Could AI Benefit Access to Testing for Childhood Leukemia in Resource-Limited Settings?
Identifying the molecular subtype of childhood B-cell acute lymphoblastic leukemia (ALL) is important for risk assessment and treatment planning, but comprehensive molecular testing can be costly and difficult to access, particularly in resource-limited settings.
Researchers in Mexico evaluated a machine-learning approach that uses targeted RNA sequencing to classify B-cell ALL into molecular subtypes—potentially offering a more accessible alternative to approaches that rely on whole-transcriptome sequencing.
Joaquín García-Solorio, PhD, of the Instituto Nacional de Medicina Genómica in Mexico City, presented findings in a poster presented at the Association for Molecular Pathology (AMP) Annual Meeting & Expo in Seattle (Abstract H011).1
Study Background and Methods
According to the researchers, Hispanic/Latino children have a particularly high incidence of B-cell ALL, including high-risk subtypes such as Philadelphia chromosome–like (Ph-like) ALL. However, molecular classification remains largely inaccessible in many parts of Latin America because of cost and laboratory infrastructure limitations. Existing classification methods also frequently rely on whole-transcriptome sequencing datasets derived from predominantly European-ancestry populations, potentially limiting their applicability to underrepresented populations and targeted sequencing platforms.
To address these challenges, the investigators developed a machine-learning classifier combining four computational approaches trained on gene-expression data from 1,664 samples available through St. Jude Cloud. The classifier was subsequently evaluated in an independent cohort of 170 Mexican pediatric patients using a targeted RNA sequencing panel.
The analysis focused on 1,385 genes shared between the reference dataset and the targeted panel and evaluated classification across 12 molecular subtype categories. The investigators compared the classifier's predictions with molecular subtype assignments established through fusion-detection analysis and additional molecular or cytogenetic testing.
Key Findings
The combined classifier achieved an overall accuracy of 82.3% and a Cohen's kappa of 0.80, outperforming the individual computational approaches, which achieved accuracies ranging from 66.1% to 79.6%.
Performance varied by molecular subtype. The classifier demonstrated particularly strong results for TCF3::PBX1 (F1 score = 0.98), ETV6::RUNX1/ETV6::RUNX1-like (0.94), KMT2A-rearranged (0.88), and BCR::ABL1/Ph-like (0.86) ALL. However, classification was less reliable for several other subtypes, including hypodiploidy (0.20), iAMP21 (0.57), and PAX5-altered ALL (0.53).
Notably, the researchers reported that gene-expression signatures could potentially identify Ph-like ALL or cases with cryptic genetic drivers even when the underlying abnormality was not detected through fusion analysis. This finding suggests that the approach may help identify patients whose disease warrants additional molecular investigation.
Clinical Implications
The findings suggest that machine-learning applied to targeted RNA-sequencing data could provide a potentially less costly approach to molecular classification of pediatric B-cell ALL, particularly in settings where comprehensive genomic testing is not routinely available.
However, the investigators acknowledged limitations of the study, including the small number of patients with certain rare molecular subtypes and the absence of direct comparisons with existing classification tools. Additional validation will be needed to establish the approach's reliability and clinical utility across a broader range of patients and molecular subtypes.
DISCLOSURES
Dr. García-Solorio reported no conflicts of interest.
REFERENCES
García-Solorio J, Flores Espino F, Campos Perez I, et al: An ensemble machine learning classifier for B-ALL molecular subtyping via targeted RNA-seq: A cost-effective approach for resource-limited settings.” Abstract H011. Presented at the Association for Molecular Pathology. November 14, 2026.