold3 and introduces a novel Design View that integrates self-conditioning features to guide the generative process.
- Left: Design View (Generative model)
- Right: Folding View (Input representation)
Unconditional monomer benchmark. Left: Design Success Rate vs Length. Right: Number of Unique Clusters vs Length.
Project Lead
Contributors
Yiming Ma†
Fei Ye
Chan Lu
Yi Zhou
Kexin Zhang†
Lan Wang
Minrui Gui†
†Work done during internship at ByteDance Seed
Technical Lead
@misc{qu2025seedproteoaccuratenovoallatom,
title={SeedProteo: Accurate De Novo All-Atom Design of Protein Binders},
author={Wei Qu and Yiming Ma and Fei Ye and Chan Lu and Yi Zhou and Kexin Zhang and Lan Wang and Minrui Gui and Quanquan Gu},
year={2025},
eprint={2512.24192},
archivePrefix={arXiv},
primaryClass={q-bio.BM},
url={https://arxiv.org/abs/2512.24192},
}