Lysosome-derived biomarkers for predicting survival outcome in acute myeloid leukemia

Abstract

Lysosomes have a tight connection to cancer and can eliminate cancer cells. The dismal prognosis of acute myeloid leukemia (AML) patients may thus be improved by a thorough examination of the function of lysosome-related genes (LRGs). By using a variety of machine learning methods including random forest approach, LASSO-COX regression, and extreme gradient boosting (XGBoost), we create a prognostic six-LRGs-related signature (HPS1, BCAN, SLC2A8, DOC2A, CHMP4C, and SLC29A3), which categorized AML patients into two groups with significant survival and tumor microenvironment (TME) differences. Data from the ICGC and TARGET cohorts were used as test cohorts for the validation of the prognostic LRGs-related signature. We also discovered that chemotherapeutic susceptibility was connected to the LRGs-related signature. Finally, we evaluated gene expression levels in the LRGs-related signature between normal and AML samples and confirmed the elevation of CHMP4C expression in 90 clinical samples. In summary, a six-LRGs-related signature was developed to predict the prognosis of AML patients, and more research is necessary to determine whether this signature has therapeutic promise as an anti-AML target.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03302-8.

Introduction

The most prevalent form of acute leukemia in adults is acute myeloid leukemia (AML), which is characterized by a variety of genetic abnormalities and the buildup of immature myeloid progenitors in the bone marrow and peripheral circulation. The 5-year survival rate of patients with AML is still inadequate despite new treatment choices in recent years [1]. More than half of AML patients experience relapses and develop drug resistance, which eventually results in treatment failure and even death [2]. Therefore, the creation of new prognostic indicators is urgently required to track the prognosis of AML patients.

The internal pH of lysosomes, which are vesicular structures with a single membrane and a significant number of acidic hydrolases, ranges between 4.5 and 5.5 [3]. Lysosomes release acid hydrolases to degrade damaged proteins and organelles absorbed by endocytosis to preserve metabolic balance as a dynamic regulator of cellular and organismal equilibrium [4, 5]. Lysosomes have a tight connection to cancer and can eliminate cancer cells [6]. Lysosomes are important regulators of proliferation signals and proliferative chemicals, controlling growth factor signals and supplying nutrients to control tumor cell proliferation [7]. Lysosomes have also been linked to tumor invasion, metastasis, and migration [8]. Radiation therapy is one of the most commonly used treatments for AML, but tumor cells usually develop reactive adaptations to radiation therapy, leading to radioresistance, and lysosomes promote the development of radioresistance [9]. In addition, lysosomes have been demonstrated to have a role in tumor angiogenesis, immunosuppression, chemoresistance to chemotherapy, and the control of polarization in tumor-associated macrophages (TAMs) [10]. Lysosome-related genes (LRGs) constitute a unique and targetable vulnerability in AML, fundamentally differing from metabolic or epigenetic regulators in terms of both mechanism and clinical impact. In contrast, metabolic genes (e.g., IDH1/2) are primarily involved in energy production, while epigenetic modifiers (e.g., DNMT3A, TET2) are responsible for reprogramming transcriptional landscapes. Overall, lysosomes play an unparalleled role in the dynamic control of cellular and tissue homeostasis, and therapeutic interventions for cancer targeting lysosomes have great potential and deserve further exploration. Despite previous studies focusing on the role of lysosomes in human health and disease, there remains a lack of comprehensive research specifically examining the relationship between LRGs and the prognosis of patients with AML.

In this study, we employed a multi-cohort analytical methodology, integrating clinical data from prominent databases, including The Cancer Genome Atlas (TCGA), the International Cancer Genome Consortium (ICGC) and The Children’s Oncology Research Group (TARGET), and utilized machine learning techniques for data analysis. The primary objective was to develop a prognostic signature based on LRGs and to investigate its association with survival outcomes and tumor microenvironment characteristics in patients with AML. By conducting a comprehensive analysis of LRGs, we aim to offer novel insights for the personalized treatment of AML, thereby advancing research progress in this domain.

