What is leukemia? Leukemia is a blood cancer in which the body produces abnormal white blood cells that crowd out healthy blood cells in the bone marrow. There are four main types - acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), chronic myeloid leukemia (CML), and chronic lymphocytic leukemia (CLL) - each with different behavior, prognosis, and treatment requirements.
Why accurate classification matters: Getting the leukemia subtype correct is critical for treatment decisions. AML, for example, is most common in adults over 70 and requires urgent treatment due to its rapid progression. ALL is the most common leukemia in children and has achieved cure rates of 80 to 90% in pediatric cases with intensive modern therapy. CML is defined by a specific chromosomal abnormality (the BCR-ABL fusion oncogene), while CLL is primarily a disease of older adults with a slower course.
The gene expression approach: Each type of leukemia has a characteristic pattern of gene activity - which genes are turned up or down compared to healthy cells. By measuring gene expression across many samples, researchers can identify genetic signatures that reliably distinguish one leukemia subtype from another and from normal bone marrow.
The challenge of fragmented data: A major problem in this field is that biological data is spread across many studies using different platforms - some using Microarray technology, others using RNA sequencing (RNA-seq). Until now, most studies addressed only two classes at a time (such as comparing one leukemia type against healthy tissue), limiting the broader picture. This paper attempts a five-class analysis combining all four leukemia types and healthy subjects simultaneously.
Microarray technology: Microarrays were the dominant gene expression measurement technology for two decades. They work by attaching DNA spots to a surface, allowing RNA from a sample to bind to its matching sequence, and then reading expression levels with a laser. While RNA-seq has largely surpassed microarrays in capability, there is an enormous archive of existing microarray data that still holds value.
RNA-seq technology: RNA sequencing directly reads the RNA molecules present in a cell and counts how many copies of each gene's message exist. It offers higher sensitivity, a broader measurement range, and can detect single-nucleotide variations. However, it is more expensive and there are fewer publicly available leukemia RNA-seq datasets compared to microarray datasets.
Why integrate both? The study collected 13 publicly available datasets - 11 Microarray and 2 RNA-seq - totaling 658 samples after quality control. This included samples from AML (352 samples), ALL (197), CML (53), CLL (29), and healthy bone marrow (26), from laboratories in Spain, the USA, Germany, Italy, South Korea, Norway, Belgium, and Canada. Integrating both technologies multiplied the available data, especially for AML where RNA-seq samples existed.
The normalization challenge: Combining data from different platforms is technically difficult because each technology and manufacturer introduces its own systematic biases. The study applied careful normalization steps - including quantile normalization and batch correction - to make samples from different sources comparable before any analysis was done.
The standard approach and its limitation: The most commonly used method for identifying differentially expressed genes (DEGs) is called limma. However, limma was designed for comparing two groups (such as cancer vs. healthy). When applied to a five-class problem like this one, it only compares one pair at a time and misses whether a gene is useful for distinguishing multiple leukemia types simultaneously.
Introducing COVERAGE (COV): To solve this, the researchers introduced a new parameter called COVERAGE (COV). COV measures how many out of all possible pairs of classes a given gene can distinguish. With five classes, there are ten possible pairs. A gene with a COV of 5 can tell apart at least 5 of those 10 pairs, making it far more useful for multiclass classification than a gene that only distinguishes one pair.
Combining COV with Log-Fold Change: Genes were selected as candidate biomarkers only if they showed a large enough difference in expression (log-fold change of 2.5 or more) AND had a COV of at least 5, along with a very stringent statistical threshold (p-value of 0.001 or less). This combination filtered from thousands of genes down to just 42 candidate DEGs. All 42 DEGs achieved COV values of 6 or 7 - meaning each gene could distinguish six or seven out of ten possible leukemia type pairings.
ANOVA to optimize parameters: Before settling on these thresholds, the researchers ran an extensive statistical analysis (ANOVA test) across hundreds of combinations of parameters - varying the log-fold change threshold, COV value, number of genes, and choice of classifier. This rigorous testing confirmed that COV was the single most important factor for achieving high classification accuracy, and that a COV of 5 with an LFC of 2.5 was the optimal combination.
Ranking genes by relevance: From the 42 identified DEGs, the researchers applied a feature selection algorithm called minimum-Redundancy Maximum-Relevance (mRMR) to rank them. This algorithm prioritizes genes that carry the most information about the leukemia class while minimizing overlap with each other - ensuring that each additional gene in the signature adds new, non-redundant diagnostic value.
