Deep-learning-based prediction of AMD and its progression with GWAS and fundus image data

视网膜相关性黄斑变性（AMD）是全世界不可逆失明的主要原因。成功 AMD的全基因组关联研究（GWAS）已经鉴定了许多疾病易感基因。通过 国际GWAS联盟和大型合作项目的巨大努力，大量数据集 包括高质量的GWAS数据和良好表征的临床表型， dbGaP和UK Biobank等数据库。在临床上，彩色眼底图像已被广泛使用， 眼科医生诊断AMD及其严重程度。丰富的GWAS数据和眼底 图像数据为研究人员提供了前所未有的机会来测试新的假设， 原始项目的目标。其中，AMD发展及其进展的预测模型基于 对GWAS和眼底图像数据的影响尚未研究。现有的大多数预测模型只关注 经典的统计方法，通常是具有有限数量的预测因子的回归模型（例如，SNP）。 此外，大多数预测只给出静态风险，而不是随时间变化的动态风险轨迹，其中 后者对于像AMD这样的进行性疾病更有信息量。机器学习技术的最新进展， 特别是深度学习，已经被证明可以通过结合 当具有明确表型的大规模训练数据集被 available.尽管深度学习在许多领域取得了成功，但在AMD和其他公司的眼中， 疾病受AMD发展或进展的多项大规模研究的启发，其中GWAS和/或 纵向眼底图像数据已经收集，我们提出了新的深度学习方法来预测 AMD状态及其进展，并确定具有显著不同风险特征的亚组。特别是，在Aim 1，我们将构建一个新的局部卷积神经网络来预测疾病的发生（AMD与否）， 严重性（例如，轻度AMD、中度AMD、晚期AMD），基于（1a）：35，000+个体的大型队列， GWAS数据和（1b）：具有GWAS和眼底图像数据的4，000多名个体的较小队列。在目标2中， 我们将开发一种新的深度神经网络生存模型，用于预测个体疾病进展， 轨迹（例如，到晚期AMD的时间）。在这两个目标中，我们将使用局部线性逼近技术来识别 有助于预测个体风险状况和识别具有不同风险的亚组的重要预测因子 数据区.在目标3中，我们将使用独立的队列验证和校准我们的方法，并实现提出的方法。 将这些方法转化为用户友好的软件和易于访问的Web界面。最近FDA批准了 Beovu是一种治疗湿性AMD（晚期AMD的一种）的新型注射剂，通过抑制VEGF，从而抑制 异常血管的生长，它使我们的研究更有意义，因为它将提供最前沿的 和AMD的综合预测模型，其具有促进早期诊断的巨大潜力， 定制治疗和疾病的临床管理。

Age-related macular degeneration (AMD) is a leading cause of irreversible blindness worldwide. Successful genome-wide association studies (GWAS) of AMD have identified many disease-susceptibility genes. Through great efforts from international GWAS consortium and large-scale collaborative projects, massive datasets including high-quality GWAS data and well-characterized clinical phenotypes are now available in public repositories such as dbGaP and UK Biobank. Clinically, color fundus images have been extensively used by ophthalmologists to diagnose AMD and its severity level. The combination of wealthy GWAS data and fundus image data provides an unprecedented opportunity for researchers to test new hypotheses that are beyond the objectives of original projects. Among them, predictive models for AMD development and its progression based on both GWAS and fundus image data have not been explored. Most existing prediction models only focus on classic statistical approaches, often regression models with a limited number of predictors (e.g., SNPs). Moreover, most predictions only give static risks rather than dynamic risk trajectories over time, of which the latter is more informative for a progressive disease like AMD. Recent advances of machine learning techniques, particularly deep learning, have been proven to significantly improve prediction accuracy by incorporating multiple layers of hidden non-linear effects when large-scale training datasets with well-defined phenotypes are available. Despite its success in many areas, deep learning has not been fully explored in AMD and other eye diseases. Motivated by multiple large-scale studies of AMD development or progression, where GWAS and/or longitudinal fundus image data have been collected, we propose novel deep learning methods for predicting AMD status and its progression, and to identify subgroups with significant different risk profiles. Specially, in Aim 1, we will construct a novel local convolutional neural network to predict disease occurrence (AMD or not) and severity (e.g., mild AMD, intermediate AMD, late AMD) based on (1a): a large cohort of 35,000+ individuals with GWAS data and (1b): a smaller cohort of 4,000+ individuals with both GWAS and fundus image data. In Aim 2, we will develop a novel deep neural network survival model for predicting individual disease progression trajectory (e.g., time to late-AMD). In both aims, we will use the local linear approximation technique to identify important predictors that contribute to individual risk profile prediction and to identify subgroups with different risk profiles. In Aim 3, we will validate and calibrate our methods using independent cohorts and implement proposed methods into user-friendly software and easy-to-access web interface. With the very recent FDA approval for Beovu, a novel injection treatment for wet AMD (one type of late AMD) by inhibiting VEGF and thus suppressing the growth of abnormal blood vessels, it makes our study more significant, as it will provide most cutting-edge and comprehensive prediction models for AMD which have great potential to facilitate early diagnosis and tailored treatment and clinical management of the disease.