AML-related gene expression profiles and clinical data were extracted from the TCGA (https://portal.gdc.cancer.gov/) and GEO (https://www.ncbi.nlm.nih.gov/gds) databases. A total of 151 AML patients with RNA-seq data were enrolled from the TCGA database, which were downloaded by the TCGAbiolinks R package (17). As suggested by the tutorial, data filtering was performed using function rpkm in edgeR package (18).

After searching from the GEO database, another five datasets with survival outcomes were downloaded using R package GEOquery in this study (GSE37642, GSE76009, GSE16432, GSE12417, and GSE71014) (19). The number of patients in these five datasets was as follows: 562 in GSE37642, 534 in GSE76009, 436 in GSE16432, 405 in GSE12417, and 104 in GSE71014, respectively. As a result, a total of 2041 patients from the GEO database were included for subsequent analysis.

To improve the accuracy of network construction, the genes used for WGCNA were filtered. The R package WGCNA was used to conduct WGCNA between the gene expression data profiles of TCGA-LAML and GEO datasets, grouping highly co-expressed genes into modules (14). We use pickSoftThreshold to establish a scale-free network. A similarity matrix was built after performing the Pearson correlation of all gene pairs. The adjacency matrix was then transformed into a topological overlap matrix (TOM) and the TOM-based dissimilarity matrix for hierarchical clustering. Then, we correlated previously computed module features with clinical characteristics to identify the co-expression network’s functional modules.

To figure out the differentially expressed genes (DEGs) between AML patients with different survival status (alive or dead), the R package limma was applied in the TCGA and GEO databases. The DEGs were screened with the criteria |logFC| ≥ 1.0 and adj. P-value < 0.05. The R package ggplot2 and VennDiagram were used to draw a volcano map and Venn diagram (20).

To determine the functional relevance of the modules, we tested whether the selected genes from the modules were enriched for specific functions or signaling pathways. Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were performed by the clusterProfiler R package (21). p<0.05 was considered statistically significant.

We used the online database STRING (https://string-db.org/) to construct a protein-protein interaction (PPI) network. The PPI network was analyzed and visualized using Cytoscape software (v3.7.2) with an extraction score ≥ 0.4. In the PPI network, the nodes stand for proteins, and the edges represent the interactions of proteins. To identify the hub genes from the PPI network, we used the maximal clique centrality (MCC) algorithm calculated by a plugin in Cytoscape named CytoHubba. Genes with top 10 MCC values were set as hub genes.

To verify the reliability of the hub genes, we analyzed the expression level of hub genes in AML patients from the GEO database with different survival status (alive or dead). The expression level was represented with a boxplot for each gene. Univariate Kaplan-Meier survival analysis was done with R package survival to identify the prognostic value of hub genes further. Overall survival (OS) was defined as the time from diagnosis to death, regardless of the causes or last follow-up. Relapse-free survival (RFS) was defined as the time from complete remission to recurrence. Statistical significance was determined by Student’s t-test for two groups and one-way ANOVA was used for comparison among multiple groups. p<0.05 was regarded as statistically significant.

Human myeloid leukemia cell lines (MV4-11, MOLM13, KASUMI-1, OCI-AML3, SKM1, THP-1) were obtained from the American Tissue Culture Collection (ATCC). All cells were cultured in RPMI 1640 (Gibco, Detroit, MI, USA) medium, supplemented with 10% fetal bovine serum (Gibco, Detroit, MI, USA), 1% penicillin-streptomycin (Gibco, Detroit, MI, USA) at 37 ˚C with 5% CO₂.

Bone marrow samples were obtained from 46 patients with newly diagnosed AML and 11 healthy donors (HD) from the First Affiliated Hospital of Soochow University between April 2020 and March 2021. The induction chemotherapy is a standard first-line treatment including an IA/DA regimen (Either idarubicin 8-12 mg/m² or daunorubicin 60-90 mg/m² on days 1-3 and cytarabine at 100mg/m² on days 1-7). For some older patients or patients with organ dysfunction, CAG or revised CAG (IAG, HAG) were administered with or without hypomethylating agents. Bone marrow mononuclear cells were isolated by Ficoll density gradient centrifugation. Basic clinical information, including age, gender, white blood cell count, hemoglobin level, platelet count, bone marrow morphology, karyotype, and molecular information was acquired from each patient. Patient risk stratification was based on 2017 ELN genetic categories. Receiver operating characteristics (ROC) curve analysis was performed based on the patient survival status to determine the optimal cut-off value of TRIM32. Informed consent was available from all patients and approved by the Medical Ethics Committee of the First Affiliated Hospital of Soochow University. The study was carried out following the Declaration of Helsinki.

