A systems approach points to a therapeutic role for retinoids in asparaginase-associated pancreatitis.

To identify factors that could impact AAP, we used a systems approach ( Fig. 1 ) that began with Connectivity Map (CMap) analysis ( 8 ), which is a method to compare transcriptional expression data from a disease or phenotype (using differentially expressed genes [DEGs] or “gene signatures”) with transcriptional data from in silico treatment with perturbagens, here a library of both FDA-approved and investigational small molecules. We identified a set of DEGs from asparaginase-treated leukemic cells ( GSE94289 ) ( 9 ), as well as from the pancreas of mice that were chemically induced caerulein-induced pancreatitis (CIP) ( GSE109227 ) ( 10 ). We performed a CMap analysis ( Fig. 2A ) against the National Institutes of Health (NIH) Library of Integrated Network-based Cellular Signatures (LINCS) L1000 Touchstone reference dataset ( 11 ) and identified several candidate compounds with a summary connectivity score ≤90 ( Fig. 2B – C , Data file S1 ). These hits included AC-55649 (4’-octyl-[1,1’-biphenyl]-4-carboxylic acid), TTNPB (tetrahydro-5,5,8,8-tetramethyl-2-naphthalenyl)-1-propenyl]benzoic acid), acitretin, tretinoin, retinol, fenretinide, AM-580 (tetrahydro-5,5,8,8-tetramethyl-2-naphthalenyl)carboxamido]benzoic acid), alitretinoin, and retinyl. Among the top-ranked molecules were compounds belonging to the retinoids ( Fig. 2D ). Other classes of compounds included glucocorticoid receptor agonists, calcium channel blockers, sirtuin activators, serotonin receptor antagonists, estrogen receptor agonists, and transforming growth factor beta receptor inhibitors ( Data file S1 ). Several of these classes of compounds have been shown to have a protective role in AP models ( 12 ); however, whether these compounds could also be protective in AAP is unclear. There is a potential discrepancy between the finding of an effect of glucocorticoid receptor agonists alongside asparaginase and the lack of observed differences in pancreatitis risk. Although most ALL trial protocols use both drugs, the role of glucocorticoid receptor agonists in AAP requires further investigation. Because the plurality of hits from the perturbagen analysis was of vitamin A and its analogs, we performed an additional systems analysis to validate the findings using two large real-world datasets: the E-HRs from more than 67 healthcare organizations across the US, and drug-induced AE reports data from the FDA. E-HRs were derived from TriNetX, a global federated health research network providing access to statistics on more than 105 million E-HRs ( 13 ). Using E-HRs, we performed a comparative analysis of longitudinal outcomes of 7,689 patients: 6,557 patients had taken asparaginase without vitamin A and 1,132 patients had taken asparaginase concomitantly with vitamin A. The vitamin A medications and their routes of administration are provided in table S1 . The two patient cohorts underwent propensity score matching for age, sex, race, and ethnicity as covariates. The rationale for ensuring that the cohorts were age-matched is that older age is a risk factor for AAP ( 5 , 7 ). After matching, there were 1,132 patients each in the two cohorts (with or without vitamin A). Patient characteristics of the two cohorts are provided in Table 1 . Pancreatitis was identified using ICD-10 (International Classification of Disease 10 th revision) diagnosis codes ( table S2 ) and counted as an outcome in the dataset if diagnosed within five years after the first dose of asparaginase was administered ( Fig. 3A ). Because asparaginase treatment is given over 2–3.5 years, depending on the protocol ( 14 ), we wanted to ensure the capture of the full duration of asparaginase exposure plus a tail of 1.5 more years. To control for lead time bias, AP monitoring began at the time of the first asparaginase exposure in both groups. The asparaginase without vitamin A cohort had 38 patients (3.44%) diagnosed with AP, which is within the reported range of pancreatitis (2–10%) ( 6 ). There was a 60% reduction in AP among the asparaginase with vitamin A cohort, with only 15 (1.4%) reported patients (OR: 2.43; 95% CI: 1.33–4.44; P < 0.005; Fig. 3B ). Pancreatitis is defined by the presence of two of the following three clinical features: typical abdominal pain, biochemical elevations in serum amylase or lipase greater than three times the upper limit of normal, and characteristic changes on imaging of the pancreas ( 15 ). TriNetX allowed us to investigate the absolute values of serum amylase and lipase. However, the upper limit of normal reference values for each of the two indices is specific to each institution’s laboratory, and this information is not available. Therefore, we were unable to calculate the fold increase