To ensure the accuracy of this experiment, volunteers with other malignancies and metabolic disease such as diabetes, and liver and kidney dysfunction were excluded. All DLBCL patients were diagnosed and reviewed by experienced hematopathologists and received an anti-CD20-based chemotherapy regimen. Finally, a total of 65 volunteers including 22 healthy controls (Ctrl), 25 DLBCL patients newly diagnosed (ND), and 18 DLBCL patients achieving complete remission (CR) were enrolled between Feburary 2018 and March 2019 in this analysis. This study was approved by the Medical Ethical Committee of Nanjing Drum Tower Hospital. Informed consent for plasma samples was obtained from all volunteers in accordance with the Helsinki Declaration.

Newly diagnosed DLBCL patients were further classified according to Han’s algorithm (the germinal center B cell subtype (GCB) vs. non-GCB) and the Ann Arbor stage (stage I-II vs. stage III-IV). Han’s algorithm was performed by senior pathologists based on immunohistochemistry (the expression of immune molecule markers including CD10, Bcl-6, and MUM1) (19). Efficacy of CR was evaluated based on the standards revised by the Lugano meeting (2014), including the length of target lesions (lymph nodes) was less than 1.5 cm, and no extranodal lesions were found; unmeasurable lesions disappeared; enlarged organs returned to normal; no new lesions; and bone marrow morphology was normal. Clinical characteristics of DLBCL patients and healthy volunteers involved are summarized in Table 1 .

The plasma was extracted from peripheral blood which was collected using EDTA tubes and then centrifuged at 1000 g for 10 min. The collected plasma samples were immediately frozen at -80°C until metabolomics analysis. All plasma samples were collected from volunteers after overnight fasting (for at least 10 h), and newly diagnosed patients were new diagnosis and prior to any medications.

The metabolomics method based on GC/MS was performed as previously reported (20). Briefly, 50 µL of plasma supernatants was extracted with 200 µL of methanol which contains 5 µg/mL of [¹³C₂]-myristic acid internal standard (IS). The mixture was vortexed for 5 min and centrifuged (20,000×g, 4°C) for 10 min to remove protein. An aliquot of 100 µL of supernatant was collected and concentrated to dryness. The extract was then oximated with 30 µL of pyridine containing 10 mg/mL methoxyamine for 16 h at room temperature, followed by trimethylsilylation with 30 µL of MSTFA + 1% TMCS for another 1 h. Subsequently, 30 µL of n-heptane containing methyl myristate (15.0 µg/mL) as the quality control reference standard was added and thoroughly mixed. The quality control samples (QC) were pooled with the remaining upper layers of samples in the study set and mixed.

A 0.5 μL portion of the prepared samples was injected into Shimadzu GC/MS QP2010Ultra/SE equipped with a 30 m × 0.25 mm ID, fused silica capillary column, which was chemically bonded with 0.25 m DB1-MS stationary phase. The column temperature was programed as follows: kept at 80°C for 3.0 min, then increased to 300°C at 20°C/min and held for 5.0 min. The transfer line temperature was set at 220°C and the ion source temperature at 200°C. Ions were generated by a 70-eV electron beam at a current of 3.2 mA. The mass spectra were acquired over the mass range of 50-700 m/z at a rate of 25 spectra/s after a solvent delay of 160 s. To diminish the systematic variation during instrumental analysis, all the samples were randomly selected for analysis by GC/MS and QC samples spaced evenly among the injections.

The metabolites were identified by comparing the mass spectrum and retention indexes for the analyte with the corresponding values from Wiley 9, the National Institute of Standards and Technology (NIST) library 2.0, and an in-house mass spectra library database. A match factor greater than 80% is considered to be reliable.