We used lentivirus vector pLKO.1-TRC-EGFP for carrying TRIM32 target shRNA and for non-silencing control. The sequence for TRIM32 shRNA was: GGUGGAAAGCUUUGGUGUU. Cells were transfected with the indicated plasmids by using calcium chloride.

RT-qPCR measured the mRNA expression of TRIM32. Total RNA was extracted from Ficoll-separated mononucleated bone marrow samples in AML patients and healthy donors using TRIzol reagents (Invitrogen, USA). Reverse transcription was performed at 25°C for 10 minutes, 42°C for 15 minutes, and 85°C for 5 minutes. qPCR was performed at 95°C for 10 min, followed by 40 cycles of amplification at 95°C for 15 s and 60°C for 60 s. The GAPDH mRNA was used as an internal control. The sequences of the oligonucleotides used in this study were as follows: TRIM32, forward, 5’-CTCGG-AAGTTCTTCACAGGCTC-3’ and reverse, 5’-CTCCAGTAGTGCTA-CATCTGCC-3’; GAPDH, forward, 5’-GAGAAGGCTGGGGCTCATT-T-3’ and reverse, 5’-ATGACGAACATGGG-GGCATC-3’.

Cells were lysed on ice for 30 min with RIPA buffer (Beyotime) containing protease and phosphatase inhibitors (Beyotime). After centrifuging at 12,000g for 15 min at 4°C, the supernatant was re-suspended in buffer and heated at 100°C for 10 min. Protein quantification was performed by a BCA protein assay kit (Thermo). Equal amounts of proteins were electrophoresed in 10% sodium dodecyl sulfate (SDS) polyacrylamide gels and then transferred onto polyvinylidene difluoride (PVDF) membranes (Pierce, USA). After being blocked with 5% non-fat dried milk in Tris-buffered saline containing 0.1% Tween 20. The membranes were incubated at 4°C overnight with antibody TRIM32 (Abcam) and β-actin (Proteintech). After being incubated with appropriate secondary antibodies, proteins were visualized by a detection system of enhanced chemiluminescence (ECL).

Proliferation assay was performed using a Cell Counting Kit-8 (CCK-8, Bimake). Appropriately 5 ×10³ cells/well were plated in a 96-well plate and cultured in complete medium. 10 μL of CCK8 reagent was added into each well and incubated for 2 hours at 37 °C with different time points of 0, 24, 48, and 72h. The absorbance was read at 450nm.

The flowchart of this study is shown in Figure 1 . To identify key gene sets for AML, we performed a WGCNA from the TCGA database as well as the GSE37642, GSE76009, GSE16432, GSE12417, and GSE71014 datasets based on the WGCNA package ( Table S1 ). Each color represents a module. Here, 13 unique modules in TCGA-LAML ( Figure 2A ) and 8 modules in the above five GEO datasets ( Figure 2C ) were identified. Figures 2B, D showed heat maps of module-trait relationships, indicating that the green modules in TCGA and the brown modules in the GEO dataset were highly distinguishable between tumors and normal individuals (green modules: r=0.27, p=8e-04, brown modules: r=0.034, p=0.2).

A total of 624 DEGs in TCGA ( Figure 3A ) and 338 DEGs in the five GEO datasets ( Figure 3B ) were screened by the limma package with |logFC|≥1.0 and adj. p<0.05 as cut-off criteria. 1122 co-expressed genes were extracted from these significant modules based on WGCNA analysis (850 genes in the green module genes and 272 genes in the brown module). Finally, 20 overlapping genes were obtained at the intersection of the DEGs list and the two co-expression modules ( Figure 3C ).

To comprehensively understand the potential function of these genes in the co-expression modules, the GO ( Figure 4A ) and KEGG ( Figure 4B ) enrichment analyses for the genes in the brown module were performed using the R package “ClusterProfiler.” GO analysis divides genes into three categories: biological processes (BP), cellular components (CC), and molecular functions (MF). In BP, genes were enriched in phospholipid metabolic processes, neutrophil degranulation, and activation. In the CC category, many genes were involved in secretory granule membranes, endocytic vesicles, and glutamatergic synapses. In addition, protein serine/threonine kinase activity and DNA-binding transcription factor binding were mainly involved in MF. KEGG showed that genes were enriched in choline metabolism in cancer.

