Related Experiment Video
Updated: Oct 9, 2026
Selecting Multiple Biomarker Subsets with Similarly Effective Binary Classification Performances
Published on: October 11, 2018
Clinical Evaluation of the Utility of Machine Learning-Based Subtype Classifiers of Acute Lymphoblastic Leukemia in
Seoyoung Lim1, Yu Jeong Choi2, Yu Jin Park3
1Department of Laboratory Medicine, Graduate School of Medical Sciences, Brain Korea 21 PLUS Project, Yonsei University College of Medicine, Seoul, Korea.
Background:
B-acute lymphoblastic leukemia (B-ALL) is a heterogeneous disease comprising more than 20 molecular subtypes defined by distinct gene expression profiles (GEPs) and genetic alterations. Subtypes primarily defined by GEPs remain challenging to identify using conventional cytogenetic and targeted molecular assays. Although GEP-based classifiers can predict B-ALL subtypes, their clinical performance has not been validated using independent cohorts, and reported accuracies may be inflated due to overlapping training and test datasets. Therefore, we evaluated the clinical performance of four open-source GEP-based classifiers using an independent cohort.
Methods:
We retrospectively analyzed RNA-sequencing data from 92 patients with B-ALL to evaluate the performance of four open-source GEP-based classifiers (ALLCatchR, ALLSorts, ALLIUM GEX, and Allspice). To ensure fair comparisons across models with heterogeneous subtype definitions and output formats, prediction results were unified using a multilabel encoding framework, followed by label-based performance evaluation.
Results:
Among the 92 patients, 71 (77.2%) were assigned to a single definitive molecular subtype, whereas 11 (12.0%) exhibited co-occurring subtype-defining features. Overall, the evaluated classifiers achieved high accuracy (>97.0%), although performance varied by model and subtype; the highest accuracy reached 99.2% in single-subtype cases and 98.5% when mixed cases were included. Integration of GEP-based classifiers increased the overall subtype assignment rate from 83.7% to 89.1% by enabling the identification of cryptic subtypes. In two cases previously classified as not otherwise specified, classifier-guided analyses revealed TCF3::IGK::DBP and PVT1::IGH fusions, neither of which has been previously reported in B-ALL.
Conclusions:
GEP-based classifiers provide valuable support for B-ALL subtype classification and show potential as adjunct tools in routine diagnostics.
