Leveraging Deep Learning to Construct a Programmed Cell Death-Driven Prognostic Signature in Acute Myeloid Leukemia.

Area of Science:

• Hematology
• Oncology
• Computational Biology

Background:

• Acute myeloid leukemia (AML) presents significant clinical challenges due to molecular heterogeneity and high relapse rates.
• Programmed cell death (PCD) pathways are crucial in leukemogenesis and treatment response, but a unified prognostic framework is lacking.
• Existing prognostic models do not fully integrate the complex interplay of diverse PCD modalities in AML.

Purpose of the Study:

• To develop a comprehensive prognostic framework for AML by integrating multi-modal PCD pathways.
• To identify novel PCD-related biomarkers for AML subtyping and risk stratification.
• To create a clinically applicable tool for predicting patient outcomes and guiding personalized treatment strategies.

Main Methods:

• Systematic transcriptomic analysis of 1624 genes across 13 PCD forms.
• Application of a novel computational pipeline using variational autoencoder (VAE) and multilayer perceptron (MLP) for biomarker discovery.
• SHapley Additive exPlanations (SHAP) for interpreting biomarker significance and unsupervised consensus clustering for molecular subtyping.

Main Results:

• Identification of 48 candidate genes distinguishing AML from normal bone marrow.
• Delineation of two molecular subtypes with distinct clinical outcomes and immune microenvironment profiles based on PCD genes.
• Development and validation of an 8-gene prognostic signature (SORL1, PIK3R5, RIPK3, ELANE, GPX1, VNN1, CD74, IL3RA) and a prognostic nomogram.

Conclusions:

• An integrative model connecting multi-modal PCD pathways to AML prognosis has been established.
• A novel molecular subtyping system for AML based on PCD pathways offers new insights into disease heterogeneity.
• The developed prognostic signature and nomogram provide a clinically applicable tool for improved risk assessment and personalized treatment in AML.