Transfer learning and uncertainty quantification in epigenetic clocks

项目摘要/摘要 拟议研究的目标是开发一个迁移学习框架，以 改进现有的DNA甲基化(DNaM)时钟并提供不确定性量化 以及年龄预测。越来越多的证据表明，dNaM水平 在特定的年龄相关CpG位点代表稳定和可重现的年龄生物标志物。 虽然已经构建了许多表观遗传学时钟来预测时间年龄， 年龄加速，或与年龄相关的疾病，它们可能缺乏普遍性或适应性，因为 它们大多是从特定的亚群中开发和验证的。几乎没有什么存在 探索如何将从现有表观遗传学中学到的知识转移的研究 当从不同的亚群收集新数据时，与成年人一起训练的时钟 比如儿童和青少年群体。该项目将开发一种创新的 无需重新访问即可更新现有表观遗传时钟的迁移学习方法 原始训练数据中的个别级数据。此外，大多数现有的表观遗传学 时钟只提供点预测，没有不确定性量化。然而， 在固定的验证集上，合理的准确性是不够的，而且这个项目还将 扩展保角推理框架以构造预测区间 有保证的置信度。这样一个具有不确定性量化的预测范式 很重要，因为它通过提供一个 一组看似合理的预测结果，用一定的方式覆盖了基本事实 概率。此外，这项拟议的工作将允许整合顺序 更新和自适应校准程序，利用 从不同人群收集的丰富的新dNaM数据集的可用性。这个 该方案的完成将极大地提高系统的通用性、可靠性和可扩展性 现有表观遗传时钟对儿童等新目标人群的适应性 和青春期队列。

PROJECT SUMMARY / ABSTRACT The objective of the proposed research is to develop a transfer learning framework to refine existing DNA methylation (DNAm) clocks and provide uncertainty quantification along with age prediction. A growing body of evidence has shown that the DNAm levels at specific age-related CpG sites represent stable and reproducible biomarkers of age. While numerous epigenetic clocks have been constructed to predict chronological age, age acceleration, or age-related diseases, they may lack generality or adaptability as they are mostly developed and validated from a certain subpopulation. There exists little research that explores how to transfer knowledge learned from the existing epigenetic clocks trained with adults when new data are collected from a different subpopulation such as the children and adolescent cohorts. This project will develop an innovative transfer learning approach to update existing epigenetic clocks without re-accessing individual-level data in the original training data. In addition, most existing epigenetic clocks provide only point predictions without uncertainty quantification. However, reasonable accuracy on a fixed validation set is not enough, and this project will also extend the conformal inference framework to construct prediction intervals with a guaranteed confidence level. Such a predictive paradigm with uncertainty quantification is important because it communicates better with the science community by providing a set of plausible predicted outcomes that covers the ground truth with a certain probability. Furthermore, this proposed work will allow for an integration of sequential updating and adaptive calibration procedures that construct prediction intervals with the availability of rich new DNAm datasets collected from diverse populations. The completion of this proposal will greatly improve the generalizability, reliability, and adaptability of existing epigenetic clocks to a new target population such as the children and adolescent cohort.