Research map / Vision-language and multimodal / Cross-modal learning
BioBridge: Bridging Biomedical Foundation Models via Knowledge Graphs
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 BioBRIDGE, a parameter-efficient framework that trains a small transformer bridge over frozen unimodal biomedical foundation models (ESM2-3B, UniMol, PubMedBERT) using PrimeKG triples to enable cross-modal retrieval, zero-shot cross-species matching, and retrieval-augmented generation.
Problem
Biomedical foundation models are mostly unimodal, and joint multimodal training scales poorly in compute and paired data across more than two modalities.
Approach
Keep all unimodal foundation models frozen and learn only relation-aware transformation modules (plus modality projection heads) with contrastive loss on biomedical knowledge graph triples, so embeddings of one modality can be mapped into another modality's space without the KG at inference time.
Key results
- On seven PrimeKG cross-modal retrieval tasks, BioBRIDGE achieves the best MRR on all tasks (e.g., Protein to MF 0.326 vs 0.119 for best KGE baseline RotatE), reportedly about 76.3% better on average than the best baseline.
- On GO semantic similarity inference, BioBRIDGE averages 0.81 Spearman versus 0.41 for the best baseline (KeAP).
- In zero-shot cross-species matching, BioBRIDGE outperforms a supervised dual-encoder on mouse protein to mouse phenotype Rec@3 (0.185 vs 0.145) and far exceeds it on mouse protein to human phenotype (e.g., Rec@20 0.168 vs 0.051).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BioBRIDGE instead of the strongest KGE baseline, RotatE | Cross-modal KG retrieval · PrimeKG cross-modality retrieval tasks | MRR: 0.034 vs 0.079 (Protein → BP); 0.326 vs 0.119 (Protein → MF); 0.319 vs 0.107 (Protein → CC); 0.189 vs 0.150 (Drug → Dise | Helped |
| Transform protein embeddings into BP, MF, and CC spaces with BioBRIDGE | Protein semantic similarity inference · Semantic similarity (Avg) | Spearman correlation: 0.41 -> 0.81 | Helped |
| Inject the ppi relation into ESM2 protein embeddings and transform back to protein space | Protein PPI prediction · SHS27K | F1 (B+D): 0.732 -> 0.739 | Helped |
| Use BioBRIDGE instead of the supervised paired-data model | Cross-species protein-phenotype matching · Mouse protein → mouse phenotype | Rec@1: 0.061 -> 0.067 | Helped |
| Use BioBRIDGE instead of the supervised paired-data model | Cross-species protein-phenotype matching · Mouse protein → human phenotype | Rec@1: 0.001 -> 0.022 | Helped |
Surprising findings
- Zero-shot BioBRIDGE beats a supervised dual-encoder trained on paired mouse protein-phenotype data, even though neither mouse proteins nor phenotypes were in its training data.
- The supervised baseline failed to extrapolate to mouse protein to human phenotype matching, while BioBRIDGE achieved non-trivial retrieval.
Limitations
- Proof-of-concept limited to six node types and three modalities (protein, molecule, text) from a simplified PrimeKG.
- Some baselines are weak or not directly comparable; the cross-species task has no prior work for comparison.
- Generation and QA results are qualitative case studies without systematic quantitative evaluation.
Experimental setup
Datasets: PrimeKG UniProt Swiss-Prot sequences DrugBank MGI/MPI human phenotype data SHS27K SHS148K STRING GO semantic similarity test matrices
Benchmarks: PrimeKG cross-modality retrieval (7 triple types) GO semantic similarity inference SHS27K SHS148K STRING PPI MGI cross-species protein-phenotype matching DrugBank multimodal Q&A and generation case studies
Metrics: MRR Hit@1 Hit@3 Hit@10 Spearman correlation F1 Recall@K Precision@K nDCG@K Tanimoto similarity
Baselines: TransE TransR TransH TransD ComplEx DistMult RotatE KeAP OntoProtein ESM2-3B
Builds on: PrimeKG InfoNCE ESM2 UniMol PubMedBERT
Closest papers by meaning
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.85)
- Retrieval-Augmented Language Model for Knowledge-aware Protein Encoding (ICML 2025 · similarity 0.85)
- MedMax: Mixed-Modal Instruction Tuning for Training Biomedical Assistants (NeurIPS 2025 · similarity 0.84)
- Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences (ICLR 2025 · similarity 0.84)
- ProtGO: Function-Guided Protein Modeling for Unified Representation Learning (NeurIPS 2024 · 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.