LEOPARD: missing view completion for multi-timepoint omics data via representation disentanglement and temporal knowledge transfer.

Summary

LEOPARD introduces a representation-disentanglement approach to complete missing omics views in longitudinal datasets and consistently outperforms common imputation methods. Validated across four real-world cohorts spanning 3 days to 14 years, it improves detection of age-associated metabolites, eGFR-associated proteins, and CKD prediction.

Key Findings

• Introduces LEOPARD, which disentangles content and temporal representations to impute missing omics views.
• Validated on four real-world datasets (MGH COVID, KORA) covering 3 days to 14 years, outperforming missForest, PMM, GLMM, and cGAN.
• LEOPARD-imputed data showed highest agreement with observed data for age-related metabolites, eGFR-associated proteins, and CKD prediction.

Clinical Implications

While methodological, LEOPARD can enhance biomarker discovery, patient stratification, and trajectory modeling in diabetes, obesity, and endocrine disorders by maximizing use of incomplete longitudinal omics.

Limitations

• No prospective clinical outcome validation directly tied to imputed features.
• Generalizability to omics modalities beyond those tested remains to be established.

Future Directions

Prospectively validate LEOPARD-driven biomarkers in endocrine cohorts (e.g., T2D, obesity), integrate EHR and imaging, and extend to multimodal (omics-clinical) harmonization.