Machine Learning and Augmented Intelligence Enables Prognosis of Type 2 Diabetes Prior to Clinical Manifestation.

BACKGROUND: The global incidence of type 2 diabetes (T2D) persists at epidemic proportions. Early diagnosis and/or preventive efforts are critical to attenuate the multi-systemic clinical manifestation and consequent healthcare burden. Despite enormous strides in the understanding of pathophysiology and on-going therapeutic development, effectiveness and access are persistent limitations. Among the greatest challenges, the extensive research efforts have not promulgated reliable predictive biomarkers for early detection and risk assessment. The emerging fields of multi-omics combined with machine learning (ML) and augmented intelligence (AI) have profoundly impacted the capacity for predictive, preventive, and personalized medicine. OBJECTIVE: This paper explores the current challenges associated with the identification of predictive biomarkers for T2D and discusses potential actionable solutions for biomarker identification and validation. METHODS: The articles included were collected from PubMed queries. The selected topics of inquiry represented a wide range of themes in diabetes biomarker prediction and prognosis. RESULTS: The current criteria and cutoffs for T2D diagnosis are not optimal nor consider a myriad of contributing factors in terms of early detection. There is an opportunity to leverage AI and ML to significantly enhance the understanding of the underlying mechanisms of the disease and identify prognostic biomarkers. The innovative technologies being developed by GATC are expected to play a crucial role in this pursuit via algorithm training and validation, enabling comprehensive and in-depth analysis of complex biological systems. CONCLUSION: GATC is an emerging leader guiding the establishment of a systems approach towards research and predictive, personalized medicine. The integration of these technologies with clinical data can contribute to a more comprehensive understanding of T2D, paving the way for precision medicine approaches and improved patient outcomes.

A sample selection strategy to boost the statistical power of signature detection in cancer expression profile studies.

In case-control profiling studies, increasing the sample size does not always improve statistical power because the variance may also be increased if samples are highly heterogeneous. For instance, tumor samples used for gene expression assay are often heterogeneous in terms of tissue composition or mechanism of progression, or both; however, such variation is rarely taken into account in expression profiles analysis. We use a prostate cancer prognosis study as an example to demonstrate that solely recruiting more patient samples may not increase power for biomarker detection at all. In response to the heterogeneity due to mixed tissue, we developed a sample selection strategy termed Stepwise Enrichment by which samples are systematically culled based on tumor content and analyzed with t-test to determine an optimal threshold for tissue percentage. The selected tissue-percentage threshold identified the most significant data by balancing the sample size and the sample homogeneity; therefore, the power is substantially increased for identifying the prognostic biomarkers in prostate tumor epithelium cells as well as in prostate stroma cells. This strategy can be generally applied to profiling studies where the level of sample heterogeneity can be measured or estimated.

Expression changes in the stroma of prostate cancer predict subsequent relapse.

Biomarkers are needed to address overtreatment that occurs for the majority of prostate cancer patients that would not die of the disease but receive radical treatment. A possible barrier to biomarker discovery may be the polyclonal/multifocal nature of prostate tumors as well as cell-type heterogeneity between patient samples. Tumor-adjacent stroma (tumor microenvironment) is less affected by genetic alteration and might therefore yield more consistent biomarkers in response to tumor aggressiveness. To this end we compared Affymetrix gene expression profiles in stroma near tumor and identified a set of 115 probe sets for which the expression levels were significantly correlated with time-to-relapse. We also compared patients that chemically relapsed shortly after prostatectomy (<1 year), and patients that did not relapse in the first four years after prostatectomy. We identified 131 differentially expressed microarray probe sets between these two categories. 19 probe sets (15 genes overlapped between the two gene lists with p<0.0001). We developed a PAM-based classifier by training on samples containing stroma near tumor: 9 rapid relapse patient samples and 9 indolent patient samples. We then tested the classifier on 47 different samples, containing 90% or more stroma. The classifier predicted the risk status of patients with an average accuracy of 87%. This is the first general tumor microenvironment-based prognostic classifier. These results indicate that the prostate cancer microenvironment exhibits reproducible changes useful for predicting outcomes for patients.

Diagnosis of prostate cancer using differentially expressed genes in stroma.

More than one million prostate biopsies are performed in the United States every year. A failure to find cancer is not definitive in a significant percentage of patients due to the presence of equivocal structures or continuing clinical suspicion. We have identified gene expression changes in stroma that can detect tumor nearby. We compared gene expression profiles of 13 biopsies containing stroma near tumor and 15 biopsies from volunteers without prostate cancer. About 3,800 significant expression changes were found and thereafter filtered using independent expression profiles to eliminate possible age-related genes and genes expressed at detectable levels in tumor cells. A stroma-specific classifier for nearby tumor was constructed on the basis of 114 candidate genes and tested on 364 independent samples including 243 tumor-bearing samples and 121 nontumor samples (normal biopsies, normal autopsies, remote stroma, as well as stroma within a few millimeters of tumor). The classifier predicted the tumor status of patients using tumor-free samples with an average accuracy of 97% (sensitivity = 98% and specificity = 88%) whereas classifiers trained with sets of 100 randomly generated genes had no diagnostic value. These results indicate that the prostate cancer microenvironment exhibits reproducible changes useful for categorizing the presence of tumor in patients when a prostate sample is derived from near the tumor but does not contain any recognizable tumor.