Impact of Age on Pharmacogenomics and Treatment Outcomes of B-Cell Acute Lymphoblastic Leukemia.

PURPOSE: Acute lymphoblastic leukemia (ALL) can occur across all age groups, with a strikingly higher cure rate in children compared with adults. However, the pharmacological basis of age-related differences in ALL treatment response remains unclear. METHODS: Studying 767 children and 309 adults with newly diagnosed B-cell ALL enrolled on frontline trials at St Jude Children's Research Hospital, MD Anderson Cancer Center, the Alliance for Clinical Trials in Oncology, and the ECOG-ACRIN Cancer Research Group, we determined the ex vivo sensitivity of leukemia cells to 21 drugs. Twenty-three ALL molecular subtypes were identified using RNA sequencing. We systematically characterized the associations between drug response and ALL genomics in children, adolescents and young adults, and elderly adults. We evaluated the effect of age-related gene expression signature on ALL treatment outcomes. RESULTS: Seven ALL drugs (asparaginase, prednisolone, mercaptopurine, dasatinib, nelarabine, daunorubicin, and inotuzumab ozogamicin) showed differential activity between children and adults, of which six were explained by age-related differences in leukemia molecular subtypes. Adolescents and young adults showed similar patterns of drug resistance as older adults, relative to young children. Mercaptopurine exhibited subtype-independent greater sensitivity in children. Transcriptomic profiling uncovered subclusters within CRLF2-, DUX4-, and KMT2A-rearranged ALL that were linked to age and cytotoxic drug resistance. In particular, a subset of children had adult-like ALL on the basis of leukemia gene expression patterns across subtypes, despite their chronological age. Resistant to cytotoxic drugs, children with adult-like ALL exhibited poor prognosis in pediatric ALL trials, even after adjusting for age and minimal residual diseases. CONCLUSION: Our results provide pharmacogenomic insights into age-related disparities in ALL cure rates and identify leukemia prognostic features for treatment individualization across age groups.

Blinatumomab for MRD-Negative Acute Lymphoblastic Leukemia in Adults.

BACKGROUND: Many older adults with B-cell precursor acute lymphoblastic leukemia (BCP-ALL) have a relapse despite having a measurable residual disease (MRD)-negative complete remission with combination chemotherapy. The addition of blinatumomab, a bispecific T-cell engager molecule that is approved for the treatment of relapsed, refractory, and MRD-positive BCP-ALL, may have efficacy in patients with MRD-negative remission. METHODS: In a phase 3 trial, we randomly assigned patients 30 to 70 years of age with BCR::ABL1-negative BCP-ALL (with :: indicating fusion) who had MRD-negative remission (defined as <0.01% leukemic cells in bone marrow as assessed on flow cytometry) after induction and intensification chemotherapy to receive four cycles of blinatumomab in addition to four cycles of consolidation chemotherapy or to receive four cycles of consolidation chemotherapy alone. The primary end point was overall survival, and relapse-free survival was a secondary end point. RESULTS: The data and safety monitoring committee reviewed the results from the third efficacy interim analysis and recommended that they be reported. Complete remission with or without full count recovery was observed in 395 of 488 enrolled patients (81%). Of the 224 patients with MRD-negative status, 112 were assigned to each group. The characteristics of the patients were balanced between the groups. At a median follow-up of 43 months, an advantage was observed in the blinatumomab group as compared with the chemotherapy-only group with regard to overall survival (at 3 years: 85% vs. 68%; hazard ratio for death, 0.41; 95% confidence interval [CI], 0.23 to 0.73; P = 0.002), and the 3-year relapse-free survival was 80% with blinatumomab and 64% with chemotherapy alone (hazard ratio for relapse or death, 0.53; 95% CI, 0.32 to 0.87). A higher incidence of neuropsychiatric events was reported in the blinatumomab group than in the chemotherapy-only group. CONCLUSIONS: The addition of blinatumomab to consolidation chemotherapy in adult patients in MRD-negative remission from BCP-ALL significantly improved overall survival. (Funded by the National Institutes of Health and others; E1910 ClinicalTrials.gov number, NCT02003222.).

Biologic and clinical analysis of childhood gamma delta T-ALL identifies LMO2/STAG2 rearrangements as extremely high-risk.

