Research map / Vision-language and multimodal / Cross-modal learning
Structure-Aware Fusion with Progressive Injection for Multimodal Molecular 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 MuMo, a Mamba-based molecular model that fuses 2D topology and 3D geometry into a unified structural prior and asymmetrically injects it into the SMILES stream via progressive injection, evaluated on 29 TDC/MoleculeNet/Reaxtica tasks.
Problem
Multimodal molecular property prediction suffers from unreliable RDKit conformers and modality collapse when 2D/3D/SMILES inputs are naively fused.
Approach
A Structured Fusion Pipeline builds a unified 2D+3D graph (bond lengths and angles as geometric linkages, BRICS-based geometry-aware partitioning with gated multiscale fusion) as a structural prior, which is injected asymmetrically into late Mamba layers of the SMILES stream through cross-attention and a global token, while the prior itself evolves independently via state space propagation.
Key results
- MuMo averages 2.7% improvement over the best baseline across tasks and ranks first on 22 of 29 tasks, with a reported 27% gain on LD50.
- Ablation: removing 3D geometry (no SUG) drops performance by 7.46%, and using sequence only drops 13.11% on MoleculeNet tasks.
- Progressive injection from layer 9 outperforms full, early, and late injection, and fixed injection is 6.28% worse than progressive injection on BBBP/ESOL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the SUG structural unified graph representation (including 3D geometry) (ablation) | MuMo SFP · MoleculeNet BACE, BBBP, ESOL, LIPO | Avg.Drop: -7.46% | Helped |
| Add GSP geometry-aware substructure partitioning (ablation) | MuMo SFP · MoleculeNet BACE, BBBP, ESOL, LIPO | Avg.Drop: -3.00% | Helped |
| Use 2D topology or SMILES alongside the SFP structural modalities (ablation) | MuMo SFP · MoleculeNet BACE, BBBB, ESOL, LIPO | Avg.Drop: -13.11% | Helped |
| Inject structural priors rather than using sequence-only input (ablation) | MuMo Progressive Injection · BBBP and ESOL | Avg.Drop: -18.91% | Helped |
| Inject structural priors progressively at layers 9–16 instead of fixed injection (ablation) | MuMo Progressive Injection · BBBP and ESOL | Avg.Drop: -6.28% | Helped |
| Add explicit torsion features (ablation) | MuMo · QM9 alpha | MAE: 0.283 -> 0.281; +0.71 | Helped |
| Add explicit torsion features (ablation) | MuMo · QM9 C_v | MAE: +10.3 | Helped |
| Add explicit torsion features (ablation) | MuMo · QM9 ZPVE | MAE: 0.0005 -> 0.0005; 0.0 | No clear effect |
| Add explicit torsion features (ablation) | MuMo · QM9 HOMO/LUMO/GAP | MAE: 0.0030 -> 0.0032; -6.67 | Hurt |
| Add explicit torsion features (ablation) | MuMo · Pretraining and QM8 SFT | Step time (ms): 987/731 -> 2623/1316 | Hurt |
Surprising findings
- Explicit torsion angle features give minimal accuracy gains on QM9 (HOMO/LUMO/GAP worsens by 6.67%) while raising runtime up to 2.6x.
- Full or early injection causes modality collapse, so injecting structure only in later layers works better.
Limitations
- Designed as a task-specific model rather than a general-purpose backbone, as stated by the authors.
- Unified graph cannot distinguish enantiomers because it uses unsigned lengths and angles without chirality.
- Conformer ensembles are not modeled, so conformational flexibility may be missed.
Experimental setup
Datasets: ChEMBL-1.6M TDC MoleculeNet QM7 QM8 QM9 Reaxtica ZINC
Benchmarks: TDC ADMET tasks MoleculeNet QM7 QM8 QM9 Reaxtica BHC Reaxtica CPA Reaxtica HTE
Metrics: AUROC MAE RMSE R^2
Baselines: AttentiveFP FPGNN DMPNN AttrMasking ContextPred TranFoxMol DeepMol MolBERT GROVER Uni-Mol MoLFormer ChemBERTa-2
Builds on: Mamba BRICS Uni-Mol GROVER
Closest papers by meaning
- Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine (ICLR 2025 · similarity 0.86)
- Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.86)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.85)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.84)
- Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.84)
- Multimodal 3D Genome Pre-training (NeurIPS 2025 · similarity 0.84)
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.