clinTALL: machine learning-driven multimodal subtype classification and treatment outcome prediction in pediatric...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning in Oncology

Background:

• Childhood T-lineage acute lymphoblastic leukemia (T-ALL) is aggressive with poor prognosis.
• Current risk stratification for T-ALL is insufficient compared to B-cell precursor ALL.
• Interpreting complex multi-omics data for clinical use in T-ALL is challenging.

Purpose of the Study:

• To develop a deep learning pipeline (clinTALL) for pediatric T-ALL subtype classification and treatment outcome prediction.
• To integrate multimodal data (clinical, genomic, transcriptomic) for enhanced precision.
• To provide a user-friendly tool for clinical application.

Main Methods:

• Developed clinTALL, a deep learning multi-task pipeline.
• Utilized a neural network architecture for shared latent embedding and multi-task prediction.
• Employed a competing risk model for event-specific outcome prediction.
• Trained on a public dataset of 1309 pediatric T-ALL samples.

Main Results:

• Transcriptomic-only model achieved 92.2% accuracy for subtype prediction and 65.9% C-index for event-free survival (EFS).
• Integrating all data modalities improved EFS C-index to 67.5% while maintaining high subtype accuracy (91.7%).
• Accurate prediction of induction failure (96.0% C-index) and second malignant neoplasm (62.1% C-index).
• Validated on an internal dataset with 81.8% accuracy for subtype prediction.

Conclusions:

• The clinTALL framework enables automated, accurate T-ALL subtype classification and outcome inference.
• Multimodal data integration advances precision risk stratification for pediatric T-ALL.
• clinTALL is available as a Docker application for broad accessibility.