Acute lymphoblastic leukemia expressing the gamma delta T cell receptor (yo T-ALL) is a poorly understood disease. We studied 200 children with yo T-ALL from 13 clinical study groups to understand the clinical and genetic features of this disease. We found age and genetic drivers were significantly associated with outcome. yo T-ALL diagnosed in children under three years of age was extremely high-risk and enriched for genetic alterations that result in both LMO2 activation and STAG2 inactivation. Mechanistically, using patient samples and isogenic cell lines, we show that inactivation of STAG2 profoundly perturbs chromatin organization by altering enhancer-promoter looping, resulting in deregulation of gene expression associated with T-cell differentiation. High throughput drug screening identified a vulnerability in DNA repair pathways arising from STAG2 inactivation, which can be targeted by Poly(ADP-ribose) polymerase (PARP) inhibition. These data provide a diagnostic framework for classification and risk stratification of pediatric yo T-ALL.

Investigation of inherited noncoding genetic variation impacting the pharmacogenomics of childhood acute lymphoblastic leukemia treatment.

Defining genetic factors impacting chemotherapy failure can help to better predict response and identify drug resistance mechanisms. However, there is limited understanding of the contribution of inherited noncoding genetic variation on inter-individual differences in chemotherapy response in childhood acute lymphoblastic leukemia (ALL). Here we map inherited noncoding variants associated with treatment outcome and/or chemotherapeutic drug resistance to ALL cis-regulatory elements and investigate their gene regulatory potential and target gene connectivity using massively parallel reporter assays and three-dimensional chromatin looping assays, respectively. We identify 54 variants with transcriptional effects and high-confidence gene connectivity. Additionally, functional interrogation of the top variant, rs1247117, reveals changes in chromatin accessibility, PU.1 binding affinity and gene expression, and deletion of the genomic interval containing rs1247117 sensitizes cells to vincristine. Together, these data demonstrate that noncoding regulatory variants associated with diverse pharmacological traits harbor significant effects on allele-specific transcriptional activity and impact sensitivity to antileukemic agents.

Single-cell systems pharmacology identifies development-driven drug response and combination therapy in B cell acute lymphoblastic leukemia.

Leukemia can arise at various stages of the hematopoietic differentiation hierarchy, but the impact of developmental arrest on drug sensitivity is unclear. Applying network-based analyses to single-cell transcriptomes of human B cells, we define genome-wide signaling circuitry for each B cell differentiation stage. Using this reference, we comprehensively map the developmental states of B cell acute lymphoblastic leukemia (B-ALL), revealing its strong correlation with sensitivity to asparaginase, a commonly used chemotherapeutic agent. Single-cell multi-omics analyses of primary B-ALL blasts reveal marked intra-leukemia heterogeneity in asparaginase response: resistance is linked to pre-pro-B-like cells, with sensitivity associated with the pro-B-like population. By targeting BCL2, a driver within the pre-pro-B-like cell signaling network, we find that venetoclax significantly potentiates asparaginase efficacy in vitro and in vivo. These findings demonstrate a single-cell systems pharmacology framework to predict effective combination therapies based on intra-leukemia heterogeneity in developmental state, with potentially broad applications beyond B-ALL.

Epigenomic mapping reveals distinct B cell acute lymphoblastic leukemia chromatin architectures and regulators.

B cell lineage acute lymphoblastic leukemia (B-ALL) is composed of diverse molecular subtypes, and while transcriptional and DNA methylation profiling has been extensively examined, the chromatin landscape is not well characterized for many subtypes. We therefore mapped chromatin accessibility using ATAC-seq in primary B-ALL cells from 156 patients spanning ten molecular subtypes and present this dataset as a resource. Differential chromatin accessibility and transcription factor (TF) footprint profiling were employed and identified B-ALL cell of origin, TF-target gene interactions enriched in B-ALL, and key TFs associated with accessible chromatin sites preferentially active in B-ALL. We further identified over 20% of accessible chromatin sites exhibiting strong subtype enrichment and candidate TFs that maintain subtype-specific chromatin architectures. Over 9,000 genetic variants were uncovered, contributing to variability in chromatin accessibility among patient samples. Our data suggest that distinct chromatin architectures are driven by diverse TFs and inherited genetic variants that promote unique gene-regulatory networks.

SINGLE CELL DISSECTION OF DEVELOPMENTAL ORIGINS AND TRANSCRIPTIONAL HETEROGENEITY IN B-CELL ACUTE LYMPHOBLASTIC LEUKEMIA.

