Research map / Probabilistic and scientific modeling / Molecules and proteins
GeoAB: Towards Realistic Antibody Design and Reliable Affinity Maturation
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 GeoAB, a geometry-aware antibody CDR co-design model (GeoAB-Designer/Refiner) with a von Mises torsion initializer and mechanics-informed atom updates, plus GeoAB-Optimizer that jointly predicts ΔΔG and mutant structures for affinity maturation.
Problem
Generated antibody CDR structures often violate physical internal geometry constraints, and affinity maturation relies on ΔΔG predictors that need mutant structures unavailable at inference, making maturation unreliable.
Approach
Combine a generative von Mises mixture torsion initializer with NeRF reconstruction and a hierarchical residue/atom graph whose atom-level layers model bond, angle, and torsion mechanics with energy-informed geometric constraints; for affinity maturation, a twin weight-shared refiner predicts both mutant structures and ΔΔG from sequence and context, trained jointly on paired and unpaired data.
Key results
- GeoAB-R reduces A-RMSD and UA-RMSD by 10.71% and 10.02% over C-DyMEAN on RAbD CDR-H3 (A-RMSD 1.3978, UA-RMSD 1.5852).
- GeoAB-D improves AAR to 0.3939 on RAbD CDR-H3 versus DiffAB's ~0.27-0.28, with much lower RMSD (1.40 vs 5.9).
- On SKEMPI2 cross-validation, GeoAB reaches overall PCC 0.6324 and SRCC 0.4972 with RMSE 1.4758, competitive with RDE-PPI and DDG-Pred.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GeoAB-R instead of C-DyMEAN | CDR-H3 co-design · RAbD CDR-H3 | UA-RMSD: 1.7641 -> 1.5852 | Helped |
| use GeoAB-D instead of DiffAB | generative CDR-H3 design · RAbD CDR-H3, 20% | UA-RMSD: 9.4393 -> 1.5038 | Helped |
| use GeoAB for mutation-effect prediction | SKEMPI2 mutation-effect prediction · SKEMPI2 cross-validation, overall | MAE: 1.0821 -> 1.0611 | Helped |
| remove the binary interaction layer (ablation) | GeoAB-R · RAbD CDR-H3 | UA-RMSD: 1.5852 -> 1.6695 | Hurt |
| remove the ternary interaction layer (ablation) | GeoAB-R · RAbD CDR-H3 | UA-RMSD: 1.5852 -> 1.6101 | No clear effect |
| remove the quaternary interaction layer (ablation) | GeoAB-R · RAbD CDR-H3 | UA-RMSD: 1.5852 -> 1.6779 | Hurt |
| remove the bond-length geometric constraint (ablation) | GeoAB-R · RAbD CDR-H3 | JSD, Cα-C: 0.5286 -> 0.6136 | Hurt |
| remove the angle geometric constraint (ablation) | GeoAB-R · RAbD CDR-H3 | JSD, Cα-C-N: 0.6031 -> 0.6941 | Hurt |
| remove the torsion-angle geometric constraint (ablation) | GeoAB-R · RAbD CDR-H3 | JSD, C-N-Cα-C: 0.3180 -> 0.5696 | Hurt |
| remove ESM pretraining embedding (ablation) | GeoAB-R · RAbD CDR-H3 | AAR: 0.4053 -> 0.4096 | No clear effect |
| remove dynamic loss weights (ablation) | GeoAB-R · RAbD CDR-H3 | JSD, Cα-C: 0.5286 -> 0.5428 | No clear effect |
Surprising findings
- Removing the ESM pretraining embedding slightly improves AAR (0.4096 vs 0.4053) on RAbD, suggesting ESM brings little gain for co-design.
Limitations
- Interaction layers add computational cost relative to vanilla EGNN.
- ΔΔG prediction accuracy trails structure-aware pretraining methods such as RDE-PPI.
- Affinity maturation relies on enumerating mutations at a few CDR-H3 positions (1 or 2 points).
Experimental setup
Datasets: SAbDab RAbD SKEMPI2 PDB-REDO
Benchmarks: RAbD SAbDab 10-fold cross-validation SKEMPI2
Metrics: AAR A-RMSD UA-RMSD lddt TM-score JSD MAE MCE PCC SRCC
Baselines: RefineGNN C-RefineGNN MEAN DyMEAN C-DyMEAN DiffAB Rosetta ddG FoldX DDG-Pred RDE-PPI
Builds on: EGNN DiffAB MEAN Graph Mechanics Networks NeRF
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.87)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.86)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 2025 · similarity 0.86)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.85)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.85)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.85)
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.