Implicit generative modeling for computational genomics

点击翻译按钮获取中文摘要

Genomics is the study of DNA sequences in our cells and how they drive function or dysfunction. Genomics data also reveals the flow of information within cells. By comparing genomics data captured under different conditions (mutations or drug treatments) we can observe how they affect cell function. Machine learning can do more than merely observe: it can learn the complex "rules" by which DNA drives function, directly from data. Once learned, these rules can predict which mutations are likely to cause a gene to stop functioning and cause disease. Genomics data can increasingly be generated by automated laboratories, including remote cloud laboratories, which offer better accuracy and consistency than traditional laboratory work. This combination of information-rich data and automated experimentation is finally setting the stage for "self-driven" laboratories. In a self-driven laboratory, experiments are designed by machine learning systems to systematically interrogate an outcome of interest, such as the sequence-function relationships in human cells. Data and models are improved in a 24/7 feedback loop. The proposed research aims to make tractable progress towards this vision. This vision is important because it represents an opportunity for drug discovery. Today, 90% of all drug candidates fail at the clinical stage, after years of development and tens or hundreds of millions of dollars have been spent. Better predictive models of how cells work will help to identify and design compounds for genetic disorders up front, with a higher final success rate. The proposed research focuses on four objectives. The first is to make it easier for machine learning scientists to participate in advancing this vision. Genomics data is notoriously difficult to understand and to work with, so the goal is to build a code library and data repository that enables the machine learning community to contribute to genomics research. The second objective is to improve the accuracy with which genomics data is interpreted, specifically data from RNA sequencing, or RNA-seq. RNA-seq captures a "snapshot" of the RNA molecules in a sample, and is the penultimate step in countless protocols for interrogating cells. However, the snapshot is fragmented. Algorithms are needed to infer (to reconstruct) what RNA molecules were actually present in the sample. The proposed research will investigate a new, more accurate approach to inference based on deep learning. The third objective is to build more accurate models of how RNA is processed in cells by new deep learning techniques. Models of RNA processing are important for predicting the effects of mutations and of potential therapies. (Such models are also based on RNA-seq data, and may benefit from advances in the second objective.) The fourth and final objective is to create algorithms that can automatically design experiments, generating the most valuable data possible from which to build models of RNA biology.