Sequencing of bulk tumor populations has improved genetic classification and risk assessment of B-ALL, but does not directly examine intratumor heterogeneity or infer leukemia cellular origins. We profiled 89 B-ALL samples by single-cell RNA-seq (scRNA-seq) and compared them to a reference map of normal human B-cell development established using both functional and molecular assays. Intra-sample heterogeneity was driven by cell cycle, metabolism, differentiation, and inflammation transcriptional programs. By inference of B lineage developmental state composition, nearly all samples possessed a high abundance of pro-B cells, with variation between samples mainly driven by sub-populations. However, ZNF384- r and DUX4- r B-ALL showed composition enrichment of hematopoietic stem cells, BCR::ABL1 and KMT2A -r ALL of Early Lymphoid progenitors, MEF2D -r and TCF3::PBX1 of Pre-B cells. Enrichment of Early Lymphoid progenitors correlated with high-risk clinical features. Understanding variation in transcriptional programs and developmental states of B-ALL by scRNA-seq refines existing clinical and genomic classifications and improves prediction of treatment outcome.

Biologic and clinical features of childhood gamma delta T-ALL: identification of STAG2/LMO2 γδ T-ALL as an extremely high risk leukemia in the very young.

PURPOSE: Gamma delta T-cell receptor-positive acute lymphoblastic leukemia (γδ T-ALL) is a high-risk but poorly characterized disease. METHODS: We studied clinical features of 200 pediatric γδ T-ALL, and compared the prognosis of 93 cases to 1,067 protocol-matched non-γδ T-ALL. Genomic features were defined by transcriptome and genome sequencing. Experimental modeling was used to examine the mechanistic impacts of genomic alterations. Therapeutic vulnerabilities were identified by high throughput drug screening of cell lines and xenografts. RESULTS: γδ T-ALL in children under three was extremely high-risk with 5-year event-free survival (33% v. 70% [age 3-<10] and 73% [age ≥10], P =9.5 x 10 -5 ) and 5-year overall survival (49% v. 78% [age 3-<10] and 81% [age ≥10], P =0.002), differences not observed in non-γδ T-ALL. γδ T-ALL in this age group was enriched for genomic alterations activating LMO2 activation and inactivating STAG2 inactivation ( STAG2/LMO2 ). Mechanistically, we show that inactivation of STAG2 profoundly perturbs chromatin organization by altering enhancer-promoter looping resulting in deregulation of gene expression associated with T-cell differentiation. Drug screening showed resistance to prednisolone, consistent with clinical slow treatment response, but identified a vulnerability in DNA repair pathways arising from STAG2 inactivation, which was efficaciously targeted by Poly(ADP-ribose) polymerase (PARP) inhibition, with synergism with HDAC inhibitors. Ex-vivo drug screening on PDX cells validated the efficacy of PARP inhibitors as well as other potential targets including nelarabine. CONCLUSION: γδ T-ALL in children under the age of three is extremely high-risk and enriched for STAG2/LMO2 ALL. STAG2 loss perturbs chromatin conformation and differentiation, and STAG2/LMO2 ALL is sensitive to PARP inhibition. These data provide a diagnostic and therapeutic framework for pediatric γδ T-ALL. SUPPORT: The authors are supported by the American and Lebanese Syrian Associated Charities of St Jude Children's Research Hospital, NCI grants R35 CA197695, P50 CA021765 (C.G.M.), the Henry Schueler 41&9 Foundation (C.G.M.), and a St. Baldrick's Foundation Robert J. Arceci Innovation Award (C.G.M.), Gabriella Miller Kids First X01HD100702 (D.T.T and C.G.M.) and R03CA256550 (D.T.T. and C.G.M.), F32 5F32CA254140 (L.M.), and a Garwood Postdoctoral Fellowship of the Hematological Malignancies Program of the St Jude Children's Research Hospital Comprehensive Cancer Center (S.K.). This project was supported by the National Cancer Institute of the National Institutes of Health under the following award numbers: U10CA180820, UG1CA189859, U24CA114766, U10CA180899, U10CA180866 and U24CA196173. DISCLAIMER: The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funding agencies were not directly involved in the design of the study, gathering, analysis and interpretation of the data, writing of the manuscript, or decision to submit the manuscript for publication.

Stromal-induced epithelial-mesenchymal transition induces targetable drug resistance in acute lymphoblastic leukemia.

