Lovastatin Targets LIPA to Induce ER Stress-Mediated Apoptosis in Acute Myeloid Leukemia: A Multi-Omics Study

Abstract

Background

Acute myeloid leukemia (AML) remains a therapeutic challenge, necessitating the identification of novel targets and repurposable drugs. This study integrates multi‑omics and Mendelian randomization (MR) to investigate the role of lysosomal acid lipase (LIPA) in AML and to explore the potential of lovastatin as a LIPA‑targeting agent.

Methods

We combined transcriptomic data from TCGA and GTEx with MR analysis using eQTLs to assess the causal relationship between LIPA expression and AML risk. A prognostic signature was constructed via LASSO and validated in external GEO cohorts. Network pharmacology, molecular docking, and molecular dynamics simulations were employed to identify drugs targeting LIPA. In vitro, AML cell lines (THP‑1, K562) were treated with lovastatin and/or the ER stress inhibitor 4‑PBA; apoptosis, ER stress markers, and ultrastructure were assessed by flow cytometry, qPCR, Western blot, and transmission electron microscopy.

Results

MR established a causal link between elevated LIPA expression and increased AML risk (OR=1.32, p=0.003). A 16‑gene prognostic signature including LIPA effectively stratified patients (p<0.0001). Lovastatin was identified as a potential high‑affinity LIPA inhibitor. In vitro, lovastatin induced marked apoptosis in AML cells, which was accompanied by downregulation of ER stress markers (ATF6, CHOP, IRE1) and constricted ER morphology. Notably, the ER stress inhibitor 4‑PBA phenocopied these effects, consistent with lovastatin exerting its anti‑AML activity through suppression of ER stress.

Conclusion

This multi-omics study establishes LIPA as a causal prognostic biomarker in AML and reveals that lovastatin triggers apoptosis by inhibiting ER stress, providing a mechanistic rationale for repurposing lovastatin in AML therapy.

Graphical Abstract

Introduction

Acute myeloid leukemia (AML) is a clonal disorder of hematopoietic stem and progenitor cells characterized by abnormal proliferation, impaired differentiation, and accumulation of immature myeloblasts.¹ While therapeutic advances—including targeted agents and allogeneic hematopoietic stem cell transplantation—have improved outcomes for some patients, AML remains challenging to treat. Many patients, particularly the elderly, exhibit primary resistance or acquire resistance to standard therapies, and relapse after transplantation remains a significant concern.^2,3 These limitations underscore the urgent need for novel therapeutic targets and strategies.

Lipid metabolism reprogramming has emerged as a critical driver of tumorigenesis, progression, and drug resistance in various cancers, including AML.^4–6 Dysregulated cholesterol homeostasis, in particular, contributes to chemoresistance and supports leukemic cell survival.⁷ Among the key regulators of lipid metabolism, lysosomal acid lipase (LIPA)-the sole enzyme responsible for hydrolyzing cholesteryl esters and triglycerides in lysosomes-has recently gained attention as a potential therapeutic target in solid tumors.^8,9 Elevated LIPA expression has been reported in ovarian and breast cancers, where it promotes tumor progression and predicts poor prognosis.^10,11 However, the role of LIPA in AML remains unexplored.

Statins, competitive inhibitors of HMG-CoA reductase, are among the most widely prescribed lipid-lowering drugs. Beyond their cardiovascular benefits, statins have shown anti-tumor activity in multiple cancer types, including AML.^12,13 Clinical trials have demonstrated that adding pravastatin to standard chemotherapy can improve response rates in AML patients, particularly those with poor-risk features.^14,15 However, the underlying mechanism remains poorly understood, and whether statins exert their anti-leukemic effects by modulating LIPA activity has not been investigated.

Endoplasmic reticulum (ER) stress and the unfolded protein response (UPR) play complex, context-dependent roles in cancer. While severe ER stress triggers apoptosis, a basal level of UPR activation is often essential for tumor cell survival-a phenomenon termed “ER stress addiction”.^16,17 Emerging evidence links lipid metabolism to ER stress regulation, suggesting that LIPA-mediated lipid handling may influence ER homeostasis and cell fate decisions.^18,19

In this study, we integrate Mendelian randomization, transcriptomics, and machine learning to: (1) establish LIPA as a causal prognostic biomarker in AML; (2) identify lovastatin as a potential LIPA-targeting drug through network pharmacology and molecular dynamics simulations; and (3) experimentally validate that lovastatin induces apoptosis in AML cells by suppressing ER stress. Our findings provide a mechanistic rationale for repurposing lovastatin in AML therapy and highlight LIPA as a novel therapeutic target.

Materials and Methods

Raw Data

Transcriptomic and clinical data of 151 AML patients were obtained from the TCGA-AML project, while RNA-seq data of 70 normal whole blood samples were downloaded from the GTEx database. RNA-seq expression data were obtained in FPKM format and log2-transformed as log2(FPKM+1) for downstream analysis. To minimize systematic bias between TCGA and GTEx cohorts, batch effect correction was performed using the ComBat function in the “sva” R package. The batch variable was defined as data source (TCGA vs GTEx), while biological condition (AML vs normal) was retained in the design matrix to preserve biological differences. Principal component analysis (PCA) was conducted before and after batch correction to evaluate residual batch effects. Differential expression analysis was performed using the “limma” package with linear modeling and empirical Bayes moderation. Genes with |log2 fold change| > 1 and false discovery rate (FDR) < 0.05 were considered differentially expressed genes (DEGs). For prognostic model construction, GSE12417 (n = 242) and GSE37642 (n = 553) datasets were retrieved from GEO. Raw CEL files (Affymetrix platform) were processed using the “affy” package for background correction and quantile normalization. Probe-level data were log2-transformed and mapped to gene symbols according to platform annotation. Batch effects between the two GEO datasets were further corrected using ComBat prior to merging. Survival analysis and model training were conducted on the merged GEO cohort. The number of observed events (deaths) was recorded to ensure adequate event-per-variable ratio for Cox regression modeling.

