Lipidomics: LPC Signature Predicts Asparaginase Toxicity
Asparaginase cures leukemia, and it sometimes inflames the pancreas on the way. Between 5% and 20% of children treated for acute lymphoblastic leukemia (ALL) develop asparaginase-associated pancreatitis, and no validated test says in advance who is at risk. A JCI Insight study ran lipidomics and proteomics on banked blood from 161 pediatric ALL patients in two Dana-Farber trial cohorts, sampled before any asparaginase dose and again at the end of induction. What separated the groups was a missing response rather than a new lipid: children who later developed pancreatitis failed to raise lysophosphatidylcholine (LPC) the way matched controls did.
Key Takeaways
A blunted LPC rise during induction, not a low baseline level, marked the children who developed asparaginase-associated pancreatitis.
Plasma lipidomics can flag susceptibility to a drug toxicity before symptoms appear, a different problem from diagnosing disease after it presents.
The IL-18 to LPC(18:0) ratio separated cases from controls only inside the very high-risk leukemia subgroup, reaching an AUC of 0.81.
Network structure carried information that mean concentrations missed; only the coregulation analysis reproduced the finding in the validation cohort.
Lipidomics Before the Injury, Not After
The discovery cohort was small and tightly matched: 26 children who developed pancreatitis within nine months, paired with 26 who did not, matched on sex, age, asparaginase formulation, and initial ALL risk. Plasma came from two timepoints, one before asparaginase and one on day 32. Lipid profiling detected 944 species across 14 classes, 853 of which survived quality filtering. The whole lipidome shifted between timepoints, which corticosteroids and vincristine alone would explain, yet cases overlapped controls almost completely at baseline. Separation appeared only after induction, and only in specific classes.
Key Findings
Eight LPC species carried the discovery signal: Mixed-effects modeling found 48 lipid species differing after induction, 46 of them lower in cases, with LPC the dominant class (q < 0.0001).
Higher postinduction LPC meant lower odds of pancreatitis: The class-level odds ratio was 0.22 (95% CI 0.06 to 0.58), with LPC(18:0) strongest at 0.21.
Coregulation broke down before the drug was given: Network analysis placed 11 of 12 LPC species in one module that was not preserved in cases at the pre-asparaginase timepoint.
Cytokine pathways dominated the proteomic side: Of 2,538 Olink proteins passing quality control, 103 changed differently in cases, and all 10 enriched pathways were cytokine-related.
Figure 1. Plasma lipidome of the discovery cohort across asparaginase-containing induction therapy. Panel A shows the sampling scheme for 26 pancreatitis cases and 26 matched controls. Panels B and C summarize the 14 lipid classes and the per-class species counts for the 853 lipids retained after filtering. Panels D through G show class concentrations and principal component analysis at the initial and postinduction timepoints. Panel H compares class-level concentrations, with LPC, lysophosphatidylethanolamine, and cholesteryl esters lower in cases. Adapted from Tsai et al. (2026), JCI Insight.
Why the Validation Cohort Looked Negative
The second cohort had 109 patients, 53 cases and 56 controls, and on a first pass it failed. Neither mixed-effects nor multivariable models found significant case-control differences. Three things had changed: serum instead of plasma, nonfasting collection, and a different trial protocol. Any one of those adds enough variance to bury a mean shift. Network analysis then recovered an LPC-enriched module that was not preserved after induction, with the eight discovery species inside it. That is a replication of structure rather than of effect size, and the distinction matters for anyone planning a prospective assay.
Lipid Biomarkers That Only Work Inside a Subgroup
Across the full validation cohort, IL-18 to LPC(18:0) ratios were flat. Stratifying by protocol-defined leukemia risk changed that: within the very high-risk group, cases had clearly elevated ratios and ROC analysis reached an AUC of 0.81. That subgroup held 16 patients. Leukemic-cell RNA-seq covered 10 of them, and four of seven cases carried Philadelphia-like kinase lesions while no control did, hinting that the signal tracks tumor biology as much as pancreatic vulnerability. Conditional performance like this is familiar from single-species work such as a CSF lipid marker of multiple sclerosis progression, and it means the marker needs a stratification rule attached.
When Sample Volume Decides the Study Design
Proteomics ran only in the validation cohort, because too little discovery-cohort plasma remained to do both, so the LPC-cytokine link could never be tested where the lipid signal was strongest. That tradeoff is routine in pediatric trial biobanks, where aliquot volume gets fixed years before anyone knows what to measure. Pulling lipids and proteins from one small aliquot, which is what the Omni-MS workflow at Dalton was built for, keeps lipid profiling and protein readouts anchored to the same specimens rather than split across cohorts, and that single-injection design is what multi-omics biomarker discovery programs need once samples get this precious.
Frequently Asked Questions
What is lipidomics used for in drug development? Lipidomics measures hundreds to thousands of individual lipid species per sample, usually by liquid chromatography coupled to mass spectrometry. Teams use it to characterize mechanism, track pharmacodynamic response, and find toxicity or response markers that protein and transcript panels miss.
Can lipidomics predict drug side effects before they occur? Sometimes, when the sampling design supports it. Here the predictive variable was the change in LPC between a pre-dose and a post-dose sample, not the level in any single draw. That works only if paired specimens were banked before anyone suspected a toxicity signal.
How much plasma is needed for combined lipid and protein profiling? Most lipid panels run on tens of microliters of plasma or serum, and affinity proteomics panels need a similar amount. Splitting one aliquot across both platforms is usually feasible; the harder constraint is matched timepoints from every participant.
Conclusion
The direction holds up: children who fail to mount an LPC rise during induction are the ones who go on to pancreatitis, and the pattern survived a change of cohort, sample matrix, and analytical method. It extends the same group's earlier metabolomic work in these trials (Tsai et al., 2023), which flagged retinoid depletion in the same graded way. The IL-18 ratio isn't a test yet; 16 very high-risk patients cannot validate anything. For toxicity biomarker programs, the practical lesson is to bank paired pre-dose and post-dose specimens from day one, because the informative quantity was a delta.
Related Reading
See how we run these analyses in one lab: Dalton's multi-omics CRO services.
Citation
Tsai, C.-Y., Bo, N., Tran, T. H., Abu-El-Haija, M., Swaminathan, G., Lee, B., Ghandikota, S., Wen, L., Théorêt, Y., Mittelman, S. D., Ladas, E. J., Jegga, A. G., Silverman, L. B., Ding, Y., & Husain, S. Z. (2026). A lipid-immune network signature defines susceptibility to asparaginase-associated pancreatitis. JCI Insight, 11(12), e202662. https://doi.org/10.1172/jci.insight.202662
Note
This blog post summarizes findings from the above-cited research. Figures are adapted from the original publication. For full details, please refer to the source article.
By Seungjun Yeo, CEO at Dalton Bioanalytics. Specializing in multi-omics mass spectrometry for drug discovery and biomarker research.