The bone marrow microenvironment (BME) drives drug resistance in acute lymphoblastic leukemia (ALL) through leukemic cell interactions with bone marrow (BM) niches, but the underlying mechanisms remain unclear. Here, we show that the interaction between ALL and mesenchymal stem cells (MSCs) through integrin ß1 induces an epithelial-mesenchymal transition (EMT)-like program in MSC-adherent ALL cells, resulting in drug resistance and enhanced survival. Moreover, single-cell RNA sequencing analysis of ALL-MSC co-culture identifies a hybrid cluster of MSC-adherent ALL cells expressing both B-ALL and MSC signature genes, orchestrated by a WNT/ß-catenin-mediated EMT-like program. Blockade of interaction between ß-catenin and CREB binding protein impairs the survival and drug resistance of MSC-adherent ALL cells in vitro and results in a reduction in leukemic burden in vivo. Targeting of this WNT/ß-catenin-mediated EMT-like program is a potential therapeutic approach to overcome cell extrinsically acquired drug resistance in ALL.

Obesity in adult acute myeloid leukemia is not associated with inferior response or survival even when dose capping anthracyclines: An ECOG-ACRIN analysis.

BACKGROUND: Obesity (body mass index [BMI] ≥30 kg/m2 ) is an important epidemiological risk factor for developing acute myeloid leukemia (AML). Therefore, the authors studied the association of obesity with clinical and genetic phenotype and its impact on outcome in adults with AML. METHODS: The authors analyzed BMI in 1088 adults who were receiving intensive remission induction and consolidation therapy in two prospective, randomized therapeutic clinical trials of the Eastern Cooperative Oncology Group-American College of Radiology Imaging Network: E1900 (ClinicalTrials.gov identifier NCT00049517; patients younger than 60 years) and E3999 (ClinicalTrials.gov identifier NCT00046930; patients aged 60 years or older). RESULTS: Obesity was prevalent at diagnosis (33%) and, compared with nonobesity, was associated with intermediate-risk cytogenetics group (p = .008), poorer performance status (p = .01), and a trend toward older age (p = .06). Obesity was not associated with somatic mutations among a selected 18-gene panel that was tested in a subset of younger patients. Obesity was not associated with clinical outcome (including complete remission, early death, or overall survival), and the authors did not identify any patient subgroup that had inferior outcomes based on BMI. Obese patients were significantly more likely to receive <90% of the intended daunorubicin dose despite protocol specification, particularly in the E1900 high-dose (90 mg/m2 ) daunorubicin arm (p = .002); however, this did not correlate with inferior overall survival on multivariate analysis (hazard ratio, 1.39; 95% confidence interval, 0.90-2.13; p = .14). CONCLUSIONS: Obesity is associated with unique clinical and disease-related phenotypic features in AML and may influence physician treatment decisions regarding daunorubicin dosing. However, the current study demonstrates that obesity is not a factor in survival, and strict adherence to body surface area-based dosing is not necessary because dose adjustments do not affect outcomes.

Dynamic phosphatase-recruitment controls B-cell selection and oncogenic signaling.

Initiation of B-cell receptor (BCR) 1 signaling, and subsequent antigen-encounter in germinal centers 2,3 represent milestones of B-lymphocyte development that are both marked by sharp increases of CD25 surface-expression. Oncogenic signaling in B-cell leukemia (B-ALL) 4 and lymphoma 5 also induced CD25-surface expression. While CD25 is known as an IL2-receptor chain on T- and NK-cells 6-9 , the significance of its expression on B-cells was unclear. Our experiments based on genetic mouse models and engineered patient-derived xenografts revealed that, rather than functioning as an IL2-receptor chain, CD25 expressed on B-cells assembled an inhibitory complex including PKCδ and SHIP1 and SHP1 phosphatases for feedback control of BCR-signaling or its oncogenic mimics. Recapitulating phenotypes of genetic ablation of PKCδ 10 - 12 , SHIP1 13,14 and SHP1 14, 15,16 , conditional CD25-deletion decimated early B-cell subsets but expanded mature B-cell populations and induced autoimmunity. In B-cell malignancies arising from early (B-ALL) and late (lymphoma) stages of B-cell development, CD25-loss induced cell death in the former and accelerated proliferation in the latter. Clinical outcome annotations mirrored opposite effects of CD25-deletion: high CD25 expression levels predicted poor clinical outcomes for patients with B-ALL, in contrast to favorable outcomes for lymphoma-patients. Biochemical and interactome studies revealed a critical role of CD25 in BCR-feedback regulation: BCR-signaling induced PKCδ-mediated phosphorylation of CD25 on its cytoplasmic tail (S 268 ). Genetic rescue experiments identified CD25-S 268 tail-phosphorylation as central structural requirement to recruit SHIP1 and SHP1 phosphatases to curb BCR-signaling. A single point mutation CD25 S268A abolished recruitment and activation of SHIP1 and SHP1 to limit duration and strength of BCR-signaling. Loss of phosphatase-function, autonomous BCR-signaling and Ca 2+ -oscillations induced anergy and negative selection during early B-cell development, as opposed to excessive proliferation and autoantibody production in mature B-cells. These findings highlight the previously unrecognized role of CD25 in assembling inhibitory phosphatases to control oncogenic signaling in B-cell malignancies and negative selection to prevent autoimmune disease.