above the upper limit of normal, which is required for a biochemical diagnosis of pancreatitis. Nonetheless, as described in a previous report ( 16 ), we assessed what has been described as “pancreatic cellular injury” by using thresholds of hyperamylasemia (≥140 U/L) and hyperlipasemia (≥160 U/L), which are thresholds shared by many laboratories in the US ( 17 ). Using these thresholds, serum amylase was elevated in 28 patients (2.5%) in the asparaginase group compared to 10 patients (0.9%) in the asparaginase and vitamin A group (reduced by 64%; OR: 2.73; 95% CI: 1.32–5.66; P < 0.005). Lipase was elevated in 61 patients (5.6%) in the asparaginase group compared to 22 patients (2.1%) in the asparaginase and vitamin A group (reduced by 62.5%; OR: 2.74; 95% CI: 1.67–4.49; P < 0.005; Fig. 3B ). The elevations in serum amylase and lipase were not used for AP diagnosis in the current study, as they did not meet the definition of three times the upper limit of normal. It is also conceivable that enzyme elevations may occur without clinical pancreatitis because each of these enzyme elevations is known to occur in non-specific scenarios independent of pancreatitis. To address this important question, we adjusted for several of these potentially confounding diagnoses (intestinal obstruction/ischemia; peptic ulcers; appendicitis; and pancreatic cancer) and found that our overall AP results did not change after adjustment (OR: 1.996, P < 0.05). To complement the E-HR signals, we analyzed the FAERS database for evidence of pancreatitis risk among AE reports of pancreatitis with asparaginase exposure, vitamin A exposure, or combined asparaginase and vitamin A exposure. The baseline characteristics of the three cohorts and medications used to define the inclusion and exclusion criteria for each cohort are provided in table S3 and S4 , respectively. We performed a comparative analysis of the differential risk of pancreatitis signature events ( table S5 ) calculated as safety signals. Safety signals > 0 represent a strong association between a drug and AEs, whereas signals < 0 represent anticorrelation and may signify a protective role ( 18 ). The groups were divided as follows: asparaginase without vitamin A exposure ( n = 20,186), vitamin A without asparaginase exposure ( n = 281,917), and asparaginase with vitamin A exposure ( n = 15). It is important to note that unless vitamin A was explicitly stated in the patient report, we excluded patients reporting multivitamin intake because the presence of several vitamins could potentially make our results uninterpretable. In the asparaginase only group, the average safety signal among the seven AEs for pancreatitis combined was 3.75 ( Fig. 3C – D ). The safety signals for each AE were as follows: pancreatitis (3.06); acute pancreatitis (3.71); necrotizing pancreatitis (5.29); relapsing pancreatitis (2.74); hemorrhagic pancreatitis (5.47); elevated amylase (2.37); and elevated lipase (3.63). However, the risk of these pancreatitis signature events was lower in the vitamin A group, in which the safety signals were as follows: pancreatitis (−0.28); acute pancreatitis (−0.94); necrotizing pancreatitis (−1.48); relapsing pancreatitis (−1.05); hemorrhagic pancreatitis (−3.03); elevated amylase (−1.25); and elevated lipase (−0.98) ( Fig. 3C – D ). Statistical analysis was not feasible for the group of identified patients with asparaginase with vitamin A due to the small sample size ( n = 15). However, none of these 15 patients in the third descriptive dataset reported pancreatitis signature events ( table S5 ). Both the E-HRs and FAERS data analyses suggested vitamin A as a protective factor in AAP. To investigate the relationship of vitamin A metabolism in patients receiving asparaginase who developed pancreatitis, we performed a case-control metabolomic study from a discovery cohort of 50 children and adolescents with ALL enrolled in the Dana-Farber Cancer Institute (DFCI) ALL clinical trial protocols 05–001 ( 19 ) and 11–001 ( 20 ). The exact dosing and route of administration of asparaginase differs among clinical protocols ( 14 ). Four asparaginase formulations were used in the two DFCI protocols from which the subjects were derived ( table S6 and Supplementary Materials and Methods ). Twenty-four case subjects were selected based on the development of pancreatitis within 9 months from the start of induction therapy (termed as cases). The median time to pancreatitis was 3.68 months (interquartile range: 3.58 months, Fig. 4A ). Twenty-six control subjects were selected because they did not develop pancreatitis during the same evaluation period (termed as controls). The controls were matched for sex, asparaginase formulation, age, and initial ALL risk group. The latter two variables are risk factors for AAP in