III: Small: Computational Infrastructure for the Identification of Copy Number Variations from SNP Microarrays

近年来发现人类基因组中拷贝数变异（copy number variations，CNVs）十分普遍，并对表型产生重要影响。目前，用于大规模检测和表征CNV的主要平台是SNP（单核苷酸多态性）微阵列。从微阵列数据中计算鉴定CNV的当前技术水平主要依赖于基于模型的方法（例如，隐马尔可夫模型）。然而，这种方法需要大量的训练数据，这些数据可能并不总是可用的。此外，由于这些方法使用常见的CNV来训练它们的模型，因此它们在识别罕见的CNV方面并不成功，据信罕见的CNV构成了人群中所有CNV的相当大的比例。本项目的目标是开发基于优化的CNVs识别和基因分型算法和软件，以便快速准确地识别不同类型的CNVs（罕见和常见），而不需要训练数据。所提出的框架开发了一种新的计算方法，通过明确制定CNV识别为一系列优化问题，包括多个因素，包括对噪声的敏感性、CNV的稀有性/共同性、基因型特异性和简约性。该公式使得能够开发有效的算法，将罕见和常见CNVs的识别视为具有不同目标函数的不同问题。由此产生的软件向社区的可用性将能够更有效和准确地识别大样本中的CNVs，促进理解CNVs在一系列复杂表型中的作用，包括HIV，自闭症，精神分裂症，精神发育迟滞等。此外，该项目引入的计算创新可能会在下一代测序中找到应用。

It was recently discovered that copy number variations (CNVs) in human genome are quite common, and have important implications on phenotype. Currently, the primary platforms for large-scale detection and characterization of CNVs are SNP (single nucleotide polymorphism) microarrays. The current state-of-the-art in computational identification of CNVs from microarray data relies mostly on model-based approaches (e.g., Hidden Markov Models). However, such methods require extensive training data, which may not be always available. Furthermore, since these methods use common CNVs to train their models, they are not as successful in identifying rare CNVs, which are believed to make up a substantial proportion of all CNVs in the human population. The objective of this project is to develop optimization based algorithms and software for the identification and genotyping of CNVs, with a view to enabling fast and accurate identification of different types of CNVs (rare and common), without the requirement of training data.The proposed framework develops a novel computational approach by explicitly formulating CNV identification as a series of optimization problems that incorporate multiple factors, including sensitivity to noise, rarity/commonality of CNVs, genotypic specificity, and parsimony. This formulation enables development of efficient algorithms that treat identification of rare and common CNVs as different problems with different objective functions. Availability of the resulting software to the community will enable more efficient and accurate identification of CNVs in large samples, facilitating advances in understanding the role of CNVs in a range of complex phenotypes, including HIV, autism, schizophrenia, mental retardation, and many others. Furthermore, the computational innovations introduced by this project are likely to find applications in next generation sequencing.