Epigenomic mapping in B-cell acute lymphoblastic leukemia identifies transcriptional regulators and noncoding variants promoting distinct chromatin architectures.

B-cell lineage acute lymphoblastic leukemia (B-ALL) is comprised of diverse molecular subtypes and while transcriptional and DNA methylation profiling of B-ALL subtypes has been extensively examined, the accompanying chromatin landscape is not well characterized for many subtypes. We therefore mapped chromatin accessibility using ATAC-seq for 10 B-ALL molecular subtypes in primary ALL cells from 154 patients. Comparisons with B-cell progenitors identified candidate B-ALL cell-of-origin and AP-1-associated cis-regulatory rewiring in B-ALL. Cis-regulatory rewiring promoted B-ALL-specific gene regulatory networks impacting oncogenic signaling pathways that perturb normal B-cell development. We also identified that over 20% of B-ALL accessible chromatin sites exhibit strong subtype enrichment, with transcription factor (TF) footprint profiling identifying candidate TFs that maintain subtype-specific chromatin architectures. Over 9000 inherited genetic variants were further uncovered that contribute to variability in chromatin accessibility among individual patient samples. Overall, our data suggest that distinct chromatin architectures are driven by diverse TFs and inherited genetic variants which promote unique gene regulatory networks that contribute to transcriptional differences among B-ALL subtypes.

Functional investigation of inherited noncoding genetic variation impacting the pharmacogenomics of childhood acute lymphoblastic leukemia treatment.

Although acute lymphoblastic leukemia (ALL) is the most common childhood cancer, there is limited understanding of the contribution of inherited genetic variation on inter-individual differences in chemotherapy response. Defining genetic factors impacting therapy failure can help better predict response and identify drug resistance mechanisms. We therefore mapped inherited noncoding variants associated with chemotherapeutic drug resistance and/or treatment outcome to ALL cis-regulatory elements and investigated their gene regulatory potential and genomic connectivity using massively parallel reporter assays and promoter capture Hi-C, respectively. We identified 53 variants with reproducible allele-specific effects on transcription and high-confidence gene targets. Subsequent functional interrogation of the top variant (rs1247117) determined that it disrupted a PU.1 consensus motif and PU.1 binding affinity. Importantly, deletion of the genomic interval containing rs1247117 sensitized ALL cells to vincristine. Together, these data demonstrate that noncoding regulatory variation associated with diverse pharmacological traits harbor significant effects on allele-specific transcriptional activity and impact sensitivity to chemotherapeutic agents in ALL.

DNA polymerase θ protects leukemia cells from metabolically induced DNA damage.

Leukemia cells accumulate DNA damage, but altered DNA repair mechanisms protect them from apoptosis. We showed here that formaldehyde generated by serine/1-carbon cycle metabolism contributed to the accumulation of toxic DNA-protein crosslinks (DPCs) in leukemia cells, especially in driver clones harboring oncogenic tyrosine kinases (OTKs: FLT3(internal tandem duplication [ITD]), JAK2(V617F), BCR-ABL1). To counteract this effect, OTKs enhanced the expression of DNA polymerase theta (POLθ) via ERK1/2 serine/threonine kinase-dependent inhibition of c-CBL E3 ligase-mediated ubiquitination of POLθ and its proteasomal degradation. Overexpression of POLθ in OTK-positive cells resulted in the efficient repair of DPC-containing DNA double-strand breaks by POLθ-mediated end-joining. The transforming activities of OTKs and other leukemia-inducing oncogenes, especially of those causing the inhibition of BRCA1/2-mediated homologous recombination with and without concomitant inhibition of DNA-PK-dependent nonhomologous end-joining, was abrogated in Polq-/- murine bone marrow cells. Genetic and pharmacological targeting of POLθ polymerase and helicase activities revealed that both activities are promising targets in leukemia cells. Moreover, OTK inhibitors or DPC-inducing drug etoposide enhanced the antileukemia effect of POLθ inhibitor in vitro and in vivo. In conclusion, we demonstrated that POLθ plays an essential role in protecting leukemia cells from metabolically induced toxic DNA lesions triggered by formaldehyde, and it can be targeted to achieve a therapeutic effect.