Methods

Collection of clinical AML cohorts and lysosome-related genes

In all, 129 AML samples from the TCGA cohort (TCGA-LAML, https://xenabrowser.net/), 143 AML samples from the ICGC cohort (LAML-US, https://dcc.icgc.org/), and 187 AML samples from the TARGET cohort ( https://ocg.cancer.gov/programs/target/)were employed in our analysis. Case inclusion requirements: comprehensive clinical information and survival statistics. Table S1 provides a summary of the patient’s complete clinical features. Additionally, 894 lysosome-relevant genes those are detailed in Table S3 (LRGs) were obtained from the GO (Gene Ontology) and GSEA (Gene Set Enrichment Analysis) databases, which may be found online at https://geneontology.org/ and http://www.gsea-msigdb.org/, respectively.

Recognition of molecular subtypes by the non-negative matrix factorization (NMF)

The AML samples were categorized using the NMF algorithm with the criteria “brunet” and 50 iterations based on these LRGs. The best number of clusters (K) was established based on cophenetic, dispersion, and profile, with K ranging from 2 to 6. Further Kaplan-Meier survival analysis was done to compare the overall survival outcomes of the various NMF subtypes.

Prognostic LRGs-related signature establishment and validation

As previously mentioned [11], using univariate Cox regression, the relationship between LRGs and survival outcomes for HCC patients was determined. Next, the hub genes were investigated using the random forest approach, LASSO-COX regression, and extreme gradient boosting (XGBoost). The intersecting genes from the three machine learning techniques were ultimately subjected to a multivariate Cox regression analysis to develop a prognostic LRGs-related signature in the TCGA cohort. Score =∑iCoefficient (LRGi)*Expression (LRGi). The samples in the TCGA, ICGC, and TARGET cohorts were then separated into score^high and score^low groups based on the mid-value of the scores. We utilized Kaplan-Meier survival analysis and time-dependent receiver operating characteristic (ROC) curve analysis, respectively, to examine the variations in overall survival outcomes between the two groups and the prediction power of the LRGs-related signature.

Evaluation of genetic variations and functional enrichment analysis

The mutation data from AML patients were analyzed by the “TCGAmutations” package to explore variations in genetic abnormalities in various subgroups. The GSEA was used to explore GO and KEGG items that were substantially different across score^high and score^low groups. Statistical significance was determined using the Banjamini-Hochberg (BH) adjusted P less than 0.05.

Analysis of tumor microenvironment (TME)

The stromal and immunological scores were computed using the Expression data (ESTIMATE) algorithm to assess the infiltrative status of immunocytes and stromal cells [12]. The abundance of tumor-infiltrating immune cells (TIICs) was evaluated using the CIBERSORT [13], xCELL [14], and MCPcounter [15] databases.

Sensitivity testing for drugs

The NCI-60 database was examined using CellMiner (https://discover.nci.nih.gov/cellminer/) [16]. To determine the association between target gene expression and drug sensitivity, the website’s target gene expression status and z-score for cell sensitivity data were downloaded and subjected to Pearson correlation analysis. If a drug’s adjusted P-value was less than 0.01 and its Pearson correlation coefficient was more than 0.3, it was categorized as tumor-sensitive. The variations in half maximal inhibitory concentration (IC50) for several categories of tumor-sensitive medicines were then compared across score^high and score^low groups.

Evaluation of gene expression in clinical samples

We obtained peripheral blood samples from 45 healthy volunteers and 45 AML patients. Then, we used quantitative real-time PCR (qRT-PCR, Sangon Biotech) to determine the mRNA expression levels of CHMP4C in all samples, as previously reported [17]. The primer synthesized sequences were as follows: CHMP4C forward: 5ʹ-TGGTCCGACTTCGGGAGAC-3ʹ, reverse: 5ʹ-GCCAGGGCGATTTCTCTCTG-3ʹ; β-ACTIN forward: 5ʹ-CGTGGGCCGCCCTAGGCACCA-3ʹ, reverse: 5ʹ-TTGGCTTAGGGTTCAGGGGGG-3ʹ.

