Research map / Probabilistic and scientific modeling / Molecules and proteins
ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids
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.
Fine-tunes Multiflow with 3D ellipsoid layout conditioning (via invariant cross attention and classifier-free guidance) to control protein backbone generation and to sample diverse, novel, and compositional designable proteins from a statistical ellipsoid prior.
Problem
Protein structure generative models offer little control over higher-level spatial layout of secondary structure elements, and unconditional samples are often over-helical and low in compositional complexity, limiting diversity and novelty.
Approach
Represent protein layouts as sets of 3D Gaussian ellipsoids annotated with secondary structure and residue count, inject them into Multiflow via a novel SE(3)-invariant cross-attention, apply classifier-free guidance across translation, rotation and discrete flows, and sample synthetic ellipsoid layouts from a simple repulsive statistical model to boost diversity and novelty.
Key results
- On PDB validation layouts, ProtComposer (lambda=1.6) reaches coverage 0.86 and accuracy 0.93 versus 0.49 and 0.59 for unconditioned Multiflow.
- Synthetic-ellipsoid conditioning expands the designability-diversity/novelty Pareto frontier over Multiflow, Chroma and RFDiffusion, with lambda=2.0 giving diversity 48 and helicity 47% vs PDB 41 and 42%.
- Compositionality (effective number of components) rises from 1.9 for Multiflow to 3.2 at lambda=2.0, close to the PDB value of 4.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition Multiflow on PDB-derived ellipsoid layouts with guidance λ=1.6 | Multiflow · PDB validation ellipsoid layouts | Coverage: 0.49 -> 0.86 | Helped |
| Condition Multiflow on PDB-derived ellipsoid layouts with guidance λ=1.6 | Multiflow · PDB validation ellipsoid layouts | Accuracy: 0.59 -> 0.93 | Helped |
| Condition Multiflow on PDB-derived ellipsoid layouts with guidance λ=1.6 | Multiflow · PDB validation ellipsoid layouts | Misplacement: 0.40 -> 0.13 | Helped |
| Increase classifier-free guidance from λ=0.2 to λ=1.6 | ProtComposer · PDB validation ellipsoid layouts | Coverage: 0.55 -> 0.86 | Helped |
| Use statistical synthetic ellipsoid layouts for conditional generation | ProtComposer · Generated proteins | Diversity: 29 -> 48 | Helped |
| Use statistical synthetic ellipsoid layouts for conditional generation | ProtComposer · Generated proteins | Helicity: 73% -> 47% | Helped |
| Use statistical synthetic ellipsoid layouts for conditional generation | ProtComposer · Generated proteins | Compositionality: 1.9 -> 3.2 | Helped |
| Generate sequences with ProteinMPNN rather than use ProtComposer's jointly generated sequences (ablation) | ProtComposer structures · Sequence-structure co-generation | 1-seq Designability: 0.75 -> 0.81 | Helped |
| Generate sequences with ProteinMPNN rather than use ProtComposer's jointly generated sequences (ablation) | ProtComposer structures · Sequence-structure co-generation | Median scRMSD: 1.76 -> 1.65 | Helped |
| Generate sequences with ProteinMPNN rather than use ProtComposer's jointly generated sequences (ablation) | ProtComposer structures · Sequence-structure co-generation | Mean scRMSD: 2.15 -> 2.41 | Hurt |
| Interpolate unconditional and ellipsoid-conditioned self-conditioning inputs for both models (ablation) | ProtComposer under classifier-free guidance · Guided generation | Designability and Resegment JSD | Helped |
| Use synthetic ellipsoid conditioning instead of inference-time parameter adjustment alone | Multiflow · Designability–diversity, designability–novelty, and designability–helicity tradeoffs | Pareto frontier | Helped |
Surprising findings
- Ellipsoid adherence peaks at guidance strengths above 1 (lambda=1.6), and extreme values eventually break the model down.
- Jointly generated Multiflow sequences have lower designability than ProteinMPNN-designed sequences (0.75 vs 0.81 at 1-seq).
Limitations
- Designability decreases as ellipsoid layouts become more novel and diverse (a cost of the diversity gains).
- Only secondary structure (alpha and beta) annotations are supported; other functional annotations are not modeled.
- Extreme hand-constructed layouts are not always designable.
Experimental setup
Datasets: PDB proteins (Multiflow training set) Multiflow synthetic designable proteins
Benchmarks: PDB validation set (post-2021 proteins)
Metrics: coverage misplacement accuracy likelihood soft accuracy resegment JSD designability Vendi diversity novelty helicity
Baselines: Multiflow Chroma RFDiffusion
Builds on: Multiflow Invariant Point Attention classifier-free guidance ProteinMPNN
Closest papers by meaning
- ProtPainter: Draw or Drag Protein via Topology-guided Diffusion (ICLR 2025 · similarity 0.88)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- Proteus: Exploring Protein Structure Generation for Enhanced Designability and Efficiency (ICML 2024 · similarity 0.86)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.86)
- SurfPro: Functional Protein Design Based on Continuous Surface (ICML 2024 · similarity 0.86)
- Generating Highly Designable Proteins with Geometric Algebra Flow Matching (NeurIPS 2024 · similarity 0.86)
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.