The datasets were obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). We searched the GEO database using the keywords “acute myeloid leukemia “ [MeSH Terms] AND “Homo sapiens” [porgn: txid9606] and “Expression profiling by array” [All Fields]. The criteria for screening included the following: the microarray dataset referred to the genome-wide gene expression profiles in blood. The microarray dataset contained samples from AML and samples from healthy conditions. The included samples were not associated with any other diseases. The number of AML samples needed to be greater than 10. Based on the above conditions, we screened GSE9476 (including 38 normal samples and 26 AML samples) and GSE24395 (including 5 normal samples and 12 AML samples). These datasets were merged to form a new dataset which eliminated batch effects to form the experimental dataset. GSE30029 (comprising 31 normal and 90 AML samples) served as the validation dataset.

We used the “limma” package in R to identify DEGs, with a threshold of P < 0.05 and |log2 FC| > 2. Subsequent analyses included gene ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses, and immune-related gene set enrichment analysis (GSEA). Statistical significance was defined by a P value of 0.05, with a threshold of |log2 FC| > 1 applied for GSEA.

We clustered the samples and removed outliers. The optimal power value was determined to be 11, which was used to assess the fit index and average connectivity. Based on this optimal power value, a scale-free network was constructed. The efficacy of this construction was evaluated by plotting the topology of the scale-free network, which allowed for the generation of a distance matrix for gene clustering. Subsequently, dynamic module identification was conducted, focusing on modules containing at least 30 genes. Highly correlated modules were clustered and merged. Heatmaps illustrating module-clinical trait relationships and gene significance were generated to identify key modules, with parameters set at GS > 0.5 and MM > 0.8 to ultimately determine the hub genes.

We further employed the JSVM-RFE algorithm for feature gene selection. The results intersected with the genes from the key modules identified in WGCNA and the DEGs, ultimately yielding the hub genes.

Firstly, we constructed a receiver operating characteristic (ROC) curve to validate the hub genes. Gene expression was then compared between the two groups using box plots for both experimental and validation datasets. Additionally, LASSO regression was utilized for cross-validation. We also utilized the GEPIA database (http://gepia.cancer-pku.cn/) to perform survival analysis, evaluating the diagnostic accuracy.

GSEA was performed using gene sets that synergistically interact with the hub genes, allowing for the identification of enriched pathways associated with these gene sets.

Immune-related single-sample GESA (ssGSEA) enrichment analysis and CIBERSORT immune infiltration analysis were conducted. These two methods provided complementary insights into the immune landscape.

The IEU database (https://gwas.mrcieu.ac.uk/) served as the source for this portion of the data. The IL-2 dataset (GWAS ID: prot-c-3070_1_2) include 501,428 SNPs from a European population. The EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/) were source of the data on AML and immune cells, with the AML accession number GCST90435652. Immune cell data were collected under accession numbers GCST90274758 to GCST90274848, encompassing 728 immune cell types along with their corresponding GWAS IDs ( Supplementary Table S1 ), all derived from a European population.

Genome-wide significant SNPs with a threshold of P < 5×10−8 were included. In the absence of such SNPs, we considered those with P < 5×10−6 as potential instruments. We clustered SNPs based on linkage disequilibrium (window size = 10,000 kb and r² < 0.001), excluding weak instrumental variables (F-statistics < 10).

We used inverse variance weighting (IVW) and MR-Egger methods as the primary methods for assessing causal relationships. Both methods needed to achieve a significance threshold of P < 0.05, and if neither method achieved this level, the IVW results were prioritized. IVW combines the causal effects represented by the Wald ratio of each SNP through meta-analysis, relying on the assumption that all SNPs are valid instruments. Therefore, this approach could be applied only after excluding SNPs exhibiting pleiotropy.

Firstly, the causal relationship between IL-2 and AML was evaluated using a two-sample bidirectional MR. Directionality was assessed using P values for either IVW or MR-Egger. If this condition was not met, the direction of overall effect was derived from the cumulative steps in the decomposition process, ensuring P values remained below 0.05 in each step, while also calculating the total utility.

Bulk MR analyses were performed using 728 immune cell types as exposure and AML as the outcome, identifying immune cells that yielded significant results. IL-2 was treated as the exposure, with the selected immune cells as outcomes, identifying the double-positive immune cells. Refer to the method in the literature (21). We conducted a three-step MR analysis for mediation assessment. In the first step, IL-2 was used as the exposure, with the AML as the outcome to calculate the effect (beta_all). In the second step, IL-2 was used as the exposure, with the identified double-positive immune cells as the outcome to calculate the effect (beta1). In the third step, these double-positive immune cells served as the exposure, with AML as the outcome to calculate the effect (beta2). Different SNPs were utilized in each step to investigate whether immune cells mediate the association between IL-2 and AML. The overall effect of IL-2 on AML included its direct effect on AML and an indirect effect mediated through immune cells. The mediation effect was assessed as the indirect effect divided by the overall effect. Additionally, the delta method was employed to calculate the 95% confidence intervals (CI).