Four classifiers tested: Four machine learning classification methods were applied: Support Vector Machines (SVM), k-Nearest Neighbor (k-NN), Naive Bayes (NB), and Random Forest (RF). Each was tested on subsets of 10, 20, 30, and 40 genes from the mRMR ranking. A 5-fold cross-validation approach was used throughout to ensure results would generalize to new, unseen samples.
k-NN as the top performer: k-NN - a method that classifies a sample based on the class of the most similar examples in the training data - achieved the highest accuracy across all gene subset sizes. Using just the top 10 genes, k-NN reached 96.40% accuracy and a 96.28% f1-score across all five classes. Performance improved further with 20 genes (98.56% accuracy, 98.75% f1-score) and continued incrementally with 30 and 40 genes.
Why these numbers are remarkable: Achieving over 96% accuracy in distinguishing five classes - four leukemia subtypes plus healthy controls - using only 10 genes from a dataset spanning 13 studies across multiple countries and two different technologies represents a significant result. It demonstrates that the COV-based gene selection method identifies genuinely informative biological signals rather than platform-specific noise.
The ten genes identified: The top 10 DEGs ranked by mRMR were BLK, DOCK2, LAPTM4B, EEF1A1, RPS15, RPS24, AZU1, PABPC1, C11orf58, and BLNK. Nine of the ten have established connections to leukemia or to the blood cell types that leukemia affects.
B-cell genes at the top: Several of the top genes are normally expressed in B lymphocytes - the immune cells affected by many leukemia subtypes. BLK is a proto-oncogene involved in B-cell proliferation and acts as a tumor suppressor in CML stem cells. BLNK is critical for B-cell development and its loss has been linked to pre-B ALL. DOCK2 promotes CLL cell proliferation through Wnt5a signaling and has been proposed as a drug target specific to blood cancers.
T-cell and neutrophil genes: LAPTM4B (expressed in T-cells) and EEF1A1 contribute to leukemogenesis through the PI3K/AKT signaling pathway and promotion of leukemia cell survival, respectively. The neutrophil gene AZU1 (Azurocidin 1) has been used to predict whether CML patients will respond to tyrosine kinase inhibitor therapy, and its downregulation correlates with poor outcomes in AML.
Ribosomal proteins as cancer markers: RPS15 and RPS24 encode components of the ribosome - the cell's protein-making machinery. RPS15 is found mutated in CLL patients, where it impairs a key tumor-suppressor protein called p53. RPS24 is mutated in Diamond-Blackfan anemia, a condition associated with elevated leukemia risk. The appearance of ribosomal protein genes in the leukemia signature reflects the known role of ribosome dysfunction in blood cancers.
A practical diagnostic tool: The results suggest that a gene expression panel of as few as 10 carefully chosen genes could, in principle, help clinicians distinguish between the four main types of leukemia and healthy bone marrow with high confidence. This kind of compact genetic signature could be developed into a diagnostic test, potentially supplementing or validating current laboratory methods.
The value of technology integration: The study demonstrates that combining Microarray and RNA-seq data - despite their technical differences - is both feasible and beneficial. For leukemia subtypes where very few RNA-seq datasets exist publicly (such as CML and CLL), being able to leverage the larger archives of Microarray data dramatically increases the number of samples available and strengthens the statistical reliability of identified biomarkers.
Multiclass analysis is the key advance: Most prior gene expression studies in leukemia are binary - they compare one subtype against healthy controls, or against one other subtype. This five-class framework reveals genes that are simultaneously informative across multiple leukemia types, which is closer to the real clinical scenario where a patient's diagnosis is unknown at the time of testing.
COV as a reusable framework: The COVERAGE parameter is not specific to leukemia and could be applied to any multiclass genomic classification problem - for example, distinguishing subtypes of breast cancer, lung cancer, or other diseases with multiple distinct molecular forms. The ANOVA-driven optimization of COV and LFC thresholds provides a principled method for adapting the approach to new datasets.
Main contributions: This study delivered three main advances: a novel multiclass gene selection parameter (COVERAGE), a validated pipeline for integrating heterogeneous genomic datasets across multiple platforms, and a compact 10-gene leukemia classifier that achieves over 96% accuracy across five classes using publicly available data.
42 DEGs with broad leukemia relevance: All 42 differentially expressed genes identified passed stringent statistical thresholds and were capable of distinguishing six or seven out of ten possible leukemia type pairings. This broad discriminatory power suggests these genes capture fundamental biological differences between leukemia subtypes, not just platform-specific artifacts.
Biological grounding of the results: Nine of the top ten genes selected by the mRMR algorithm have documented roles in leukemia biology or in the specific blood cell lineages that leukemia affects. This biological validation independently confirms that the computational pipeline is identifying real disease-relevant signals.
Next steps: Future work could expand the dataset further - particularly with more RNA-seq samples for CML and CLL subtypes - and validate the gene signature prospectively in clinical samples. The DOCK2 gene in particular, identified here as one of the top discriminators, has already been proposed as a therapeutic target in CLL, suggesting this computational work may have direct translational implications.