According to the Declaration of Helsinki, the study was approved by the ethics review board of Peking University People’s Hospital, and enrolled patients provided written informed consent. All the enrolled patients were diagnosed and treated at Peking University People’s Hospital, Beijing, China.

The inclusion criteria are as follows: (i) newly diagnosed de novo AML (non-acute promyelocytic leukemia, APL), age 16–70 years; (ii) no previous antineoplastic therapy; (iii) ECOG performance score ≤ 2, without other major coexisting illnesses such as severe kidney and liver insufficiency.

BM samples were acquired before the first induction chemotherapy. Cell sorting and mass spectrometry measurements were followed as published study (8). The diagnosis, classifications, and risk stratification were based on European Leukemia Net (ELN) recommendations (9). Induction, consolidation, and allogeneic hematopoietic stem cell transplant (allo-HSCT) followed our previous report (10).

The human AML cell lines K562 were purchased from the American Type Culture Collection (ATCC) agent in China. All the cell lines have passed the STR authentication. Cell lines were cultured in RPMI 1640 medium (Thermo Fisher Scientific, Cat#11875093, USA) containing 10% FBS (Thermo Fisher Scientific, Cat#12483020, USA) and 1× antibiotic–antimycotic (Thermo Fisher Scientific, Cat#15240096, USA). Cells were grown at 37°C in a humidified atmosphere with 5% CO₂.

The clinical and transcription matrix of AML was downloaded from the TCGA database (https://portal.gdc.cancer.gov/) and the Beat AML database (http://www.vizome.org/aml/). We retrieved 154 non-APL AML BM mRNA sequence data (excluded 16 APL and 3 incomplete data) and corresponding clinicopathological characteristics from the TCGA. Additionally, 197 BM mRNA sequence data (excluded 9 APL and 25 incomplete data) were retrieved from the Beat AML database (BEAT-AML). The GSE52919 database (https://www.ncbi.nlm.nih.gov/geo/) was performed to analyze the correlation of ECHDC3 expression with cytarabine (Ara-C) sensitivity in AML samples.

Lasso regression was constructed by the glmnet R package according to the method previously published by our institute (11). In brief, 249 differentially expressed genes (DEGs) were identified by the comparison of ECHDC3 high and low expression samples according to the criteria ∣log₂FC∣>1 and adjusted p < 0.05. LASSO regression was then applied to remove redundant prognostic genes based on the coefficient and partial likelihood deviance for developing the prognostic model. ECHDC3 and other Nine hub genes (RPS6KL1, RELL2, FAM64A, SPATS2L, MEIS3P1, CDCP1, CD276, IL1R2, and OLFML2A) were ultimately retained. Then, the hub genes were used to construct a risk signature based on the coefficient for each patient according to the following formula: R S   =   ∑ i = 1 n   C o e f   ( i )   X   ( i ) , where Coef(i) denotes the coefficient, and X(i) is the z-score transformed relative expression level for each DEG. To validate the established model, the prognostic risk score of each patient was calculated using the formula, and, according to the median score, patients were classified into high- and low-risk groups. Then, time-dependent receiver operating characteristic (ROC) curve (tdROC) and Kaplan-Meier survival curve analysis were performed to verify this risk score system ( Figures 4C, D ).

Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis of differentially expressed ECHDC3 were performed using the R package “cluster profile”. GO terms were classified as molecular functions (MFs), biological processes (BP), and cellular components (CCs). Both p- and q-values less than 0.05 were considered to indicate statistically significant differences. Gene set enrichment analysis (GSEA) was performed to find enriched term differences.

The gene transcriptome data were used to estimate the content of multiple immune cells infiltrated in the BM microenvironment. CIBERSORT in combination with the LM22 signature matrix was used to estimate the fractions of 22 human immune cell phenotypes between AML samples with high or low ECHDC3 expression. The CIBERSORT analysis was conducted in R software by the CIBERSORT R script (version 1.03). The correlation analysis of the index was completed using the Spearman’s test.

ECHDC3 siRNA was designed and synthesized by GenePharma (Shanghai, China). The plasmid was then transfected into the K562 cell lines. Opti-MEM medium and Lipofectamine L300015 reagent were used for the transfection of siRNA according to the manufacturer’s instructions. After transfection, whole-cell cDNAs were synthesized, and quantitative real-time PCR was performed to verify the ECHDC3 expression level.

