Genes selection using deep learning and explainable artificial intelligence for chronic lymphocytic leukemia...

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Gene expression profiling (GEP) analysis using AI offers insights into cancer.
• Accurate prognostic markers are crucial for chronic lymphocytic leukemia (CLL) patient management.

Purpose of the Study:

• To introduce DeepSHAP Autoencoder Filter for Genes Selection (DSAF-GS), a novel deep learning and explainable AI method for feature selection in GEP data.
• To identify informative genes for CLL prognosis using DSAF-GS on a GEP database.
• To evaluate the predictive performance of DSAF-GS and selected genes on patient outcomes.

Main Methods:

• Developed DSAF-GS, a deep learning approach leveraging autoencoders and explainable AI for feature selection in GEP data.
• Applied DSAF-GS to a GEP database of 217 CLL cases to identify influential genes.
• Validated the prognostic significance of selected genes using univariate and multivariable analyses, assessing impact on time to first treatment (TTFT).

Main Results:

• The DSAF-GS model achieved 86.4% accuracy, 85.0% sensitivity, and 87.5% specificity for CLL prognosis.
• CEACAM19, PIGP, MKL1, and GNE were identified as highly influential genes.
• IGF1R, COL28A1, and QTRT1 were significantly associated with TTFT in multivariable analysis, enhancing the prognostic model's predictive power (Harrell's c-index increased to 78.6%).

Conclusions:

• DSAF-GS is an effective tool for feature selection in GEP data, identifying significant genes for CLL prognosis.
• The identified genes, particularly IGF1R, COL28A1, and QTRT1, hold potential as novel biomarkers for predicting CLL progression and treatment needs.
• This study highlights the utility of explainable AI in uncovering biologically relevant markers for cancer prognostication.