Acute myeloid leukemia (AML) is an aggressive blood cancer in which immature myeloid cells multiply uncontrollably in the bone marrow. Treatment typically involves intensive multi-drug chemotherapy, often followed by a stem cell transplant. While AML can be cured in roughly 35 to 40 percent of adults under 60 with standard therapy, chemotherapy-resistant disease and relapse remain major obstacles to survival.
Doctors currently classify AML patients into risk groups primarily based on cytogenetic abnormalities - visible chromosomal rearrangements detected under a microscope - and specific gene mutations such as FLT3 and NPM1. The European LeukemiaNet (ELN) risk classification system organizes patients into favorable, intermediate, and adverse risk categories. However, 50 to 70 percent of AML patients have normal or risk-indeterminate chromosomes, making this classification unreliable for a large fraction of patients.
Critically, even within each ELN risk category, patient outcomes vary enormously. A reanalysis of data from 1,540 AML patients showed that one-third have survival predictions that deviate more than 20 percent from what their ELN category would predict. This means the current system frequently misclassifies patients, potentially denying aggressive treatment to those who need it or over-treating those who would do well with standard therapy.
This study from Nottingham Trent University and international collaborators applied artificial neural network (ANN) machine learning to gene expression data from 593 adults with AML, with the goal of discovering a simple, highly predictive gene signature that could refine risk stratification. Rather than seeking dozens of genes, the approach aimed for a minimal but powerful predictor - a concept called parsimony in statistics.
The study used gene expression data from five independent AML patient datasets with a total of 1,643 patients. The primary discovery cohort was the HOVON series - 593 adults from Dutch and Swiss clinical trials - analyzed on Affymetrix gene chips measuring the activity of 54,675 gene transcripts simultaneously. Four independent validation cohorts were then used to test whether the discovered signature predicted outcomes in entirely different patient populations.
The machine learning approach used was a multilayer perceptron artificial neural network (ANN) with back-propagation learning - a type of algorithm that mimics how biological neural networks learn by adjusting connection weights based on errors. Unlike traditional statistical approaches such as linear regression, ANNs can detect complex non-linear patterns. The ANN first assessed each gene individually, then tested combinations of genes sequentially, identifying which combinations provided the strongest prediction of whether a patient survived beyond 20 months - the inflection point in the survival curve.
From 54,675 measured genes, the ANN identified three genes - CALCRL, LSP1, and CD109 - as the combination providing the most powerful prediction of patient survival. Their expression levels were used to compute a prognostic index (PI) using a mathematical formula: PI = (1.734 x CALCRL) + (1.092 x LSP1) + (0.826 x CD109), where each gene's expression was normalized between 0 and 1. Patients were divided into low PI (below 1.0), intermediate PI (1.0 to 1.5), and high PI (above 1.5) groups.
Validation was performed in three independent adult AML cohorts (the German, TCGA, and Beat AML series, comprising 905 patients total) and one pediatric cohort (the TARGET series, 145 children). The researchers also confirmed the results using real-time PCR in 38 additional patients from a German clinical center (the SAL series), validating that the gene expression measurements could be performed in a standard clinical laboratory setting.
Each of the three genes individually predicted patient survival with high statistical significance. When combined into the PI formula, the prognostic power was substantially greater. In the discovery cohort, patients in the PIhigh group had a median overall survival of just 0.74 years (about 9 months), compared with 1.49 years for intermediate-PI patients and an undefined (more than 8 years) median survival for low-PI patients. These differences were extraordinarily statistically significant (P less than 0.0001).
The 3-gene PI performed as well or better than existing molecular risk tools. Comparing against the 17-gene leukemia stem cell score (LSC17) - a validated prognostic tool - the 3-gene PI achieved a similar area under the receiver operating characteristic curve (AUC of 0.72 vs 0.69), while outperforming a 66-gene expression score from bone marrow cells (PI AUC of 0.69 vs 0.61). Critically, the PI remained a significant independent predictor of survival in multivariate analyses that included age, cytogenetic risk group, FLT3 mutation, NPM1 mutation, and other established prognostic factors - meaning its predictive value could not be explained by these known factors alone.
Particularly striking was the PI's ability to refine prognosis within each ELN cytogenetic risk category. Among patients classified as ELN favorable risk - expected to do relatively well - those with a high PI had a median overall survival of only 0.96 years, compared with undefined survival for low-PI patients in the same favorable-risk category. In other words, the PI identified a subgroup of supposedly favorable-risk patients who had outcomes as poor as adverse-risk patients.
At the other extreme, in the ELN adverse-risk group, PIhigh patients had event-free and overall survival rates of just 1.8% and 7%, respectively - identifying the highest-risk subgroup within what is already a poor-prognosis category. The PI also added prognostic value within specific molecular subgroups including patients with NPM1 mutations, FLT3 mutations, and wild-type variants of both, demonstrating broad applicability across all molecular subsets of AML.
The 3-gene PI was validated in the German series (535 adults, including 223 with cytogenetically normal AML - the largest and most difficult-to-classify subset). In this cohort, PIhigh patients had a median overall survival of 0.7 years versus a median undefined for PIlow patients (P less than 0.0001). Even within the cytogenetically normal AML subgroup, where no chromosomal abnormality provides prognostic information, the PI separated patients into meaningfully different survival groups.
In the TCGA and Beat AML series - both using RNA sequencing rather than the microarray technology of the discovery cohort - the PI continued to stratify survival significantly. This cross-platform validation is particularly important: it demonstrates that the PI is not an artifact of a specific gene measurement technology but captures genuine biological differences in the cancer. In the Beat AML cohort, the 3-gene PI predicted treatment response (AUC of 0.707, sensitivity 91%) while standard cytogenetic risk classification failed to predict response in this dataset.