Ara-C (cytarabine) and doxorubicin (DOX) were purchased from Selleck Chemicals (Selleck, USA). Cell viability was determined by CCK-8 assay (Bimake, Cat#B34034, USA). The cells were seeded in a density of 2 × 10⁴ cells/100 μl in a 96-well plate and incubated at 37°C for 48 h. Ten microliters of CCK-8 was added to each well of the 96-well plate, followed by further incubation for 1 h. The absorbance was measured at 450 nm. The cell viability was calculated using GraphPad Prism software.

Whole-cell cDNAs were synthesized, and quantitative real-time PCR was performed using the FastStart Universal SYBR Green Master mix (Millipore Sigma, MA, USA) with a StepOnePlus real-time PCR system (ABI Prism 7900HT; Applied Biosystems, USA). The threshold cycle (Ct) values of target genes were normalized over the Ct of β-actin in whole-cell lysate. For comparison, the siCT group was set as 1 unless otherwise indicated. Primers used for real-time PCR and quantitative real-time PCR are listed in Table S1 .

R software 4.05 and GraphPad Prism 8.0.2 were mainly used for laboratory data statistical analysis. Comparisons between the two groups were performed using the Mann-Whitney U-test for continuous variables and the X2 test for categorical data. P < 0.05 marked with *, P < 0.01 marked with **, and P < 0.005 marked with *** are considered significant. The survival functions were estimated by the Kaplan-Meier method using the log-rank test with asymmetric 95% confidence intervals (CIs). Variables included in the multivariate analysis (MVA) were selected by the backward elimination process with a criterion of P < 0.10 for retention. Age is linear with estimates of HRs for 10-year difference.

The total study design is illustrated in Figure 1 .

In the training cohort, 24 AML (APL) patients were enrolled between September 2019 and September 2020, followed up to July 2022. According to the median protein level of ECHDC3 in enriched CD34⁺ cells at diagnosis, patients were divided into “ECHDC3^high” and “ECHDC3^low” groups, and the baseline characteristics were similar between the two groups ( Table S2 ). By median follow-up of 23.7 months, 11 patients experienced relapse, which corresponds to a 2-year probability of relapse at 45.8% (95% CI, 24.1–65.1), in which 5 patients achieved CR2 and preceding to allo-HSCT. None of these patients had non-relapse mortality. OS was significantly lower in the ECHDC3^high group than in the ECHDC3^low (2-year OS, 55.6% vs. 100%; 95% CI, 23.1–79.0 vs. 100%; p = 0.011, Figure 2A ). There was a trend that ECHDC3^high conferred inferior leukemia-free survival (2-year LFS, 41.7% vs. 66.7%, p = 0.178, Figure S1A ) but without statistical significance.

In the validation set, 351 BM RNA-seq data of non-APL AML were obtained by two independent online datasets (TCGA-LAML and BEAT-AML). Patients were divided into “ECHDC3^high” and ECHDC3^low” groups according to the median expression level of ECHDC3 in BM, and the baseline characteristics were listed in Table 1 . OS was significantly lower in the ECHDC3^high group than in the ECHDC3^low (5-year OS, 9.6% vs. 24.3%; p = 0.002; Figure 2B ). Correspondingly, the alive patients showed a lower ECHDC3 expression level compared to dead patients (P = 0.0009; Figure 2C ). Of the RNA-seq data of 180 peripheral blood (PB) samples of non-APL AML in BEAT-AML, ECHDC3^high also showed inferior 5-year OS (3.5% vs. 14.8%; 95% CI, 23.1–79.0 vs. 100%; p = 0.045; Figure S1B ).

In the validation set, the expression of ECHDC3 was analyzed by age and ELN risk groups. Patients more than 60 showed higher ECHDC3 levels compared to those less or equal to 60 (p = 0.023; Figure 2D ). In different risk subgroups, patients of int risk or adv risk showed a higher level of ECHDC3 compared with those of favorable (fav) risk (p < 0.001; p < 0.001; respectively; Figure 2E ); meanwhile, there was no significant difference in ECHDC3 level between int- and adv-risk groups (p=0.23).

