Acute myeloid leukemia (AML) is one of the most aggressive blood cancers, with a mostly fatal prognosis. Despite decades of research into its molecular underpinnings, the standard pathway to diagnosis still relies on a combination of microscopic cell assessment, immunophenotyping, cytochemistry, and genetic testing - all procedures that require significant investment in specialized equipment, trained pathologists, and laboratory infrastructure.
A key limitation of this traditional diagnostic approach is that it does not scale economically. Each additional patient sample requires roughly the same amount of expert time and physical infrastructure. This creates a global accessibility problem: high-quality AML diagnosis is available in well-resourced hospitals but is difficult to access in rural areas, lower-income countries, or settings with limited specialist availability.
Gene expression profiling (GEP) measures the activity of thousands of genes simultaneously by quantifying messenger RNA levels - essentially a snapshot of which genes are switched on and how strongly. Leukemia cells have distinctive gene expression patterns that differ markedly from healthy blood cells and from other leukemia types. Researchers recognized this potential 20 years ago, but translating it into scalable, robust clinical tools has remained elusive. This study takes on that challenge directly.
The foundation of this study is an unprecedented dataset assembled from public repositories. The researchers systematically searched the Gene Expression Omnibus (GEO) database, the world's largest public repository of gene expression data, starting from over 153,000 datasets. After rigorous filtering for human peripheral blood and bone marrow samples, exclusion of duplicates and specialized cell subset experiments, they arrived at 105 studies containing 12,029 patient samples - one of the largest AML transcriptomics meta-datasets ever assembled.
The dataset included samples measured on three different platforms: two different Affymetrix microarray types (HG-U133A with 2,500 samples and HG-U133 2.0 with 8,348 samples) and Illumina RNA sequencing (1,181 samples). This diversity was deliberate: real-world clinical deployment requires models that work across different measurement technologies, not just within a single laboratory's preferred system.
The 12,029 samples included 4,145 AML samples spanning diverse disease subtypes, plus 7,884 controls including healthy donors, patients with other leukemias (ALL, CML, CLL), myelodysplastic syndrome (MDS), and other non-leukemic diseases. This breadth allowed rigorous testing of the classifier's ability to distinguish AML from conditions that might be encountered in real differential diagnosis scenarios.
With 12,708 genes measured per sample, AML prediction is a high-dimensional classification problem - there are far more potential predictors (genes) than there are samples in any single study. Standard statistical approaches fail in this regime, but modern high-dimensional machine learning methods can handle it effectively.
The primary classifier used was the LASSO (Least Absolute Shrinkage and Selection Operator), a type of regularized logistic regression that automatically selects a small subset of the most predictive genes from the full genome while setting all other genes' contributions to exactly zero. This sparsity is clinically valuable: it identifies a compact gene signature that could potentially be measured with a targeted assay rather than requiring full transcriptome sequencing.
The researchers compared the lasso against eight other classification methods including k-nearest neighbors, linear support vector machines, linear discriminant analysis, random forests, and deep neural networks with 5 and 10 layers. The lasso was ultimately preferred not only for its strong performance (comparable to deep neural networks) but also for its interpretability - the lasso explicitly identifies which genes drive each prediction, whereas deep neural networks function as black boxes. This interpretability is essential for clinical adoption.
The results were striking. In the standard setting (training and test samples randomly mixed from the same pool of studies), the classifiers achieved 97-99% accuracy across all three platforms when distinguishing AML from all other samples. Importantly, accuracy was similarly high in the differential diagnosis setting - where controls were not healthy people but other leukemia patients (ALL, CLL, CML, MDS) - suggesting the model captures AML-specific biology rather than just the difference between sick and healthy.
The learning curves (how accuracy changes with the number of training samples) showed that even relatively small training sets produced highly accurate classifiers, though accuracy continued to improve with more data. This has practical implications: even in a new clinical setting with only moderate historical data, a useful classifier can be deployed, and it will improve as more data accumulates - a property the authors call incremental learning.
Accuracy was consistent across all major AML subtypes (classified by the French-American-British system: M0 through M7). No single subtype was dramatically harder to classify than others, suggesting the gene expression signature captures a common AML biology shared across subtypes. The algorithm also correctly classified acute promyelocytic leukemia (M3, also defined by the PML-RARA translocation) against other AML subtypes with 98.9% accuracy, 97.6% sensitivity, and 99.5% specificity in the M3-specific subtype experiment.
