Research map / Image generation / Autoregressive and latent generation
Floating Anchor Diffusion Model for Multi-motif Scaffolding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FADiff, a diffusion model that treats motifs as rigid bodies that float freely during denoising, enabling multi-motif scaffolding without specifying relative motif positions.
Problem
Motif scaffolding for multiple motifs requires known relative sequence and structure positions of motifs, which are usually unavailable; existing inpainting and conditional generation methods either fix motif positions or cannot guarantee motif presence.
Approach
Treat each motif as a rigid anchor that translates and rotates independently during forward and reverse diffusion (updates averaged over each motif's residues), so motif positions are learned automatically and motif presence is guaranteed; sequence order is set by a TSP-like greedy chain over residue distances.
Key results
- FADiff reaches 73.05% average in silico success rate on scaffolding two motifs, versus about 16-24% for conditional generation and 51-80% for inpainting across lengths 160-360.
- Trained on two motifs, FADiff achieves 62.38%, 58.40%, and 46.67% success for 3, 4, and 5 motifs, and 80.00% for two huge domains.
- Removing the TSP sequence construction (random connection) lowers average success rate from 73.05% to 60.67%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FADiff's floating rigid-anchor diffusion instead of inpainting with randomly assigned motif positions | Two-motif scaffolding · MS Benchmark, across protein lengths | in silico success rate: 76.67, 71.67, 81.67, 65.00, 86.67 vs 51.58, 56.84, 62.11, 60.00, 79.47 | Helped |
| Use FADiff's floating rigid-anchor diffusion instead of conditional generation | Two-motif scaffolding · MS Benchmark, across protein lengths | in silico success rate: 76.67, 71.67, 81.67, 65.00, 86.67 vs 23.75, 22.50, 21.25, 23.75, 16.25 | Helped |
| Replace TSP sequence construction with random connection (ablation) | FADiff · MS Benchmark, 2 motifs, average across lengths | in silico success rate: 73.05 -> 60.67 | Hurt |
| Reduce translation noise scale by half during sampling (ablation) | FADiff · MS Benchmark, 2 motifs, average across lengths | in silico success rate: 73.05 -> 72.56 | No clear effect |
| Double translation noise scale during sampling (ablation) | FADiff · MS Benchmark, 2 motifs, average across lengths | in silico success rate: 73.05 -> 72.22 | No clear effect |
| Remove the increased translation noise scale during training (ablation) | FADiff · MS Benchmark, 2 motifs | average scTM: 0.213 | Hurt |
| Remove the increased translation noise scale during training (ablation) | FADiff · MS Benchmark, 2 motifs | average in silico success rate: close to $0\%$ | Hurt |
| Perform TSP multiple times during the first 100 sampling steps instead of once (ablation) | FADiff · MS Benchmark, 2 motifs | scTM | Helped |
Surprising findings
- Random sequence connection still achieves 60.67% average success, suggesting the motif positions are steered by the diffusion process rather than the sequence order alone.
- Without increasing translation noise scale during training, the model's average scTM drops to 0.213 and success rate is close to 0%.
Limitations
- Success rate drops as the number of motifs increases (46.67% for 5 motifs) due to fewer scaffold residues.
- Motif RMSD metric is not applicable since it is always zero, limiting direct comparison with conditional methods.
- Inpainting baselines were adapted by randomly assigning relative motif positions, which may understate their potential.
Experimental setup
Datasets: VM dataset (from PDB) MS Benchmark (from PROSITE)
Benchmarks: MS Benchmark VM dataset evaluation set
Metrics: in silico success rate scTM motif RMSD pdbTM diversity
Baselines: Inpainting (RFdiffusion-style) Conditional generation (SMCDiff-style) FrameDiff VFN-Diff
Builds on: FrameDiff VFN-Diff RFdiffusion SMCDiff
Closest papers by meaning
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.83)
- 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance (AAAI 2025 · similarity 0.83)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.82)
- Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design (ICLR 2025 · similarity 0.82)
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models (ICML 2024 · similarity 0.82)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.