Research map / Probabilistic and scientific modeling / Molecules and proteins
Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer
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 ADesigner, a one-shot model that co-designs antibody CDR sequences and 3D backbone structures using a protein complex invariant embedding over full backbone atoms and a cross-gate MLP for sequence-structure co-learning.
Problem
Antibody CDR design requires modeling sequence and 3D structure jointly, but prior methods use limited geometry (C-alpha only or orientations) and iterative refinement or diffusion sampling, which makes inference inefficient.
Approach
A protein complex invariant embedding that encodes intra- and inter-component interactions among all backbone atoms (CA, N, C, O), plus a cross-gate MLP that lets sequence and structure representations gate and refine each other, decoding CDRs in one shot without iterative refinement.
Key results
- On SAbDab 10-fold CV, ADesigner reaches 64.34% AAR on CDR-H1, 55.52% on CDR-H2, and 37.37% on CDR-H3, outperforming MEAN (58.29%, 47.15%, 36.38%).
- On the RAbD CDR-H3 benchmark, ADesigner achieves 40.94% AAR and 1.55 RMSD, versus MEAN's 36.77% and 1.81.
- In affinity optimization, ADesigner reaches an average ΔΔG of -10.78 kcal/mol, compared with -5.33 for MEAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ADesigner instead of MEAN | CDR-H1 sequence and structure modeling · SAbDab CDR-H1 | AAR: 58.29±7.26% -> 64.34±3.37% | Helped |
| use ADesigner instead of MEAN | CDR-H2 sequence and structure modeling · SAbDab CDR-H2 | AAR: 47.15±3.09% -> 55.52±3.36% | Helped |
| use ADesigner instead of MEAN | CDR-H3 sequence and structure modeling · SAbDab CDR-H3 | AAR: 36.38±3.08% -> 37.37±2.33% | Helped |
| use ADesigner instead of MEAN | CDR-H3 design · RAbD CDR-H3 | AAR: 36.77% -> 40.94% | Helped |
| use ADesigner instead of MEAN | binding affinity optimization · Binding affinity optimization | ΔΔG: -5.33 -> -10.78 | Helped |
| use ADesigner instead of MEAN | antibody CDR design · SAbDab training set, one epoch | training time: 22s -> 14s | Helped |
| use ADesigner instead of MEAN | antibody CDR design · SAbDab testing set | inference time: 28s -> 24s | Helped |
| add PIE (ablation) | ADesigner · SAbDab CDR-H1 | AAR: 60.27±6.69% -> 64.34±3.37% | Helped |
| add PIE (ablation) | ADesigner · SAbDab CDR-H2 | AAR: 49.14±2.96% -> 55.52±3.36% | Helped |
| add PIE (ablation) | ADesigner · SAbDab CDR-H3 | AAR: 35.78±2.43% -> 37.37±2.33% | Helped |
| add CGMLP instead of equivariant graph neural networks (ablation) | ADesigner · SAbDab CDR-H1 | RMSD: 1.07±0.17 -> 0.82±0.12 | Helped |
| add CGMLP instead of equivariant graph neural networks (ablation) | ADesigner · SAbDab CDR-H2 | AAR: 52.50±3.78% -> 55.52±3.36% | Helped |
Limitations
- Limited to in silico design; no wet-lab validation.
- Affinity estimates rely on a pre-trained predictor rather than experiments.
Experimental setup
Datasets: SAbDab RAbD benchmark SKEMPI
Benchmarks: SAbDab RAbD SKEMPI
Metrics: amino acid recovery RMSD TM-score binding free energy change (ΔΔG)
Baselines: LSTM C-LSTM RefineGNN C-RefineGNN DiffAB MEAN RosettaAD
Builds on: RefineGNN MEAN DiffAb E(n)-equivariant GNN
Closest papers by meaning
- Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.89)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.88)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.88)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.87)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.87)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.87)
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.