A machine learning approach to predict pancreatic islet grafts rejection versus tolerance

Abstract

The application of artificial intelligence (AI) and machine learning (ML) in biomedical research promises to unlock new information from the vast amounts of data being generated through the delivery of healthcare and the expanding high-throughput research applications. Such information can aid medical diagnoses and reveal various unique patterns of biochemical and immune features that can serve as early disease biomarkers. In this report, we demonstrate the feasibility of using an AI/ML approach in a relatively small dataset to discriminate among three categories of samples obtained from mice that either rejected or tolerated their pancreatic islet allografts following transplant in the anterior chamber of the eye, and from naïve controls. We created a locked software based on a support vector machine (SVM) technique for pattern recognition in electropherograms (EPGs) generated by micellar electrokinetic chromatography and laser induced fluorescence detection (MEKC-LIFD). Predictions were made based only on the aligned EPGs obtained in microliter-size aqueous humor samples representative of the immediate local microenvironment of the islet allografts. The analysis identified discriminative peaks in the EPGs of the three sample categories. Our classifier software was tested with targeted and untargeted peaks. Working with the patterns of untargeted peaks (i.e., based on the whole pattern of EPGs), it was able to achieve a 21 out of 22 positive classification score with a corresponding 95.45% prediction accuracy among the three sample categories, and 100% accuracy between the rejecting and tolerant recipients. These findings demonstrate the feasibility of AI/ML approaches to classify small numbers of samples and they warrant further studies to identify the analytes/biochemicals corresponding to discriminative features as potential biomarkers of islet allograft immune rejection and tolerance.

Parallel Multi-Omics in High-Risk Subjects for the Identification of Integrated Biomarker Signatures of Type 1 Diabetes

“…Notably, predictions of the same functions, pathways, and networks were consistently made at several, less stringent, cutoff values in our current dataset. We recently showed that the prediction accuracy in a small dataset was higher based on complete peak patterns in whole electropherograms rather than a few discriminative peaks selected based on significance, despite the potentially higher noise (non-significantly different peaks) in the compared electropherograms [ 94 ]. For instance, inflammatory response, proliferation of immune cells including T-lymphocytes, especially CD4+, and their increased motility, and the activation of macrophages and ROS generation were only indicated at the three lower cutoff values ( Figure 3 ).…”

Integrated Metabolomics and Proteomics Analyses in the Local Milieu of Islet Allografts in Rejection versus Tolerance

“…Capitalizing on these technical capabilities of the ACE-platform, we previously identified significant changes in the local proteome and metabolome in association with intraocular islet allografts rejection [25,26]. Here, we further examine specific changes in local metabolites and proteins involved in immunometabolism in the context of immune tolerance.…”

“…The migration of each peak, which correspond to a putative metabolite, is dictated by the electro-osmotic properties of the metabolite [27,28]. Therefore, peaks in different EPGs with the same migration time likely correspond to the same metabolite [26]. The first group of peaks (group A) migrated between 6 and 15 min.…”

The Eye as a Transplantation Site to Monitor Pancreatic Islet Cell Plasticity

“…Microliter-size aqueous samples can be obtained in the immediate vicinity of the engrafted tissue, allowing for the analysis of islet-related metabolites and proteins. In particular, this strategy has been used with the aim to define early predictive markers of type 1 diabetes by detecting changes in the metabolic profile (50), and to predict the risk of allograft rejection to allow for a timely therapeutic intervention (51,52). As a whole, these methodologies using the ACE platform are perfectly suited for the study of islet plasticity in health and disease under in vivo conditions.…”