Pancreatic ductal adenocarcinoma (PDAC) is the most common form of pancreatic cancer and is characterized by late diagnosis and poor prognosis. By identifying transcriptomic and metabolic features we wanted to further our understanding of the altered metabolism in PDAC and identify potential candidates for further research into diagnostic tests and therapeutic targets.
In our analysis we included tumour and normal samples from 52 patients. Euclidean distance was used to assess changes in metabolic subsystems defined in the Human1 model, while potential features were identified using DESeq2 analysis, the area under the ROC curve (AUC), and the paired Wilcoxon test. LASSO logistic regression was used to determine combinations of potential features and to assess their ability to correctly classify normal and tumour samples.
With DESeq2 analysis we discovered 6897 genes with statistically significant differential expression. TPM analysis provided us 402 genes with a high ability to discriminate between normal and tumour samples (AUC>0,8), while flux analysis identified only eight reactions with AUC>0,8. Integration of the results from all three analyses identified four reactions associated with five genes as the most promising candidate features. These candidates were associated with transport of retinol and acetate, steroid metabolism and reduction of hydrogen peroxide. With logistic regression LASSO we assessed accuracy and AUC of gene-based, reaction-based and combined models. The gene-based model, containing three transcriptomic features, achieved the highest values for both measures, with a median AUC value of 0,99 and a median accuracy of 0,941.
The results of this thesis suggest that integrating transcriptomic data with metabolic fluxes can facilitate the identification of PDAC-associated features. The combined LASSO model showed lower predictive performance and stability than the gene-based and reaction-based models.
|
