AbstractBackgroundTau protein abnormalities have been attracting more and more attention in Alzheimer’s disease (AD) pathophysiology. Tau spreads in a hierarchical pattern in the AD brain, likely by trans‐synaptic propagation between neurons. Discovering genetic risk factors associated with tau pathology could be key for improving our ability to predict and treat AD. In this study, we leveraged the non‐linear analytic approach of deep learning to identify genes that are associated with cerebrospinal fluid (CSF) total tau levels.MethodWe selected 384 subjects [118 normal control and 266 mild cognitive impairment (MCI)] from ADNI cohort. We designed a deep neuron network model that predicted CSF total tau levels using peripheral blood transcriptomic data. More specifically, we developed an occlusion map approach to rank 2,364 transcripts, which were initially selected from 49,386 transcripts by dimension reduction. The occlusion map aimed to rank the transcripts and identified the critical ones.ResultOur deep neural network successfully predicted CSF total tau levels with an accuracy of 67% and the occlusion map identified the critical genes for the prediction. The list included genes like TTL5, AKIRIN2, CPEB1, PTPN7, CIRBP and RPS23, which have been previously reported to be related to AD. For example, TTL5 relates to missorting of tau protein. AKIRIN2 is essential for the formation of the cerebral cortex. CPEB is found to be an interacting factor of amyloid‐precursor protein (APP). PTPN7 belongs to protein tyrosine phosphatases family, among which PTPN1& PTPN5 has been reported to tau pathology and amyloid pathology. CIRBP is reported to exert protective effects against neuronal amyloid toxicity. RPS23 regulates amyloid level and tau phosphorylation.ConclusionUncovering the genes that have critical association with tau protein abnormalities is critically important for improving our understanding of the drivers of neurodegeneration in prodromal AD. Our deep learning algorithm has shown promising results in identifying the critical genes for tau pathology. This study has demonstrated the effectiveness of deep learning in genetic analysis for AD research.1. Bahureksa, L., et al. (2017)2. Wilcockson, T.D.W., et al. (2019)3. Tomoeda, C. K. (2001)4. Feldman, H. H., et al. (2008)5. Winblad, B., et al. (2016)6. Thomann, A. E., et al. (2018)