children and adolescents with ALL ( 5 , 7 ). The body-mass index (BMI) and final ALL risk were similar between the groups ( table S7 ). Figure 4A shows a schematic study design of the discovery cohort, the two collection time points for the plasma samples and the food frequency questionnaires (FFQs), and the histogram of time to develop pancreatitis. The initial plasma sample was collected within the first week of induction therapy (day 1 for subjects enrolled in the DFCI 05–001 protocol or day 7 for subjects enrolled in the DFCI 11–001 protocol). The second plasma sample was collected on day 32 (post-induction). We also evaluated the dietary intake of vitamin A ( table S8 ) in subjects enrolled in the DFCI 05–001 protocol ( 21 ). The time points were chosen to ascertain plasma samples and dietary information both before the first dose of asparaginase and within four weeks of a single dose of asparaginase. The latter time point should be able to identify the lingering effect of asparaginase (25 days after administration) on reducing asparagine in the circulation ( 22 ). In most cases, these time points were before the development of pancreatitis. Plasma samples were subjected to global metabolomics screening. When comparing the cases with matched controls at each of the two time points, there were no significant differences among the metabolites prior to induction therapy using a threshold of significance of q < 0.1 ( fig. S1A ). Consistent with the lack of differences among the metabolites, there were also no differences in the dietary intake of vitamin A and carotenoids between both groups over the 30 days before induction ( table S9 ). After induction therapy, there were still no differences observed among the metabolites ( fig. S1B ). Next, we examined changes in metabolites over the two time points paired with each subject ( Fig. 4B ). The three carotene diol isomers were significantly reduced in the cases ( q < 0.1). The exact identities of the three carotene diol isomers were not determined because of a lack of pure reference standards. However, based on their abundance in humans, the possibilities include lutein, meso-zeaxanthin, and zeaxanthin. The changes in each of the individual carotenoids and retinol between the initial and post-induction time points are shown in Fig. 4C . The data demonstrated that the change in the three plasma carotene diol isomers increased, on average, in the controls after induction therapy, whereas there was no increase among the cases. The beta-cryptoxanthin ratio did not change among the controls but was reduced among the cases. The retinol ratio increased after induction in both groups, but there was less of an increase in the cases. Pairwise Spearman’s rank correlation analysis of carotenoids and retinol showed a high degree of positive correlation with each other ( Fig. 4D ). Pathway enrichment analysis confirmed carotenoid and retinoid metabolism as the top hit ( Fig. 4E ). Carotene diol isomers and beta-cryptoxanthin were positively correlated with each other in both groups in the network correlation ( Fig. 4F ). In the control group, (14 or 15)-methylpalmitate (a17:0 or i17:0) was positively correlated with carotene diol ( 1 ), and carnitine was negatively correlated with beta-cryptoxanthin. In the case group, 2-hydroxyphytanate, 5-methyluridine (ribothymidine), 4-methyl-2-oxopentanoate, and allantoin were positively correlated with several carotenoids, whereas quinolinate was negatively correlated with beta-cryptoxanthin and retinol. Using carotene diol ( 3 ) as the index metabolite, we performed additional metabolite correlations of the four control subjects with ‘low’ carotene diol, who, however, did not develop pancreatitis (termed ‘controls low’). Conversely, we examined the seven case subjects with ‘high’ carotene diol, who, however, still went on to develop pancreatitis (termed ‘cases high’) ( fig. S2 ). Tridecenedionate (C13:1-DC) was reduced in the controls low subjects ( P < 0.0001). 4-Methyl-2-oxopentanoate was increased ( P < 0.01), whereas C-glycosyltryptophan was reduced ( P < 0.01) in the cases high subjects. The role of these correlated metabolites in AAP remains to be elucidated. The dietary intake of vitamin A and beta-cryptoxanthin was significantly reduced in cases compared to controls during induction therapy ( P < 0.05, Fig. 5 , table S9 ). The median retinol activity equivalents (RAE) intake during the 4 weeks of induction was 656.92 mcg/day among the controls, and was reduced by 34.6% to a median of 429.40 mcg/day among the cases, which is just above the recommended dietary allowance (RDA) of 400 mcg/day for the 4–8 year-old age group ( 23 ). This age group constituted the bulk of the subjects in the case-control study. Plasma retinol at the end of induction correlated with dietary