PDAC-ANN: an artificial neural network to predict pancreatic ductal adenocarcinoma based on gene expression

Abstract

Background: Although the pancreatic ductal adenocarcinoma (PDAC) presents high mortality and metastatic potential, there is a lack of effective therapies and a low survival rate for this disease. This PDAC scenario urges new strategies for diagnosis, drug targets, and treatment.

Methods: We performed a gene expression microarray meta-analysis of the tumor against normal tissues in order to identify differentially expressed genes (DEG) shared among all datasets, named core-genes (CG). We confirmed the CG protein expression in pancreatic tissue through The Human Protein Atlas. It was selected five genes with the highest area under the curve (AUC) among these proteins with expression confirmed in the tumor group to train an artificial neural network (ANN) to classify samples.

Results: This microarray included 461 tumor and 187 normal samples. We identified a CG composed of 40 genes, 39 upregulated, and one downregulated. The upregulated CG included proteins and extracellular matrix receptors linked to actin cytoskeleton reorganization. With the Human Protein Atlas, we verified that fourteen genes of the CG are translated, with high or medium expression in most of the pancreatic tumor samples. To train our ANN, we selected the best genes (AHNAK2, KRT19, LAMB3, LAMC2, and S100P) to classify the samples based on AUC using mRNA expression. The network classified tumor samples with an f1-score of 0.83 for the normal samples and 0.88 for the PDAC samples, with an average of 0.86. The PDAC-ANN could classify the test samples with a sensitivity of 87.6 and specificity of 83.1.

Conclusion: The gene expression meta-analysis and confirmation of the protein expression allow us to select five genes highly expressed PDAC samples. We could build a python script to classify the samples based on RNA expression. This software can be useful in the PDAC diagnosis.

Keywords: Artificial neural network; Meta-analysis; Pancreatic ductal adenocarcinoma.

Fig. 1

Artificial neural network architecture. A graphical representation of a fully connected artificial intelligence algorithm (PDAC-ANN). PDAC-ANN is a set of mathematical equations; in each layer, it transforms expression values up to the last layer. The expression values from AHNAK2, KRT19, LAMB2, LAMC2, and S100P genes are data inserted in the input layer (green neurons), the hidden layers (blue neurons) process the expression values, and the output layer (red neurons) give the classification in normal or PDAC sample as a probability

Fig. 2

Variation in protein expression data from the GC list retrieved from immunohistochemical staining images in HPA. The protein expression data shows that 14 genes have more than 75% of images with high plus medium expression in pancreatic cancer, evidencing the expression of predicted core-genes in the pancreatic tissue. The genes with protein expression confirmed in IHC staining images were highlighted in red. Data credit: Human Protein Atlas

Fig. 3

Representative immunohistochemistry staining of AHNAK2, KRT19, LAMB2, LAMC2, and S100P in Pancreatic Ductal Adenocarcinoma (Tumor) and normal pancreatic tissue (Normal). The proteins presented more than 75% of images with high plus medium expression in HPA. Scales bars represent 400 μm. Image courtesy of Human Protein Atlas

Fig. 4

PCA and hierarchical analysis of the merged data set into one data. a. PCA analysis clearly showed two distinct groups corresponding to normal and tumor samples. b. Clustering analysis. The red band indicates the PDAC samples with similar gene expression on 40-core-gene, and the blue band indicates the normal samples