Furthermore, to test whether ECHDC3 would be a useful biomarker in re-stratification of AML patients, subgroup analysis was carried out in fav-, int-, and adv-risk AML, as well as in patients with or without NPM1/FLT3 mutation. ECHDC3^high predicted inferior 5-year OS in the subgroup of patients with ELN int-risk (9.5% vs. 26.3%, p = 0.039; Figure 2F ) but not in the fav-risk or adv-risk groups ( Figures S1C, D ). In addition, ECHDC3^high predicted inferior 5-year OS in the subgroup of FLT3+NPM1− adv-risk AML (6.4% vs. 31.8%, p = 0.003; Figure 2G ). There was a trend that ECHDC3^high related to inferior 3-year OS of FLT3−NPM1+ fav-risk AML (31.7% vs. 54.8%, p = 0.077, Figure S1E ) but without significance. Meanwhile, ECHDC3 did not affect OS in FLT3−NPM1− or FLT3+NPM1+ AML patients ( Figures S1F, G ).

As an expression of ECHDC3 in BM was associated with a high percentage of blast cells in PB and a low percentage of blast cells in BM ( Figures S1H ; Table 1 ), while there was a trend that a high percentage of blast cells in PB rather than BM might be associated with inferior OS ( Figure S1I, J ), subgroup analysis was carried out addressing the percentage of blast cells in PB, which indicated that ECHDC3^high would be a useful biomarker predicting 5-year OS in the subgroup with a high (p < 0.001) rather than a low percentage of blast cells (p = 0.537) in PB ( Figures S1K, L ).

Among the confounding factors in univariate analysis, ECHDC3, age, and risk stratification level were chosen for MVA ( Figure 3A ). ECHDC3 was confirmed as an independent prognostic factor for OS (HR 1.159, 95% CI 1.013–1.326, p = 0.032; Figure 3B ). Age is another independent prognostic factor (HR 1.291, 95% CI 1.114− 1.496, p < 0.001). Meanwhile, ELN risk might be an independent prognostic factor without statistical significance (HR 1.365, 95% CI 0.972−1.915, p = 0.072; Figure 3B ).

As ECHDC3, age, and ELN risk were independent risk factors for OS, then we combined these three factors as a new prognostic model and tested it in non-APL AML patients of TCGA-LAML datasets (excluding four cases due to missing data). Adding ECHDC3 expression to age and ELN risk improved the area under the curve (AUC) from 0.657 to 0.671 of ROC ( Figure 4A ). A risk-score model combining ECHDC3, age, and ELN risk was developed based on a linear combination of each parameter (Coef) weighted by the regression coefficient derived from the univariate Cox regression analysis (β). The equation was calculated as R S   =   ∑ i   =   1 154   C o e f   ( i )   *  β  ( i ) .According to the median risk score, patients with high-risk scores have inferior 5-year OS (7.6% vs. 27.7%, p < 0.001; Figure 4B ).

In non-APL AML patients of TCGA-LAML datasets, a total of 249 survival correlated genes (according to the criteria ∣log2FC∣ > 1 and adjusted p < 0.05) performed 6-fold cross-validation and 9 genes in addition to ECHDC3 were finally found with regression coefficients, namely, RPS6KL1, RELL2, FAM64A, SPATS2L, MEIS3P1, CDCP1, CD276, IL1R2, and OLFML2A ( Figures S2A–C ). Each gene was an independent prognostic factor by two-way interactions checked between the factors with the main effect in chi-square analysis ( Figure S2D ). Next, the prognostic model was established with Lasso regression to improve the predicted accuracy for overall survival in AML. The equation is calculated as R S   =   ∑ i = 1 154   C o e f   ( i )   *   g e n e   e x p r e s s i o n   ( i ) , and that is Y = ECHDC3 × 0.0127-RPS6KL1 × 0.0006 + RELL2 × 0.0256-FAM64A × 0.0823 + SPATS2L × 0.0611-MEIS3P1 × 0.0284 + CDCP1 × 0.0296 + CD276 × 0.0099 + IL1R2 ×0.0033 + OLFML2A × 0.1069 in this study. Y ≥ 3.090 is categorized as “high risk” ( Figures S2E–F ). The AUC of the ROC curve following this risk model is 0.792 ( Figure 4C ). Patients with high-risk score have inferior 5-year OS (9.3% vs. 23.5%, p < 0.001; Figure 4D ).