intake of retinol and RAE during the induction period ( fig. S3 ). Overall, these results suggest that lower vitamin A intake during the induction period could predispose children and adolescents with ALL to pancreatitis. Next, we examined whether the cases had reductions in other fat-soluble vitamin metabolites identified by the global metabolomic screening. Three metabolites were detected in the vitamin E group, but no vitamin D or K metabolites were detected. To evaluate the overall micronutrient changes, we also detected 12 water-soluble vitamin metabolites, three in the vitamin C group and nine in the vitamin B group. We found no differences in the changes in these fat-soluble or water-soluble vitamin metabolites between the cases and controls after induction therapy ( fig. S4 ). Whereas most cases developed pancreatitis three months after the start of induction (median of 4.47 months), five subjects developed “early pancreatitis,” defined as the onset of pancreatitis within the period of induction therapy ( Fig. 4A ). The timing of pancreatitis in these subjects ranged from 21 to 29 days after the start of induction. These subjects were also likely to have had an acute flare of pancreatitis or were in the immediate recovery phase of pancreatitis during the second blood collection. Comparing the cohorts now as three groups—controls, non-early cases, and early cases—there was a greater relative reduction in carotenoids and retinol in early cases ( fig. S5A ). Although the early cases were the major contributors to these differences, after excluding the early cases, carotenoids were still among the top differentially expressed metabolites ( fig. S5B ), and the carotenoid and retinoid metabolism remained among the top pathway hits in the enrichment analysis ( fig. S5C ). These results indicate that non-early cases also contributed to these differences. We constructed logistic regression models to differentiate between the cases and controls. The model incorporated BMI, final ALL risk, and the top five significant metabolites identified from each sampling condition, including the initial time point, the post-induction time point, and the ratio of change between the two time points ( table S10 ). Age, sex, and initial ALL risk were not entered into the models because the controls were matched to cases based on these three variables. The adjusted R 2 for the models in each of the three sampling conditions was > 0.5. The importance of the top five significant metabolites ( P < 0.05) in these models was underscored by the finding that, without their input, the use of BMI and final ALL risk alone yielded an adjusted R 2 of 0.042. In addition to the logistic regression models, we performed a second analysis to predict the future development of pancreatitis by excluding the five early cases and using the remaining 19 non-early cases in the comparison scenario. We ran a 5-fold cross-validation and discovered that the average area under the curve (AUC) was ≥ 0.75 ( table S11 ). To validate the findings from the discovery cohort, a larger external validation cohort of 109 subjects enrolled in DFCI protocol 16–001 was identified. Fifty-three cases who went on to develop pancreatitis and 56 controls who did not develop pancreatitis were included. Four serum samples were obtained from each subject in the validation cohort. The first two samples of the validation cohort were obtained at the same time points as those of the discovery cohort (time point 1 and 2). The next two samples were obtained later in the clinical trial to monitor changes in metabolites over a 6-month period. Time point 3 samples were obtained prior to the second asparaginase dose, and time point 4 samples were obtained 14 or 28 days after the second asparaginase dose. The second asparaginase dose was administered at the beginning of the central nervous system (CNS) phase. We examined the changes in carotenoids and retinol over time for each subject. There was no difference in carotenoids between cases and controls between time points 2 and 1. However, carotene diol (1) decreased significantly between time points 3 and 1 ( P < 0.05). Furthermore, carotene diol (1) and carotene diol (2) decreased significantly between timepoints 4 and 1 ( P < 0.05, table S12 – 14 ). By pairwise Spearman’s rank correlation analysis of carotenoids and retinol, similar to the findings in the discovery cohort, each of the carotenoids in the validation cohort had a strong positive correlation with one another ( fig. S6A ). There was, however, less of a correlation between the carotenoids and retinol. Using the metabolite information obtained from time points 1 and 4, pathway enrichment analysis confirmed carotenoid and retinoid metabolism as the top hit ( fig. S6B ). The discrepancy between the discovery and