Screening of Expression Quantitative Trait Locus (eQTL) Exposure Data and AML Outcome Data

The single nucleotide polymorphisms (SNPs) associated with 19,942 eQTLs were obtained by utilizing the function “extract_instruments” from the TwoSampleMR R package. The eQTL summary statistics were derived from the eQTLGen Consortium, a large-scale meta-analysis of whole-blood expression quantitative trait loci, comprising over 31,000 individuals of predominantly European ancestry (https://www.eqtlgen.org/). Eligible SNPs were filtered following criteria of p < 5e-08, clump_r2 = 0.001, and clump_Kb =10000. In this study, the AML outcome data encompassed 10,534,735 SNPs obtained from 3301 AML H samples from European (GWAS ID: prot-a-235).

Two-Sample MR Analysis Between Exposure and Outcome Data

The two-sample Mendelian randomization (MR) analysis was conducted under three core assumptions: (1) the genetic variants are robustly associated with the exposure (relevance assumption); (2) the variants are independent of confounders (independence assumption); and (3) the variants influence the outcome exclusively through the exposure (exclusion restriction assumption). SNPs associated with the exposure were selected at genome-wide significance (p < 5×10⁻⁸) and clumped based on linkage disequilibrium (r² < 0.001, window size = 10,000 kb). F-statistics were calculated to evaluate instrument strength, with F > 10 considered indicative of sufficiently strong instruments. Harmonization of exposure and outcome datasets was performed to align effect alleles. Five complementary MR methods were applied using the “TwoSampleMR” R package: inverse variance weighting (IVW), MR-Egger, weighted median, weighted mode, and simple mode. The primary causal estimate was derived from the IVW method. Heterogeneity among SNPs was assessed using Cochran’s Q test. Horizontal pleiotropy was evaluated using the MR-Egger intercept test. Sensitivity analyses included leave-one-out analysis to assess the influence of individual SNPs. Statistical significance was defined as p < 0.05 for the IVW estimate, while a non-significant MR-Egger intercept (p > 0.05) indicated absence of directional pleiotropy.

Construction and External Validation of the AML Prognostic Model

Differentially expressed genes (DEGs) were first subjected to univariate Cox proportional hazards regression to identify genes associated with overall survival (OS) in AML. Multiple testing correction was performed using the Benjamini–Hochberg method, and genes with adjusted p < 0.05 were considered prognostically relevant. To reduce dimensionality and minimize overfitting, candidate genes were further analyzed using LASSO Cox regression implemented in the “glmnet” R package. Ten-fold cross-validation was applied to determine the optimal penalty parameter (λ). Genes with non-zero coefficients at the optimal λ were retained for subsequent modeling. These genes were then entered into a multivariate Cox proportional hazards regression model. The proportional hazards assumption was evaluated using Schoenfeld residuals. Model selection was guided by the Akaike Information Criterion (AIC), and the final prognostic signature was established based on the model with the lowest AIC value. The risk score for each patient was calculated as a linear combination of gene expression levels weighted by their corresponding multivariate Cox regression coefficients. External validation was performed using the TCGA cohort. Time-dependent receiver operating characteristic (ROC) curves were generated using the “timeROC” R package to evaluate predictive performance at 1-, 3-, and 5-year time points.

Machine Learning and SHAP-Based Feature Interpretation

In our study, utilizing the merged cohort dataset, we employed ten mainstream machine learning algorithms with the caret package (https://cran.r-project.org/package=caret) to construct diagnostic prediction model for identifying the efficacy of multiple key genes in discriminating AML. Specific algorithms were Partial Least Squares, Random Forest (RF), Decision Tree, Support Vector Machine, Logistic Regression, K-Nearest Neighbors, eXtreme Gradient Boosting, Gradient Boosting Machine, Artificial Neural Network (NeuralNet), and Generalized Linear Model Boosting. Using the createDataPartition function, 70% of the TCGA-GTEx were sampled randomly as the training set, and the remaining 30% were classified into the testing set. ROC curves integrated the predicted probabilities of the testing set, and the pROC package was adopted for calculating the AUC value, with a purpose of clarifying the model’s classification performance.

We selected the machine learning model with the highest AUC as the optimal model to interpret features. The SHapley Additive exPlanations (SHAP) value was computed with the permshap function, so as to quantify the contribution of each feature gene to the model output. In addition, SHAP bar plots, bee-swarm plots, and waterfall plots (the shapviz package) illustrate the representative samples.

Revealing and Characterization of Hub Genes Along with Their Expression, Prognosis and Prediction of Drug Targets

An intersection analysis of machine learning diagnostic model genes and prognostic model genes was adopted to identify the core genes, with the construction of a Venn diagram. Our subsequent comprehensive analyses focused on the expression of core genes, prognosis, and ROC curves. Besides, the Drug-Gene Interaction database (DGIdb, https://www.dgidb.org/) was visited to conduct an analysis on potential drug targets of core genes, with result visualization using CytoScape.

Analysis of the Expressions of Hub Genes and Immune Cell Infiltration in AML Patients

GEO, TCGA, and GTEx cohorts were integrated to establish a combined cohort, comprising 70 normal samples and 935 AML samples with comprehensive survival data. Then, the CIBERSORT method was adopted to calculate the relative proportions of 22 immune cell types in each AML sample, coupled with the elucidation of relevance of hub genes to individual immune cell populations through Spearman correlation analysis. Additionally, this analysis also focused on the examination of the impact of the immune cell infiltration extent on the AML prognosis.

Molecular Docking

After obtaining the 3D structure of the protein in Uniprot, we first acquired the sdf structure of the ligand from Pubchem database, and then converted it into PDB file by OpenBabel. After that, the protein target underwent dehydrogenation and hydrogenation with AutoDock Tools 1.5.6, coupled with the conversion of the formats of the active ingredient and the target protein into pdbqt. Subsequently, molecular docking was performed by AutoDock. Finally, Pymol and Ligplot served for visualizing the protein-ligand complexes in 3D and 2D modes.

