利用深度學習尋找基因交互作用及其在阿茲海默症的應用

Abstract

阿茲海默症是一種會隨時間不斷惡化、症狀會從記憶力減退到失去行動能力並最終導致死亡的疾病。目前阿茲海默症的發病原因只有可能的推測，真正的原因仍不明瞭，且目前的治療方法只能延緩惡化速度，而無法治癒。儘管過去許多全基因組關聯研究已找出該疾病的數個危險因子，但也只能用來推測個體的患病風險，現有方法最大的缺點在於只考慮單一因子的效應而忽略了基因組合的效應。然而目前對於組合效應，也就是上位作用的研究方法，無論是基於統計線性方法或是傳統機器學習方法，都需要較大的記憶體容量來檢測所有可能的組合，且這種組合通常侷限在兩兩的組合。因此本研究選擇使用深度學習的多層感知器方法去從個體的基因型資料預測其表現型，並從模型中解讀上位作用。本研究的材料來自阿茲海默症腦神經影像計畫的364名個體的資料，這些個體分成病例與對照組兩種類別。在本研究中，我們依序建立了單基因模型與跨基因模型，並從兩種模型分別找出單基因與跨基因內的上位作用。找到的遺傳特徵包括一些已知的重要危險因子，例如最著名的APOE-ε4。這佐證了本研究的深度學習模型確實能找到一些真實的上位作用，除此之外，本研究方法也找到了不只兩兩的組合，解決了目前研究上位方法的限制。

Alzheimer's disease (AD) is a condition that worsens over time, and symptoms can gradually deteriorate from memory loss to loss of mobility and ultimately death. At present, the cause of AD is only speculated. The real cause is still unclear, and the current treatment can only delay the rate of deterioration, but the disease cannot be cured. Although previous genome-wide association studies have identified several risk factors for this disease, these factors can only be used to predict the risk of an individual. The biggest disadvantage of such approaches is that only the effects of a single factor are considered and the effects of gene combinations are ignored. The current research method for the discovery of compound effect, that is, Epistasis, whether based on statistical linear methods or machine learning methods, requires a large memory capacity to detect all possible combinations, and the combination is usually limited to two elements. In this regard, this study aims at replacing current methods for epistasis with multi-layer perceptron, which is one kind of deep learning methods, to predict individual phenotypes from its genotype data, as well as interpreting certain epistasis from our model. The material for this study was derived from 364 individuals in Alzheimer’s Disease Neuroimaging Initiative (ADNI). These individuals were diagnosed as AD or cognitively normal (CN) control. This study established single-gene model and cross-gene model in sequence, and found within-gene and cross-gene epistasis from these two models. The genetic features found included some known important risk factors, such as APOE-ε4. This proves that the deep learning model of this study can indeed find some important combinations in real data. In addition, this research method has also found combinations with more than two elements, which solves the limitation of the current methods.