Transforming dbGaP genetic and genomic data to FAIR-ready by artificial intelligence and machine learning algorithms

dbGaP是NIH资助项目的存储库，它包含许多遗传和基因组数据。 然而，那里的数据还没有为人工智能和机器学习应用做好准备。本申请 提出了解决这一问题的方法。我们有两个目标：1）。发展和规范 将遗传和基因组数据转换为图像状对象和标记化定制的过程 词汇表，以便数据可以被先进的人工智能算法（如CNN、自动编码器）利用。 和Transformer。为了将遗传数据转化为图像，我们将等位基因剂量值重新编码为像素 将诸如SNP和CNV遗传标记的集合强度化并排列成人工 图像对象，以便可以通过CNN算法进行分析。也可以使用遗传标记 定义单倍型，可以标记成自定义词汇表，用于NLP模型。 2）。使用阿尔茨海默病和精神分裂症作为案例研究，以证明 用于发现和鉴定两种病症的风险变体/基因的转换数据。 我们计划使用脑特异性eQTL来估算遗传控制的基因表达， AD数据集的个体基因型，并将表达数据转换为图像对象 用具有自注意机制的CNN模型进行分析。对于精神分裂症，我们计划使用k- mer标记化器将单倍型分解成小的单倍型块的集合，并将它们视为 用于NLP模型分析的令牌。我们使用CNN和NLP模型作为屏幕工具， 使用注意力权重选择有希望的候选者，然后直接测试这些候选者 使用logistic回归分析其与AD/精神分裂症的关联。由于选择效应， 我们可以大大减少测试的数量，显著提高我们的统计能力， 检测AD/精神分裂症的风险变体/基因。

dbGaP is a repository for NIH funded projects and it contains many genetic and genomic data. However, data there are not ready for AI and machine learning applications. This application proposes methods to address this issue. We have two aims: 1). Develop and standardize procedures to transfer genetic and genomic data into image like objects and tokenized custom vocabulary so that the data can be utilized by advanced AI algorithms such CNN, autoencoder and transformer. To transform genetic data into image, we recode allele dosage value as pixel intensity and arrange a collection of genetic markers such as SNPs and CNVs into an artificial image object so that it can be analyzed by CNN algorithms. Genetic markers can also be used to define haplotypes, which can be tokenized into custom vocabularies for use in NLP models. 2). Use Alzheimer's disease and schizophrenia as case studies to demonstrate the utilities of transformed data for the discovery and identification of risk variants/genes for both conditions. We plan to impute genetically controlled gene expression using brain specific eQTLs and individual genotypes for an AD dataset, and transform the expression data into image objects for analyses by CNN model with self attention mechanism. For schizophrenia, we plan to use k- mer tokenizer to break haplotypes into a collection of small haplotype blocks and treat them as tokens for analyses by NLP models. We use both CNN and NLP models as screen tools to select promising candidates using the attention weights, and then directly test these candidates for their association with AD/schizophrenia using logistic regression. Due to the selection effect, we can dramatically reduce the number of testing, significantly increase our statistical power to detect risk variants/genes to AD/schizophrenia.