Molecular Dynamic (MD) Simulations of Protein-Ligand Complex

The protein-ligand complex were subjected to MD simulations with GROMACS 2020.3 software, aiming at exploring the receptor-ligand interaction.^15,16 The parameter and topology of proteins and ligands were created by the amber99sb-ildn force field and the general Amber force field (GAFF), respectively. We set the size of the simulation box to enable the distance between each atom of the protein and the box to be >1.0 nm. Subsequently, the box was filled with an explicit solvent-simple point charge model (SPC216 water molecules), with the water molecules replaced by Na+ and Cl- counterions to guarantee an electrically neutral simulation system. The steepest descent method was adopted for optimizing the entire system to attenuate inappropriate contact or atom overlap. Furthermore, to enable the simulation system to be sufficiently pre-equilibrated, the study performed NVT and NPT ensemble for 100 ps at 300 K and 1 bar, respectively, followed by 50 ns-MD simulation with periodic boundary conditions, with the temperature (300 K) and pressure (1 bar) controlled via the V-rescale and Parrinello-Rahman methods, respectively.¹⁷ Calculation of the Newton equation of motion was conducted by utilizing the leapfrog integration (time step: 2 fs). After that, using Fourier spacing of 0.16 nm, the Particle Mesh-Ewald (PME) method was adopted for calculating the long-range electrostatic interaction, and the LINCS method served for constraining all bond lengths. The visual molecular dynamics (VMD) software (version 1.9.3) and PyMOL (version 2.4.1) served for visually displaying, analyzing, and animating trajectories.¹⁸ The gmx_mmpbsa (http://jerkwin.github.io/gmxtool) was applied to estimate the binding free energy of the compound finally.

Validation Experiments

Experimental Materials

Anti-LIPA polyclonal antibody (Cat No. AC60669, Acmec), eIF2α/EIF2S1 mouse monoclonal antibody (Cat No. AG1813, Beyotime), XBP1 (U) rabbit polyclonal antibody (Cat No. AF8367, Affinity), IRE1A mouse monoclonal antibody (Cat No. AG1876, Beyotime), Phospho-PERK (Thr982) rabbit polyclonal antibody (Cat No. AF5902, Affinity), ATF6 rabbit polyclonal antibody (Cat No. AF6243, Affinity), DDIT3/CHOP rabbit polyclonal antibody (Cat No. AF6684, Affinity), Vinculin rabbit monoclonal antibody (Cat No. A2752, ABclonal), HRP goat anti-rabbit IgG (Cat No. AS014, Abclonal), and HRP goat anti-rabbit IgG (Cat No. AS003, Abclonal).

Cell Culture

The human leukemia cell lines K562 (ATCC^® CCL-243™) and THP-1 (ATCC^® TIB-202™) were commercially obtained from Sangon Biotech (Shanghai, China). Specifically, the K562 cells were maintained in IMDM (PRICELLA, PM150510) supplemented with 10% fetal bovine serum (FBS, E600050, Bioengineering), while the THP-1 cells in RPMI-1640 medium (E600028, Bioengineering) with the same serum concentration. Cells, upon reaching the density of 1.5×10⁶ cells/mL, were passaged, with subsequent density adjusted to 7×10⁵ cells/mL after each passage. Next, the processed cells received indicated period of incubation in a humidified atmosphere containing 5% CO₂ at 37°C, with the culture medium refreshed each 2–3 days. THP‑1 (acute monocytic leukemia) and K562 (chronic myeloid leukemia in blast crisis) were selected as representative myeloid leukemia cell lines commonly used in preclinical studies of statins and lipid metabolism. While K562 is not strictly an AML line, it serves as a complementary model for evaluating potential anti‑leukemic effects. Pilot experiments with longer exposures (48 h and 72 h) showed increased cytotoxicity but similar relative trends; therefore, the 24 h time point was selected for detailed mechanistic analysis to capture early apoptotic events while minimizing non‑specific effects.

Cell Apoptosis Assay

Flow cytometry (FCM), with an Annexin V-FITC/Propidium Iodide (PI) apoptosis detection kit (BD Biosciences, 556547), served for cell apoptosis assessment. Briefly, harvested cells underwent 5 min of centrifugation (1000 rpm, 4°C). After two times of washes in cold PBS, cells were re-suspended in 100 μL of Binding Buffer, added with 5 μL of Annexin V-FITC and 5 μL of PI to undergo gentle vortexing before 15 min of incubation at room temperature (RT) in the dark. The mixture was then added with 400 μL of Binding Buffer, with the obtained samples measured via a BD LSRFortessa flow cytometer within 1 h.

Cell Cycle Analysis

FCM also determined the cell cycle distribution. After harvest and 5 min of centrifugation (1000 rpm, 4°C), the yielded cells underwent two times of PBS washes, followed by one night of fixation of the cell pellets in 70% ethanol at 4°C. After another centrifugation and one time of PBS wash, the obtained cells were resuspended in 500 μL of PBS that contained 50 μg/mL PI, 100 μg/mL RNase A, and 0.2% Triton X-100. The suspension received 30 min of incubation at 4°C in the dark. At last, cell cycle distribution was analyzed via a BD LSRFortessa flow cytometer, with data processing using cell cycle fitting software.

