MIT-6874-Deep Learning in the Life Sciences Week5

AI Application in regulatory omics

1 Fundamentals of Biology ： Regulatory building blocks

1.1 Gene regulation ：Cell diversity, Epigenomics, Regulators(TFs), Motifs, Disease role

Genes are regulated , Express different cell types , So as to produce different organs in the human body .
DNA packaging The meaning of ：DNA Highly spiralized , Realize the purpose of chromatin assembly
dna After forming a double chain by double helix , It binds to histones to form nucleosomes , In the formation of supercoiles by continuous spiraling and curling , Finally, they do not short aggregate to form chromosomes / Chromatin . If this kind of highly dense structure is not available, the space required will be very large , It also hinders various metabolic reactions in cells

（DNA What are the main steps of compression ？ - Ryan Answer - You know
https://www.zhihu.com/question/381570484/answer/1103039550）

Nucleosome assembly is the first step in chromosome assembly ,DNA Packed into nucleosomes , About compressed 7 times .
Chromatin is packed and compressed gradually with nucleosome as its basic structure , the 30nm Chromatin fibers 、 Super spiral ring 、 Finally, it is compressed and packaged into chromosomes , A total of four levels of packaging .
● DNA To the nucleosome
● From nucleosome to solenoid (solenoid)
● From solenoid to super solenoid (supersolenoid)
● From super solenoids to chromosomes

DNA Compression in cells is very important .DNA Inner covenant 147 Base pairs loop Form nucleosome （nucleosome）, Each nucleosome consists of four Histone form , Probably 50 Nucleosomes form a chromatin filament （chromatin fiber）

Epigenomics – Programming methods for each cell type in the body , By adjusting the compression specific DNA Characteristics of sequence promoter region .

Starter ：RNA Polymerase can bind to , Concurrent transcription DNA Part of . The area where cell expression begins to activate

Transcription factor utilization DNA Binding domains identify specific genes in the genome DNA Sequence

Each protein has a different structure , The binding between proteins is not like DNA,RNA The same complementarity , So they need to go through binding motifs

DNA The representation method of ： Pile up all the joints , Then build the matrix , expression DNA The specificity of . The height of the letter indicates the amount of information

1.2 Probing gene regulation: TFs/histones: ChIP-seq, Accessibility: DNase/ATAC-seq

• DNA Transcription ： Histone ：ChIP-seq
• DNA accessibility （ Accessibility ）： Understand chromatin accessibility and ATAC-seq https://zhuanlan.zhihu.com/p/166500744

---

2 Traditional methods of regulatory omics and motif discovery

2.1 be based on Enrichment-based Motif discovery :EM Algorithm 、Gibbs Sampling

2.2 An experimental approach ：PBMs,SELEX. Comparative genomics:Evolutionary conservation

3 Basic application of convolutional neural network in transcriptional regulation

3.1 Low dimensional features ： Main idea ：pixels<->DNA letters. Patches/filters<->Motifs. Higher<->combinations

take DNA The sequence is expressed as a two-dimensional matrix


correct = Ignore signals below certain thresholds .
Pooling = Press max or average Summarize each channel .

Input the extracted features into the network for prediction

3.2 High dimensional features ： Learn convolution kernel <-> Motif discovery. Applying them<-> Motif matches

---

4 CNNS/RNNs Practical application in regulatory omics ： Various architectures

4.1 DeepBind: learn motifs, use in(shallow) fully-connected layer, mutations impact

link ： http://www.nature.com/nbt/journal/v33/n8/full/nbt.3300.html

DeepBind summary

The key deep learning techniques:

• Convolutional learning
• Representational learning
• Back-propagation and stochastic gradient
• Regularization and dropout
• Parallel GPU computing especially useful for hyper-parameter search

Limitations in DeepBind:

• Require defining negative training examples, which is often arbitrary
• Using observed mutation data only as post-hoc evaluation
• Modeling each regulatory dataset separately

4.2 DeepSea: Train model directly on mutational impact prediction

4.3 Baseet: Multi-task DNase prediction in 164 cell types, reuse/learn motifs

Basset: Learning the regulatory code of the accessible genome with deep convolutional neural networks

4.4 ChromPuter: Multi-task prediction of different TFs, reuse partner motifs

4.5 DeepLIFT: Model interpretation based on neuron activation properties

---

5 Lecture ： Stanford Anshul Kundaje,Deep Learning for Reg. Genomics

Deep learning at base-resolution revels cis-regulatory motif syntax

Traditional approach ：
Using traditional sequencing methods ：ATAC-seq / DNase-seq

Machine learning methods ：
Using the accessibility of transcription factors or chromatin , Expression peak, etc. as a label for supervised learning

BPNet: DNA sequence to base-pair resolution profile regression

ChIP-exo/nexus: High resolution TF binding footprints