Statistical analysis

The correlation between two variables was determined using the Pearson correlation technique. The Kaplan-Meier technique was used to produce survival graphs. The Student’s t-test was used to assess two sets of qRT-PCR data. Univariate or multivariate Cox regression analysis was carried out to investigate the connection between OS and LRGs-score as well as clinical characteristics. To find the genes linked to OS, the hazard ratio (HR) and 95% confidence interval (CI) were computed. Statistics were judged significant at a P value less than 0.05.

Results

Recognition and validation of LRGs-related molecular subtypes

Three clusters were found to be the ideal number based on co-seismicity, dispersion, and profile (Figure S1). Unfortunately, at this time, there was no discernible difference in patients’ chances of survival across the three subgroups. As a result, we determined that two clusters were the ideal number, and we separated the TCGA dataset’s AML samples into two subclasses based on LRGs (Fig. 1A). Significant survival variations between patients in the two categories were seen (Fig. 1B). Additionally, the two subgroups’ TME features were contrasted. Figure 1C demonstrates that as compared to samples from cluster 2, samples from cluster 1 had lower immune, stromal, and ESTIMATE scores (Fig. 1C). The infiltration abundance of several TIICs, including naïve B cells, memory B cells, plasma cells, CD8 T cells, resting memory CD4 T cells, activated memory CD4 T cells, Tregs, resting NK cells, monocytes, resting mast cells, activated mast cells, and eosinophils varied between the cluster 1 and cluster 2 groups, according to CIBISORT (Fig. 1D). Finally, using data from the ICGC, we verified this grouping. We discovered that the ICGC dataset’s AML samples could also be split into two subgroups (Fig. 1E), and the individuals in each grouping had distinct survival characteristics (Fig. 1F). Compared to samples in cluster 2, samples in cluster 1 exhibited lower immune, stromal, and ESTIMATE scores, as illustrated in Fig. 1G. The infiltration abundance of some TIICs, such as naïve B cells, plasma cells, CD8 T cells, naïve CD4 T cells, resting memory CD4 T cells, activated memory CD4 T cells, follicular helper T cells, resting NK cells, activated NK cells, monocytes, macrophages M2, resting mast cells, activated mast cells, and eosinophils, varied across the Cluster 1 and Cluster 2 subgroups, according to CIBISORT (Fig. 1H). All of the aforementioned findings imply that there were differences in the prognosis, TME features, and TIIC infiltration between the two patient categories, which may have an impact on how well immunotherapy works for AML patients.

Fig. 1.

Recognition and validation of LRGs-related molecular subtypes. (A) We separated the TCGA dataset’s AML samples into two subclasses based on the LRGs. (B) Significant survival variations between patients in the two categories were seen. (C) The two subgroups’ TME features were contrasted. (D) The infiltration abundance analysis by CIBISORT. (E) We separated the ICGC dataset’s AML samples into two subclasses based on the LRGs. (F) Significant survival variations between patients in the two categories were seen in the ICGC cohort. (G) The two subgroups’ TME features were contrasted in the ICGC cohort. (H) The infiltration abundance analysis by CIBISORT in the ICGC cohort. ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001

Establishment of a prognostic LRGs-related signature for the TCGA cohort

By univariate Cox regression analysis, 103 of 894 LRGs were identified as prognostic LRGs at P values less than 0.01, including 70 hazardous LRGs and 33 protective LRGs (Fig. 2A). The LASSO-COX, RF, and XGBoost analyses were used to further filter the 103 prognostically important LRGs. From these machine-learning techniques, we selected 30 genes from LASSO-COX (Fig. 2B), 15 genes from XGBoost (Fig. 2C), and 52 genes from RF (Fig. 2D), respectively. Finally, a prognostic LRGs-related signature was built using six of these genes that overlapped (Fig. 2E). Score = (0.8404886×HPS1) + (0.2070149×BCAN) + (0.2210732×SLC2A8) - (0.1985767×DOC2A) + (0.1630757×CHMP4C) + (0.4966097×SLC29A3). The scores of the AML patients and their distribution are shown in Fig. 3A. Patients were divided into score^high and score^low groups according to their mid-value LRGs-scores, and those with higher LRGs-scores had considerably worse survival rates (Fig. 3B). The AUCs for this signature at 1, 2, and 3 years were 0.828, 0.795, and 0.782, respectively, demonstrating its good predictive potential (Fig. 3C). Furthermore, the outcomes of univariate and multivariate Cox regression analysis suggested that this LRGs-related signature might be employed as an independent prognostic factor for AML patients (univariable HR = 2.675, 95%CI 1.989–3.598, P < 0.001; multivariate HR = 2.282, 95%CI 1.669–3.121, P < 0.001).