We assessed heterogeneity and horizontal pleiotropy by calculating P values. P > 0.05 indicated no significant heterogeneity or pleiotropy. Outliers were removed when detected, and causal estimates were recalculated. If significant heterogeneity persists following removal, a random effects model would be applied to assess result stability, as this model is less sensitive to weak SNP-exposure associations. We also conducted a leave-one-out analysis to evaluate the impact of each SNP on the overall causal estimate.

This study protocol was reviewed and approved by Ethics Committee of Gaomi Maternity and Child Health Hospital, approval number 20230206-09. For studies in which human tissues, fluid, or cell lines were used, written informed consent was obtained from the donors’ parents to participate in the study. Donors’ parents signed an informed consent according to the principles outlined in the Declaration of Helsinki.

The human myeloid leukemia cells (KG-1a) were obtained from CELLCOOK (Guangzhou, Guangdong, China) and validated via STR analysis ( Supplementary Figure S1 ). The control group consisted of human umbilical cord blood stem cells sourced from children at our institution who had consent from their legal guardians. A portion of these stem cells was sent to the Qilu Stem Cell Bank for testing, while the remainder was stored in liquid nitrogen in our laboratory. Successful testing by the Qilu Stem Cell Bank indirectly confirms the usability of the stem cells stored in our laboratory. The culture, cryopreservation, and passaging of KG-1a cells were conducted according to the product manual ( Supplementary Figure S2 ). The cells (1 × 105) were cultivated in each well of six-well plates. In the experiment, cells (mRECs) from passages 3 to 6 are used.

The Homo sapiens CFD gene sequence was retrieved and downloaded from NCBI ( Supplementary Table S2 ). Primers were designed using the coding sequence (CDS) of the target gene, excluding the stop codon, and using XbaI and Eco53KI restriction sites, at both ends. Perform double enzyme digestion using XbaI restriction endonuclease (Biosharp, Shanghai, China) and Eco53KI restriction endonuclease (KALANG, Shanghai, China). Plasmids were constructed by restriction‐enzyme double digestion and ligation. Plasmid pBI121(HonorGene, Changsha, Hunan, China) was selected as the expression vector ( Supplementary Figure S3 ).

The digested fragments and vectors were ligated to construct recombinant plasmids ( Supplementary Figure S4 ). The ligation product was then transformed into the competent Escherichia coli DH5α cell (Whenzhou KeMiao Biological Technology Co.,Ltd, Wenzhou, Zhejiang, China), and the competent bacterial strain was revived on blasticidin-free media. A portion of the cells was plated on plates with kanamycin (Eta Biology,Beijing,China) resistance. After single colonies emerged, several were randomly selected for qPCR analysis following plasmid transfection into target cells. This process allowed for the assessment of hub gene expression in the target cells and confirmed the successful construction of the plasmids.

A total of three groups of cells were analyzed, including the normal group (human umbilical cord blood stem cells), the control group (KG-1a cells), and the experimental group (human umbilical cord blood stem cells transfected with plasmids). Each group had three compound holes. All groups were supplemented with 4500 µL of 20% DMEM culture medium (absin,Shanghai,China), 500 µL of fetal bovine serum(opcel,Shanghai,China), and 200 µL of P/S (penicillin and streptomycin) dual antibiotics (absin,Shanghai,China). Cells in the logarithmic growth phase were selected and transfected with plasmid DNA using Lipofectamine 2000 (Invitrogen,Hangzhou,Zhejiang,China),with three compound holes for each group.

We extracted the total RNA from each cell group, and reverse transcription was performed to synthesize cDNA using specific primers ( Supplementary Table S3 ). Data were analyzed using the 2-ΔΔCt method for quantification. GAPDH was the internal reference gene. The detection was performed with the Gentier 96E fluorescence quantitative PCR instrument made by TIANLONG, a company in China.

We used the RT-qPCR to compare the hub gene expression between the normal and control groups. After culturing cells for 72h, we utilized RT-qPCR to assess the hub gene expression in both the experimental group and the normal group. Subsequently, we replaced the fetal bovine serum in the culture medium with human serum, which was derived from residual blood collected post-transfusion in neonates with coagulation disorders at our institution. After an additional 72h, we performed RT-qPCR to evaluate the hub gene expression and the associated JAK-STAT and PI3K-Akt signaling pathways across all groups following induced overexpression. The experiment was repeated twice. Moreover, perform cell proliferation assays using the MTT Cell Proliferation and Cytotoxicity Detection Kit - 500T (Wanlei Biotechnology, Shanghai, China) and the Multiskan™ FC microplate reader (ThermoFisher, USA). Three parallel holes were set in each group, and the experiment was repeated three times. Referring to the literature (22), the concentration of MTT is 0.1 mg/mL; the wavelength of transmitted light is 565 nm. In 96-well plates, 5000 cells are seeded in each well.