Biologically informed neural network for subphenotype classification in septic AKI

Sammanfattning: Sepsis is a life threatening condition where the body’s reaction to an infection results in a dysregulated immune response - ultimately causing damage to tissues and organs. The syndrome is diverse, both in underlying biology, disease manifestation and severity, and is therefore divided into endotypes and further into subphenotypes. Further understanding of the biological pathways of the various sepsis types is required in order to develop targeted diagnostic and therapeutic tools necessary to combat the disease. In this thesis, the plasma proteome of patients suffering from two subphenotypes of septic acute kidney injury with varying severity were analyzed. The proteomic data was combined with the Reactome pathway database, and leveraged to generate and train a biologically informed neural network in classifying the two subphenotypes. The network was able to distinguish between the subphenotypes, achieving an accuracy of $98.2 \pm 0.02\%$ when created with four hidden layers. The informed nature of the network allows for introspection into the network's decision making - allowing us to utilize feature importance values to interpret which proteins and biological pathways the network deemed important for classification. Ultimately, this identified several biomarkers for the subphenotypes including apolipoproteins, histones and known inflammatory markers such as CD14 and osteopontin. The algorithm generating the biologically informed network was generalized and is publicly available as a Python package: https://github.com/InfectionMedicineProteomics/BINN.