Research map / Probabilistic and scientific modeling / Molecules and proteins
Steering Protein Family Design through Profile Bayesian Flow
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 ProfileBFN, a discrete Bayesian Flow Network extended to MSA profiles that trains on single sequences but samples family proteins from profiles, outperforming DPLM and EvoDiff on family generation, enzyme function, and representation tasks.
Problem
Protein family design needs generative models that use family (MSA) information, but constructing MSA-profile training sets is computationally expensive and MSAs are variable-sized sets that are hard to scale.
Approach
Extend the discrete Bayesian Flow Network to accept generalized profiles (distributions over amino acids per position) as inputs, so the model can be trained on single sequences (one-hot degenerate profiles) and then sample family proteins conditioned on an MSA-derived profile.
Key results
- On CAMEO family generation, ProfileBFN-Profile (650M) reaches LR P@L 0.173 vs 0.186 for searched MSA baseline and 0.100 for ESM-2 650M, with diversity 0.293 and novelty 0.641.
- On enzyme function (accuracy x uniqueness), ProfileBFN-Profile (650M) scores 95.19% on P40925, 98.98% on Q7X7H9 and 42.67% on Q15165, versus EvoDiff-MSA at 27.93%, 88.69%, 1.39%.
- On protein representation tasks, ProfileBFN (650M) achieves thermostability 0.710 Spearman and EC Fmax 0.887, outperforming reproduced DPLM (650M) at 0.698 and 0.881.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an MSA-derived profile instead of a single sequence as input (ablation) | ProfileBFN · CAMEO | LR P@L/5: 0.422 -> 0.474 | Helped |
| Use an MSA-derived profile instead of a single sequence as input (ablation) | ProfileBFN · CAMEO | Div.: 0.421 -> 0.293 | Helped |
| Increase model size from 150M to 650M parameters (ablation) | ProfileBFN-Profile · CAMEO | LR P@L/5: 0.384 -> 0.474 | Helped |
| Increase model size from 150M to 650M parameters (ablation) | ProfileBFN-Profile · CAMEO | Div.: 0.283 -> 0.293 | Hurt |
| Use ProfileBFN instead of DPLM | 650M protein family generation · CAMEO | LR P@L: 0.102 -> 0.173 | Helped |
| Use ProfileBFN instead of EvoDiff-MSA | protein family generation · CAMEO | LR P@L/5: 0.168 -> 0.474 | Helped |
| Use ProfileBFN instead of EvoDiff-MSA | protein family generation · CAMEO | Div.: 0.225 -> 0.293 | Mixed |
| Generate enzymes with ProfileBFN instead of EvoDiff-MSA | enzyme family design · P40925 | Accuracy × Uniqueness: 27.93% -> 95.19% | Helped |
| Generate enzymes with ProfileBFN instead of EvoDiff-MSA | enzyme family design · Q7X7H9 | Accuracy × Uniqueness: 88.69% -> 98.98% | Helped |
| Generate enzymes with ProfileBFN instead of EvoDiff-MSA | enzyme family design · Q15165 | Accuracy × Uniqueness: 1.39% -> 42.67% | Helped |
| Use ProfileBFN instead of reproduced DPLM | 650M protein representation model · HumanPPI | ACC(%): 77.77 -> 82.22 | Helped |
| Use ProfileBFN instead of reproduced DPLM | 650M protein representation model · DeepLoc Binary | ACC(%): 93.17 -> 93.58 | Helped |
Surprising findings
- Generated family sequences achieve better contact-prediction precision (LR P@L) than sequences from the searched MSA that AlphaFold2 relies on, on the CAMEO set.
Limitations
- Comparisons with DPLM rely partly on the authors' own reproduction, which the authors note may be unstable.
- Structural evaluation uses non-parametric contact prediction and ESMFold, which the authors acknowledge can hallucinate.
- Protein-protein interaction (HumanPPI) lags SaProt, which uses structure information.
Experimental setup
Datasets: UniRef CAMEO targets (61 proteins) enzyme families (P40925, Q7X7H9, Q15165) phage lysozyme families thermostability HumanPPI metal ion binding EC GO DeepLoc
Benchmarks: CAMEO-derived family generation set CLEAN enzyme function evaluation thermostability HumanPPI metal ion binding EC GO DeepLoc
Metrics: sequence diversity (identity) novelty (1 - max identity) LR P@L contact precision (CCMPred) pLDDT max TM-score accuracy x uniqueness Spearman's rho ACC Fmax
Baselines: ESM-2 PoET EvoDiff-MSA DPLM SaProt MIF-ST ESM-1
Builds on: Bayesian Flow Networks DPLM ESM-2 PoET
Closest papers by meaning
- Variable-Length Generative Protein Design via Generalized Poisson Flow (NeurIPS 2026 · similarity 0.86)
- CarbonNovo: Joint Design of Protein Structure and Sequence Using a Unified Energy-based Model (ICML 2024 · similarity 0.86)
- Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design (ICML 2024 · similarity 0.86)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.86)
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.85)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.84)
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.