validation cohorts, at the exact time point during which carotene diols were reduced, could be due to regional differences in dietary habits and the use of modified protocols, albeit from the same consortium. Nonetheless, the analysis from the validation cohort supports the key findings from the discovery cohort, which is that, compared with the controls, the AAP cases had reduced circulating carotene diols, and the carotenoid and retinoid pathway was highly enriched. Induction chemotherapy for ALL consists of several drugs in combination ( table S15 ) ( 24 ). As asparaginase is not administered as a single agent in any standard protocol, it is unclear whether asparaginase alone reduces circulating retinoids. For this reason, we used a murine model of asparaginase exposure. Mice received a daily intraperitoneal injection of either native E. coli L-asparaginase or phosphate buffered saline as the vehicle for eight days. Neither weight nor oral intake was affected by this method of asparaginase administration ( 25 ). Serum retinol and retinyl ester, the most abundant retinoid forms in the body, were measured using HPLC ( 26 ). Serum retinyl ester was mostly undetectable. However, serum retinol was reduced by 50% in asparaginase-exposed mice compared with that in vehicle-treated controls ( P < 0.05; Fig. 6A ). Next, we assessed the effects of asparaginase exposure on retinol in the liver, the main storage site of retinol. After eight days of asparaginase exposure, liver retinol was reduced by 37.4% in asparaginase-exposed mice compared with that in vehicle-treated controls ( P < 0.05; Fig. 6B ). Asparaginase as a single agent, reduces retinol not only in the circulation, which provides retinol to the pancreas, but also in the liver. In AP, proinflammatory cytokines such as interleukin 6 and interleukin 1 beta are expressed by pancreatic acinar cells via nuclear factor-kB (NF-kB) activation ( 27 ). We examined the effects of all-trans retinoic acid (ATRA) and 9-cis retinoic acid (9cRA), two naturally occurring retinoic acids, on the mRNA abundance of interleukin 6 ( Il6) and interleukin 1 beta ( Il1b) in primary mouse acinar cells. Treatment with a high concentration of caerulein, which induces pancreatitis in vivo , increased the mRNA abundance of Il6 and Il1b ( Fig. 7A ). Pretreatment with retinoic acid prevented the upregulation of Il6 and Il1b . In contrast, Il6 and Il1b abundance increased when retinoic acid signaling was blocked by BMS-493, a pan retinoic acid receptor inverse agonist. A similar protective pattern of retinoic acid treatment was observed with the exposure of pancreatic acinar cells to asparaginase ( Fig. 7B ). Here, we demonstrate the protective effect of retinoic acids in reducing acinar inflammatory signals, whereas blocking the retinoic acid receptor amplifies inflammatory gene expression.

The study design for the retrospective case-control metabolomics screen and animal experiments are provided below, according to the STROBE ( 53 ) and ARRIVE ( 54 ) guidelines, respectively. The objective of case-control metabolomic screening was to profile the early changes in plasma metabolites in subjects with ALL who developed pancreatitis. Sample sizes were determined by power calculations that accounted for the false discovery rate (FDR) using the R package (ssize.fdr). We assumed the following: This distribution of the differential effects (on a log2-transformed scale) between the cases and controls (across all measured metabolites) follows a normal distribution, with a mean of 1.3 and a standard deviation of 1; the variance of metabolites follows an Inverse Gamma ( 3 , 1 ) distribution; approximately 5% of the total measured metabolites are differential; and the FDR (threshold of significance) is set at q < 0.1; hence, the identified sample size of 50 would yield a power of 80%. The objective of animal experiments was to examine whether asparaginase alone reduces circulating and hepatic retinol. All animal experiments and procedures were approved by the Rutgers University Institutional Animal Care and Use Committee, with IACUC approval number PROTO201702605. The age of the mice in this study was restricted to > 8 weeks to minimize the differential effects of asparaginase, which is known to occur in younger mice ( 55 ). To determine whether asparaginase reduces serum and hepatic retinol, 16 and 12 wild-type C57BL/6J adult mice, respectively, were used. The treatment effect size was based on our previously published experiments using asparaginase ( 25 ). Animals were randomized into treatment groups by weight. The method used for asparaginase administration in mice has been previously published ( 25 , 56 ). Wild-type C57BL/6J mice were administered eight daily intraperitoneal injections of 3 IU/g body weight of native E. coli L-asparaginase (Elspar, Merck). The dosing regimen was consistent with the FDA Guidance of drugs between human and animal studies ( 57 ). Empirically, this dosing in mice led to the desired effect expected in clinical practice of non-detectable serum asparagine ( 56 ). Mice were euthanized approximately 8 h after the final injection to obtain serum and liver samples. Serum retinol concentrations were measured using reversed-phase HPLC analysis ( 26 ). Liver tissue samples were sent to Metabolon for global metabolomic analyses ( 58 ). The investigators were not blinded to the treatments during the animal experiments but were blinded during the laboratory analyses. No animals were excluded from the data analysis. TriNetX is a global federated health research network that provides access to E-HR statistics. The differential rates of AP and elevated amylase and lipase were compared between patients taking asparaginase without vitamin A ( n = 6,557) and those taking asparaginase and vitamin A concurrently ( n = 1,132). Propensity score-based matching was used to balance the cohorts, using age, sex, race, and ethnicity at the start of therapy as covariates. The diagnosis of AP was based on ICD-10: K85 during the monitoring period ( table S2 ). Cases before 2015 were evaluated after converting the records from ICD-9 to ICD-10. The time window to monitor for AP and elevations in serum amylase or lipase was inclusive of five years from the first dose of asparaginase in both groups. A total of 107 patients with an history of AP were excluded from this analysis to limit confounding due to potential exacerbation of underlying pancreatic injury. These patients were diagnosed with AP less than 1 year prior to the first dose of asparaginase. AEs data reported to the FAERS and normalized within AERS Mine ( 18 ) were used to identify differential rates of AP correlated with FDA-approved drugs and to validate the E-HR findings. We analyzed approximately 17 million patient reports from the FAERS database regarding the use of asparaginase ( n = 20,186), vitamin A ( n = 281,917), and their combination ( n = 15). If a patient was reported to FAERS more than once, only the last report was considered. Three mutually exclusive cohorts were analyzed for AP frequency. Safety signals were used to determine the strength of the correlation between the drugs and AP. Patients with a history of pancreatitis were excluded from analysis. Benjamini, Hochberg, and Yekutieli test was used for FDR correction, with a threshold of significance set at q < 0.05. Subjects enrolled in the DFCI 05–001 were eligible to participate in the diet and acute lymphoblastic leukemia treat (DALLT) study ( 21 , 49 ). Informed consent for participation in the DALLT was obtained from parents or guardians at the time of enrollment. The dietary data collection and classification of dietary intake are described in the DALLT study. Initial induction, post-induction, and the ratio between the two time points for each metabolite were log2 transformed to better satisfy the normality assumption. For each metabolite sampling condition, a multivariable linear regression model was used, adjusting for age, BMI (age-adjusted percentile), and initial risk for ALL. Both P and q values were obtained ( 59 ). In pathway enrichment analysis, the numbers of significantly ( P = 0.05) between the cases and controls in a specific pathway compared to all other pathways were analyzed using a 2 × 2 contingency table (one-tailed Fisher’s exact test). We used a one-tailed test because our hypothesis is that there are more, not less, significantly changed metabolites. Correlation network analysis was performed using pairwise Spearman’s rank correlations from the R package ‘Hmisc’ (v4.1.1) and weighted, undirected networks were plotted with ‘igraph’ (v1.2.6) ( 60 ). Correlations among carotenoids, retinol, and other metabolites with q values < 0.05 were included and displayed via the Fruchterman-Reingold method. The nodes were color-coded according to the directionality of the change (increased or decreased). Dietary intake data were analyzed using the one-tailed Wilcoxon Rank-Sum test and Spearman’s rank correlation. The rationale for using a one-tailed test to analyze metabolomic analytes and dietary data is first that the current study is an exploratory post-hoc analysis for hypothesis generation. Second, in the dietary data, there was prior knowledge of one-tailed directionality because we previously reported that patients with ALL had reduced micronutrient intake ( 49 , 61 ). We performed a similar statistical analysis in the external validation cohort where in addition to the time points 1 and 2, we also analyzed time points 3 and 4 (for example, we analyzed the ratio between each post-induction timepoint and the initial-induction point).