After identification, statistical analyses were performed using normalized peak area (normalized by IS). The normalized peak area was introduced in SIMCA software for multivariate statistical analysis. The goodness of fit for a model was evaluated using three quantitative parameters as follows: R2X (the explained variation in X), R2Y (the explained variation in Y), and Q2 (the predicted variation in Y based on the model using cross-validation). Statistical analyses were performed using the GraphPad Prism software. Metabolomics pathway analysis of the metabolic biomarkers was carried out using MetaboAnalyst (21). Fold change (FC) was calculated by comparing average level between groups. PET-derived metrics include: tumor maximum standardized uptake value (SUVmax); metabolically active tumor volume (TMTV), calculated as the total volume of tumor with FDG uptake; and total lesion glycolysis (TLG), the sum of the tumor volume weighted by the intensity of FDG uptake. The correlation analysis was performed by R software. Relative distance values (RDV) were calculated as previously reported (22). Survival functions were estimated using the Kaplan-Meier method and compared with the log-rank test. All data are expressed as the means ± SDs. Differences among groups were evaluated by two-tailed Student’s t-test or one-way ANOVA. P values less than 0.05 were considered to indicate statistical significance.

The demographics of all volunteers included in this study are summarized in Table 1 . All participants were age‐matched and sex‐matched in the Ctrl, ND, and CR groups. Compared with healthy controls, no obvious liver and kidney dysfunction were observed in newly diagnosed DLBCL patients according to the biochemistry analysis. In contrast, abnormal cholesterol metabolism was observed in the ND group as indicated by significantly decreased cholesterol, high-density lipoprotein cholesterol, and apolipoprotein A. The level of lymphocyte, red blood count (RBC), and hemoglobin was also significantly reduced in the ND group, while these laboratory characteristics were restored in the CR group. Moreover, a multivariate statistical model was constructed based on clinical characteristics, showing obvious cluster classification among collected samples ( Supplementary Figure S1A ), and the variable of importance (VIP, Supplementary Figure S2 ) values indicated that lymphocyte, RBC, and hemoglobin contributed the most to distinguishing these three groups. This result suggested that lymphocyte, RBC, and hemoglobin may be the candidate clinical indicators related to diagnosis in DLBCL patients.

Deconvolution of the plasma chromatograms produced a total of 128 distinct peaks and 87 were authentically identified. The overall process variability was calculated based on the intensity of methyl myristate; the RSD was 6.8%. Moreover, pooled QC samples were clustered well in a PCA plot ( Supplementary Figure S3 ), indicating stable instrument operation and good reproducibility of the assay throughout the experiment. To explore the metabolic characteristics among the Ctrl, ND, and CR groups, a supervised PLS-DA model was created to obtain an overview of the data set. The score plot ( Figure 1A ) showed that samples within each group tended to cluster closely, indicating that the metabolic features within each group were similar. Furthermore, the ND group was well separated from the control group, while the CR group moved much closer to the control group, suggesting the metabolic profile could well characterize the disease states of DLBCL. To further quantitatively evaluate the perturbed metabolism, the relative distance values (RDV) were calculated. The RDV between the Ctrl group and the ND group, the CR group was decreased from 7.1 ± 2.1 to 3.3 ± 2.2, indicating the metabolic fluctuation was restored after achieving complete remission. Compared with clinical characteristics ( Supplementary Figures S1B, C ), the OPLS-DA score plots of metabolomics presented better fitness and higher predictability of models for the Ctrl group vs. the ND group ( Figure 1B ) and the CR group vs. the ND group ( Figure 1C ). This indicated the metabolic profiles illustrated stronger recognition among the Ctrl, ND, and CR groups.

To recognize the differential metabolites, volcano plots were processed to display metabolites changed greatly (fold change > 1.5 or < 0.667) and significantly (p <0.05). Compared with the control group, 3 metabolites were identified with 2 metabolites decreased (glucose and aspartate) and one metabolite increased (2-hydroxy-2-methyl-butyric acid) in the ND group ( Figure 1C ). While compared with the CR group, 3 metabolites were identified with 2 metabolites decreased (glucose and aspartate) and 1 metabolite increased (malonic acid) in the ND group ( Figure 1C ). These metabolites screened above suggested that glucose and aspartate may be the candidate metabolic biomarkers for the diagnosis in DLBCL patients ( Figure 1D ).

