Skip to content

LIPID Imputation

Summary

Dakic, Wu, Wang, et al. (2024) present an elastic-net-based framework for harmonising plasma lipidomic datasets acquired at different chromatographic resolutions, so that lipid species unmeasured in an older, lower-resolution study can be imputed from a newer, higher-resolution reference cohort. The method was demonstrated by imputing 413 lipid species into the LIPID trial (342 measured species) using models trained on the AusDiab cohort (747 species), then validated in an ethnically distinct third cohort, the San Antonio Family Heart Study (SAFHS).

Motivation

Lipidomic platforms have improved chromatographic and mass-spectrometric resolution over time, so a lipid species reported as a single measurement on an older platform may resolve into two or more isomeric species on a newer one. This makes it difficult to align datasets built at different times or on different platforms, forcing analysts to either restrict analysis to the lipid species common to all datasets (losing power/resolution) or discard older datasets entirely (losing sample size and long-term follow-up).

Pipeline

  1. Nomenclature alignment: each target-dataset lipid is mapped to either a single best-matching reference lipid or, when chromatographic resolution differs, a composite measure (linear sum) of two or more reference lipid species. In the LIPID-trial/AusDiab pairing, 225 lipids mapped one-to-one, 82 lipids mapped to composite sums of 176 AusDiab lipids, and 35 species had no adequate match, yielding 307 lipid measures usable as predictors.
  2. Discrepancy detection: partial-correlation vectors (each lipid's conditional correlation with every other lipid, controlling for the rest) are compared between reference and target datasets. Lipids whose reference-vs-target partial-correlation distance exceeds ~2–2.5 median absolute deviations from the median are sequentially removed as unreliable predictors (13 discordant lipids removed between AusDiab and LIPID; 12 between AusDiab and SAFHS).
  3. Model construction: elastic-net regression (glmnet, α tuned over {0, 0.1, 0.25, 0.5, 0.75}, λ selected by cross-validation) predicts each unmeasured lipid from the remaining matching lipid species plus age, sex, BMI, and lipid-lowering treatment status.
  4. Transferability assessment: for lipids measured in both datasets, models trained in the reference are applied to both reference and target data, and prediction accuracy (Pearson correlation between observed and predicted concentration) is compared across datasets to confirm the model generalizes rather than overfitting the reference cohort.
  5. Target imputation and accuracy filtering: models are first cross-validated within the reference cohort; only lipids achieving reference-study prediction correlation > 0.6 are then applied to impute the missing lipid in the target study (26 of the candidate "missing" lipids were excluded on this basis for LIPID; a corresponding filter left 304 usable predictors and 384 imputable lipids for SAFHS).

Results

  • AusDiab → LIPID trial: after removing discordant predictors, 413 of the missing lipid species were imputed into the LIPID trial (n = 5,991, two timepoints) using 294 concordant matching lipid species as predictors. Prediction accuracy in the LIPID trial closely tracked accuracy achieved in the AusDiab reference for most lipids.
  • Robustness to differential treatment exposure: imputation accuracy was comparable between placebo- and pravastatin-randomised LIPID participants, and between baseline and follow-up timepoints, despite pravastatin having lipidome-specific effects — indicating the lipid correlation structure itself, not clinical covariates, drives prediction accuracy.
  • Robustness to reduced predictor/sample size: reducing the predictor set to 75–90% of matching lipids caused only small declines in accuracy/precision; ≥1,000 training observations were sufficient for stable models, and even 200 observations produced reasonably stable predictions for most lipids.
  • Cross-ethnicity validation (SAFHS): despite SAFHS being an entirely Mexican-American cohort (versus predominantly European-ancestry AusDiab), imputation models transferred with strong accuracy across the two populations, supporting the claim that plasma lipid correlation structure is stable across studies with substantially different demographics, medication exposure, and cardiovascular risk profiles.

Availability

Source code: github.com/BakerMetabolomics/LIPID_imputation. AusDiab and LIPID summary statistics are provided as supplementary files; SAFHS anthropometric/genomic data are available via dbGaP, with lipidomic data available on request under a material transfer agreement.

See Also

  • Human Lipidome Genetics — genetic mapping of the circulating lipidome that this cross-platform harmonisation approach helps integrate across cohorts of differing lipidomic resolution.
  • IWAE Cross-Platform Metabolomics Imputation — a deep-learning approach to the analogous cross-platform harmonisation problem for broader (non-lipid) metabolomics platforms.

Citations

[1] Dakic, A., Wu, J., Wang, T., Huynh, K., et al. (2024). Imputation of plasma lipid species to facilitate integration of lipidomic datasets. Nature Communications, 15, 1540.