A lipid-related transcriptomic signature for distinguishing Kawasaki disease from febrile infections: a Mendelian randomization and machine-learning study.
Wang S, Luo G, et al. • Frontiers in immunology • 2026
A lipid-related 12-gene transcriptomic signature derived via LASSO machine learning achieved an external AUC of 0.851 for distinguishing Kawasaki disease from febrile infections, though no lipid or immune trait survived Mendelian randomization correction and prospective validation is required before clinical use.
Key Findings
Results
No lipid, metabolic, or immune-cell exposure survived multiplicity correction in bidirectional two-sample Mendelian randomization analysis against two independent KD GWAS cohorts.
422 candidate exposures were screened, comprising 419 lipid/metabolic traits and three immune-cell phenotypes.
The minimum Benjamini-Hochberg q-value across all 422 tests was 0.979, indicating no causal relationship survived correction.
Nine discovery-stage nominal hits were evaluated in a second KD GWAS; eight had directionally concordant point estimates but none reached unadjusted P < 0.05 in the replication cohort.
Results indicate no statistically robust causal link between lipid/metabolic traits and KD was established by MR.
Results
A frozen LASSO-derived 12-gene penalized model achieved moderate external discrimination between KD and febrile controls with an AUC of 0.851.
The model was trained on differentially expressed genes from a training cohort comparing KD with rigorously classified febrile controls, intersected with a lipid-metabolism-related GeneCards gene set, yielding 459 candidates.
459 candidates were evaluated across 130 machine-learning combinations; a full-training LASSO fit at lambda.1se yielded the 12-gene panel.
The frozen penalized model was applied unchanged to the independent cohort GSE68004 without cross-cohort batch correction.
External AUC was 0.851 (95% CI 0.768–0.934).
Authors note the 12-gene set 'represents one candidate realization of a partition-sensitive modeling pipeline' and that recalibration and prospective validation are required before clinical use.
Results
CIBERSORT immune deconvolution indicated a higher estimated neutrophil fraction in KD compared to febrile controls.
Immune microenvironment estimation was performed using CIBERSORT.
A higher estimated neutrophil fraction in KD was concordant with a mature-neutrophil marker score derived from the same expression matrix.
This finding is consistent with the known innate immune-driven vascular damage characteristic of KD vasculitis.
Results
CETP, KREMEN1, and CCL23 expression levels were significantly lower in convalescent KD samples after IVIG treatment compared to acute disease.
Longitudinal tracking of acute-to-convalescent expression changes was performed.
All three genes — CETP, KREMEN1, and CCL23 — showed reduced expression in convalescent samples after IVIG treatment at FDR < 0.0001.
The decline in expression post-IVIG suggests these genes may be involved in the acute inflammatory process of KD.
Results
TNF-α stimulation of HUVECs induced transcriptional upregulation of CETP, KREMEN1, and CCL23, along with canonical endothelial activation markers.
Selected markers were assessed in TNF-α-stimulated human umbilical vein endothelial cells (HUVECs).
CETP, KREMEN1, and CCL23 were induced in TNF-α-treated HUVECs.
Canonical endothelial activation markers VCAM-1, ICAM-1, and IL-8 were also induced concurrently.
Authors explicitly characterize these findings as 'transcriptional and hypothesis-generating.'
Results
KREMEN1 was identified as a gene warranting separate biological investigation based on its marginal, longitudinal, and immune-association findings.
KREMEN1 appeared in the 12-gene panel and showed reduced expression post-IVIG (FDR < 0.0001).
KREMEN1 was induced in TNF-α-stimulated HUVECs alongside endothelial activation markers.
Authors recommend KREMEN1 be studied on the basis of 'marginal, longitudinal, and immune-association findings rather than its selection stability,' acknowledging the model's partition sensitivity.
Methods
The study identified 459 lipid-metabolism-related differentially expressed gene candidates by intersecting KD versus febrile control transcriptomic data with a GeneCards lipid-metabolism gene set.
Differentially expressed genes were derived from a training cohort comparing KD patients with rigorously classified febrile controls.
Intersection was performed with a 'broad lipid-metabolism-related GeneCards set.'
These 459 candidates served as the input for the 130 machine-learning pipeline combinations evaluated.
What This Means
Kawasaki disease (KD) is a serious inflammatory condition in children that can damage blood vessels and the heart, but it is difficult to diagnose because there is no definitive test and it looks similar to common childhood febrile illnesses. This study tried to determine whether known disturbances in fat (lipid) metabolism seen during KD are causally connected to the disease, and whether a gene-based blood test could help tell KD apart from other fevers. Using a genetic analysis technique called Mendelian randomization — which uses inherited genetic variants to test for cause-and-effect relationships — the researchers found no convincing evidence that lipid or immune traits causally drive KD. This means the lipid changes seen in KD may be a consequence of inflammation rather than a cause.
Despite the negative causal findings, the researchers used machine learning to build a 12-gene signature that could distinguish KD from febrile infections with moderate accuracy, achieving an AUC (a measure of diagnostic performance, where 1.0 is perfect) of 0.851 in an independent dataset. Three genes in particular — CETP, KREMEN1, and CCL23 — dropped significantly in patients after they received standard KD treatment (IVIG), and were also activated when blood vessel cells were exposed to TNF-α, a key inflammatory molecule. This suggests these genes may play a role in KD's vascular inflammation, though the authors caution that these cell-culture findings are preliminary and hypothesis-generating.
This research suggests that a lipid-related gene panel could potentially help doctors distinguish KD from other childhood fevers, but the authors are explicit that the model needs further refinement and real-world prospective testing before it could be used clinically. The gene KREMEN1 is highlighted as particularly interesting for future biological study. Overall, this work advances understanding of KD's molecular biology while honestly acknowledging the limitations of the modeling approach and the lack of proven causal links between lipid metabolism and KD.
Wang S, Luo G, Ji Z, Zhou G, Pan S. (2026). A lipid-related transcriptomic signature for distinguishing Kawasaki disease from febrile infections: a Mendelian randomization and machine-learning study.. Frontiers in immunology. https://doi.org/10.3389/fimmu.2026.1851332