GSEA was performed to identify numbers of hallmark terms enriched in the high ECHDC3_exp and low ECHDC3_exp groups. The top 6 Hallmark terms identified in the low ECHDC3_exp group were mainly immune-related ( Figure S3A ), namely, complement system, IL6 signaling, IFN-α, IFN-γ, P53, and xenobiotic metabolism. Additionally, Figure 5A showed the KEGG term enriched in high ECHDC3_exp and low ECHDC3_exp groups.

The top 40 DEGs, including 20 upregulated genes and 20 downregulated genes in ECHDC3^high and ECHDC3^low groups, were demonstrated in the heatmap ( Figure S3B ). By comparison of the top KEGG pathways enriched in DEGs in ECHDC3^high and ECHDC3^low groups, it showed that these marker genes mainly enriched in the phagosome, the intestinal immune network for IgA production, cell adhesion molecules, viral myocarditis, viral myocarditis, allograft rejection, antigen presentation, type I diabetes mellitus, and staphylococcus aureus infection ( Figure S3C ). GO analysis further declared the potential function of the DEGs ( Figure S3D ). The DEGs were most significantly enriched in the BP term of “positively regulation of cytokine production” and “regulation of leukocytes proliferation”, in addition to the CC term of “collagen-containing extracellular matrix” and “external side of plasma membrane’’. In MF term, genes were enriched of “amide binding” and “peptide binding”. These results indicated that the DEGs were distributed in the immune system–related processes and pathways, which have been proved to play a pivotal role in the immune microenvironment. The top 8 enriched GO terms and demonstrated represents genes were demonstrated in Figure 5B .

In the BM microenvironment, immune infiltrating cells controlled the fate of AML cells. In this study, the composition of significant tumor-infiltrating immune cells was assessed by the CIBERSORT algorithm in the AML BM RNA-seq dataset. The histogram of immune infiltrating cells was shown in Figure S4A . Moreover, resting NK cells (p = 0.009), monocytes (p = 0.041), naïve CD4⁺ T cells (p = 0.065), and resting mast cell (p= 0.039) had significant differences in the immune cell fractions between the two groups ECHDC3^high and ECHDC3^low ( Figure 6 ). Next, the co-expression analysis using prognostic tumor-infiltrating immune cells in the AML cohort with ECHDC3^high and ECHDC3^low was performed. From Figures S4B–C , we can see that monocytes have a negative correlation with resting CD4⁺ T memory cells and CD8⁺ T cells in the immune phenotype profiles, whereas resting NK cells were positively correlated with T regulatory cells and resting CD4⁺ T memory cells. Considering all the interpreted results, higher ECHDC3 expression may induce NK, mast cell resting, and monocyte differentiation, and ECHDC3 may directly alter the BM microenvironment cell infiltration.

The human leukocyte antigen (HLA) system is an important part of the immune system, and HLA encodes cell surface molecules specialized to present antigenic peptides to the T-cell receptor (TCR) on T cells (12). Immune checkpoints are regulators of the immune system. These pathways are crucial for self-tolerance, which prevents the immune system from attacking cells indiscriminately. However, some cancers can protect themselves from attack by stimulating immune checkpoint targets (13). The expression information of the HLA family and immune checkpoints was shown ( Figures 7A, B ), which indicated that the HLA family and immune checkpoints were significantly different between the ECHDC3^high and ECHDC3^low groups. As can be seen in Figure 7B , TNF superfamily member 15 (TNFSF15) and CD40 were significantly upregulated in the ECHDC3^high subgroup with p < 0.001, whereas TNF superfamily member 4 (TNFSF4) were significantly downregulated (p = 0.003).

In the GSE52919 database, patients with Ara-C resistance showed higher ECHDC3 expression compared with sensitive patients (p = 0.0136; Figure 8A ).

To verify the role of ECHDC3 in mt-DNA transcriptome and chemoresistance by experiments, ECHDC3 was knocked down by RNAi (SiECHDC3-351 and SiECHDC3-417) in the K562 cell lines ( Figure 8B ). The mt-DNA transcriptomes (MT-CO1, MT-CO2, ND1, ND2, and ND3) did not alter by inhibiting ECHDC3 ( Figure 8C ), whereas knocking down of ECHDC3 promoted the death of leukemia cells in either Ara-C at 0.1–1 µM or DOX at 0.04–0.16 µg/ml ( Figures 8D–G ).