Accelerating biological discovery with automated machine learning for single-cell data analysis

点击翻译按钮获取中文摘要

Biological research is becoming increasingly quantitative, with the analysis and interpretation of the vast quantities of data produced being key steps to extracting insight. In particular, the advent of single-cell RNA-sequencing (scRNA-seq) has revolutionized many aspects of research, with transcriptome-wide gene expression data routinely being generated from hundreds of thousands of cells. However, analyzing and interpreting the resulting data remains a key bottleneck to researchers. In particular, the large number of possible ways to analyze the data including choices of algorithms for quality control, normalization, clustering (assigning cells to discrete cell types or states) and dimensionality reduction (for visualization or lineage inference) leads to a combinatorial number of possible workflows that is far too large for the computational biologist analyzing the data to explore in practice. Furthermore, different workflows can have dramatically different results on the same dataset, with few metrics to assess how well a given workflow has performed and no data set -specific workflow recommendation methods to--date. Consequently, ad--hoc manual workflow choice and optimization remains a major bottleneck in biological data analysis that can significantly alter downstream results and interpretation. To address these issues, this program aims to investigate and develop novel statistical machine learning approaches to automatically learn optimal workflows for single--cell data analysis. This will be the first research that uses machine learning to predict which workflow will perform optimally for a given scRNA-seq dataset, and will create the first models to intelligently perform real--time recommendation of workflows in a dataset--specific manner. Furthermore, it will include significant and novel research into automated machine learning (AutoML) models, an expanding research domain that aims to automatically select optimal models and hyperparameters for machine learning tasks. In contrast with existing work we will research AutoML models for unsupervised learning problems, of which most single--cell analysis workflows are comprised. Consequently, we aim to establish the field of BioAutoML and automate aspects of single--cell data analysis while contributing novel machine learning research, leveraging both advances in artificial intelligence research and the large number of single--cell datasets currently available. In the longer term, our vision is to develop a suite of models and algorithms for intelligent, machine-learning driven workflow recommendation and execution in the biological sciences, reducing the bottleneck of human-intensive data analysis and interpretation.