To investigate the correlation between DEGs and co-expression modules, we performed protein-protein interaction (PPI) analysis through the STRING database, and a PPI network containing 15 nodes and 15 edges was created in Figure 4C . In addition, the top 10 hub genes with MCC scores in the PPI network were calculated using Cytoscape’s CytoHubba plugin, which included TLE1, GLI2, HDAC9, MICALL2, DOCK1, PDPN, RAB27B, SIX3, TRIM32, and TBX1 ( Figure 4D ).

To further determine the expression levels of hub genes in AML patients with different survival status (alive or dead), we mapped the expression in the GEO database. In addition to TLE1, four hub genes (DOCK1, GLI2, RAB27B, and TRIM32) were highly expressed in those who died, while the remaining five hub genes (HDAC, MICALL2, PDPN, SIX3, and TBX1) were expressed at lower levels ( Figure 5 ). Then, to further explore the prognostic value of these ten hub genes, we performed a Kaplan-Meier survival analysis via the survival R package of the GEO dataset and was illustrated in Figure 6 . The increased expression levels of DOCK1, GLI2, and TRIM32 were highly correlated with poor prognosis in AML (p<0.05).

To date, the role of TRIM32 has not been reported in AML previously. Thus, we validated the role of TRIM32 with the clinical data of patients from our center. The mRNA levels of TRIM32 in 46 AML and 11 HD bone marrow samples were also detected by RT-qPCR. As shown in Figure 7A , TRIM32 expression was significantly upregulated in AML compared with that in HD (p<0.05). All patients were divided into two groups based on the optimal cut-off value of TRIM32 determined by the ROC curve (AUC: 0.682, sensitivity:91.7%, specificity: 43.3%) ( Figure S1 ). The clinicopathological characteristics of these patients were summarized in Table 1 . The expression of TRIM32 was correlated with biallelic CEBPA and PTPN11 mutation. Patients with lower expression of TRIM32 were more likely to achieve complete remission (CR) after the first induction therapy. Additionally, patients with elevated expression of TRIM32 were more frequently found in the adverse-risk group based on 2017ELN. The results of univariate and multivariate analyses of risk factors for OS were shown in Table 2 . Older age, not achieving CR after first induction therapy, not receiving HSCT, gene mutation status (PTPN11^mt, RUNX1^mt and IDH1^mt) and high expression of TRIM32 were associated with poor OS in univariate analysis ( Figure 7B ). However, TRIM32 did not play an independent role in OS while RUNX1 mutation proved to be an independent risk factor for worse OS after multivariate analysis (HR:7.441, 95%CI: 1.595-34.712, p=0.011). Receiving HSCT was an independent protective factor (HR:0.104, 95%CI: 0.020-0.531, p=0.007). Similar results were observed for RFS shown in Table 3 . Patients with overexpression of TRIM32 experienced a remarkably worse RFS in Figure 7C (p=0.025), although it was not an independent prognostic factor. Taken together, both the external database and clinical data in our center show that TRIM32 does have a certain influence on the prognosis of AML, whose role needs to be further studied.

We investigated the relative expression of TRIM32 mRNA levels in six AML cell lines (MV4-11, MOLM13, KASUMI-1, OCI-AML3, SKM1, THP-1) and bone marrow of HD ( Figure 8A ). TRIM32 was highly expressed in AML cells compared with HD. We also measured the protein level of TRIM32 in these cell lines ( Figure 8B ). To assess the potential role of TRIM32 in proliferation in AML cell lines, we chose THP-1 and KASUMI-1with higher expression of TRIM32 to stably knockdown TRIM32 using a lentiviral delivery system. The efficiency of the shRNA in THP-1 was verified by RT-qPCR ( Figure 8C ) and western blot ( Figure 8D ). Compared with the empty vector (pLKO.1), shRNA knockdown of TRIM32 significantly inhibited proliferation in vitro as indicated by CCK-8 assay ( Figure 8E ). Similar results were obtained in another AML cell line KASUMI-1 shown in Figure S2 .