Transmission Electron Microscopy (TEM)

For TEM analysis, cells underwent 2–4 h of fixation in 2.5% glutaraldehyde at 4°C, followed by centrifugation and removal of the supernatant. The pellet was resuspended in 3% low-melting-point agar for another centrifugation repeatedly. After solidification of the agar, the cell-agar block was cut and placed back into 2.5% glutaraldehyde for storage. The block underwent three times of washes in 0.1 M phosphate buffer (15 min for each wash) before 4 h or one night of post-fixation in 1% osmium tetroxide at 4°C. Subsequent to gradual ethanol dehydration (30%–50%–70%–80%–95%–100%–100%, 40 min each), the samples received three 30-minute treatments with 100% propylene oxide. After infiltration with a mixture of propylene oxide, the samples were embedded in embedding medium (2:1) for 4 h at RT, and then received one night of treatment in propylene oxide mixed with embedding medium (1:2). With another infiltration of the samples with pure embedding medium for 24 h, the resin blocks were trimmed for ultrathin sectioning after polymerization at 60°C for 48 h. After slicing by Leica UC7 ultramicrotome (Leica Microsystems), the sliced sections (70 nm thick) were mounted on 150-mesh copper grids (AZH150, Zhongjing Keyi). Afterwards, the processed sections underwent 8 min of staining with 3% uranyl acetate in ethanol, three times of washes in 70% ethanol and three times of washes in ultrapure water, followed by another staining with 2.7% lead citrate for 8 min. After three times of washes in ultrapure water, the grids were blotted dry with filter paper. The samples were finally examined and imaged using a JEM1400 transmission electron microscope.

Validation of the Expressions of Potential Pathway Genes

The expression levels of potential genes were validated by real-time PCR assay firstly. In a brief statement, an RNA extraction kit was employed to separate cellular RNA as instructed. Subsequently, the extracted RNA underwent reverse transcription into cDNA with a universal real-time fluorescence quantitative reagent. The QuantStudio 6 Flex real-time fluorescence quantitative PCR system was adopted for assessing the mRNA expressions of relevant genes as per the product’s experimental protocol. The gene expression level (F) was calculated via F = 2^(-ΔΔCt), where ΔΔCt = (Ct of target gene - Ct of internal reference) in the experimental group - (Ct of target gene - Ct of internal reference) in the control group. Each experiment was conducted in triplicate to ensure reliability. For further validation, Western blot (WB) was conducted using K562 cells intervened with lovastatin at 50, 100, and 200 μM for 24 h, with the untreated group as the control group. Following the addition of RIPA protein lysate for sufficient grinding, the supernatant, ie, the total tissue protein, was obtained by separation in an ultra-high-speed centrifuge pre-cooled at 4°C. The BCA protein quantification kit determined the concentration of the extracted protein, followed by SDS-PAGE electrophoresis and membrane movement. Subsequent steps included incubation using the primary and secondary antibodies diluted as instructed, and ECL development.

Statistical Analysis

Statistical analysis relied on R software (Version 4.3.1). A KM survival analysis evaluated the differential survival. Wilcoxon and Spearman correlation tests served for difference analysis and correlation analysis, respectively. P-value < 0.05 denoted statistical significance. For comparisons involving multiple groups or multiple genes, p-values were adjusted using the Benjamini–Hochberg method to control the FDR. Adjusted p < 0.05 was considered statistically significant. Each experiment was conducted in triplicate.

Results

DEGs Between AML and Normal Samples

Initially, the presence of batch effects between the TCGA and GTEx datasets was excluded by performing principal component analysis (PCA). Before batch correction, there was separate cluster of samples from different sources, indicating significant batch effects (Figure 1A). Critically, the inter-batch variation was effectively mitigated after applying the ComBat algorithm, with better integration of samples across datasets (Figure 1B), suggesting successful batch effect removal. The differential expression analysis of TCGA cohort revealed 866 DEGs between AML and normal samples totally, with 414 DEGs down-regulated and 452 up-regulated in AML (Figure 1C and D).

Figure 1.

Batch effect correction and differential gene expression analysis. (A) PCA plot before batch effect correction; (B) PCA plot after batch effect correction; (C) Heatmap of the top-50 DEGs between AML and normal samples; (D) Volcano plot of DEGs between AML and normal samples.

Abbreviations: PCA, principal components analysis; DEGs, differentially expressed genes; and AML, acute myeloid leukemia.

MR Analysis Between eQTLs and AML

This study performed MR analysis on 19,942 eQTLs and 3301 AML cases. Following the SNP screening criteria, we identified 5430 eQTLs that contained 25,107 SNPs. The IVW method (p<0.05) together with heterogeneity analysis (p>0.05) helped to identify 246 eQTLs (all of dataset can obtain https://zenodo.org/records/19390059). MR analysis suggested a causal relationship between elevated LIPA expression and increased AML risk (OR=1.32, p=0.003), although this inference is based on genetic instruments and requires further experimental validation.

Development and External Validation of a Prognostic Model for AML

Using univariate Cox survival analysis and KM survival analysis, this study identified 39 prognostic-related genes from a pool of 866 DEG (p<0.05) ((all of datasset can obtain https://zenodo.org/records/19390059)). Subsequently, the lasso analysis identified the most informative 31 genes, alongside constructing a prognostic model (Figure 2A and B). Finally, a AML prognostic model, encompassing 16 genes, was established according to the results of the multivariate Cox regression analysis (Figure 2C).

Figure 2.

Construction and external validation of a prognostic model for AML. (A) Lambda plot of the lasso regression analysis; (B) Cross plot of the lasso regression analysis; (C) Forest plot of that prognostic model involving the protective genes (MYH4, DSCR4, SYT5, and ITGA4), and the adverse genes (HIST3H2A, DIRAS3, DSG2, HIST1H4J, PSAT1, CFH, SCD, LIPA, and IGHM) (p<0.05); (D) Survival plots depicting the overall survival of the high- and low-risk groups in the training cohort (GEO); (E) The ROC curve of the training cohort (GEO), with the OS at 1, 3, and 5 years of 0.751, 0.783, and 0.792 in patients with AML, respectively; (F) Survival plots depicting the overall survival of the the high- and low-risk groups in the validation cohort (TCGA); (G) The ROC curve of the validation cohort (TCGA), with the OS at 1, 3, and 5 years of 0.757, 0.71, and 0.778 in patients with AML, respectively.