Fig. 2.

Establishment of a prognostic LRGs-related signature. (A) Univariate Cox regression analysis. (B) LASSO-Cox regressive analysis. (C) RF analysis. (D) XGBoost analysis. (E) Overlapped genes

Fig. 3.

Verification of the LRGs-related signature. (A) The scores of the AML patients and their distribution in the TCGA cohort. (B) Patients with higher LRGs-scores had considerably worse survival rates. (C) ROC analysis in the TCGA cohort. (D) The scores of the AML patients and their distribution in the ICGC cohort. (E) Patients with higher LRGs-scores had considerably worse survival rates. (F) ROC analysis in the ICGC cohort. (G) The scores of the AML patients and their distribution in the TARGET cohort. (H) Patients with higher LRGs-scores had considerably worse survival rates. (I) ROC analysis in the TARGET cohort

Verification of the LRGs-related signature in the ICGC and TARGET cohorts

Patients in the ICGC cohort were classified as belonging to score^high and score^low groups based on their mid-value LRGs-scores, which were calculated in the same way (Fig. 3D). Patients with higher LRGs-scores had considerably worse survival rates (Fig. 3E). The AUCs for this signature at 1, 2, and 3 years were 0.676, 0.739, and 0.674, respectively, demonstrating its good predictive potential (Fig. 3F). Furthermore, the outcomes of univariate and multivariate Cox regression analysis suggested that this LRGs-related signature might be employed as an independent prognostic factor for AML patients (univariable HR = 1.202, 95%CI 1.089–1.327, P < 0.001; multivariate HR = 1.206, 95%CI 1.092–1.332, P < 0.001). Similarly, patients in the TARGET cohort were classified as belonging to score^high and score^low groups based on their mid-value LRGs-scores, which were calculated in the same way (Fig. 3G). Patients with higher LRGs-scores had considerably worse survival rates (Fig. 3H). The AUCs for this signature at 1, 2, and 3 years were 0.627, 0.581, and 0.591, respectively, demonstrating its good predictive potential (Fig. 3I). Unfortunately, LRGs-related signature is not an independent factor influencing patient survival due to individual patient variability in the TARGET cohort.

Correlation analysis between the LRGs-related signature and TME

Compared to patients in the score^low group, patients in score^high group exhibited higher stromal and ESTIMATE scores, as illustrated in Fig. 4A, showing some variations in TME between the two groups. The infiltration abundance of some TIICs, such as monocyte and macrophage monocyte, varied across the score^high and score^low subgroups, according to the MCPcounter database (Fig. 4B). The infiltration abundance of some TIICs, such as B cells plasma, Tregs, monocytes, macrophage M2, and activated mast cells varied across the score^high and score^low subgroups, according to CIBISORT (Fig. 4C). The infiltration abundance of some TIICs, such as non-regulatory CD4 T cell, common lymphoid progenitor, common myeloid progenitor, cancer-associated fibroblast, macrophage, macrophage M1, macrophage M2, monocyte, Neutrophil, plasmacytoid dendritic cell, T cells gamma delta, varied across the score^high and score^low subgroups, according to xCELL database (Fig. 4D). According to Table S2, individuals in the score^high subgroup had greater immune cell infiltrates in their TME than patients in the score^low subgroup when the three aforementioned algorithms are combined. The most common kind of differential immune cell was the monocyte.

Fig. 4.

