Integration of real world evidence, cell line assays, and single cell omics via deep reinforcement learning for Alzheimer's disease drug discovery

About This Grant

PROJECT SUMMARY Alzheimer’s disease (AD) is an aging-related incurable neurodegenerative disorder, which a slow but irreversible cognitive decline leaves patients unable to take care of themselves and eventually leads to death in 5-10 years after the diagnosis. About 40 million people worldwide have developed AD and this number is expected to double in the next 20 years. it is imperative that novel solutions are found for the prevention and treatment of AD. Although the disease mechanism remains unknown, through large cellular atlases and the single cell molecular profiling, several types of disease associated cell subtypes, such as disease associated microglia and astrocytes have been found to contribute to the disease etiology or progression. However, despite all these years of research, there is still no effective treatment for AD. With the advance of real-world data (RWD) analysis from large databases of electronic medical records, drug repurposing has become a viable option to address this urgent need for AD drugs. Our project addresses these needs by combining all three of these approaches into a single proposal so that the molecular profiles of high-risk cellular subpopulations, drug repurposing studies, and cell line based mechanistic studies can be leveraged together to find repurposable drug(s) on targetable cell populations in the brain. Specifically, we will apply deep transfer learning models developed by our group to identify the most high-risk subpopulations of cells that are related to clinical and pathological features of AD and use these subpopulations to pick the most similar induced pluripotent stem cell (iPSC)-derived neural cell lines for validation. Simultaneously, based on large RWD cohorts of AD patients, repurposable drug candidates will be identified and tested on human iPSC-derived neural cell lines to validate their causal effects. Specifically, by using the most high-risk cell types and drug candidates identified in this study, these combinations will be tested using cell culture and co-culture assays to detect the effects of these drugs on identified cell types and their effects on neuron loss. When an effect is found, these cultures will be sequenced at single cell resolution to phenotype the cellular response and further pinpoint the cell subtypes that are responsible for the phenotype. Finally, the results of the iPSC culture assays will be used to inform our deep reinforcement transfer learning models of single cell data and our drug repurposing algorithms of RWD via hidden confounder debiasing, resulting in improved future predictions. With this feedback architecture of the reinforcement learning approach, we will be able to develop predictive deep learning models with higher accuracy on cell subtypes that are tightly associated to AD-related clinical and pathological features, while at the same time identify repurposable drugs with direct effect on brain cell subtypes that are linked to AD.