Fig. 1: UBTF::ATXN7L3 gene fusion defines a BCP-ALL molecular subtype.

Bone marrow or peripheral blood samples of n = 568 adult patients at first diagnosis of BCP-ALL with at least 20% blast infiltration were analyzed by transcriptome sequencing. Gene expression profiles of previously established definitions [1] were used to train a machine learning algorithm (Extreme Gradient Boosting) with feature selection (LASSO) to allocate samples to 15 established and 2 novel BCP-ALL subtypes. A Uniform manifold approximation and projection (UMAP) plot depicts distribution of molecular ALL subtypes based on expression of the top informative genes for subgroup allocation. B Unsupervised clustering of gene expression specific for the novel UBTF::ATXN7L3 subgroup. Subtype-specific gene sets were identified by multi-comparison ANOVA using the variance stabilizing transformation for normalizing expression values. Genes had to be differentially expressed to at least 11 other subtypes with an FDR-corrected p value ≤ 0.01 (Supplementary Table S1,2). C Expression of selected UBTF::ATXN7L3-specific oncogenes is shown in comparison to other molecular subtypes with at least three samples. The complete list of Cosmic cancer gene census genes differentially expressed by multi-comparison ANOVA in UBTF::ATXN7L3 ALL is shown in Supplementary Table S2. Expression of cancer-associated genes upregulated in UBTF::ATXN7L3 in comparison to other molecular subtypes is shown in Supplementary Figure S6. D Subtype-specific driver fusions and the novel candidate driver gene fusion UBTF::ATXN7L3 were called from RNA-Seq data (FusionCatcher) [15] Virtual karyotypes were obtained from whole-exome sequencing or SNP-Arrays (Illumina Infinium Global Screen array v3.0) to support classification of aneuploid subtypes. Subtype-specific hotspot driver mutations (eg. PAX5) were called from RNA-Seq data. Heatmap depicts probabilities for allocation of samples to the molecular subtypes (class probabilities) obtained by the gene expression-based machine learning classifier together with corresponding genomic driver events. * indicates one UBTF::ATXN7L3 sample with 20% blast content, where the driver gene fusion was not called by FusionCatcher but was confirmed by break-point specific PCR and sanger sequencing (Supplementary Fig. S7B).