SurvIAE: Survival prediction with Interpretable Autoencoders from Diffuse Large B-Cells Lymphoma gene expression data

Area of Science:

• Bioinformatics
• Computational Biology
• Oncology

Background:

• Novel biomarkers are needed for risk assessment in Diffuse Large B-Cell Lymphoma (DLBCL).
• Current risk stratification methods require refinement.
• Autoencoders (AE) and Explainable Artificial Intelligence (XAI) offer potential for biomarker discovery.

Purpose of the Study:

• To develop and validate a computational pipeline (SurvIAE) for prognostic risk stratification in DLBCL.
• To derive a gene-based signature using AE and XAI for improved DLBCL risk assessment.
• To compare the performance of SurvIAE against the Revised International Prognostic Index (R-IPI).

Main Methods:

• Gene expression data from three public DLBCL datasets were processed using AE for unsupervised representation learning.
• A Multi-layer Perceptron (MLP) model classified prognosis based on the latent representation.
• XAI techniques identified a gene signature from the best-performing AE-MLP model (SurvIAE-Small-PFS36).
• The derived signature was validated on an independent DLBCL dataset.

Main Results:

• AE models effectively reduced batch effects in gene expression data.
• SurvIAE significantly outperformed R-IPI in predicting progression-free survival (PFS36 and PFS60).
• A three-risk group stratification was established based on GAB1 and GPR132 expression levels, refining R-IPI categories.

Conclusions:

• SurvIAE demonstrates potential for deriving a clinically relevant gene signature in DLBCL.
• The developed pipeline offers a reusable framework for prognostic biomarker discovery in other diseases.
• This approach enhances DLBCL risk stratification and has translational implications.