Acute myeloid leukemia (AML) is one of the most aggressive blood cancers. Although initial treatment achieves remission in most patients, relapse remains the leading cause of treatment failure. Among adults, 40-60% relapse within 3 years; among children, the rate is 30-40%. The 5-year survival rate is only 28% for adults and 70% for children.
Current tools for predicting relapse rely on cytogenetic and genetic features identified at diagnosis. However, for many patients - especially those without known causative mutations - risk stratification remains difficult. There is an urgent need to identify molecular markers that predict which patients are at highest risk of relapse, and to understand what biological changes drive disease progression.
RNA sequencing (RNA-seq) allows researchers to measure the activity level of every gene in a cell simultaneously, providing a comprehensive snapshot called the transcriptome. By comparing gene activity in tumor samples collected at diagnosis versus at relapse, this study aimed to identify genes whose expression changes as AML progresses and becomes resistant to treatment.
This study analyzed longitudinal samples - meaning samples from the same patients collected at multiple time points (diagnosis, relapse, and sometimes a second relapse). This paired design is powerful because each patient serves as their own control, making it easier to detect true disease-driven changes rather than differences between patients.
The study analyzed 122 samples from 70 patients with relapsed or treatment-resistant AML, including 47 adults and 23 pediatric patients from Nordic countries. Samples were collected at diagnosis, first relapse, and in some cases second or third relapse, making this one of the largest longitudinal AML transcriptomic datasets assembled.
All patients had previously been characterized by whole-genome sequencing (WGS) or whole-exome sequencing (WES), providing a detailed map of each patient's mutations. This allowed the researchers to integrate gene expression data with known genetic changes - a more complete picture than gene expression alone can provide.
Gene expression analysis was performed using RNA-seq and differential expression analysis tools. The team then applied machine learning using the Monte Carlo Feature Selection (MCFS) algorithm, which identifies which genes are most informative for distinguishing diagnosis from relapse. These top genes were used to build rule-based classifiers that express predictions as interpretable IF-THEN rules rather than black-box outputs.
To validate their findings, the researchers compared results against two independent large datasets: TCGA (The Cancer Genome Atlas, for adults) and TARGET (Therapeutically Applicable Research to Generate Effective Treatments, for children). Genes showing the same trends in independent cohorts provide much stronger evidence of true biological relevance.
When comparing gene expression at diagnosis between patients who relapsed quickly (short event-free survival) versus those who stayed in remission longer (long event-free survival), three genes stood out: GLI2, IL1R1, and ST18.
GLI2 is a mediator of the Sonic Hedgehog (Shh) signaling pathway, which helps maintain stem cells. Overexpression of GLI2 was strongly associated with shorter event-free survival (p=0.0001) and shorter overall survival. GLI2 overexpression was especially common in patients with FLT3-ITD mutations, but even after excluding FLT3-ITD cases, high GLI2 remained a marker of poor outcome. This suggests GLI2 may support leukemia stem cells that resist chemotherapy.
IL1R1 encodes a receptor for interleukin-1, a key inflammatory signaling molecule. High IL1R1 expression was linked to both shorter event-free survival and shorter overall survival. IL1R1 activates inflammatory pathways including NF-kB and MAPK, and promotes cancer cell survival by inhibiting programmed cell death. Drugs targeting IL1R1 signaling have already been shown to increase sensitivity to chemotherapy in AML models.
ST18 encodes a zinc finger transcription factor that normally suppresses inflammation and promotes cell death. Low ST18 expression was associated with shorter event-free survival and shorter overall survival. ST18 was first identified as a tumor suppressor in breast cancer, and this study shows its downregulation is also a negative prognostic marker in AML.
Comparing matched diagnosis and relapse samples identified genes that change as AML evolves. Two genes were significantly altered in both adult and pediatric patients: CR1 (complement receptor 1) and DPEP1 (dipeptidase 1).
CR1 encodes a receptor found on circulating blood cells that links the innate and adaptive immune systems by helping clear antibody-coated bacteria. CR1 expression was significantly lower at relapse than at diagnosis in both adults and children. Since CR1 normally acts as a brake on the complement system - a cascade that can promote inflammation and tumor growth - its loss at relapse may contribute to a more pro-tumor inflammatory environment in the bone marrow.
