B-cell precursor acute lymphoblastic leukemia (BCP-ALL) is one of the most common blood cancers in children and also affects adults. It is not a single disease but a collection of up to 26 molecularly distinct subtypes, each with different genetic drivers, clinical behavior, and optimal treatments.
Accurate molecular subtype identification is essential because risk stratification and targeted therapy selection depend on it. For example, patients with Philadelphia chromosome-positive ALL may benefit from tyrosine kinase inhibitors, while those with KMT2A rearrangements require different treatment approaches.
Currently, subtype identification requires combining multiple diagnostic methods - chromosome analysis, fluorescence in situ hybridization (FISH), PCR-based tests, and SNP arrays - a complex, expensive, and time-consuming process. RNA sequencing (RNA-Seq) can theoretically identify all subtypes with a single test, but robust computational tools for systematic gene expression analysis were lacking.
The researchers developed ALLCatchR, a machine learning classifier that uses RNA-Seq gene expression data alone to classify BCP-ALL samples into all 21 gene expression-defined molecular subtypes recognized by current international classification systems.
ALLCatchR was developed using a dataset of 3,532 BCP-ALL transcriptome profiles assembled from six independent international cohorts spanning both pediatric (64.5%) and adult (35.5%) patients. This breadth was critical for training a classifier that would generalize across different age groups and clinical settings.
The classifier integrates two complementary approaches: a linear support vector machine (SVM) trained to make deterministic subtype predictions based on 2,802 highly discriminative genes selected by LASSO regression, and a nearest-neighbor model using subtype-specific gene set enrichment scores. Combining these two methods overcame the limitations of either approach alone.
The data was split into a training set of 1,869 samples and three independent hold-out cohorts (1,018 samples with confirmed subtypes plus 111 previously unassigned samples) not seen during training. Validation in entirely new cohorts - rather than just a random split of the same dataset - is the gold standard for testing real-world applicability.
Raw gene expression count data from 15,728 protein-coding genes were normalized and scaled to remove batch effects between different sequencing platforms and laboratories. This preprocessing was essential for achieving consistent performance across the diverse technical conditions of real-world diagnostic laboratories.
In the three independent hold-out validation cohorts, ALLCatchR successfully allocated 98.8% of samples to one of 21 subtypes, with an overall accuracy of 95.7%. High-confidence predictions - where the classifier's confidence exceeded a pre-defined threshold - were achieved in 83.7% of samples with 98.9% accuracy.
Specificity exceeded 99% for all 21 subtypes, meaning ALLCatchR almost never misidentified a sample as belonging to a subtype it did not have. Average sensitivity across subtypes was 91.1%, though a few rare subtypes with very few samples (such as NUTM1, CEBP, iAMP21, and near haploid) achieved lower sensitivities due to their scarcity in training data.
Performance was consistent across pediatric and adult patients, across different sequencing platforms, and across samples from clinical trials and routine diagnostic laboratories. The real-world diagnostic lab cohort (MLL) achieved 91.4% accuracy despite less stringent patient preselection criteria - confirming ALLCatchR's robustness outside of controlled research settings.
When compared to three other gene expression classifiers (ALLSorts, ALLIUM, and ALLspice), ALLCatchR outperformed all competitors in accuracy, number of unclassified samples, and coverage of rare subtypes. Other tools covered only 14-19 of the 21 subtypes, while ALLCatchR classified all 21.
A particularly valuable feature of ALLCatchR is its ability to classify samples that previously could not be assigned to any known subtype. Among 111 samples defined as 'unassigned' or 'B-other' in the original studies, ALLCatchR provided high-confidence or candidate predictions for 81.9%, with only 18% remaining unclassified.
In 72.1% of high-confidence predictions for previously unassigned samples, corresponding genomic driver aberrations were confirmed by RNA-Seq fusion calling, cytogenetic analysis, or whole genome sequencing. This validation demonstrates that ALLCatchR was identifying real biological subtypes, not generating spurious classifications.
