A Hiplot-Based Web Application for the Rapid Prediction of Adult BCR::ABL1-Positive and BCR::ABL1-like B-Cell Precursor Acute Lymphoblastic Leukemia

Background The BCR::ABL1-positive/-like molecular subtype is characterized by similar gene expression profiles in adults with B-cell precursor acute lymphoblastic leukemia (BCP-ALL). Due to the limitations of gene fusions-based calling, gene expression-based models have emerged as the gold standard approach for BCR::ABL1-like subgroup, particularly for patients lacking gene fusion events. However, existing prediction models based exclusively on adult patients and contemporary web-based applications remain inadequate for the rapid prediction of BCR::ABL1-positive/-like group.

Method We retrospectively analyzed 477 newly diagnosed adult patients with BCP-ALL treated in our institute between May 2019 and April 2024. The BCR::ABL1-like group was predicted using two published prediction models (AllSorts and ALLCatchR). Concurrently, the tSNE, UMAP and heatmap visualization utilizing the 2,000 top variance genes were employed to corroborate the accuracy of the subtyping. In this phase, any potential unstable gene clusters and gender-associated patterns were excluded (n=634). The genomic markers of BCR::ABL1-positive/-like were detected using RNA sequencing (RNA-Seq, n=477), 100-gene targeted exome sequencing (n=469), and whole-genome sequencing (WGS, n=41) data, respectively. The copy number variations (CNVs) were called by using the tumor-only mode of HMFtools' purple (v4.0.2) software and ExomeDepth (v1.1.6) R package. Subsequently, an AutoML-based machine learning model (autogluon) and the HIPLOT-based web plugin were constructed for the prediction of adult BCR::ABL1-positive/-like group.

Result Based on a standard RNA-Seq workflow, the BCR::ABL1 gene fusions validated by RT-qPCR were identified in 211 patients (44%), and 60 patients (13%) were classified as BCR::ABL1-like BCP-ALL. The genomic marker analysis revealed that over 63% of adult BCR::ABL1-like patients exhibited gene fusion events, involving JAK-STAT (28%) [CRLF2-r (13%), JAK2-r (8%), and EPOR-r (7%), etc.] and ABL-class (13%). In the subset of BCR::ABL1-like patients who are negative for gene fusion, kinase mutations could be identified in at least 54% of cases. The CNV analysis of WGS data indicated that the most common copy number variations of BCR::ABL1-positive/-like patients were PAX5, CDKN2A/B, IKZF1, ETV6, BTG1, RB1, ATP10A, EBF1, RUNX1, ERG, RAG2, and HBS1L, which occurred in over 10% of cases at least, respectively. Survival data from our cohort confirmed that the BCR::ABL1-like patients exhibited a trend towards a worse survival compared to BCR::ABL1-positive counterparts. To further simplify the prediction of adult BCR-ABL1-like patients, we developed the AutoML-based prediction model. The internal test of the training process revealed that the value of area under the curve for the prediction model exceeded 0.99, indicating high prediction accuracy. The performance of the prediction model was also successfully validated in two independent BCP-ALL cohorts: the Ph-only cohort (n=31) and the PNAS-2018 cohort (n=258). Moreover, an easy-to-use web application has been developed based on the HIPLOT (ORG) website, which allows users to obtain a BCR::ABL1-positive/-like prediction in one-click way (https://hiplot.cn/clinical-tools).

Conclusion The high-performance prediction model and its web interface of adult BCR::ABL1-positive/-like subtype have been developed. The toolkit has been incorporated into routine practice at our center, and may facilitate the accurate and rapid diagnosis of the adult BCP-ALLs based on RNA-Seq.

No relevant conflicts of interest to declare.