Acute myeloid leukemia (AML) is a fast-growing cancer of the blood and bone marrow in which the body produces abnormal white blood cells called blasts. One of the most difficult aspects of treating AML is predicting which patients will do well and which will have poor outcomes - a process called risk stratification.
Currently, doctors use cytogenetics (the study of chromosome abnormalities in cancer cells) and specific genetic mutations to classify AML patients into risk groups. The leading system for this is the European LeukemiaNet (ELN) classification, which sorts patients into favorable, intermediate, and adverse risk categories.
However, within each risk category there is still enormous variation. Some patients classified as intermediate-risk do extremely well, while others fare poorly. This means many patients may be receiving treatment that is either too aggressive or not aggressive enough for their actual disease biology.
Researchers have tried adding individual genetic markers, epigenetic changes, or multi-gene signatures to improve predictions, but most existing models were developed only in specific AML subsets - particularly the 40-50% of patients whose leukemia cells have chromosomally normal karyotypes (CN-AML). There remained an unmet need for a tool that works across all AML subtypes and is simple enough for routine clinical use.
The researchers used a genome-wide expression analysis approach - examining the activity levels of tens of thousands of genes across large groups of AML patients - to identify which genes best predicted survival. This unbiased approach avoids assumptions about which genes might matter.
Four independent patient datasets were used: two served as training sets (to build the model) including 163 patients from The Cancer Genome Atlas (TCGA) and 79 patients from the German AML Cooperative Group (AMLCG) 1999 trial. Two additional datasets served as validation sets to test whether the model worked in independent patient populations.
The analytical pipeline was rigorous: genes had to show significant impact on overall survival in both univariate and multivariate analyses (accounting for age as a competing factor), and the effect had to point in the same direction across both training sets. This stringent filtering narrowed 15,939 genes down to just 30 candidate genes.
From those 30 candidates, the researchers systematically calculated every possible combination of 1, 2, 3... up to 30 genes to find the optimal balance between predictive power and clinical simplicity. Predictive gain plateaued after about 8 genes, and a 3-gene combination was chosen as the most practical for routine clinical use while still delivering strong predictive power.
The best 3-gene combination consisted of CAP1 (adenylyl cyclase-associated protein 1), FAM124B (family with sequence similarity 124B), and CXCR6 (C-X-C chemokine receptor type 6). Together these form the Tri-AML Score (TriAS).
The score is calculated simply: add 1 point if CAP1 expression is above the median for that dataset; add 1 point if FAM124B expression is above the median; subtract 1 point if CXCR6 expression is above the median; then add 2 points to keep values positive. This yields a score ranging from 1 to 4 points.
Patients are then grouped into three risk categories: low risk (1 point), intermediate risk (2-3 points), and high risk (4 points). In the training set of 242 patients, these categories showed dramatically different median survival times: low-risk patients survived a median of 1,413 days while high-risk patients survived only 210 days.
The score performed equally well in the two independent validation sets. In the Taiwan hospital validation set of 224 patients, low-risk patients had a median survival that was not reached (meaning most were still alive at follow-up), while high-risk patients survived only 17.4 months on average.
A critical test for any new prognostic tool is whether it adds information beyond what doctors already know. The researchers performed multivariate Cox regression analysis - a statistical method that tests whether a factor predicts survival even when accounting for all other known factors simultaneously.
TriAS remained independently significant when analyzed alongside age over 65, cytogenetic risk group, gender, and molecular mutations including FLT3 and NPM1 - two of the most well-known AML-relevant gene mutations. This means TriAS provides genuinely new prognostic information beyond these established factors.
The score also predicted outcomes in the particularly challenging population of elderly patients over 65. Among 166 patients in this age group, the three TriAS categories showed distinct median survival times: 26.8 months for low-risk, 8.3 months for intermediate-risk, and only 4.1 months for high-risk patients.
Additionally, TriAS predicted not just overall survival but also relapse-free survival - how long patients remained in remission without the leukemia returning - in the Taiwan validation dataset.
