Research map / Probabilistic and scientific modeling / Molecules and proteins
Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property Prediction
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.
Identifies a structure embedding bias when predicted (AlphaFold2) protein structures replace experimental ones in property prediction, and proposes SAO, a bootstrap-style embedding alignment pretraining framework with mask-view augmentation and denoising guidance that improves robustness on predicted and experimental structures.
Problem
Structure-based protein property prediction models trained on experimental structures degrade notably when applied to accurately predicted structures (e.g., AlphaFold2), due to a structure embedding bias between predicted and experimental structure representations.
Approach
Pretrain a protein structure encoder by aligning the embedding of predicted structures to that of paired experimental structures using a bootstrap (EMA teacher) mechanism, augmented with a masked-sequence view and a backbone frame denoising loss, then fine-tune on downstream tasks; the method is encoder-agnostic.
Key results
- On the TMscore>=0.5 predicted-structure setting, SAO improves REInet EC Fmax from 0.584 (vanilla) to 0.731 and AUPR from 0.506 to 0.699.
- SAO achieves best results on all 8 metrics across EC and GO tasks in the main comparison, e.g., GO-MF Fmax 0.547 on TMscore>=0.5 vs 0.519 vanilla.
- SAO also improves experimental-structure inference, e.g., EC Fmax 0.784 vs 0.746 vanilla.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAO rather than TonP | REINet · EC (TMscore ≥ 0.5) | AUPR: 0.693 -> 0.699 | Helped |
| Use SAO rather than TonP | REINet · GO-MF (TMscore ≥ 0.5) | Fmax: 0.537 -> 0.547 | Helped |
| Use SAO rather than TonP | REINet · GO-CC (TMscore ≥ 0.5) | AUPR: 0.266 -> 0.225 | Hurt |
| Use SAO rather than Vanilla | REINet · EC (Experimental Structure) | Fmax: 0.746 -> 0.784 | Helped |
| Use SAO rather than Vanilla | REINet · GO-CC (Experimental Structure) | AUPR: 0.251 -> 0.241 | Hurt |
| Add structure denoising guidance (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | Fmax: 0.429 -> 0.442 | Helped |
| Add structure denoising guidance (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | AUPR: 0.206 -> 0.226 | Helped |
| Add mask view augmentation (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | Fmax: 0.388 -> 0.442 | Helped |
| Add mask view augmentation (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | AUPR: 0.201 -> 0.226 | Helped |
| Add embedding alignment (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | Fmax: 0.389 -> 0.442 | Helped |
| Add embedding alignment (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | AUPR: 0.203 -> 0.226 | Helped |
| Add mask language modeling (ablation) | SAO on REINet · GO-CC (TMscore ≥ 0.5) | Fmax: 0.422 -> 0.442 | Helped |
Surprising findings
- Structure refinement (ATOMRefine-based RefthenPred) only marginally helps or does not fix the performance gap on predicted structures.
- Training directly on predicted structures (TonP) degrades experimental-structure performance substantially (e.g., EC Fmax 0.543 vs 0.746 vanilla).
Limitations
- Limited exploration of non-classification tasks.
- Limited sources of predicted protein structures (mostly AlphaFold2).
Experimental setup
Datasets: EC GO (MF, BP, CC) AlphaFold2 predicted structures
Benchmarks: EC GO-MF GO-BP GO-CC
Metrics: Fmax AUPR
Baselines: Vanilla TonP Mixed RefthenPred RefandPred
Builds on: BYOL GearNet Uni-Mol ATOMRefine EquiPPIS
Closest papers by meaning
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.89)
- Enhancing Graph Contrastive Learning for Protein Graphs from Perspective of Invariance (ICML 2025 · similarity 0.87)
- Atom-level Protein Representation Learning Improves Protein Structure Prediction (NeurIPS 2026 · similarity 0.86)
- Multi-Scale Representation Learning for Protein Fitness Prediction (NeurIPS 2024 · similarity 0.86)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.86)
- Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure (NeurIPS 2024 · 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.