Abbreviations: ROC, receiver operating characteristic; AML, acute myeloid leukemia; GEO, gene expression omnibus; OS, overall survival; and TCGA, the cancer genome atlas program.

In the created prognostic model, the median value of risk was taken into account to divide patients into high- and low-risk groups. According to KM survival analysis results, patients from the high-risk group in the training group exhibited remarkably worse prognosis (Figure 2D). The AUC values for the prognostic model were determined to be 0.751, 0.783, and 0.792 in predicting 1-, 3-, and 5-year OS rates of AML patients, respectively, indicating excellent diagnostic efficacy (Figure 2E). In addition, as depicted in Figure 2F and G, external validation cohort demonstrated the robustness of the AML prognostic model. Unlike most existing AML prognostic signatures that rely solely on transcriptomic data, our model integrates MR‑supported causal genes, thereby prioritizing targets with genetic evidence for a potential role in disease etiology.

Machine Learning-Based AML Diagnostic Model with SHAP Interpretation

Through MR analysis, an intersection of the DEGs from the TCGA-GTEx cohort with eQTL‑related genes was conducted to identify candidate diagnostic genes for AML, with the identification of 6 genes. The obtained data were subsequently used for machine learning model construction and SHAP interpretability assessment (Figure 3A).

Figure 3.

Machine learning-based AML diagnostic model with SHAP interpretation. (A) Identification of 6 feature genes (PCSK6, MYEOV, RAP1GAP, LIPA, FES, and CTTN) through the intersection of DEGs and MR genes outputs). (B) ROC curves of ten classifiers in the testing cohort, with the highest AUC achieved by the RF model (AUC = 0.894). (C) SHAP summary bar plot for the NeuralNet model, ranked by mean absolute SHAP value. (D) SHAP bee‑swarm plot illustrating gene impact direction and magnitude across individual patients. (E) SHAP waterfall plot for a representative patient to illustrate gene‑specific contributions to the NeuralNet prediction threshold.

Abbreviations: ROC, receiver operating characteristic curve; AML, acute myeloid leukemia; and SHAP, shapley additive exPlanations.

After integrating the 6 genes into 10 different machine learning algorithms, the RF model achieved the largest AUC value, indicating its superior discriminative ability in diagnosing AML (Figure 3B). SHAP summary bar plots revealed that MYEOV had the greatest impact on the output of the model, followed by LIPA and RAP1GAP (Figure 3C).

The SHAP bee-swarm plot further demonstrated the association of higher expression of all 6 genes with increased SHAP values, suggesting their positive contribution to the prediction of AML risk (Figure 3D).

Furthermore, to get into the bottom of the gene-specific effects more intuitively, a SHAP waterfall plot was generated for a representative sample (Figure 3E). The contribution of each gene was taken into account to adjust the baseline prediction (E[f(x)] = 0.689). In this specific case, the expression levels of PCSK6 (0.84), MYEOV (1.57), RAP1GAP (1.96), LIPA (2.73), FES (4.46), and CTTN (2.53) all contributed to decreased predicted probability, resulting in a final prediction value of f(x) = 0.0649. The RF model accurately classified sample as normal, given that this value was below the classification threshold.

Identification of the Hub Genes and Their Expression, Prognostic Significance, and Potential as Drug Targets

As presented in Figure 4A, a core gene called LIPA, was identified by intersecting RF diagnostic model genes and prognostic model genes. Notably, LIPA exhibited obviously higher expression in AML than that of normal samples (Figure 4B). Furthermore, the ROC curve analysis demonstrated an exceptional diagnostic efficiency of LIPA based on its AUC value of 0.925 for the diagnostic performance of LIPA in AML (Figure 4C). The KM survival analysis revealed an obvious association of high LIPA expression with unfavorable prognosis in AML patients (Figure 4D and E). Additionally, the drug target prediction analysis revealed that LIPA exhibited potential as a druggable target for compounds LIOTHYRONINE, TRIFLUOPERAZINE, THERAPEUTIC HORMONE, LOVASTATIN, GLUCAGON (RDNA) and SEBELIPASE ALFA (Figure 4F).

Figure 4.

Identification of hub genes and their expression, prognostic significance, and potential as drug targets. (A) Intersection Venn diagram depicting the overlapping of genes from the diagnostic and prognostic models, with LIPA identified to be the only intersection gene; (B) Boxplot of LIPA expression between tumor and normal samples, AML samples with higher expression levels (p<0.05); (C) ROC curve of the accuracy of LIPA in the diagnosis of AML, with the result of 0.925; (D) KM survival analysis of LIPA expression and its association with prognosis (GEO), with higher expression level leading to adverser outcome (p<0.032); (E) KM survival analysis of LIPA expression and its association with prognosis (TCGA), with higher expression level leading to adverser outcome (P=0.044); (F) Network diagram of potential drug targets (lovastatin, glocugan, seblipase alfa, therapeutic hormone, liothyronine, and trifluoperazine) for LIPA.

Abbreviations: LIPA, liposomal acid lipase; ROC, receiver operating characteristic; AML, acute myeloid leukemia; GEO, gene expression omnibus; KM, kaplan-meier; and TCGA, the Cancer Genome Atlas Program.

MR Analysis of LIPA and AML

As indicated by the results of MR analysis of LIPA (exposure) versus AML (outcome), high expression of LIPA was related to an increased AML incidence (Figure 5A). The forest plot in Figure 5B demonstrated the utilization of 11 SNPs in MR analysis, while the funnel plot indicated no apparent bias in this analysis (Figure 5C). Besides, upon elimination of these 11 SNPs individually, the leave-one-out analysis illustrated minimal impact on the overall MR effect (Figure 5D).

Figure 5.

MR analysis of LIPA and AML. (A) Scatter plots for MR analysis of the exposure and outcome factors; (B) Forest plots of SNPs included in the MR analysis; (C) Funnel plot of MR analysis; (D) Leave-one-out analysis of forest plots for MR analysis.

