Research map / Probabilistic and scientific modeling / Molecules and proteins
PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design
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 PPDiff, a joint sequence-structure DDPM built on the SSINC network (interleaved self-attention, kNN equivariant graph layers, causal attention), pretrained on a 706k-complex PDB dataset (PPBench) and finetuned for mini-binder and antigen-antibody design.
Problem
Designing high-affinity protein binders for arbitrary protein targets on demand, where existing methods often fail because designed sequences do not fold into the intended structure or structures do not bind the target.
Approach
A discrete-continuous diffusion model that jointly denoises residue types (uniform categorical diffusion) and C-alpha coordinates (Gaussian diffusion) with a SSINC network that interleaves ESM-2-initialized self-attention with kNN equivariant graph layers, followed by a causal attention layer for sequence dependencies, plus energy-based and DSSP-fragment informative initialization at inference.
Key results
- On general complex design (top-1 of 100 candidates), PPDiff reaches a 50.00% success rate versus 47.36% for the SSINC Network and 45.61% for InterleavingDiff.
- On target-protein mini-binder design, PPDiff averages 23.16% success rate versus 19.32% for InterleavingDiff and 10.96% for SSINC Network; RF+MPNN averages 21.46%.
- On antigen-antibody design, PPDiff averages 16.89% success rate, the highest among baselines, with the highest ipTM (0.541 average) and novelty scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PPDiff instead of SSINC Network | PPBench complex design · PPBench test, top-1 | success rate: 47.36% -> 50.00% | Helped |
| use PPDiff instead of SSINC Network | PPBench complex design · PPBench test, top-5 | success rate: 43.85% -> 45.71% | Helped |
| use PPDiff instead of SSINC Network | PPBench complex design · PPBench test, top-10 | success rate: 36.84% -> 37.68% | Helped |
| use PPDiff instead of RF+MPNN | Mini-binder design · Mini-binder design, average | success rate: 21.46% -> 23.16% | Helped |
| use PPDiff instead of SSINC Network | Antigen-antibody design · Antigen-antibody design | success rate: 16.74% -> 16.89% | Helped |
| add one causal attention layer (ablation) | PPDiff · PPBench complex design | design quality | Helped |
| increase causal attention layers above one (ablation) | PPDiff · PPBench complex design | design quality | No clear effect |
| increase diffusion steps to 1,000 (ablation) | PPDiff · PPBench complex design | design quality | Helped |
| increase model scale from 164M to 692M parameters (ablation) | PPDiff · PPBench complex design | ipTM and pTM | Helped |
| add Swiss-Prot pretraining (ablation) | PPDiff · PPBench complex design | success rate: 50.00% -> 53.57%; 45.71% -> 35.36%; 37.68% -> 27.68% | Mixed |
| increase candidate pool from 30 to 100 (ablation) | PPDiff · PPBench complex design, top-1 | success rate: 39.28% -> 50.00% | Helped |
| increase candidate pool from 30 to 100 (ablation) | PPDiff · PPBench complex design, top-5 | success rate: 30.36% -> 45.71% | Helped |
Surprising findings
- Pretraining on additional Swiss-Prot pseudo-complexes gave no significant benefit.
- All models failed to design successful binders for EGFR in the mini-binder task.
Limitations
- No wet-lab validation; success is judged only by AlphaFold-based metrics and docking.
- Top-5 and top-10 PPDiff results lag SSINC Network on PAE and pLDDT, and InterleavingDiff has higher diversity and novelty at larger k.
Experimental setup
Datasets: PPBench target-protein mini-binder complex dataset SAbDab Swiss-Prot (ablation)
Benchmarks: PPBench test set target-protein mini-binder dataset SAbDab antigen-antibody test split
Metrics: success rate ipTM pTM PAE pLDDT novelty diversity AF2 pAE_interaction HDOCK docking score
Baselines: SEnc+ProteinMPNN InterleavingDiff SSINC Network RFDiffusion+ProteinMPNN
Builds on: DDPM ESM-2 ProteinMPNN AlphaFold3 data pipeline
Closest papers by meaning
- Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule Generation (ICML 2024 · similarity 0.87)
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs (NeurIPS 2025 · similarity 0.87)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.86)
- Learning to design protein-protein interactions with enhanced generalization (ICLR 2024 · similarity 0.86)
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.86)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · 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.