To verify the sensitivity and specificity of the above potential biochemical indicators and metabolic biomarkers for diagnosis, receiver operating characteristic (ROC) curve was performed. As presented in Figure 2A , the area under the ROC (ROC-AUC) calculated with lymphocyte, RBC, hemoglobin, glucose, and aspartate achieved a ROC-AUC of 0.874, 0.858, 0.879, 0.742, and 0.702, respectively. To improve the performance of the diagnostic model, combined diagnosis with multiple indexes is usually used. With the metabolic biomarkers (glucose and aspartate) and the hematological indicators (lymphocyte, RBC, and hemoglobin), a logistic regression to establish a diagnostic model was applied to predict diagnosis of DLBCL. The corresponding ROC curve had an AUC of 0.936 ( Figure 2B ) which was higher than that of each candidate biomarker. This indicated that the combined diagnostic model could significantly improve the diagnostic performance with respect to DLBCL.

According to prognostic outcome after follow-up (update to September 2021, four patients were lost to follow-up), the remaining patients in the ND group were divided into the progression-free group (n=14) and the progression group (n=7). A multivariate statistical analysis showed that the progression group was well separated from the control group, while the progression-free group moved much closer to the control group ( Figure 3A ). However, the PLS-DA plots of subgroups divided based on Ann Arbor stage and Han’s algorithm displayed little difference of metabolic profiles ( Supplementary Figures S4A, C ). Furthermore, the OS (overall survival) had no significant difference between stage I-II vs. stage III-IV (p=0.998, 95% C.I): 35.5 months (30.0- 41.1 months) vs. 33.4 months (25.4-41.5 month), as well as GCB vs. non-GCB (p=0.128, 95% C.I): 38.0 months (32.4-43.6 months) vs. 30.5 months (23.1-37.9 months) ( Supplementary Figures S4B, D ). This suggested that Ann Arbor stage and Han’s algorithm displayed little prediction power for prognosis. The OPLS-DA model showed a clear and distinctive clustering between the samples of the progression-free group and the progression group ( Figure 3B ), indicating that some metabolites contributed to discriminate progression-free patients from progression patients before treatment. This suggested that the plasma metabolic profile of patients before treatment could well predict the prognosis of DLBCL.

To identify the biomarkers associated with prognosis, statistically different metabolites (p < 0.05) with a large change (fold change> 1.5 or <0.67) between the progression group and the progression-free group were screened out, including 10 metabolites (malate, fumarate, citrate, 2-hydroxy-2-methylbutyric acid, 2-hydroxybutytic acid, cysteine, methionine, leucine, isoleucine, and valine) ( Figure 4C ). The metabolites set enrichment analysis revealed that citric acid cycle, valine, leucine and isoleucine degradation, and methionine metabolism were markedly perturbed ( Figure 4D ).

Some clinic indexes or derived parameters before treatment including neutrophil/lymphocyte ratio (NLR), derived neutrophil/lymphocyte ratio (dNLR), platelet/lymphocyte ratio (PLR), lymphocyte/monocyte ratio (LMR), and monocyte were reported to have the important significance for prognosis-prediction of the patients with DLBCL (23–25). Correlation analysis was performed between the above 8 metabolites and reported indexes ( Figure 4A ), and no significant correlation was found. Apart from staging and response assessment, the baseline ¹⁸F-fluoro-2-deoxy-D-glucose positron emission tomography (FDG-PET/CT) parameters (SUVmax, TMTV, and TLG) have important prognostic significance for DLBCL patients receiving R-CHOP chemotherapy (26). We were delighted to find that the level of malate and 2-hydroxy-2-methylbutyric acid was positively related with TMTV (r=0.58, p=0.015 for malate and r=0.56, p=0.018 for 2-hydroxy-2-methylbutyric acid) and TLG (r=0.69, p=0.002 for malate and r=0.63, p=0.008 for 2-hydroxy-2-methylbutyric acid), respectively ( Figure 4B ). Furthermore, ND patients with higher levels of malate (p=0.0217, 25.57 months (15.83-35.31 months) for the high subgroup vs. 39.07 months (35.43-42.71 months) for the low subgroup) not 2-hydroxy-2-methylbutyric acid (p=0.43, 37.57 months (33.16- 41.98 months) for the high subgroup vs. 32.85 months (26.07- 39.64 months) for the low subgroup) were associated with a poorer survival ( Figures 4C, D ), which further indicated the malate was a preferential metabolic biomarker for prognosis-prediction of DLBCL patients.