In the pediatric TARGET cohort (145 children), the results were more nuanced. When all pediatric cases were included, the PI did not significantly stratify survival - likely because most children (128 of 145) were classified as low-PI, suggesting that the biology of childhood AML differs from adult AML. However, after excluding core-binding factor AML cases (which have a distinctly favorable biology even in children), the PI did significantly predict overall survival in the remaining 96 pediatric patients.
In the SAL clinical validation series of 38 adults tested by PCR in a real clinical setting, patients with a high PI were significantly more likely to have poor early treatment response - defined as more than 10% leukemia blasts in the bone marrow one week after chemotherapy ended. Furthermore, in 9 patients who developed leukemia relapse, all three genes (CALCRL, LSP1, and CD109) were significantly more highly expressed at relapse compared with diagnosis - suggesting these genes may actively contribute to treatment resistance.
CALCRL (calcitonin receptor-like receptor) is a cell surface receptor for adrenomedullin - a protein that stimulates cell growth and inhibits programmed cell death (apoptosis). Adrenomedullin signaling promotes tumor growth and has been implicated in prostate and breast cancer. In AML, high CALCRL expression may allow leukemia cells to resist the cell death signals that chemotherapy depends upon, potentially explaining why PIhigh patients respond poorly to treatment.
LSP1 (lymphocyte-specific protein 1) is an intracellular protein that binds to the cytoskeleton (specifically F-actin, the structural fiber network inside cells) and is normally expressed in immune cells including lymphocytes, neutrophils, and macrophages. Its connection to immune cell function is intriguing - abnormal LSP1 expression in AML blasts may allow leukemia cells to mimic immune cell behaviors that help them evade destruction. LSP1 had not previously been reported in AML prognostic signatures.
CD109 is a cell surface protein that normally suppresses TGF-beta and STAT3 signaling - two pathways that regulate cell growth and immune responses. In AML, CD109 was positively correlated with the percentage of leukemia blast cells in the bone marrow. Importantly, CD109 shows high expression in early hematopoietic stem cells and then declines during normal blood cell development, suggesting that AML cells with high CD109 may be arrested in or reverting to an immature stem cell-like state. CD109 is also being evaluated as a potential target for antibody-based cancer therapies.
A key finding from the pan-cancer analysis using the PRECOG database (covering 39 different tumor types and 26,000 patients) was that AML was the only cancer type in which all three genes were simultaneously upregulated and predicted shorter survival. In every other cancer type examined, the pattern did not hold. This AML-specificity suggests that the 3-gene signature is not capturing a generic cancer survival mechanism but rather something unique to AML biology - potentially making it highly specific as a diagnostic and prognostic tool.
The study identifies several categories of patients who might benefit from using the 3-gene PI to guide treatment. Patients classified as ELN favorable risk but PIhigh represent a particularly important group: current guidelines would not recommend allogeneic stem cell transplant for these patients in first remission, but their actual survival resembles intermediate-risk patients. The PI could justify offering these patients more intensive therapy including transplant if their expected relapse rate exceeds 35 to 40 percent.
Conversely, patients with ELN intermediate risk but PIlow had significantly better outcomes than expected - suggesting some patients currently recommended for aggressive treatment might safely receive less intensive approaches, sparing them chemotherapy toxicity and transplant-related risks. This ability to move patients into more accurate risk categories within the established ELN framework is the key clinical value proposition of the PI.
The PI also predicted outcomes in patients receiving different types of treatment. A high PI predicted shorter overall survival both in patients who received chemotherapy alone and in patients who received chemotherapy followed by stem cell transplant. Interestingly, allogeneic transplant improved survival for PIint and PIhigh patients but not for PIlow patients - who did well regardless of whether they received transplant. This suggests the PI could help identify which patients actually benefit from the added toxicity of transplant.
The PI can be calculated from standard PCR measurements of just three genes, a technique available in any well-equipped clinical laboratory. The fact that the PI was validated across multiple gene expression platforms (microarrays and RNA sequencing) means it is technologically versatile. Combined with its validation in over 1,600 patients across multiple countries and treatment protocols, the PI is one of the most thoroughly validated gene expression prognostic tools in AML to date.
This study demonstrates that an artificial neural network approach applied to transcriptomic data can identify a minimal but highly predictive gene signature that captures clinically meaningful variation in AML biology - variation that current cytogenetic and molecular classification systems miss. The parsimony of the approach (three genes rather than dozens) makes practical clinical implementation far more feasible.
The PI's performance in multivariate analyses - remaining significant after controlling for age, cytogenetic risk, and established molecular markers - indicates it provides genuinely independent prognostic information, not merely a repackaging of known risk factors. The ability of the PI to predict treatment response (especially in the Beat AML cohort where the standard ELN classification failed to predict response) suggests particular value for guiding initial treatment selection rather than just estimating post-treatment prognosis.
The observation that CALCRL, LSP1, and CD109 are all upregulated at relapse compared with diagnosis opens a different avenue: these genes may not only predict prognosis but may themselves be contributing to treatment resistance. If so, they represent potential therapeutic targets - blocking their activity might make AML cells more sensitive to chemotherapy. CD109 is already under evaluation as a target for antibody-based therapy, providing a direct translational path.
The study authors call for prospective clinical studies to formally test whether incorporating the PI into clinical decision-making improves patient outcomes compared with current standard risk classification. Such trials would also need to assess the PI's performance in specific subgroups underrepresented in current cohorts, including elderly patients and those with secondary AML arising from prior blood disorders. If validated in prospective trials, the 3-gene PI represents a significant step toward truly personalized AML treatment.