Correlation Analysis between the LRGs-related signature and TME. (A) Compared to patients in the score^low group, patients in score^high group exhibited higher stromal, immune, and ESTIMATE scores. (B) The infiltration abundance analysis by the MCPcounter database. (C) The infiltration abundance analysis by CIBISORT. (D) The infiltration abundance analysis by the xCELL. ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001

Evaluation of genetic variations and functional enrichment analysis

The GSEA results showed that impacted GO and KEGG components were mostly engaged in biological processes related to cell cycle, immune response, granulocyte, and chemokine signaling pathway (Figure S2). Genetic modification studies focusing on significantly altered genes showed that the two groups had extremely different mutation rates from each other (Fig. 5A). The gene with the highest mutation rate in the score^high group was NPM1, while the gene with the highest mutation rate in the score^low group was IDH2. In addition, we compared the differences in tumor mutation burden (TMB) values between the two groups. Although there was no significant difference between the two groups (Fig. 5B), patients in the score^highTMB^high group had a worse survival prognosis compared to patients in the score^lowTMB^low group (Fig. 5C).

Fig. 5.

Evaluation of genetic variations. (A) Genetic modification analysis. (B) Although there was no significant difference between the two groups, (C) patients in the score^highTMB^high group had a worse survival prognosis compared to patients in the score^lowTMB^low group. NS, not significant. (D) Expression levels of CHMP4C between AML and healthy volunteers. ***p < 0.001

Construction of a nomogram

To assess how well the coefficient prediction of this LRGs-related signature works, a nomogram was created. The findings indicated that a precise quantitative approach for predicting the 1-, 2-, and 3-year survival rates may be developed using the nomogram with a C-index of 0.809 (Figure S3A). The probability of the anticipated and actual 1-, 2-, and 3-year survival rates’ overlap on the calibration curves indicated high agreement (Figure S3B).

Sensitivity testing for drugs

We found that patients in the score^high group had a higher IC50 for RO-5,126,766, 4SC-202, Tipifarnib, ARRY-704, LY-3,009,120, and ulixertinib (Figure S4), suggesting that patients in the score^low group were more susceptible to these drugs.

Expression levels of these LRGs between AML and healthy volunteers

We started by looking at the six LRGs’ expression levels in the GEPIA database [18]. Only SLC2A8, CHMP4C, and DOC2A were screened out with the cutoff criterion of|logFC| ≥ 1 and a P value of less than 0.05 (Figure S5A). Although all six genes were associated with patient survival in the TCGA dataset (Figure S5B), DOC2A was not associated with patient survival in the ICGC dataset (Figure S5C), and only CHMP4C was associated with patient survival in the TARGET dataset (Figure S5D). Taken together, highly expressed CHMP4C is closely associated with poor prognosis in AML patients and may play an important role in AML development. The mRNA expression level of CHMP4C was examined by qRT-PCR to further confirm the differential expression between AML and healthy samples. According to the qRT-PCR experiment, CHMP4C expression was considerably greater in AML specimens than in healthy specimens, as was predicted (Fig. 5D).

Discussion

Most AML patients have benefited from improvements in AML-cell targeted molecular therapies and immunotherapies. Novel CDK8 inhibitors can regulate acute myeloid leukemia progression by inhibiting STAT-1 and STAT-5 phosphorylation [19]. Light-activatable nanoparticles encapsulating all-trans retinoic acid can effectively improve the consolidation of acute myeloid leukemia by modulating macrophage cell differentiation [20]. Dinaciclib impedes the proliferation of AML cells via the ERK1/STAT3/MYC pathways [21]. Targeting CCL2/CCR2 signaling can effectively alleviate MEK inhibitor resistance in AML [22]. TIM-4-L-directed T-cell therapy demonstrates potential anti-leukemic activity in preclinical AML models [23]. Haploid mbIL-21 in vitro expansion of NK cells significantly improves survival outcomes in AML patients [24]. Through suppression of Wnt/β-catenin signaling, the multi-CDK inhibitor dinaciclib overcomes the resistance of acute myeloid leukemia to Bromo- and extra-terminal domain (BET) inhibitors [25]. However, because of high recurrence rates, AML patient survival rates remain poor [26]. Using trustworthy molecular features to divide patients into high- and low-risk categories may assist in choosing the best treatment plans in line with precision medicine. Lysosomes are crucial in the development of tumors, according to studies [27, 28]. However, the connection between LRGs and AML prognosis is still unclear. In this work, a novel LRGs-related signature for AML patients that may effectively predict their prognosis was developed and validated in two independent cohorts.

