Research map / Vision-language and multimodal / Cross-modal learning
Unified Insights: Harnessing Multi-modal Data for Phenotype Imputation via View Decoupling
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 MPI, a dual-view graph framework that quantizes proteomics and metabolomics data with residual quantization and transfers biological-view knowledge to a phenotype-view GCN via cross-view contrastive distillation for phenotype imputation on UK Biobank ADRD cohort.
Problem
Imputing missing phenotypes in EHRs while leveraging noisy, heterogeneous, and randomly missing multi-modal biological data (proteomics, metabolomics).
Approach
Decouples biological and phenotype views into two separate graphs: biological factors from residual-quantized biological data form a patient-factor graph, while a patient-phenotype co-occurrence graph is encoded separately; a contrastive knowledge distillation loss transfers biological knowledge from the biological-view GCN (teacher) to the phenotype-view GCN (student).
Key results
- MPI reaches Hit@10 of 38.74 and MRR of 19.28 on 100% of the UK Biobank dataset, the best among all compared methods (strongest baseline GIN at 36.61 Hit@10).
- On 30% of the data MPI improves Hit@10 to 28.87 versus 26.23 for the best baseline M3Care.
- Ablation shows removing biological data (V1) drops Hit@10 from 38.74 to 36.49 on the full dataset, and the separate-graph design outperforms a single-graph variant (V4 at 37.68).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MPI instead of the strongest baseline | phenotype imputation · UK Biobank 30% | MRR: 11.87± 0.25 -> 13.22± 0.17 | Helped |
| use MPI instead of the strongest baseline | phenotype imputation · UK Biobank 50% | MRR: 13.77± 0.28 -> 14.83± 0.15 | Helped |
| use MPI instead of the strongest baseline | phenotype imputation · UK Biobank 70% | MRR: 15.75± 0.22 -> 17.44± 0.41 | Helped |
| use MPI instead of the strongest baseline | phenotype imputation · UK Biobank 90% | MRR: 17.36± 0.06 -> 18.63± 0.22 | Helped |
| use MPI instead of the strongest baseline | phenotype imputation · UK Biobank 100% | MRR: 17.99± 0.23 -> 19.28± 0.19 | Helped |
| add biological data and contrastive knowledge distillation to the patient-phenotype graph (ablation) | MPI · UK Biobank 30% | MRR: 11.98 -> 13.22 | Helped |
| add biological data and contrastive knowledge distillation to the patient-phenotype graph (ablation) | MPI · UK Biobank 100% | MRR: 17.73 -> 19.27 | Helped |
| add metabolomics alongside proteomics with contrastive knowledge distillation (ablation) | MPI · UK Biobank 30% | MRR: 13.09 -> 13.22 | Helped |
| add metabolomics alongside proteomics with contrastive knowledge distillation (ablation) | MPI · UK Biobank 100% | MRR: 19.08 -> 19.27 | Helped |
| add proteomics alongside metabolomics with contrastive knowledge distillation (ablation) | MPI · UK Biobank 30% | MRR: 13.11 -> 13.22 | Helped |
| add proteomics alongside metabolomics with contrastive knowledge distillation (ablation) | MPI · UK Biobank 100% | MRR: 19.02 -> 19.27 | Helped |
| separate the biological and phenotype views and use contrastive knowledge distillation instead of a single graph without distillation (ablation) | MPI · UK Biobank 30% | MRR: 12.32 -> 13.22 | Helped |
Surprising findings
- Graph-based multi-modal models (GRAPE, MUSE) sometimes underperform naive GraphSage/GIN that simply concatenate biological features into patient nodes.
Limitations
- Only two biological modalities (proteomics and metabolomics) are considered.
- Evaluation restricted to a single cohort of Alzheimer's disease and related dementias patients from UK Biobank.
- Uses GCN as backbone and does not explore more advanced GNNs.
Experimental setup
Datasets: UK Biobank
Benchmarks: UK Biobank ADRD cohort phenotype imputation
Metrics: Hit@10 Hit@20 Hit@50 MRR
Baselines: CMAE SMIL GraphSage GIN GRAPE M3Care MUSE
Builds on: residual quantization graph convolutional networks teacher-student knowledge distillation
Closest papers by meaning
- Causal Representation Learning from Multimodal Biomedical Observations (ICLR 2025 · similarity 0.86)
- When Are Multimodal Predictions Biologically Supported? A Diagnostic Evaluation Framework (NeurIPS 2026 · similarity 0.84)
- Prototypical Information Bottlenecking and Disentangling for Multimodal Cancer Survival Prediction (ICLR 2024 · similarity 0.84)
- DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency (AAAI 2024 · similarity 0.83)
- CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation (NeurIPS 2026 · similarity 0.83)
- BIOMEDICA: An Open Biomedical Image-Caption Archive, Dataset, and Vision-Language Models Derived from Scientific Literature (CVPR 2025 · similarity 0.83)
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.