Learning From Human Bile – A Machine Intelligence Approach

Pixabay License | Source: Gerd Altmann , Altered aspect ratio.

The digestive fluid called bile is made up of acids, salts, fats, and the pigment bilirubin, stored and concentrated in the gall bladder, and produced by the liver. Protein is a significant component of bile, 5% of the dry weight. The fluid emulsifies and aids digestion of fats in the small intestine. It is released from the gall bladder, which may contain up to 60 ml, upon secretion of the hormone pancreozymin from inclusion cells of the duodenum in response to fatty food. As bile is a complex fluid and is influenced by numerous cell types it is possible that bile composition could change in response to digestive system disorders and serve as a biomarker of disease.

Matías A. Avila, Maite G. Fernández-Barrena, Carmen Berasain, Fernando J. Corrales of Carlos III Health Institute, Madrid, Spain, and colleagues, performed parallel metabolomic and proteomic analyses of human bile from patients with benign and malignant cholangiocarcinoma (CCA) and pancreatic ductal adenocarcinoma (PDAC) biliary stenoses. Machine learning incorporating synthetic data was used to identify molecular patterns of malignant stenoses versus benign biliary strictures. The results were published in the journal cancers.

A profile of 162 metabolites was detected. The lipid profiles of serum and bile were similar, although there were clear differences in ceramide content. Compared to normal bile there was a reduction of phosphatidylcholines in bile from patients with cancer. Lysophosphatidylcholines (LPC), monoacylglycerols (MG), triacylglycerols (TG), and fatty acid amines (FAA) showed a trend towards reduced levels in cancer. Sphingomyelins (SM) and Ceramides (Cer), but not diacylglycerols (DG) were also reduced.

Total bile acids were significantly reduced in cancer patients.  Specifically, glycine-conjugated bile acids were reduced in cancer samples, but taurine-conjugated bile acid concentrations did not change significantly.


By nuclear magnetic resonance (NMR) analysis high glucose levels were detected in bile, specifically from PDAC patients. Other metabolite changes were also detected in the bile of PDAC and CCA patients by NMR.

Principal component analysis (PCA) of the lipid profile of bile could not differentiate patients with cancer from healthy controls. Interestingly a machine learning model trained with a dataset supplemented with “synthetic data” was able to classify cancer and non-cancer patients based on their bile lipid profile. This same machine learning approach could also accurately discriminate between cancer and cancer-free patients based on a proteomic dataset of limited size (n=20).

“The second aim of this work was to select molecular features (metabolites and proteins) identified in bile that could be applied for the discrimination between patients with benign and malignant strictures. However, on one hand, clinical samples tend to show high complexity and variability in their molecular composition even among same groups of patients, and on the other hand, “omics” studies are still costly to perform, a factor that limits the availability of data. It is likely that these circumstances have hampered the identification of robust biomarkers with diagnostic value for many diseases, including the discrimination between benign and malignant biliary strictures addressed in our study,” stated the authors.


  1. https://www.mdpi.com/2072-6694/12/6/1644