In this study, using a systems approach, we discovered five key findings suggesting that vitamin A has a protective effect against AAP. CMap analysis of transcriptomic data showed that the gene expression signature induced by asparaginase in leukemic cells could be reversed by vitamin A analogs. Real-world data from the E-HRs and AE reporting in FAERS demonstrated a reduction in AAP risk in patients receiving asparaginase who also had vitamin A exposure. Global metabolomic profiling of plasma samples from subjects with AAP revealed that compared to controls, carotenoids were relatively reduced among the cases, with key findings subsequently validated in a larger external cohort. The AAP cases additionally had reduced dietary intake of vitamin A and beta-cryptoxanthin specifically during induction therapy, and single-agent asparaginase administration in mice was sufficient to reduce circulating and hepatic retinol. Transcriptome-based connectivity mapping ( 8 ) is widely used in preclinical drug discovery ( 28 ). We were guided by findings from the transcriptomic data of asparaginase-treated human leukemic cell lines that vitamin A analogs were predicted to reverse the upregulated pancreatitis gene signature ( 29 ). Vitamin A analogs were the top hits in a similar CMap analysis performed using a transcriptomic dataset from experimental AP in mice (CIP). The findings of the current study are distinct from those previously reported for vitamin A and exocrine pancreatic diseases ( 30 ). Vitamin A exposure reduces the fibrogenic potential of isolated pancreatic stellate cells (PSCs), which are resident vitamin A-containing cells of the pancreas ( 31 ). In vivo , retinoic acid inhibits PSC activation to suppress pancreatic fibrosis in murine models of chronic pancreatitis (CP) ( 32 ). The current study is a previously undescribed attempt using E-HRs and AE reporting to characterize AAP. The main finding was a 60% reduction in pancreatitis with vitamin A exposure. This correlated with the FAERS safety signals for vitamin A exposure. In the primary data analysis of the E-HRs, we excluded topical vitamin A exposure because the amount of exposure and cutaneous absorption were minimal (5–7%) ( 33 ). The strengths of the TriNetX and AERS Mine frameworks that we developed ( 18 ) to examine the E-HRs and FAERS, respectively, are that each provides deep, complementary insight into real-world AE associations. We previously used a similar approach to repurpose rosiglitazone for treating gastrointestinal angioectasia-related bleeding ( 34 ). In the current study, we identified a relative reduction in carotenoids during induction therapy from plasma metabolomic screening of a case-control group. Previous metabolomic studies in patients with ALL ( 35 , 36 ) and in patients with pancreatitis ( 37 ) were neither case-control studies nor examined drug-induced pancreatitis. In a recent study, Adam et al . identified beta-carotene and cryptoxanthin among the top circulating metabolite hits in established patients with CP ( 38 ), which is consistent with the finding that patients with CP can develop malabsorption of fat-soluble vitamins, including vitamin A. Known risk factors for AAP lack accuracy. These include older age (>5 years), higher ALL risk stratification, obesity, ethnicity, particularly Hispanic ancestry, and higher dosing of asparaginase ( 5 – 7 ). There is no utility in screening patients for AAP during chemotherapy using pancreatic imaging or measuring serum pancreatic enzyme activity ( 39 ). Several gene variants have been suggested as potential risk factors for AAP ( 6 , 40 – 43 ). The models we ran using the five most significant metabolites obtained at the start and end of induction therapy had modest predictive values for AAP. Metabolomics, in combination with the above-mentioned known risk factors, could potentially offer promise as biomarkers in longitudinal studies. The mechanism of AAP remains unknown. Asparaginase causes cellular injury, particularly in leukemic cells, owing to the depletion of intracellular asparagine ( 4 ) and poor expression of the counteracting enzyme asparagine synthetase (ASNS) ( 44 ). The pancreas, on the other hand, has the highest expression of ASNS among all organs per tissue weight ( 45 ). In addition, ASNS, as well as several related amino acid transporters, are upregulated in the setting of the integrated stress response (ISR), specifically amino acid deprivation by asparaginase ( 46 ). Dysfunctional ISR renders cells more vulnerable to asparaginase, resulting in maladaptive outcomes. It is possible that individuals with reduced baseline expression of pancreatic ASNS or other ISR target genes involved in amino acid synthesis, transport, or recycling are less capable of adapting to this form of amino acid deprivation and may be at risk for AAP. Another method for replenishing amino acids is through autophagy, and genetic evidence suggests that patients with AAP have polymorphisms in the autophagy-inducing kinase, unc-51-like kinase 2 ( 47 ). As with other forms of pancreatitis, AAP can cause loss of mitochondrial function and reduced bioenergetics. Peng et al . showed that the carbohydrate monosaccharide galactose could partially rescue or improve AP by boosting the bioenergetics ( 48 ). Our current study adds to the knowledge base that the micronutrient vitamin A is important to pancreatic health and could compensate for pancreatic inflammation due to asparaginase. Whether the protective effect is at the level of amino acid deprivation or downstream in pancreatitis response needs further investigation. A major finding of the current study is that the case subjects consumed less dietary vitamin A than the controls in the