DPEP1, which encodes an enzyme involved in regulating inflammation and cell migration, was significantly higher at relapse in both adults and children. High DPEP1 has been linked to increased proliferation and survival in leukemia models, and its surface expression on AML cells makes it a potential target for emerging immunotherapy approaches such as CAR-T cell therapy.
Gene ontology enrichment analysis showed that the genes changing at relapse were concentrated in pathways related to immune response and complement activation, further supporting the idea that a pro-inflammatory tumor microenvironment is a key driver of AML progression and therapy resistance.
Using interpretable machine learning with the R.ROSETTA framework, the study built rule-based classifiers that predict whether a sample is from diagnosis or relapse. These models expressed their logic as IF-THEN rules, making them more transparent than typical black-box algorithms.
In adult AML, CD6 - a surface glycoprotein involved in cell adhesion and immune synapse formation - emerged as the feature most commonly distinguishing relapse from diagnosis, with overexpression at relapse. CD6 may allow AML cells to adhere to protected bone marrow niches that shield them from chemotherapy, a mechanism of treatment escape.
INSR (insulin receptor) was found to be downregulated at relapse in adults, often co-occurring with high CD6 expression. Low INSR may reduce cell proliferation signals, pushing leukemia cells into a more quiescent (dormant) state. Dormant cancer cells are notoriously resistant to chemotherapy, which primarily kills rapidly dividing cells. Low INSR expression at diagnosis was also associated with poor outcome in the TCGA validation cohort.
In pediatric AML, machine learning on merged local and TARGET datasets identified NFATC4 (a transcription factor associated with cell quiescence) and KATNAL2 (a microtubule-regulating enzyme) as genes downregulated at AML diagnosis whose expression recovered toward normal at relapse. This suggests the initial downregulation promotes leukemia onset, while higher expression at relapse may reflect selection for more dormant, treatment-resistant cells.
A striking theme across all analyses in this study is the central role of pro-inflammatory signaling in AML relapse. Multiple genes identified - IL1R1, CR1, DPEP1, ST18 - are directly involved in regulating inflammatory pathways. This is consistent with the growing understanding that tumors exploit inflammation to survive, evade immunity, and resist therapy.
The loss of CR1 at relapse is particularly significant because CR1 normally inhibits complement activation. When complement becomes unregulated, it can paradoxically promote tumor growth by remodeling the tumor microenvironment and suppressing anti-tumor immunity. This means AML cells at relapse may be actively disabling a brake on the cancer-promoting inflammatory cascade.
The data also suggest that Hedgehog pathway inhibitors - such as glasdegib, already approved for certain AML patients - could benefit patients with high GLI2 expression at diagnosis. Similarly, drugs targeting IL1R1 signaling and the complement system are already under clinical investigation and could be prioritized for patients whose tumors show these specific transcriptomic signatures.
BCR-ABL1 gene fusions were detected at relapse in two treatment-resistant cases, including one only identifiable by RNA-seq (not whole exome sequencing). This highlights a practical benefit of RNA-seq: it can detect structural genetic changes - like gene fusions - that DNA-based sequencing methods may miss, and such fusions may make patients eligible for tyrosine kinase inhibitor therapy.
This study provides a rich molecular map of AML progression, identifying both diagnostic biomarkers and potential therapeutic targets. By combining comprehensive RNA-seq with known genetic data and interpretable machine learning, it moves beyond single-gene analyses to reveal how networks of gene changes collectively drive relapse.
The validated finding that GLI2, IL1R1, ST18, CR1, and DPEP1 are consistently dysregulated across multiple independent cohorts gives researchers and clinicians confidence that these are genuine drivers of AML biology rather than statistical noise. Each of these genes represents a candidate biomarker for patient stratification or a drug target for clinical development.
The authors emphasize the value of longitudinal sampling - studying the same patient before and after treatment - for understanding how cancer evolves. Most studies compare patients with different outcomes at a single time point, missing the dynamic changes that drive therapy resistance. Future clinical trials should routinely collect serial biopsies to enable this kind of mechanistic insight.