Learning to learn in structural biology with deep neural networks

项目摘要/摘要 深度学习作为一种强大的工具，正在许多领域获得吸引力。在医学方面，有 最近在药物设计、预测蛋白质结构和功能基因组学方面取得了成功。 到目前为止，这些成功都发生在拥有数十万个数据点的地区 而医学上的深度学习仍然受到缺乏大型同源数据集的限制。 这项提议的重点是应用一种新的深度学习，称为元学习，它模仿 从几个例子中学习的能力就像人类一样。PI将建立一个可持续的研究 通过开发基准问题和数据集的元学习计划。PI将进一步 探索元学习，特别是多肽-蛋白质结构和核磁共振谱预测。由于 在医学中使用深度学习时迫切需要可解释性，这是一个强大的组成部分 将生物物理模型与深度学习模型联系起来。 这项工作的结果将是一种示范的新的深度学习方法，这种方法可以奏效 几乎没有数据。PI将把这些研究想法结合在一起，设计出可以结合的多肽 治疗神经退行性变的一项具有挑战性但重要的任务 疾病。这将通过元学习、分子模拟和迭代来实现 多肽设计。

Project Summary/Abstract Deep learning is gaining traction across many elds as a powerful tool. In medicine, there have been recent successes in drug design, predicting protein structure, and in functional genomics. These successes have thus far been in areas where there are hundreds of thousands of data points and deep learning in medicine is still limited by lack of large homongeous datasets. This proposal focuses on applying a new kind of deep learning called meta-learning that mimics the human-like ability to learn from few examples. The PI will establish a sustainable research program on meta-learning by developing benchmark problems and datasets. The PI will further explore meta-learning speci cally on peptide-protein structure and NMR spectra prediction. Due to the imperative need for interpretability when using deep learning in medicine, a strong component will be connecting biophysical modeling with the deep learning models. The outcome of this work will be a demonstrated new approach to deep learning that can work with little data. The PI will bring these research ideas together to design peptides that can bind to intrinsically disordred proteins, a challenging but important task for curing neurodegenerative diseases. This will be accomplished through meta-learning, molecular simulation, and iterative peptide design.