A supervised PLS-DA model revealed that the ND group was well separated from the control group, while the CR group moved much closer to the control group ( Figure 1A ), which was further confirmed by quantitative RDV. It was suggested that the plasma metabolic profile of newly diagnosed DLBCL patients deviated significantly from the healthy controls, and the metabolic fluctuation was restored after achieving complete remission. The OPLS-DA model established with metabolomics data ( Figures 1B, C ) showed better fitness and higher predictability than models with clinical characteristics ( Supplementary Figures S1B, C ), indicating the metabolic profile could well characterize disease states of DLBCL and illustrate stronger recognition among healthy controls, DLBCL patients, and patients after achieving complete remission.

Patients in the ND group were further divided into two groups based on risk stages, subtypes based on Han’s algorithm and prognosis, respectively. The PLS-DA model clearly displayed the poorer prognosis group tended to shift further away from the Ctrl group ( Figure 3A ). However, high-stage subgroups and low-stage subgroups were not well separated, as well as the GCB subgroup and the non-GCB subgroup ( Supplementary Figures S4A, C ) which was consistent with the result reported by Mi et al. (17). This further indicated that information on individual prognosis was contained in the metabolic patterns of pre-treatment bio-fluids. In summary, metabolomics study was informative for us to characterize the metabolic phenotypes, forecast risk and prognosis, and understand the metabolic mechanisms of DLBCL.

The level of glucose and aspartate in plasma was decreased in ND patients and elevated in CR patients, indicating that glucose and aspartate may be involved in the pathogenesis of DLBCL ( Figures 1C, D ). Most tumor cells depend largely on aerobic glycolysis which converts glucose into essential metabolic intermediates for energy production and cell proliferation (28). Pathological lymph nodes of DLBCL also exhibit increasing glucose uptake measured by FDG-PET/CT (12), resulting in a decrease in peripheral glucose. Thus, anti-glycolytic cancer therapy has become a rising research focus for developing anticancer drugs (29), metformin could sensitize therapeutic agents and improve prognosis in pre-clinical and clinical DLBCL (30). More and more evidence demonstrates that aspartate is an endogenous metabolic limitation for tumor growth (31), and up-regulated aspartate metabolism contributed to epithelial-mesenchymal transition in tumor (32). Moreover, the lower level of aspartate in newly diagnosed DLBCL patients may be related to the disorders of asparagine metabolism. Asparaginase catalyzes the hydrolysis of the asparagine into aspartic acid and ammonia and has been used in the treatment of leukemia. Attempts have been made on the effect of asparaginase on DLBCL, and inspiring results were obtained in both the preclinical experiment (33) and clinical trials with a low number of patients (10). More evidence with multi-center clinical trials and an enlarged number of patients are needed to validate the effect of asparaginase on DLBCL.

A clinical diagnosis of DLBCL requires the patient to undergo a pathological biopsy, which is an invasive complex operation (6). Our results suggested that glucose and aspartate were deeply involved in the occurrence and pathogenesis of DLBCL. According to the diagnostic model established based on the identified hematological indicators ( Table 1 ) and plasma metabolic indicators ( Figure 1 ), favorable performance (AUC =0.936) was achieved with the combination of five candidate indicators, which was higher than with each single indicator ( Figure 2B ).

