Research map / Label-efficient and robust learning / Recognition and long-tail
ProtGO: Function-Guided Protein Modeling for Unified Representation Learning
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 ProtGO, a teacher-student framework where a GNN teacher with a GO annotation encoder distills function-informed embeddings into a sequence-structure GNN student via a KL-based distribution alignment loss.
Problem
Protein representation learning that integrates sequence, structure, and function (GO annotation) information despite functional annotations being missing for most proteins and large modality scale differences.
Approach
A teacher with an annotation (GO) encoder produces knowledge-enhanced embeddings; the student shares the same GNN and needs no GO input, and is aligned to the teacher via a KL divergence between Gaussian approximations of the batch-level latent distributions rather than per-sample alignment.
Key results
- ProtGO student reaches 60.5% fold accuracy vs 56.7% for CDConv, and 99.8/89.4 on SuperFamily-level/reaction tasks is the best reported among compared methods.
- On GO term prediction ProtGO student achieves Fmax 0.464 (BP), 0.667 (MF), 0.492 (CC) and 0.857 on EC, exceeding CDConv (0.453, 0.654, 0.479, 0.820).
- Ablation: removing the teacher (w/o Teacher) drops fold accuracy from 60.5 to 57.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProtGO's teacher-student model instead of CDConv | sequence-structure protein representation learning · Fold Classification - Fold | accuracy (%): 56.7 -> 60.5 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | sequence-structure protein representation learning · Fold Classification - SuperFamily | accuracy (%): 77.7 -> 79.4 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | sequence-structure protein representation learning · Fold Classification - Family | accuracy (%): 99.6 -> 99.8 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | sequence-structure protein representation learning · Enzyme Reaction Classification | accuracy (%): 88.5 -> 89.4 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | protein function prediction · GO-BP | Fmax: 0.453 -> 0.464 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | protein function prediction · GO-MF | Fmax: 0.654 -> 0.667 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | protein function prediction · GO-CC | Fmax: 0.479 -> 0.492 | Helped |
| Use ProtGO's teacher-student model instead of CDConv | protein function prediction · EC number prediction | Fmax: 0.820 -> 0.857 | Helped |
| Add the annotation encoder to the teacher (ablation) | ProtGO · GO-BP | Fmax: 0.454 -> 0.464 | Helped |
| Add the annotation encoder to the teacher (ablation) | ProtGO · GO-MF | Fmax: 0.664 -> 0.667 | Helped |
| Add the teacher model and KL distillation loss (ablation) | ProtGO student · Fold Classification - Fold | accuracy (%): 57.8 -> 60.5 | Helped |
| Add the teacher model and KL distillation loss (ablation) | ProtGO student · EC number prediction | Fmax: 0.851 -> 0.857 | Helped |
Surprising findings
- Removing the teacher's annotation encoder (w/o AE-T) has only minimal impact on performance, suggesting the label-augmented component is low-sensitivity.
Limitations
- Base model is not pre-trained on large-scale protein data.
- Teacher requires GO annotations, which are unavailable for most proteins.
- Gains over CDConv on some tasks are small (e.g., 0.2% on enzyme reaction).
Experimental setup
Datasets: Fold classification dataset Enzyme reaction dataset Gene Ontology (GO) dataset Enzyme Commission (EC) dataset
Benchmarks: Fold classification (Fold, SuperFamily, Family) Enzyme reaction classification GO term prediction (BP, MF, CC) EC number prediction
Metrics: accuracy Fmax
Baselines: CDConv ProNet-Backbone CRL GearNet-Edge GearNet-Edge-IEConv GVP ESM-GearNet GearNet-ESM-INR-MC ProtST-ESM-2 ESM-2
Builds on: CDConv GearNet ProNet IEConv label-augmented learning
Closest papers by meaning
- Prot2Text: Multimodal Protein’s Function Generation with GNNs and Transformers (AAAI 2024 · similarity 0.92)
- Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment (NeurIPS 2025 · similarity 0.90)
- ProtCLIP: Function-Informed Protein Multi-Modal Learning (AAAI 2025 · similarity 0.90)
- ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention (NeurIPS 2024 · similarity 0.88)
- Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.87)
- Understanding protein function with a multimodal retrieval-augmented foundation model (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.