Machine Learning Model Validation for AD/ADRD

University Of New Mexico Health Scis Ctr, Albuquerque NM

Investigators

Linked publications, trials & patents

Abstract

PROJECT SUMMARY The primary goal of this Administrative Supplement is to study 20 candidate genes, previously not associated with Alzheimer's disease and related dementias (ADRD), in ADRD-relevant experimental models. These genes were identified by machine learning derived from meta-path-based protein knowledge graph integration. ADRD continues to be an intractable age-related neurodegenerative illness in elderly without a cure. Numerous clinical trials have failed in result in effective therapeutic management of ADRD. This unprecedented 99.6% failure rate in drug development targeting ADRD is, in part, due to lack of sufficient consideration of novel molecular pathways and mechanisms besides the ones neuropathologically associated with different dementias. The main thrust of this proposal is to use automated algorithmic network predictions based on the meta-path approach, by providing the integrated database developed at the Illuminating the Druggable Genome (IDG) Knowledge Management Center (KMC). Based on 13 distinct datasets totaling over 261 million datapoints, the AD-specific machine learning classifier model identified several key features, such as proteins that mediate inflammatory processes (JAK2, IL10 & IL2), response to oxidative stress (GSTP1), nervous system development (BDNF) and glycolysis (GAPDH). As a preliminary validation, we used human SHSY5Y cells that show robust hyper- phosphorylation of tau at S202 (itself a predictor of neurofibrillary tangle pathology relevant to ADRD) upon inflammatory stimuli (treated with cell supernatant from RAW macrophages). We used individual siRNA knockdown for the 20 predicted candidate genes in SHSY5Y cells prior to RAW macrophage media treatment, and assessed pS202-tau by Cellomics® unbiased high-content microscopy and quantitative morphometry. Four out of 20 candidate genes altered pS202 -tau levels in SHSY5Y cells, thus providing a proof-of-concept validation for the AD machine learning model. Based on this, we hypothesize that the AD-specific model is able to reliably identify novel biological datasets relevant to ADRD. We propose to test this hypothesis under three specific aims. In Aim 1, we will improve the meta-path machine-learning models and data mine novel associations. In Aim 2, we will perform in vitro biological validation for the top genes predicted by the meta-path approach. In Aim 3, we will build an interactive web platform to disseminate AD-gene associations. Successful completion of these studies will open up new avenues for ADRD research and may lead to novel therapeutic targets for ADRD.