Accepted_test

Persistent chronic inflammation or other damage to the liver causes fibrogenesis -- a condition in which normal liver tissue begins to scar. Evolution of fibrinogenic condition could be monitored by indirect biomarkers or non-invasive scanning methods (e.g. ultrasound). However, indirect measures can give ambiguous and uninformative results requiring confirmation through histological studies such as liver biopsy. Despite its invasiveness and associated risks, liver biopsy provides the opportunity for a comprehensive assessment of the molecular structure and condition of the liver tissue, ensuring high accuracy in assessing the state of the fibrogenesis in the organ. Unfortunately, the biopsy-based studies of liver fibrosis emergence and development face the problem of limited data, which complicates precise modeling of fibrinogenesis dynamics since biopsy provides only a static description of liver state at, usually, the late stages of the disease. We base our research method on system of partial differential equations that describes dynamics of biochemical properties in liver undergoing fibrogenesis. Application of physics-informed machine learning allows for faster search of solution for this system. Strict adherence of machine learning model to physics of this process is achieved by introduction of additional constraints to loss function. This approach for construction of the model can be applied in small-shot or zero-shot learning setup. In silico fibrogenesis model that we developed allows for prediction of disease process and assessment of theoretical therapeutic effect of drugs or medical interventions, thus allowing for deeper understanding of liver fibrosis emergence and optimization of drug design and treatment protocols.