Through bioinformatic analysis and validation using clinical samples, we identified the lysosome-associated gene CHMP4C as being significantly overexpressed in patients with AML and closely linked to poor prognosis in these patients. One of the charged multivesicular proteins (CHMPs), CHMP4C contributes to the formation of the endosomal sorting complex (ESCRT-III) needed for transport III, which in turn helps to separate daughter cells as needed. It has been claimed that CHMP4C has a role in the development of different cancers. CHMP4C is a novel marker that controls prostate cancer growth by modulating cyclin pathways and aiding in immunotherapy [29]. CHMP4C can be used as a T-cell exhaustion-associated gene for the prediction of survival and immunotherapy efficacy in bladder cancer [30, 31]. CHMP4C may also affect breast cancer cell growth and Adriamycin resistance [32]. However, the role and mechanism of CHMP4C in AML have not been reported and deserve further exploration in future work.

In addition to being a result of the complexity of tumor cells, the limited therapeutic efficacy of traditional treatments for tumors is also a result of the complexity of the TME, which is made up of macrophages, T cells, dendritic cells, fibroblasts, and other cells [33]. Because of their strong ties to the TME, lysosomes can influence the dystrophic, acidic, hypoxic, and ischemic characteristics of the TME, which can lead to proliferation, drug resistance, invasion, metastasis, and immunosuppression [10]. Monocytes were the common differential immune cells between the score^high and score^low groups when we specifically analyzed the relationship between LRGs-score and TME using three algorithms, including CIBERSORT, XCELL, and MCPcounter. This finding suggests that the LRGs-related signature may affect the abundance of monocytes in TME and thereby affect the prognosis of AML patients [34].

Our analysis still offers some benefits even if comparable papers have constructed traits as a prognostic factor to the prognosis of AML patients in the past. First, we evaluated the prognosis of AML patients using LRGs for the first time. The characteristic was also effectively confirmed using the TCGA ICGC, and TARGET datasets, three publicy available datasets. To validate the gene expression levels that make up this LRGs-related signature, we lastly looked at clinical data. Of course, there are certain restrictions on our study. Future research on this LRGs-related signature will require a substantial multicenter randomized controlled study with comprehensive survival data. Additional in vivo and in vitro research will be needed for future studies on the precise mechanisms of CHMP4C in AML.

The study still has certain limitations. Our study primarily uses data from the TCGA and ICGC cohorts, which underrepresent non-Caucasian populations, potentially limiting the applicability of our findings to other ethnic groups. Future research should include more diverse cohorts to explore ethnic-specific differences in LRGs in AML. While our LRG-based model offers new insights into AML prognosis, it should be used alongside traditional markers like the ELN risk stratification, which considers cytogenetic and molecular features. Combining our model with ELN could improve prognostic accuracy, and future studies should validate this integrated approach for a more robust prognostic tool. Augmenting the capability of this model to guide personalized treatment strategies for acute myeloid leukemia, particularly in predicting chemotherapy sensitivity (e.g., response to cytarabine) and assessing the efficacy of immunotherapy (e.g., the impact of PD-1/CTLA-4 inhibitors), will significantly broaden its personalized applications.

Conclusions

In summary, our study revealed a 6-gene signature associated with prognosis in AML patients. The signature can be used as a potential candidate biomarker and therapeutic a prognostic factor for the treatment of AML. The precise mechanisms of CHMP4C in AML is worthy of further study.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Gongchang Li, Yangyang Miao and Liqiang wrote the main manuscript text, Fang Yuan, Weiran Zhang, Yali Wu and Liqiang Zhu prepared figures. All authors reviewed the manuscript.

Funding

This research was supported by 2020 Henan science and technology project (LHG20200424).

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Ethics approval and consent to participate

This study was supported by the Ethics Committees of Zhengzhou University (2022-KY-0631-002). Written informed consent was obtained from all patients. All methods were performed following the relevant guidelines and regulations. The manuscript is consistent with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.