Research map / Probabilistic and scientific modeling / Molecules and proteins
HeMeNet: Heterogeneous Multichannel Equivariant Network for Protein Multi-task 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.
Builds Protein-MT, a 6-task structure-based benchmark from 4 public datasets, and proposes HeMeNet, an E(3)-equivariant heterogeneous multichannel GNN with task-aware readout for joint affinity and function prediction.
Problem
Structure-based protein function and affinity prediction is usually trained per task on small labeled datasets, limiting performance and generalization; no structure-based multi-task benchmark combining affinity and property tasks existed.
Approach
Combine affinity (LBA, PPA) and function (EC, GO-MF/BP/CC) datasets via UniProt-ID label transfer, and use a full-atom heterogeneous equivariant GNN whose readout is conditioned on learnable task-prompt queries via attention.
Key results
- Single-task HeMeNet reaches EC Fmax 0.863 and GO-MF 0.778, outperforming all listed single-task baselines on five of six tasks.
- Multi-task HeMeNet improves PPA RMSE from 6.031 (single-task) to 1.087, and LBA RMSE reaches 1.730 in the multi-task setting.
- Task-aware readout beats sum readout on most tasks; ablating it raises PPA RMSE from 1.087 to 1.970 and lowers EC Fmax from 0.810 to 0.722.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train jointly on all six tasks instead of single-task training | HeMeNet · Protein-MT PPA | RMSE: 6.031 to 1.087 | Helped |
| train jointly on all six tasks instead of single-task training | HeMeNet · Protein-MT LBA | RMSE: 1.912 to 1.730 | Helped |
| use task-aware readout instead of sum readout (ablation) | HeMeNet · Protein-MT EC | Fmax: 0.722 -> 0.810 | Helped |
| use task-aware readout instead of sum readout (ablation) | HeMeNet · Protein-MT LBA | RMSE: 1.905 -> 1.730 | Helped |
| add heterogeneous edge types and relational message-passing weights (ablation) | HeMeNet · Protein-MT EC | Fmax: 0.547 -> 0.810 | Helped |
| add full-atom geometry instead of using alpha-C atoms (ablation) | HeMeNet · Protein-MT PPA | RMSE: 0.861 -> 1.087 | Hurt |
| add full-atom geometry instead of using alpha-C atoms (ablation) | HeMeNet · Protein-MT EC | Fmax: 0.630 -> 0.810 | Helped |
| use task-aware readout instead of task-prompt weighted node features (ablation) | HeMeNet · Protein-MT EC | Fmax: 0.325 -> 0.810 | Helped |
| combine LBA, PPA, and all four property tasks rather than only LBA and PPA | HeMeNet · Protein-MT LBA | RMSE: 1.758 -> 1.730 | Helped |
| combine LBA, PPA, and all four property tasks rather than only LBA and PPA | HeMeNet · Protein-MT PPA | RMSE: 1.064 -> 1.087 | No clear effect |
Surprising findings
- Baseline models such as GearNet and SchNet degrade on property tasks when trained multi-task, while HeMeNet with task-aware readout does not.
- Removing full-atom geometry (using alpha-carbon only) improves PPA RMSE (0.861 vs 1.087), attributed to noise from large atom counts.
Limitations
- Only invariant tasks are considered; equivariant tasks are not addressed.
- Fully-labeled test set restricts adding more tasks while keeping sample size.
- Input mixes single chains and complexes, which complicates a unified setup.
Experimental setup
Datasets: PDBbind (LBA, PPA) EC/GO dataset from prior work (19201 PDB chains, 538 EC numbers) Gene Ontology UniProt
Benchmarks: Protein-MT
Metrics: RMSE MAE Fmax
Baselines: GCN GAT SchNet GearNet GearNet-fullatom CDConv EGNN GVP dyMEAN ESM2 GPT-4
Builds on: dyMEAN GearNet EGNN GVP
Closest papers by meaning
- EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction (ICML 2024 · similarity 0.87)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.86)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.86)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.85)
- Generalist Equivariant Transformer Towards 3D Molecular Interaction Learning (ICML 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.