Research map / Probabilistic and scientific modeling / Molecules and proteins
Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation
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.
Introduces FoldFlow++, a sequence-conditioned SE(3) flow matching model for protein backbone generation, trained on PDB plus filtered AlphaFold2 SwissProt structures, with ReFT fine-tuning for conditional design.
Problem
Protein structure generative models are unconditional or lack sequence conditioning, limiting de novo design for drug discovery tasks such as folding, motif scaffolding, and nanobody scaffolding.
Approach
Condition an SE(3)^N flow-matching model on sequence embeddings from a frozen ESM2-650M, fuse structure and sequence representations in a multi-modal trunk, decode with an IPA geometric decoder, train with random 50% sequence masking, and align outputs with a filter-based SFT objective (ReFT) using auxiliary rewards.
Key results
- Unconditional generation: FoldFlow++ reaches 0.976 designability (scRMSD<2Å) vs 0.969 for RFDiffusion, with novelty fraction 0.368 vs 0.116 and pairwise TM 0.205 vs 0.256.
- Motif scaffolding: fine-tuned FoldFlow++ solves 24/24 benchmark motifs and 9/25 VHH motifs vs 5/25 for RFDiffusion.
- Folding on 268 PDB proteins: FoldFlow++ RMSD 3.237 vs ESMFold 2.322 and MultiFlow 14.995.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FoldFlow++ instead of RFDiffusion | Unconditional protein generation · Unconditional generation | Designability (Frac. <2Å): 0.969 $\pm$ 0.023 -> 0.976 $\pm$ 0.010 | Helped |
| use FoldFlow++ instead of RFDiffusion | Unconditional protein generation · Unconditional generation | Novelty (Frac. TM < 0.3): 0.116 $\pm$ 0.020 -> 0.368 $\pm$ 0.031 | Helped |
| use FoldFlow++ instead of RFDiffusion | Unconditional protein generation · Unconditional generation | Novelty (avg. max TM): 0.449 $\pm$ 0.012 -> 0.363 $\pm$ 0.009 | Helped |
| use FoldFlow++ instead of RFDiffusion | Unconditional protein generation · Unconditional generation | Diversity (pairwise TM): 0.256 -> 0.205 | Helped |
| use FoldFlow++ instead of RFDiffusion | Unconditional protein generation · Unconditional generation | Diversity (MaxCluster): 0.172 -> 0.348 | Helped |
| use FoldFlow++ instead of MultiFlow | Sequence-conditioned protein folding · PDB test set (268 proteins) | RMSD: 14.995 $\pm$ 3.977 -> 3.237 $\pm$ 4.145 | Helped |
| fine-tune FoldFlow++ for motif scaffolding | FoldFlow++ · Motif scaffolding benchmark | Solved /24: 24 -> 24 | No clear effect |
| fine-tune FoldFlow++ for VHH scaffolding | FoldFlow++ · VHH scaffolding | Solved /25: 5 -> 9 | Helped |
| use FoldFlow++ instead of ESMFlow-MD | Equilibrium conformation sampling · ATLAS | Global RMSF r: 0.161 $\pm$ 0.005 -> 0.236 $\pm$ 0.004 | Helped |
| use FoldFlow++ instead of ESMFlow-MD | Equilibrium conformation sampling · ATLAS | Per-target RMSF r: 0.737 $\pm$ 0.001 -> 0.658 $\pm$ 0.001 | Hurt |
| add a skip connection from the structure encoder to the decoder (ablation) | FoldFlow++ · Protein structure generation | Performance | Helped |
| use ReFT with a diversity reward | FoldFlow++ · Secondary-structure diversity | Diversity | Helped |
Surprising findings
- Authors note that re-folding RFDiffusion with ESMFold instead of AlphaFold2 raises its solved count to 24/24, showing evaluation folding-model choice strongly affects scaffolding metrics.
- Training on the larger synthetic SwissProt dataset was not used for the main PDB-only results because PDB-only training gave the best designability.
Limitations
- Requires a competitive pretrained protein language model for sequence conditioning, which can be hard to acquire.
- ProteinMPNN used in evaluation was trained only on PDB structures, so designable outputs on the new synthetic-data distribution may be mis-processed.
- VHH scaffolding performance is low overall (9/25 solved), and motif RMSD is much higher than scaffold RMSD.
Experimental setup
Datasets: PDB (resolution < 5 Å, July 2023) SwissProt AlphaFold2 structures (532,003 filtered to ~160K) ATLAS MD dataset RFDiffusion motif scaffolding benchmark (24 motifs) Structural Antibody Database VHH nanobodies
Benchmarks: unconditional generation (designability, novelty, diversity) PDB folding test set (268 proteins) 24-motif scaffolding benchmark VHH CDR scaffolding (25 motifs) ATLAS equilibrium conformation sampling
Metrics: scRMSD < 2 Å designability TM-score novelty pairwise TM diversity MaxCluster RMSD solved motifs RMSF Pearson r PCA W2 distance
Baselines: RFDiffusion Chroma Genie FrameDiff FoldFlow MultiFlow ESMFold AlphaFlow-MD ESMFlow-MD STR2STR Eigenfold FrameFlow
Builds on: FoldFlow Riemannian flow matching invariant point attention ESM2 MultiFlow
Closest papers by meaning
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.92)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.91)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.90)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.89)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- Variable-Length Generative Protein Design via Generalized Poisson Flow (NeurIPS 2026 · similarity 0.88)
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.