early period of ALL therapy, specifically during induction. The median value for the composite intake of vitamin A constituents (RAE) was reduced by one-third to degrees approaching the borders of the RDA. This lower intake for one month would not be expected to induce vitamin A deficiency. In contrast to the reduced dietary intake of vitamin A during induction, vitamin A intake 30 days prior to induction did not differ between the two groups. This finding is not surprising, because patients with ALL are known to consume poorer diets (higher caloric intake but with lower micronutrients) when they start induction chemotherapy ( 49 ). In a prospective cohort study evaluating dietary intake and treatment-related toxicities among children and adolescents with ALL, Ladas et al . found that an increased intake of dietary antioxidant vitamins consisting of alpha- and beta-carotene, carotenoids, and vitamin A was associated with lower rates of infection and mucositis, with no increased risk of relapse or reduced survival ( 21 ). The effect of these natural antioxidants on AAP remains to be elucidated. The findings of this study must account for the following limitations. The asparaginase signature used for CMap analysis was derived from leukemic cells treated with asparaginase in lieu of asparaginase-treated pancreatic tissue. To address this limitation, we used asparaginase DEGs that were enriched in pancreatic cell types for querying ( 50 ). Another limitation is that the compound signatures in the L1000 Touchstone dataset were based on cancer cell lines. However, we and others have shown that, using a similar approach of analyzing focused expression patterns, CMap and L1000 compound signatures can be effectively leveraged for drug discovery ( 28 , 51 , 52 ). (ii) We were unable to evaluate whether there was a differential use of vitamin A in patients from the group who were excluded due to prior AP for the E-HR analysis because the number of vitamin A users in the excluded group was less than or equal to ten. Patient counts of ten or fewer were imputed to a value of ten and thus the exact number was obfuscated ( 13 ).The asparaginase dosing and duration information was also sporadic and incomplete in TriNetX, and ICD-10 billing codes were used as the primary proxy for disease documentation. Additionally, reporting to FAERS is not mandatory; there are risks of heterogeneity and under-reporting. E-HRs offer added precision over FAERS, with information on clinical laboratory values, imaging, clinical parameters, temporal parameters, and route of administration. In addition, although the metabolomic case-control design was well-matched, the limitations include its inherent retrospective nature, small sample size, and the fact that the samples were obtained in a non-fasting state. We performed a larger external metabolomic screening to validate the key findings from the discovery cohort.

Drug adverse events (AEs) are a major health problem worldwide. There are more than 24 million AEs from the US Federal Drug Administration (FDA) Adverse Events Reporting System (FAERS) public dashboard and approximately 14 million AEs are serious AEs. Governmental agencies, the pharmaceutical industry, and healthcare systems encourage AE reporting to improve product safety, protect public health, and make information available for integration into clinical care and research. Among the AEs, acute pancreatitis (AP) is a painful, life-threatening inflammatory disorder of the pancreas, and one-fifth of AP patients develop moderate to severe complications. In children, drug-induced pancreatitis is among the leading causes of AP ( 1 – 3 ). However, the mechanisms underlying drug-induced pancreatitis remain largely unknown. A prototype example of a drug that can induce pancreatitis is asparaginase. Asparaginase is a cornerstone chemotherapy ( 4 ) for the treatment of acute lymphoblastic leukemia (ALL), the most common cancer in children and adolescents. The main challenge with asparaginase, however, is its AEs, which include hypersensitivity, hepatotoxicity, thrombosis, hemorrhage, and asparaginase-associated pancreatitis (AAP). AAP occurs in 2–10% of asparaginase users, and one-third of affected patients develop severe AP with pseudocyst formation, pancreatic necrosis, and intensive care unit admissions ( 5 ). AAP leads to discontinuation of asparaginase in over a third of patients, and half of the patients who are rechallenged with asparaginase after AAP develop recurrence of pancreatitis. The risk factors for AAP are older age, higher ALL risk stratification, obesity, higher dosing of asparaginase, ethnicity (particularly Hispanic ancestry), and type of asparaginase formulation ( 5 – 7 ). However, these risk factors predict AAP with low accuracy. Furthermore, the mechanisms underlying AAP are largely unknown. Thus, there is a major unmet need to predict AAP, decipher its mechanism, and mitigate it. The objectives of the current study were to characterize interventions for AAP using a systems approach. The transcriptomic, AE, and the electronic health records (E-HRs) data analyses revealed that retinoids are likely to reverse AAP. From a metabolomic screen of case-control subjects, we demonstrated that patients with ALL who developed pancreatitis had lower plasma carotenoids after exposure to a single dose of asparaginase. This finding was validated in a larger external cohort. The AAP case subjects also had reduced dietary intake of vitamin A. In mice, asparaginase alone reduced circulating and hepatic retinol. Collectively, our findings provide crucial insights into mitigating AAP and elucidate a broader role for vitamin A and its analogs in preventing pancreatic inflammation.