Paper reading (十九)：Computational biology: deep learning

论文题目：Computational biology: deep learning

scholar 引用：20

页数：18

发表时间：2017.11

发表刊物：Emerging Topics in Life Sciences

作者：William Jones, Kaur Alasoo, Dmytro Fishman, and Leopold Parts

摘要：

Deep learning is the trendiest tool in a compuptational biologist's toolbox. This exciting class of methods, based on artificial neural networks, quickly became popular due to its competitive performance in prediction problems. In pioneering early work, applying simple network architectures to abundant data already provided gains over traditional counterpatrts in functional genomics, image analysis, and medical diagnostics. Now, ideas for constructing and training networks and even off-the-shelf models have been adapted from the rapidly developing machine learning subfield to improve performance in a range of computational biology tasks. Here, we review some of these advances in the last 2 years.

结论：

• Deep learning has already permeated computational biology research.
• we expect many more projects to make use of transfer learning in the near future.
• 为啥感觉结论里面很多观点都是模棱两可呢？
• the ultimate goal of research is understanding

Introduction：

• two main reasons why deep learning is appealing to computational biologists:

1. this powerful class of models can approximate nearly any input to output mapping if provided enough data.
2. deep neural networks can learn directly from raw input data.

正文组织架构：

1. Introduction

2. Genomics

2.1 New applications to functional genomics data

2.2 Variant calling from DNA sequencing

2.3 Recent improvements to convolutional models

        2.3.1 Incroporation elements of recurrent neural network

        2.3.2 Reverse complement parameter sharing

        2.3.3 Incorporating prior information

3. Biological image analysis

3.1 Image segmentation

3.2 Cell and image phenotyping

4. Medical diagnostics

4.1 Self-diagnosis with deep learning

4.2 Using other medical data modalities

5. Discussion

正文部分内容摘录：

• The applications of deep neural networks are not limited to genomic sequences.
• 所以有哪些non-sequence的data方面的应用呢：CODA applies a convolutional neural network to paired noisy and high-quality ChiP-seq datasets to learn a generalizable model that reduces the noise caused by low cell input, low sequencing depth, and low signal-to-noise ratio.Convolutional neural networks have also been used to predict genome-wide locations of transcription start sitesfrom DNA sequence, RNA polymerase binding, nucleosome positioning and transcriptional data, as well as gene expression from histone modifications, 3D chromatin interactions from DNA sequence and chromatin
  accessibility, DNA methylation from single-cell bisulfite sequencing data, and protein binding to RNA from the primary, secondary, and tertiary structures or other features。如果后续需要查看非序列方向的文献，maybe这段中提到的都要看一下。
• they implemented new convolutional layers that explicitly share parameters between the forward and reverse
  complement strands. This improved model accuracy, increased learning rate, and led to a more interpretable
  internal motif representation.
• The ultimate goal of much of biomedical research is to help diagnose, treat, and monitor patients.
• Short overview of computational biology deep learning papers published until the first quarter of 2017，很长（7pages）很全的一个Appendix整理。包含Name、Title、Architecture、Input、Output、Hightlight、Category。
• Appendix分了functional genomics、sequence data analysis、medical diagnostics、biomedical image analysis（image segmentation和cell and image phenotyping）四大类。