Abbreviations: LIPA, liposomal acid lipase; AML, acute myeloid leukemia; SNPs, single nucleotide polymorphisms; and MR, mendelian randomization.

Immune Infiltration in AML and Its Correlation with LIPA

In our subsequent immune cell infiltration analysis, among the 22 types of immune cells examined, compared to normal samples, 9 types exhibited high levels of infiltration, while 12 types displayed low levels of infiltration in AML. The infiltration level of Plasma cells did not show an obvious difference between AML and normal samples (Figure 6A). Furthermore, according to correlation analysis, LIPA expression level demonstrated a positive correlation with the infiltration degree of four immune cell types, but a negative correlation with ten other immune cell types. Comparatively, LIPA expression was not remarkably correlated with the infiltration degree of eight additional immune cell types (Figure 6B). Additionally, according to prognostic analysis, high levels of infiltrating B cells memory and Mast cells resting indicated improved prognosis in AML patients, whereas increased infiltration of NK cells activated was an indicative of poorer prognosis (Figure 6C–E).

Figure 6.

Analysis of immune infiltration in AML and its correlation with LIPA. (A) Violin plot of the infiltration levels of 22 immune cell types between AML and normal samples; (B) Heat map of the correlation analysis between LIPA expression level and the infiltration levels of 22 immune cell type; (C) The correlation between B cells memory infiltration level and prognosis, with higher B cells memory infiltration resulting good prognosis (p<0.003); (D) The correlation between Mast cells resting infiltration level and prognosis, with higher mast cells resting infiltration revealing good prognosis (p<0.010); (E) The correlation between NK cells activated infiltration level and prognosis, with higher B cells, and NK cells activated infiltration indicating adverse prognosis (p<0.005).

Abbreviations: LIPA, liposomal acid lipase; and AML, acute myeloid leukemia.

Network Pharmacological Analysis of Lovastatin’s Potential Targets and Molecular Pathways in AML

A total of 109 common targets between lovastatin and AML were identified through our preliminary network pharmacological analysis (Figure 7A), with key hub genes including AKT1, HSP90AA1, EGFR, and SRC (Figure 7B–D). These targets were critically involved in regulating histone deacetylase activity, IGFS signaling, and protein kinase pathways (Figure 7E). As evidenced by further enrichment analysis, these overlapping genes participated in pivotal oncogenic pathways, such as lipid metabolism and cancer-related signaling cascades, underscoring their roles in the pathogenesis and progression of AML (Figure 7F).

Figure 7.

Network pharmacology of drug-disease common targets and biological functions of common targets of lovastatin and AML. (A) The drug-disease common targets of AML and lovastatin; (B) The disease-target-component network for AML and lovastatin; (C and D) The drug-disease common target PPI network diagram and core target ranking based on PPI topology analysis (top-20 by degree); (E and F) GO and KEGG enrichment analysis of disease-drug common targets.

Abbreviations: AML, acute myeloid leukemia; GO, Gene Ontology; and KEGG, Kyoto Encyclopedia of Genes and Genomes.

Molecular docking and MD simulations

MD simulations revealed significant conformational flexibility in the ligand, with an average RMSD fluctuation of 0.0747 nm and 0.1111 nm for LIPA and the ligand, respectively (Figure 8C–E). Molecular docking demonstrated a binding energy of −6.6 kcal/mol between lovastatin and LIPA (Figure 8A). Besides, free energy landscape analysis indicated the lowest energy state at 52.44 ns, suggesting the optimal complex stability (Figure 8B and G).Binding free energy decomposition identified critical residues (ARG101, LYS109, LYS128, LYS141, GLY143, GLN144, LYS177, ARG172, and ASP257) contributing to ligand-protein stability (Figure 8F).

Figure 8.

MD simulations of protein-ligand complex of lovastatin with LIPA. (A) Molecular docking analysis of LIPA and lovastatin, with the binding energy of LIPA protein and small molecules determined to be −6.6 kcal/mol; (B) Free energy morphology diagram of the combined simulation process of LIPA and lovastatin. MD, Molecular dynamics; and LIPA, liposomal acid lipase. (C) The variation curve of RMSD of proteins over time in the LIPA and lovastatin combined simulation process, (D and E) The variation curves of the protein and ligand RMSF over time in the LIPA and lovastatin combined simulation process. (F) Protein energy decomposition diagram of the combined simulation process of LIPA and lovastatin.(G) Free energy landscape analysis between LIPA and lovastatin.

Abbreviations: MD, Molecular dynamics; and LIPA, liposomal acid lipase.

Anti-Tumor Activity and Cell Cycle Modulation

Lovastatin and 4-PBA, an ER stress inhibitor, exhibited dose-dependent cytotoxicity in THP-1 and K562 AML cells (Figure 9A and B), as measured by CCK-8 assay. For THP1 cell lines, after lovastatin intervention, there were more cells in the M phase of the cell cycle. In the K562 cell line, more cells in the S phase was accompanied by less cells in the G2-M phase (Figure 9C). After treatment, both cell lines demonstrated marked induction of apoptosis by FCM (Figure 9D).

Figure 9.

The Effect of lovastatin on the activity, cell cycle, and apoptosis in AML cell lines K562 and THP-1 in vitro. (A and B) Effects of lovastatin and 4-PBA at different concentrations on cell activities of THP-1 and K562; (C) Effects of lovastatin and 4-PBA on cell cycles of THP-1 and K562; (D) Effects of lovastatin and 4-PBA on cell apoptosis of THP-1 and K562. AML, acute myeloid leukemia. Cells were treated with lovastatin (100 μM) or 4-PBA (5 μM) for 24 hours; **** p<0.0001.

Mediating Role of ER Stress in Lovastatin-Induced Apoptosis in AML

Transmission electron microscopy revealed that control AML cells exhibited dilated and swollen ER (Figure 10), indicative of basal ER stress activity. In contrast, both lovastatin-treated and 4-PBA-treated cells displayed marked ER contraction, suggesting relief of ER stress.

