PDAC-ANN

PDAC-ANN predicts pancreatic ductal adenocarcinoma status from gene expression profiles using an artificial neural network to classify samples as normal or PDAC.

Key Features:

• Gene selection: A gene expression meta-analysis identified differentially expressed genes (DEGs) including 40 core-genes consistently upregulated in tumor samples (with one exception) linked to actin cytoskeleton reorganization and extracellular matrix receptor interactions.
• Protein validation: Protein expression of identified genes was confirmed using The Human Protein Atlas.
• Selected biomarkers: Five genes—AHNAK2, LAMB3, LAMC2, KRT19, and S100P—were selected based on high area under the curve (AUC) values for model training.
• Deep learning model: An artificial neural network was trained on gene expression data from 461 tumor and 187 normal samples to classify PDAC versus normal tissue.
• Performance metrics: Reported performance includes f1-scores of 0.83 for normal and 0.88 for PDAC (average 0.86), sensitivity of 87.6%, and specificity of 83.1%.
• Implementation: The method is implemented in Python.

Scientific Applications:

• Early detection and diagnosis: Classification of samples as PDAC or normal to support early detection efforts using gene expression data.
• Clinical decision support: Provision of molecular classification information that can inform clinical decision-making and candidate therapeutic target identification.
• PDAC pathogenesis research: Analysis of DEGs and core-genes to investigate biological processes such as cytoskeleton reorganization and ECM interactions in PDAC.
• Biomarker discovery and validation: Identification and protein-level validation of candidate PDAC biomarkers for downstream research.

Methodology:

Performed a gene expression meta-analysis comparing tumor and normal pancreatic tissues to identify DEGs (40 core-genes), validated protein expression using The Human Protein Atlas, selected five genes (AHNAK2, LAMB3, LAMC2, KRT19, S100P) based on high AUC, and trained an artificial neural network on gene expression data from 461 tumor and 187 normal samples.