Development of a deep learning approach to predict noisy biological phenotypes

机器学习长期以来一直用于预测生物学研究中的表型，例如细胞在不同培养基中的生长速率，细胞形状或大小的差异，或化学或辐射扰动下的活力。这些表型通常是复杂实验方案的结果，这些实验方案产生的数据本质上是有噪声的。虽然目前的机器学习方法解释了输入测量中不可避免的噪声-例如基因表达谱或蛋白质染色-但还没有建立有效处理噪声输出变量（表型）的方法。这一差距代表了通过考虑生物表型测量的噪声性质来提高机器学习在生物研究中的适用性的机会。该研究计划的重点是开发新的机器学习方法来预测生物研究中常见的噪声表型。我建议开发一种新的统计数据，以更好地识别分子特征和生物表型之间的关联。然后，我将使用这个统计数据来实现一个能够最大化它的神经网络，以便为感兴趣的表型开发非线性多变量预测器。我将展示我们新的计算方法的实用性，用于预测细胞对化学扰动的反应，这是我们在过去5年中策划的大型数据集汇编。最后，我的团队将创建软件工具，使科学家和公司能够在云中利用我们的新方法，从而最大限度地扩大我们研究计划的传播和影响。在我们的研究计划中开发的方法不仅限于生物数据，而且可以应用于测量连续，嘈杂表型的任何领域。这些包括工程，气候建模和物理学等。通过遵守开放科学的原则，即，通过共享数据、方法、文档代码和开源软件工具，我们希望为其他科学家创建一个平台，以测试和比较新方法，并在从噪声测量中学习是司空见惯的不同领域开辟新的研究途径。

Machine learning has long been used to predict phenotypes in biological research, such as cells' growth rate in different media, differences of cells' shape or size, or viability under chemical or radiation perturbations. These phenotypes are often the results of complex experimental protocols generating data that are intrinsically noisy. While current machine learning methods account for the inevitable noise in the input measurements -- gene expression profiling or protein staining, for instance -- there are no established methods to efficiently deal with noisy output variables (phenotypes). This gap represents an opportunity to improve the applicability of machine learning in biological research by accounting for the noisy nature of biological phenotypic measurements. This research program focuses on developing new machine learning methodologies to predict noisy phenotypes common in biological research. I propose to develop a new statistic to better identify associations between molecular features and biological phenotypes. I will then use this statistic to implement a neural network able to maximize it in order to develop nonlinear, multivariate predictors for the phenotypes of interest. I will showcase the utility of our new computational methods for predicting cells' response to chemical perturbation in a large compendium of datasets that we curated during the last 5 years. Finally, my team will create software tools to allow scientists and companies to leverage our new methods in the cloud, therefore maximizing the dissemination and impact of our research program. The methods developed in our research program are not limited to biological data but can be applied to any field where continuous, noisy phenotypes are measured. These include engineering, climate modelling, and physics for instance. By abiding by the principles of open science, i.e., sharing data, methods, documented code and open-source software tools, we hope to create a platform for other scientists to test new and compare new methods and open new avenues of research in diverse fields where learning from noisy measurements is commonplace.