Deep learning based individualized cross-platform molecular subtype classification of B-lineage acute lymphoblastic leukemia.

Molecular subtypes of B-cell acute lymphoblastic leukemia (B-ALL) are essential in modern clinical treatment. However, the fast emerging subtypes and the requirement of complex multi-omic diagnostics are continuously challenging the clinical subtypes classification in sensitivity, cost, and turnaround time. We develop B-cell Acute Lymphoblastic Leukemia Subtype Identification based on gene eXpression (BALL6), a robust deep learning framework for cross-platform B-ALL subtyping. BALL6 utilizes a recurrent neural network trained on rank-transformed expression values of feature genes, capturing subtype signals while inherently minimizing technical noise. We implement in BALL6 the rank-based augmentation framework which further enhances its performance on data-limited or imbalanced datasets. BALL6 includes two integrated models: an AL model distinguishing B-ALL, T-ALL, and AML, and a B-ALL model identifying the most updated 20 established molecular subtypes. BALL6 demonstrates robust accuracy across multiple independent datasets, achieving 99.38% (AL model) and 93.84% (B-ALL model) accuracy on previously unseen data. Notably, BALL6 is robust to missing values, which enables its cross-platform application. This is demonstrated through reliable subtype prediction by BALL6 with sparse gene expression profiles from scRNA-seq data. BALL6 is open-sourced with an accessible web tool (https://cccg.ronglian.com/#/analysis), facilitating its broad applications in leukemia research and clinical diagnostics.