Figure 10.

ER morphology of AML cells by transmission electron microscopy. Representative images from control, 4-PBA-treated, and lovastatin-treated groups. Black arrows indicate endoplasmic reticulum. Scale bar = 500 nm.

The Expressions of ER Stress Biomarkers and LIPA

To prove the regulatory effects of lovastatin and 4-PBA on ER stress at the molecular level, we further determined the expressions of ER stress and LIPA by PCR and WB. According to the PCR results, relative to the control group, the lovastatin group demonstrated remarkably downregulated LIPA and ER stress markers ATF6, eIF2α, XBP1, IRE1A, DDIT3, HSP90AA1, and PERK (p<0.0001); meanwhile ATF6, eIF2α, XBP1, IRE1A, DDIT3, HSP90AA1, and PERK were significantly down-regulated in 4-PBA group (all p<0.0001) (Figure 11). And the WB analyze the protein levels of these see in the Figure 12. As shown in Figures 11 and 12, both lovastatin and the ER stress inhibitor 4-PBA significantly downregulated the expression of ER stress markers (ATF6, CHOP, IRE1) in AML cells. The striking similarity between the effects of lovastatin and 4-PBA indicates that lovastatin functions as an ER stress suppressor. Corresponding raw WB data are shown in Supplementary Figure S1.

Figure 11.

The mRNA expression levels of LIPA, ATF6, eIF2α, XBP1, IRE1A, DDIT3, HSP90AA1, and PERK. (A-H) The ER biomarkers expression in K562 cells (I-P). The ER biomarkers expression in THP-1 cells. Cells were treated with lovastatin (100 μM) or 4-PBA (5 μM) for 24 h; **** p<0.0001.

Figure 12.

The protein expression levels of LIPA and the ER stress biomarkers. (A-H) Representative Western blot bands (upper panels) and quantitative analysis (lower panels) of LIPA, phosphorylated PERK (pPERK), IRE1A, DDIT3/CHOP, ATF6, and eIF2α in AML cells treated with vehicle (NC), 4-PBA, low-dose lovastatin (Lov-L), or high-dose lovastatin (Lov-H) for 24 h. ****p < 0.001.

Discussion

There is a gradual increase in the prevalence of AML, along with corresponding therapeutic advances, with the development of global aging. In recent decades, significant progress has been made, such as targeted therapy, immunotherapy, and allogeneic HSCT, that optimizes AML patients’ prognosis. However, some of the AML patients may exhibit low responsiveness or develop acquired resistance to specific treatment, accompanied by relapse in some AML patients following allogeneic HSCT. Elderly AML patients, with poor tolerance, can only be treated with mild medicinal regimens. In general, lipid metabolism reprogramming functions significantly in modulating tumorigenesis, development and drug resistance.^19–21 Common therapeutic agents to regulate lipid metabolism, such as statins, have been reported to improve patient prognosis and potentiation of the original treatment regimen.^22,23 Therefore, increasing the routine administration of statin may enhance the prognostic outcomes of AML patients, yet with a poor understanding of its specific mechanism. Therefore, we conducted some series of analyses to confirm whether statins have potential anti-tumor activities in AML.

Firstly, we analyzed high-risk genes for AML through MR and transcriptomic analyses, and LIPA was a risk factor for AML. The development of the risk model and the validation of external data suggested that the high LIPA expression would promote the occurrence of AML and led to deteriorated AML prognosis. However, up to now, no study has reported the function of LIPA in AML. Collier AB et al suggested high LIPA expression in ovarian cancer, which could predict poor prognosis;²⁴ and in vitro experiments showed that targeting LIPA could produce anti-tumor activity against ovarian cancer by activating ER stress.²⁴ Liu X et al also confirmed in their study that targeted LIPA had significant anti-tumor activity in vitro against triple-negative breast cancer.²⁵ It can be speculated preliminarily that high LIPA expression is a high-risk factor for cancer patients. It conformed to our study based on the risk model incorporating genes involving LIPA for high and low risk stratification, and its prediction of the risk stratification for AML patients exhibited good sensitivity and specificity.

LIPA can hydrolyze cholesterol esters and triglycerides transferred to lysosomes.²⁶ It is well known that LIPA functions as a key enzyme regulating lipid metabolism, it as may promote cancer cells including AML cell survival by maintaining lipid homeostasis and supporting basal ER stress activity. In this study, through preliminary LIPA drug target prediction, lovastatin was a potential drug target; subsequent network pharmacology identified 109 common drug-disease targets; and functional enrichment analysis revealed the involvement of the common target genes in lipid metabolism. Generally, lipids, as one of the three major nutrients, have an irreplaceable role in biological behaviors as well as basic biological functions of normal and tumor cells. The effects of lipids on tumor development are multifaceted, which can indirectly mediate the tumor microenvironment (TME) or directly regulate tumor cells. Environmental lipids offer energy for the growth and migration of tumor cells, and promote immunosuppressive TME to be well formed. In the TME, lipid accumulation increase stands for a common metabolic change. Consistently, our study ascertained the high LIPA expression in AML patients, which might promote lipid accumulation in the TME in AML cells. Tumor-induced reorganization of ER membrane lipid composition in TME regulation can maintain macrophage survival and pro-tumorigenic activity.²⁷ For instance, Mishra SK et al systematically elucidated nutrient metabolism reprogramming in AML, its impact on AML onset and progression, as well as its possible value as a potential target, and its application for therapeutic treatment.⁷ In terms of the direct action on tumor cells, Guo HZ et al showed that CD36-dependent atypical lipid metabolic programs promoted immune escape and the resistance to HMA therapy in AML.²⁸ These biological alterations may explain the insensitivity of some AML patients to demethylation therapy, confirming that targeting lipid metabolism in AML is one of the promising potential solution combined with existing treatment regimen sensitizers.