The researchers compared TriAS head-to-head against three other published gene expression scoring models for AML: a 7-gene score by Marcucci, an 11-gene score by Chuang, and a 24-gene score by Li. All three had been developed using different patient populations and different gene sets.
When all four scores, plus age, gender, and cytogenetic risk, were entered into a single multivariate model, only two factors remained independently significant in the TCGA training set: age over 65 and TriAS. The three other expression-based scores lost statistical significance when TriAS was included.
The researchers also tested whether combining multiple scores together could further improve predictions. Pairing TriAS with any of the three other scores did create more refined subgroups, and when all four scores were used together, patients could be separated into very distinct outcome groups. However, the incremental benefit of adding each additional score was relatively modest.
This result is particularly notable because TriAS achieves better independent predictive power using only 3 genes compared to models using 7, 11, or 24 genes - making it far more practical for routine clinical laboratory testing.
The European LeukemiaNet (ELN) classification divides AML patients into four risk groups based on chromosomal changes and mutations in genes such as CEBPalpha, NPM1, and FLT3. It is the most widely used risk system for adult AML. However, ELN-defined risk groups still contain considerable heterogeneity.
When TriAS was applied within each ELN risk group, it revealed substantial numbers of patients whose true prognosis differed significantly from what their ELN category suggested. For example, within ELN-favorable patients, only 58.8% were confirmed as truly favorable when TriAS was considered; the remainder had worse outcomes than expected.
Most strikingly, among ELN intermediate-risk patients (groups 1 and 2 combined), 62.1% were reclassified as actually being of adverse risk when TriAS was applied. This is clinically important because ELN intermediate patients currently may not receive the most intensive available treatments.
Overall, adding TriAS to the ELN system reclassified 44.5% of patients (167 out of 375) into a different risk category. The combined ELN+TriAS system created three well-separated groups with 3-year overall survival rates of greater than 60% (favorable), 50-60% (intermediate), and 25% or less (adverse).
CXCR6 is a receptor for the signaling molecule CXCL16, found on various immune cells and expressed in bone marrow. In many solid cancers such as breast, prostate, thyroid, and gastric cancer, the CXCR6-CXCL16 signaling axis promotes tumor growth and invasion through the ERK survival pathway. Paradoxically, in AML, higher CXCR6 expression was associated with better survival, suggesting it may function differently in blood cancers than solid tumors - a finding that requires further laboratory investigation.
FAM124B is a relatively newly identified protein that interacts with CHD7 and CHD8 - proteins involved in regulating how DNA is packaged and read (chromatin remodeling). Mutations in CHD7 cause CHARGE syndrome, a developmental disorder, and CHD8 has been linked to leukemia in animal models. In AML, higher FAM124B expression correlated with worse survival, consistent with a potential role in cancer biology.
CAP1 (adenylyl cyclase-associated protein 1) regulates the actin cytoskeleton - the internal scaffolding that allows cells to move and divide. It has been linked to tumor growth and metastasis in breast cancer, liver cancer, and glioma (brain tumor). In AML, higher CAP1 expression was associated with worse survival, consistent with an oncogenic (cancer-promoting) role.
While all three genes have known roles in other cancers, their exact mechanisms in AML remain to be fully worked out. The authors acknowledge that future laboratory studies will be needed to confirm and characterize these roles at a molecular level.
The TriAS system offers several practical advantages over existing gene expression-based scoring models. Its simplicity - just three genes - means it could potentially be implemented in clinical laboratories using a technique called quantitative real-time PCR, a standard and widely available method for measuring gene activity from a patient's leukemia sample.
The score was validated across multiple independent patient cohorts from different countries (Germany, the United States via TCGA, and Taiwan), suggesting it captures a biological signal that is robust across diverse populations and not an artifact of any single dataset.
The most immediate clinical implication is the ability to identify patients within the ELN intermediate-risk category who are actually at much higher risk. These patients might benefit from more aggressive treatments such as allogeneic stem cell transplantation during first remission, a decision currently reserved for higher-risk categories.
The authors call for prospective validation - testing the score in future clinical trials where patients are enrolled going forward, rather than using historical data - as the next critical step before TriAS could be adopted in routine clinical practice.