A crucial test for any clinical diagnostic tool is whether it works on data from institutions and studies it was not trained on - the cross-study validation. The researchers conducted a strictly disjoint cross-study analysis: training data came exclusively from a subset of studies and test data came exclusively from different studies, with no overlap. As expected, accuracy was somewhat lower than in the random mixing scenario, but in the largest dataset (8,348 samples), cross-study performance nearly caught up with the random scenario when training set size was large, with only a small residual gap.
Equally important was the cross-platform analysis: can a model trained on one measurement technology make accurate predictions using data from a different technology? This is essential for long-term clinical use, since technologies change over time. Models trained on the older HG-U133A microarray predicted well on the newer HG-U133 2.0 data. However, direct cross-platform transfer from microarray to RNA-seq failed - the two technologies produce systematically different data distributions.
The solution was elegant and simple: rank transformation to normality. Rather than using the raw gene expression values, each gene's expression level within each sample is replaced by its rank relative to all other genes in that sample. This removes the absolute magnitude information (which differs between platforms) while preserving the relative pattern (which reflects biology). With rank transformation, classifiers trained on microarray data predicted RNA-seq samples accurately, achieving good cross-platform generalization with no additional computational complexity.
The LASSO identified 141 genes as predictive for AML diagnosis in the largest dataset. Interestingly, only 46 of these genes were also differentially expressed (individually different between AML and non-AML on a gene-by-gene basis). This highlights a fundamental insight: predictive power and differential expression are different properties. A gene can be part of an accurate predictive signature without being individually elevated or reduced in cancer cells.
A stability analysis repeated the lasso selection 100 times on random halves of the dataset and identified 80 genes that appeared in at least 50% of selections - robustly chosen predictors. Among these, 35 were also differentially expressed. These 35 genes formed two clusters: one cluster of 29 genes elevated in AML versus all other conditions, and a smaller cluster of 6 genes reduced in AML. The AML-elevated cluster included well-known AML drivers such as KIT, FLT3, and RUNX2, validating the biological relevance of the identified signature.
Remarkably, removing all 155 known AML-associated genes from the training data had essentially no effect on prediction accuracy. The classifier found equally accurate signatures using genes not previously recognized as AML-related - suggesting that many genes affected by AML biology in blood cells remain biologically uncharacterized, and that the machine learning approach may help identify new aspects of AML molecular biology for future investigation.
A central argument of this paper concerns the economics of diagnostic testing. Standard AML diagnosis requires morphology assessment, immunophenotyping, cytochemistry, and cytogenetics - each requiring specialized equipment, trained personnel, and quality infrastructure. These costs scale linearly with sample volume: testing 1,000 patients costs roughly 1,000 times as much as testing 1 patient. This creates severe bottlenecks and access inequities.
In contrast, machine learning-based diagnosis has near-zero marginal cost once a model is trained: running the algorithm on a new gene expression profile is computationally trivial. The cost is dominated by the RNA measurement itself, which at the time of writing was already below $100 per sample and falling - potentially cheaper than the combined classical diagnostic workup. Low-cost whole-transcriptome protocols like BRB-seq were emerging that could further reduce costs to those comparable with measuring just four genes.
Advances in nanopore sequencing - a portable, real-time DNA and RNA sequencing technology - suggest that transcriptomic profiling could eventually be performed at the point of care, with cloud-based machine learning providing the diagnostic classification. This would enable high-quality AML screening at a primary care level, potentially identifying suspected cases earlier in the disease course and routing patients to specialist centers more quickly - before the disease becomes more difficult to treat.
This study demonstrates convincingly that combining large-scale blood transcriptomics with high-dimensional machine learning can yield robust, accurate AML classifiers that approach human expert annotation performance in a near-automated workflow. The system works across multiple studies, multiple platforms, and all major AML subtypes.
The authors position transcriptomic machine learning not as a replacement for genomic subclassification (which remains essential for treatment selection) but as a complement for the initial diagnosis step - particularly valuable in settings where hematological expertise is unavailable or costly. The vision is an integrated workflow: GEP-based ML for primary diagnosis, followed by genomic characterization for risk stratification and treatment planning.
Next steps toward clinical deployment include prospective studies specifically designed to assess diagnostic utility in real-world settings, regulatory approval processes, and the development of deployment-ready data pipelines. The Munich Leukemia Laboratory's 5,000 Genomes Project, which includes RNA-seq data, was already underway at the time of publication to develop precisely such a classifier. The authors believe that, with further development, blood transcriptomics plus machine learning could represent a fundamental shift in how AML - and potentially other blood diseases - are diagnosed worldwide.