Among the newly classified cases, the most common previously unrecognized subtype was PAX5alt - a heterogeneous group defined by diverse PAX5 gene alterations. Several cases also showed CRLF2 rearrangements consistent with Ph-like ALL. Importantly, one case classified as KMT2A-rearranged was confirmed to harbor a KMT2A partial tandem duplication - a finding more commonly associated with AML that had not previously been described in BCP-ALL.
These findings illustrate how a comprehensive gene expression classifier can not only accurately categorize known subtypes but also reveal biological insights in ambiguous cases, potentially identifying novel disease entities and improving diagnostic completeness in routine clinical practice.
Beyond subtype classification, ALLCatchR includes additional modules that predict clinical baseline characteristics directly from the gene expression data. A blast count prediction model estimates the proportion of leukemic cells in the sample, which affects classification accuracy and is needed for interpreting results from samples with lower tumor infiltration.
A patient sex prediction model achieved 99.1% accuracy from gene expression alone - an impressive demonstration that RNA-Seq captures comprehensive biological information beyond just cancer subtype. This allows cross-validation of clinical records and imputation of missing values.
An immunophenotype prediction module classified samples as pro-B vs. common/pre-B ALL with 87.1% accuracy. Immunophenotype information guides treatment decisions and is sometimes unavailable in research datasets or ambiguous in clinical settings, making this imputation capability practically useful.
Together, these additional modules mean that ALLCatchR can provide a comprehensive clinical and biological profile of a BCP-ALL sample from a single RNA-Seq test - subtype, blast count, patient sex, and immunophenotype - potentially streamlining the diagnostic workup for leukemia patients.
A novel scientific contribution of this study was the establishment of a reference map of normal human B-cell development using RNA-Seq data from seven FACS-sorted progenitor stages from healthy bone marrow donors. This provided a framework for asking where each leukemia subtype sits on the normal developmental trajectory.
Using this reference, ALLCatchR can project BCP-ALL samples onto the B-lymphopoiesis trajectory to identify which normal developmental stage each leukemia most resembles. Most BCP-ALL subtypes showed highest similarity to the pre-B I differentiation stage, but striking exceptions emerged.
KMT2A-rearranged and PAX5 P80R ALL showed patterns resembling an earlier pro-B differentiation stage, suggesting these leukemias arise from or become arrested at a more primitive developmental point. In contrast, BCL2/MYC ALL showed exclusive similarity to the mature pre-B II large stage.
These developmental insights have clinical relevance because current immunotherapy targets (such as CD19, CD22, and CD10) are expressed at different stages of B-cell development. Understanding which developmental stage each leukemia subtype resembles could inform target selection and predict which patients will respond to specific immunotherapy agents.
ALLCatchR is freely available as an R package, designed for integration into routine diagnostic RNA-Seq pipelines. Its ability to classify all 21 WHO/ICC-defined BCP-ALL subtypes with high accuracy from a single test could significantly reduce the time, cost, and complexity of current diagnostic workups that combine multiple separate tests.
The practical impact is particularly significant for rare subtypes and for adult ALL, where subtype identification is often incomplete due to technical limitations of older diagnostic methods. Comprehensive subtype assignment could open access to subtype-specific targeted therapies for patients who would otherwise receive non-optimal standard treatment.
The tool's performance in real-world diagnostic laboratory settings (as opposed to highly curated research cohorts) is an important validation step. The MLL cohort results - achieved with less stringent patient selection criteria - suggest ALLCatchR can maintain diagnostic accuracy in the messier, more variable conditions of routine clinical practice.
As RNA-Seq becomes increasingly standard in hematology diagnostic laboratories, tools like ALLCatchR provide the computational infrastructure to fully exploit the information it contains. Combined with fusion detection and other RNA-Seq-based analyses, ALLCatchR brings comprehensive molecular characterization of leukemia closer to becoming a routine diagnostic reality.