PAX5 epigenetically orchestrates CD58 transcription and modulates blinatumomab response in acute lymphoblastic leukemia.

Blinatumomab is an efficacious immunotherapeutic agent in B cell acute lymphoblastic leukemia (B-ALL). However, the pharmacogenomic basis of leukemia response to blinatumomab is unclear. Using genome-wide CRISPR, we comprehensively identified leukemia intrinsic factors of blinatumomab sensitivity, i.e., the loss of CD58 as a top driver for resistance, in addition to CD19. Screening 1639 transcription factor genes, we then identified PAX5 as the key activator of CD58. ALL with the PAX5 P80R mutation also expressed the lowest level of CD58 among 20 ALL molecular subtypes in 1988 patients. Genome editing confirmed the effects of this mutation on CD58 expression and blinatumomab sensitivity in B-ALL, with validation in patient leukemic blasts. We described a PAX5-driven enhancer at the CD58 locus, which was disrupted by PAX5 P80R, and the loss of CD58 abolished blinatumomab-induced T cell activation with global changes in transcriptomic/epigenomic program. In conclusion, we identified previously unidentified genetic mechanisms of blinatumomab resistance in B-ALL, suggesting strategies for genomics-guided treatment individualization.

Measurable residual disease by flow cytometry in acute myeloid leukemia is prognostic, independent of genomic profiling.

Measurable residual disease (MRD) assessment provides a potent indicator of the efficacy of anti-leukemic therapy. It is unknown, however, whether integrating MRD with molecular profiling better identifies patients at risk of relapse. To investigate the clinical relevance of MRD in relation to a molecular-based prognostic schema, we measured MRD by flow cytometry in 189 AML patients enrolled in ECOG-ACRIN E1900 trial (NCT00049517) in morphologic complete remission (CR) (28.8 % of the original cohort) representing 44.4 % of CR patients. MRD positivity was defined as ≥ 0.1 % of leukemic bone marrow cells. Risk classification was based on standard cytogenetics, fluorescence-in-situ-hybridization, somatic gene analysis, and sparse whole genome sequencing for copy number ascertainment. At 84.6 months median follow-up of patients still alive at the time of analysis (range 47.0-120 months), multivariate analysis demonstrated that MRD status at CR (p = 0.001) and integrated molecular risk (p = 0.0004) independently predicted overall survival (OS). Among risk classes, MRD status significantly affected OS only in the favorable risk group (p = 0.002). Expression of CD25 (α-chain of the interleukin-2 receptor) by leukemic myeloblasts at diagnosis negatively affected OS independent of post-treatment MRD levels. These data suggest that integrating MRD with genetic profiling and pre-treatment CD25 expression may improve prognostication in AML.

Variational autoencoders learn transferrable representations of metabolomics data.

Dimensionality reduction approaches are commonly used for the deconvolution of high-dimensional metabolomics datasets into underlying core metabolic processes. However, current state-of-the-art methods are widely incapable of detecting nonlinearities in metabolomics data. Variational Autoencoders (VAEs) are a deep learning method designed to learn nonlinear latent representations which generalize to unseen data. Here, we trained a VAE on a large-scale metabolomics population cohort of human blood samples consisting of over 4500 individuals. We analyzed the pathway composition of the latent space using a global feature importance score, which demonstrated that latent dimensions represent distinct cellular processes. To demonstrate model generalizability, we generated latent representations of unseen metabolomics datasets on type 2 diabetes, acute myeloid leukemia, and schizophrenia and found significant correlations with clinical patient groups. Notably, the VAE representations showed stronger effects than latent dimensions derived by linear and non-linear principal component analysis. Taken together, we demonstrate that the VAE is a powerful method that learns biologically meaningful, nonlinear, and transferrable latent representations of metabolomics data.