VL5-Advanced RNA-Seq Methods

Why?:

• sequence transcriptome to learn more about gene expression cells

Goal:

• gene/transicpt quanification

• transcript discovery

• expression profiling

How?:

• Design Experiment

• RNA Preparation

  • isolate and purify RNA

• Prepare Libraries

  • convert RNA -> cDNA, while adding sequencing adapters

• Sequence

• Analysis

  • analyze short read seqs

• single-end

  • reads produced from a single end

• paired-end

  • reads produced from a fragment and another read from the opposite side of the fragment

• multiplexing

  • adding individual barcode seq to each sample

• CPM = Counts per million

  • sum of reads divided by a million

• RPKM = Reads per kilobase million

  • CPM/RPM (read per million) divided by length

  • used for single end

• FPKM = Fragments per kilobase milllion

  • RPKM divided by 2

  • used for paired end

• TPM = Transcript per million

  • sum of reads divided by length*million

-> All of them consider sequence length and except for CPM every unit considers gene length as well

= taking the normalized read count data and performing statistical analysis to discover quantitative changes in expression levels between experimental groups

(follows after RNA-Seq)

• bias in data

• TPM does not address differences in efficiency, protocol, tissue difference

• sample size often too low to estimate variance of an individual gene’s expression with any confidence

• t-test requires normal distribution, but RNA-seq follows a negative binomial distribution

Alternative:

• Borrow information across genes

  • Option 1: transform data and then use linear model

  • Option 2: model mean/ variance relationship and shrink dispersion (e.g. DeSeq2)

• Use negative binomial distribution, which takes the inequality of mean and variance into account by using a dispersion parameter

= caused by mean and variance not being equal (variance is higher than mean) in RNA counts

Reason: transcript is present at slightly different levels in each sample

-> Instead of poisson, negative binomial distribution

• Model read count with normalization

• Median of ratio method

• Dispersion shrinkage

• Log Fold Change

• Uses negative binomial generalized linear model

• Takes size factor into account -> normalization factor

  • Typically, constant bias for all samples, but gene-sample-specific size factors exist

  • Estimated with median of ratios method

  • Models read counts for gene x in sample y

• Purpose: normalization

• 1. Creates pseudo-reference sample (row-wise geometric mean)

• 2. Calculate ratio of each sample to the reference

• 3. Caculate normalization factor for each sample (size factor)

• 1. Estimate dispersion per gene (black)

• 2. Fit smoothing curve (red)

• 3. Shrink gene-specific dispersion towards expected dispersion (blue) using empirical Bayes approach

• Strength of shrinkage depends on

  • How close dispersion values are to the fit

  • Degree of freedom (shrinkage reduced for larger samples)

Problem with data: variance of LFC estimates for genes with low read count

• A: before shrinking of LFC

• B: after shrinking

• C: counts show low dispersion for green and high for purple

• D: density plot of likelihoods (solid lines scaled to one) and posteriors (dashed) for green and purple genes and the prior (black)

-> Correction removes false positives

= different possible combination of exons, resulting in different proteins

Problem:

-> splicing makes mapping more difficult, since transcriptome might not fit exactly to genome

Solution:

-> splice aware mapping

1. Exact match to GT-AG, GC-AG, AT-AC

   -> Examples: TopHat2

2. The list of known splice sites

   -> Examples: HISAT2

3. Models (HMM, SVM)

   -> Examples: HMMSplicer, PALMapper, OLEgo

4. Unbiased from RNA-Seq

   -> Examples: CRAC, MapSplice2, STAR

= a fast RNA-seq read mapper, with support for splice-junction and fusion read detection

• Seed searching:

  • Align reads by searching for Maximal Mappable Prefix (= longest sequence that matches exactly one or more locations on reference genome)

  • Different part of reads, which are differently mapped, are called seeds

    • First part of MMP that was mapped with genome is seed 1, then rest sequence is searched in reference, if match this will be the second seed

• Extend MMPs:

  • If no exact matching seq for each part of the read, the previous MPP will be extended

    -> Reasons for no match: mismatch or indels

• Soft Clipping:

  -> If extension does not return a good alignment, then poor quality or adapter seq will be soft clipped

• Cluster, Stitching and Scoring:

  -> Create complete read by

  • first clustering the seeds together based on proximity to a set of anchor seed or seeds that are not multi-mapping

  • then stitched together based on best alignment for the read

  • get best alignment scoring based on mismatches, indels and gaps, etc.

• efficiency due to sequential search of only the unmapped portions of reads

• uses an uncompressed suffix array (SA) for efficient MMPs search

  • allows quick searching against large reference genomes

• other tools: use algorithms, which often search for entire read seq before slitting reads and perform iterative rounds of mapping -> slow

= directed graph with transcripts as nodes and edges combining consecutive transcripts

= tool for alternative splicing analysis based on RNA-Seq alignment data

Possible graphs from alternative spicing events:

Goal:

• Estimate transcript abundances

  • Classify diseases, understand expression changes, track cancer progression

Problems:

• Only an estimate, as huge number of sequences are considered

  • Nonetheless accurate, as used for diagnostics

• Traditional: Alignment based approaches

  • But: Slow, substantial computational resources

= used for fast and accurate quantification of gene expression from RNA-seq data

-> get transcript abundances for gene expression analysis

Input:

• Reference transcriptome

• RNA-Seq reads from experiment

Output:

• Kallisto

  • Index, Quantification of RNA-Seq samples (transcript abundances)

For each same sequence snippet of length k, create a node and connect each consecutive node with an edge

A)

For each read, fill cell with 1 if read is in transcript else 0

B)

1. Each row, which has the same values is part of an equivalence class

2. Counting the rows with same values results in the corresponding equivalence class count

• Characterizes and defines cell types in a heterogeneous cell population by analysing expression of cell surface and intercellular molecules

• Used to measure fluorescence produced by fluorescent antibodies detecting proteins

• Limitations:

  • Slow data acquisition

  • Cell transmissions efficiency

  • Not all measured cells are alive

• Uses antibodies to bind proteins in tissues

• Visualize location of specific cell types by using layers of coloured complexes

• Limitations:

  • High dependency on antibody quality

  • No unique reference for the stainings

  • High set up costs

= challenge of estimating the cell type or tissue composition of a heterogeneous sample based on its gene expression profile

-> Gene expression in an heterogeneous sample can be modelled as the weighted sum of the expression value for each cell type present in the mixture.

Matrix notation:

D = C x F

D = measured expression values

C = cell type specific expression values

F = relative cell type proportion

-> Reference based deconvolution: D and C is given

• Goal: minimize sum of squares between fitted (C x F) and observed values D

• Constraints:

  • Proportion in C= {0,1}

  • Sum of proportion per sample = 1

• Another approach:

  • latent variable problem, where each cell type constitutes a latent factor

  • extract latent variables from matrix with Nonnegative Matrix Factorization (NMF)

• unsupervised learning algo

• projects data into lower dimensional spaces

• reduces number of features, while retaining basic info

• decomposes a matrix, containing only negative coefficients, into two non-negative matrices with reduced ranks

1. Marker genes

   = list of enriched genes for each cell type

2. Deconvolution

   = ‘inverse’ matrix multiplication with reference profiles