Research map / Learning theory and algorithms / Prediction and decision losses
FAFE: Immune Complex Modeling with Geodesic Distance Loss on Noisy Group Frames
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 F2E, a geodesic-distance frame loss that replaces AF2's chordal-based FAPE to avoid gradient vanishing, and fine-tunes AF2-Multimer with LoRA for antibody-antigen complex modeling.
Problem
AlphaFold2's FAPE loss suffers gradient vanishing on high-rotational-error targets, which hurts accurate antibody-antigen complex structure prediction.
Approach
Shows FAPE corresponds to a chordal distance on group frames, then proposes Frame Aligned Frame Error (F2E), a geodesic frame distance that keeps stable gradients across rotation angles, plus a group-aware variant (G-F2E) that cancels the chordal term via a chordal-geodesic identity.</µcore_idea>
Key results
- F2E fine-tuned AF2 reaches 52.3% DockQ>0.23 on the evaluation set, a 182% relative improvement over AF2.3.
- On the low homology subset, success rate reaches 43.8%, a 100% relative improvement over AF2.3.
- Ablation: single-model F2E with G-aware loss reaches 0.375 DockQ>0.23 vs 0.313 for FAPE with LoRA and 0.188 for AF2 without fine-tuning; ensembling raises it to 0.438.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune AF2-Multimer with F2E instead of the original AF2.3 approach | AF2-Multimer · Full evaluation set | DockQ>0.23 correct rate: 182% improvement | Helped |
| Fine-tune AF2-Multimer with F2E instead of the original AF2.3 approach | AF2-Multimer · Low homology evaluation set | DockQ>0.23 correct rate: 100% improvement | Helped |
| Add ensemble across 5 models (ablation) | F2E + LoRA + G-aware · Ablation evaluation set (not specified) | DockQ>0.23: 0.375 -> 0.438 | Helped |
| Replace FAPE with F2E (ablation) | AF2 fine-tuned with LoRA and group-aware loss · Ablation evaluation set (not specified) | DockQ>0.23: 0.313 -> 0.375 | Helped |
| Add group-aware loss (ablation) | AF2 fine-tuned with FAPE and LoRA · Ablation evaluation set (not specified) | DockQ>0.23: 0.250 -> 0.313 | Helped |
| Add LoRA (ablation) | AF2 fine-tuned with FAPE · Ablation evaluation set (not specified) | DockQ>0.23: 0.219 -> 0.250 | Helped |
| Replace FAPE with F2E (ablation) | AF2 fine-tuned with LoRA and group-aware loss · Average DockQ | average DockQ | Helped |
| Fine-tune all model parameters rather than using the selected LoRA approach (ablation) | AF2 · Evaluation set | LDDT | Hurt |
| Fine-tune only the structure module (ablation) | AF2 · Evaluation set | performance | Helped |
| Use the LoRA setting with rank 16 instead of other tested ranks (ablation) | AF2 · Evaluation set | performance | Helped |
| Freeze the linear layer from node representations to single representations (ablation) | AF2 with LoRA · Evaluation set | performance | Hurt |
| Select predictions by ground-truth DockQ rather than the confidence score (ablation) | F2E · Evaluation set | high accuracy rate: 0.59 | Helped |
Surprising findings
- Full-model fine-tuning of AF2 overfits severely, so the authors restrict updates to LoRA on the evoformer.
- Ensembling mainly improves average DockQ of already-successful predictions rather than the success percentage.
Limitations
- Evaluation not compared against AlphaFold3 due to inability to build a fair benchmark.
- Relies on AF2-Multimer architecture and crops/MSA choices; sidechain and full-atom extension untested.
- Ensembling needed to reach the reported best success rate; the confidence score does not identify the best predictions well.
Experimental setup
Datasets: SabDab PDB
Benchmarks: antibody-antigen full evaluation set (693 samples) low homology antibody-antigen subset
Metrics: DockQ>0.23 success rate DockQ>0.8 rate average DockQ
Baselines: AlphaFold2-Multimer (AF2.3) AF2sample FAPE fine-tuning G-FAPE
Builds on: AlphaFold2 AlphaFold2-Multimer LoRA FAPE loss
Closest papers by meaning
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.86)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.84)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.84)
- AsEP: Benchmarking Deep Learning Methods for Antibody-specific Epitope Prediction (NeurIPS 2024 · similarity 0.83)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.83)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.83)
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.