ER often serves as an organelle for protein and lipid metabolism and processing, revealing that metabolic reprogramming in tumor cells can be achieved by regulating ER stress activity. Fatty acid metabolism can control the immunosuppressive phenotype of tumor-associated macrophages.²⁹ Moreover, Celik C et al systematically described the distribution of lipids, as well as their anabolism and catabolism, all of which were to some extent strictly regulated by the ER. Dysfunction and overload of lipid-related pathways, either alone or combined with ER stress, interact with the functions of other cells to promote disease development.^30,31 Lipid metabolism and ER stress were found to interact with each other to regulate liver metastasis of pancreatic cancer.³² These results showed the lipid metabolism always combine with ER stress regulation. According to tumor, ER stress has multiple functions, as tumor cells promote misfolded proteins by activating ER stress; and conversely, activating ER against some tumors can promote apoptosis, leading to anti-tumor activity.

A key finding of this study is that lovastatin exerts its anti-AML effect by suppressing ER stress, a mechanism phenocopied by the canonical ER stress inhibitor 4-PBA. While this may seem counterintuitive—as ER stress is traditionally viewed as a pro-apoptotic signal—it aligns with the emerging concept of “ER stress addiction” in cancer cells.^32,33 Malignant cells, including those in AML, often maintain elevated basal ER stress activity to sustain rapid proliferation and cope with the increased demand for protein synthesis.^34,35 This adaptive state, driven by the unfolded protein response (UPR), supports survival under oncogenic stress. Disruption of this equilibrium—whether by genetic manipulation or pharmacological intervention—can tilt the balance toward apoptotic cell death.

In our study, control AML cells exhibited dilated ER morphology and elevated expression of ER stress markers (ATF6, CHOP, IRE1), consistent with heightened basal ER stress. Lovastatin treatment reversed these features, inducing ER contraction and downregulating all markers examined. The striking similarity between the effects of lovastatin and 4-PBA strongly suggests that ER stress suppression is a primary mediator of lovastatin’s cytotoxicity, rather than a secondary or unrelated phenomenon.

This mechanistic interpretation also reconciles our findings with previous reports on LIPA targeting in solid tumors. Collier et al²⁴ and Liu et al²⁵ demonstrated that LIPA inhibition in ovarian and breast cancer cells induces ER stress, leading to apoptosis. The opposite directionality in our study—ER stress suppression leading to apoptosis—likely reflects context-dependent differences in ER stress homeostasis between AML and solid tumors, underscoring the need for disease-specific mechanistic studies.

Despite the strengths of our multi-omics approach and in vitro validation, several limitations should be acknowledged. First, while MR analysis supports a causal relationship between LIPA expression and AML risk, this inference is based on European-ancestry data and may not be generalizable to other populations. Second, the prognostic signature was derived from public datasets; independent prospective validation is warranted. Third, although molecular docking and dynamics simulations identified lovastatin as a potential LIPA inhibitor, direct biochemical evidence (surface plasmon resonance) and genetic rescue experiments (LIPA knockdown or overexpression) are needed to definitively establish LIPA as the primary target. Fourth, our in vitro studies were limited to two cell lines (THP-1 and K562); the latter is of chronic myeloid leukemia origin (it has been investigated in many research of AML as well), and findings should be interpreted with caution. Fifth, the mechanism by which ER stress suppression triggers apoptosis remains incompletely defined and requires further investigation, including in vivo models. Finally, our analysis does not account for potential confounding by environmental or lifestyle factors such as dietary lipid intake, baseline cholesterol levels, or prior statin use, which could influence LIPA expression and AML prognosis. Future studies incorporating detailed clinical metadata are needed to address these variables and the use of FITC and PI may not be optimal and that future studies should employ a preferred fluorochrome combination.

For the clinical use of lovastatin, compared to conventional chemotherapies, lovastatin offers a distinct mechanism targeting lipid metabolism and ER stress, potentially providing a well-tolerated option for combination therapy or for patients unfit for intensive treatment. As we all know lovastatin has been used clinically for decades as a lipid-lowering agent with a well-established safety profile. Common side effects include mild gastrointestinal symptoms, myalgia, and elevated liver enzymes, which are generally manageable and reversible upon dose adjustment or discontinuation. Serious adverse events such as rhabdomyolysis are rare. In the context of AML, where patients may receive statins alongside chemotherapy, liver and muscle function should be monitored. Importantly, the doses used in our in vitro studies (50–200 μM) are higher than typical clinical plasma concentrations; thus, the translational potential of lovastatin in AML would require careful dose optimization and therapeutic drug monitoring to balance efficacy and toxicity. While our multi-omics approach and in vitro experiments provide strong evidence for LIPA as a prognostic biomarker and lovastatin as a potential therapeutic agent, several steps are needed before clinical translation. These include: (1) validation in independent patient cohorts with detailed clinical annotation; (2) in vivo efficacy studies in AML animal models; (3) investigation of LIPA-specific genetic models (knockdown/overexpression) to confirm target engagement; and (4) Phase I/II clinical trials to evaluate safety and preliminary efficacy of lovastatin in AML patients, particularly those with high LIPA expression.

Conclusion

In conclusion, this study identifies LIPA as a causal prognostic biomarker in AML and reveals that lovastatin induces apoptosis by suppressing ER stress. These findings provide a mechanistic rationale for repurposing lovastatin in AML therapy, although further in vivo and mechanistic studies are needed to confirm target engagement and therapeutic efficacy.

Funding Statement

This work was supported by the Baise Scientific Research and Technology Development Program (Grant No. 20222015).

Data Sharing Statement

All data included in this study are available upon request from the corresponding authors, Rong rong Liu and Wei Jie Zhou.

Ethics Statements

This study, which involved the analysis of public databases (TCGA, GTEx, GEO) and in vitro experiments using established cell lines, was reviewed and approved by the Ethics Committee of the Baise People’s Hospital (Approval No. LW20251204003).

Disclosure

The authors report no conflicts of interest in this work.