According to the prognostic outcome, we found that patients with poor prognosis showed vigorous oxidative phosphorylation (OxPhos) metabolism indicated by enrichment of typical metabolites in a citric acid cycle (TCA), such as citrate, fumarate, and malate ( Figures 3 and 5 ). It has been reported that the subtype of DLBCL cell lines which rely on mitochondrial energy transduction called OxPhos-DLBCLs provide pro-survival benefits (34) and the intermediates in the TCA cycle have a multifaceted contribution to tumor progression (35). Citrate is a critical metabolic checkpoint involved in several important metabolic pathways, a certain level of citrate is necessary for cancer cell proliferation and tumor growth by the production of acetyl-CoA which is important for lipogenesis, membrane expansion, and acetylation (36). Fumarate is another physiological metabolic intermediate of the TCA cycle, the accumulation of fumarate which could permeate multiple subcellular compartments (37) as well as an extracellular microenvironment (38). The abnormally increased fumarate is a key activator of a variety of oncogenic cascades (39). Growing evidence supports the idea that malate accumulation may be essential for cancer growth by contributing to both redox balance (40, 41) and an elevated glycolytic flux (42, 43). Hence, it has seemed that the flux of oxidative phosphorylation was closely related to tumor growth and progression, which can indicate the prognosis of the tumor.

In addition, an obvious decline of methionine and cysteine and consistent decrease of branched-chain amino acids (BCAAs) were observed in the plasma of DLBCL patients with poor prognosis. Valine, leucine, and isoleucine are essential BCAAs whose metabolism can both influence multiple cancer phenotypes and serve as a marker of disease pathology (44). Valine was reported as a candidate biomarker for the prognosis of DLBCL by Stenson et al. (18) and Mi et al. (17). It is likely that the degradation of BCAAs to acetyl-CoA is enhanced, which further supports the generation of TCA cycle intermediates ( Figure 5 ) to fuel cancer growth. Methionine metabolism coordinates nucleotide and redox status which are relevant to cancer pathogenesis (45). The decreased methionine and cysteine may maintain cellular redox status by producing the antioxidant glutathione (GSH) against reactive oxygen species (ROS) which was generated in remarkable OxPhos metabolism ( Figure 5 ). Elevated 2-hydroxybutyric acid, a byproduct of glutathione anabolism, indirectly indicated the robust requirement to maintain the redox state ( Figure 5 ). Moreover, 2-hydroxybutyric acid has previously been identified as a biomarker of diabetes risk (46, 47) and mitochondrial disorders (48). This suggested that DLBCL patients with higher risk of diabetes were more likely to have a poor prognosis. 2-hydroxy-2-methlbutyric acid belongs to hydroxy fatty acid and little is known about its biological functions. Further study is needed to exploit its role in DLBCL. Overall, elucidation of the metabolic mechanisms underlying the DLBCL progression is still needed in our further study.

Due to the heterogeneity and different prognosis of DLBCL patients, it is urgent to exploit new biomarkers for prognosis-prediction. Candidate metabolites including malate, fumarate, citrate, 2-hydroxy-2-methylbutyric acid, cysteine, methionine, leucine, isoleucine, and valine were first screened out, correlation analysis was then performed to evaluate the talent of these markers. No significant correlation was found between the above metabolites and reported biochemical parameters which have predictive power to predict prognosis ( Figure 4A ). It is probably because these reported parameters (derived from lymphocytes, neutrophils, and monocyte) just reflect the abnormal immune responses which is an essential pathogenic factor for DLBCL (49), while metabolic markers, signs of metabolic status, may be independent risk factors and could not be simply associated with hematological parameters.

Functional imaging with FDG integrated with PET-CT provides valuable measurements of tumor metabolism and activity of DLBCL. The derived tumor metrics (SUVmax, TMTV, and TLG) not only has been affirmed in metabolic characteristics, staging, and end-of-treatment evaluation, but they have the potential to improve the accuracy of prognostication (50). Interestingly, the level of malate and 2-hydroxy-2-methylbutyric acid was positively related with TMTV and TLG ( Figure 4B ). Furthermore, ND patients with higher levels of malate but not 2-hydroxy-2-methylbutyric acid had a poorer survival ( Figures 4C, D ). Thus, malate was a preferential plasma metabolic biomarker for prognosis-prediction of DLBCL patients. Attempts should be made to get more samples to further validate our result on a larger scale, and the role of malate metabolism in DLBCL will be verified in subsequent studies.

There were still some limitations in this study. Although the interference of metabolic diseases such as diabetes was excluded as much as possible, the influence of other concomitant diseases was still unknown. Furthermore, a validation dataset with more samples is needed, which is underway by us to verify these results.