High performance sequencing and gene expression quantification

Size: px

Start display at page:

Download "High performance sequencing and gene expression quantification"

Lawrence McCoy
5 years ago
Views:

1 High performance sequencing and gene expression quantification Ana Conesa Genomics of Gene Expression Lab Centro de Investigaciones Príncipe Felipe Valencia

2 Next Generation Sequencing NGS has brought high speed not only to genome sequencing and personal medicine, but has also change the way we do genome research: Got a question on genome organization: SEQUENCE IT!!!!

3 Next Generation Sequencing Stem cell Research Cancer research Rare Diseases Ecogenomics Epigenomics Development Agrosciences Structural Genomics Chromatin organization. De novo sequencing Genome Assembly Genome Annotation Resequencing Exome Sequencing ChipSeq MethylSeq RNASeq. Sequencing of the expressed genome

4 RNA-seq protocol* total RNA purification mrna preparation oligodt RiboZ 2 nd strand synthesis fragmentation 1 st strand synthesis RNA DNA *Solexa Pair-End

5 RNA-seq protocol (II) 100bp lad adenylation 3 ends A A A A A A A A A A ligate adapters amplification library SEQUENCING!

6 Sequencing

7 Mapping RNAseq data * Mapping back to the reference genome iden6fies the expressed genes. * Gene structure (exon/intron) has to be taken into account

8 Transcriptome analysis by RNAseq

9 File formats fastq: sequence data and quali6es SAM/BAM: mapping data and quali6es

10 Some Figures How much does it cost (computa6onally) to sequence a human transcriptome? One human transcriptome: 100 Million reads 1 Solexa run ==8 lanes ==25 M reads/lane==2 x 4 G fastq/lane (PE) processor 12 cores, 48 GB RAM, 4TB disk 32 G disk space 24 hours SAM (Ascii) / BAM (Binary) output 36 G / 9 G

11 Applications of RNAseq Qualitative: * Alterna6ve splicing * An6sense expression * Extragenic expression * Alterna6ve 5 and 3 usage * Detec6on of fusion transcripts. Quantitative: * Differen6al expression * Dynamic range of gene expression. Tophat/Cufflinks Scripture Alexa edger DESeq bayseq NOISeq

specificity * Find novels splicing sites * Larger dynamic range * Detects

microarrays * Restricted to probes on array * Needs genome knowledge * Normally,

12 Advantages of RNAseq? RNAseq * Non targeted transcript detec6on * No need of reference genome * Strand specificity * Find novels splicing sites * Larger dynamic range * Detects expression and SNVs * Detects rare transcripts. microarrays * Restricted to probes on array * Needs genome knowledge * Normally, not strand specific * Exon arrays difficult to use * Smaller dynamic range * Does not provide sequence info * Rare transcripts difficult. and. are there any disadvantages?????

13 RNAseq quantification The number of reads mapped to a gene is a quantification of its expression IGV vew of algu expression in Pseudomonas aeruginosa

14 RNAseq quantification: RPKM To estimate expression value of a transcripts the number of mapped count needs to be normalized by the length of the transcript and the total number of reads, or library size. RPKM: Reads per Kilobase of exon model per Million reads 20 M. reads 27 M. reads Length CondiFon 1 CondiFon 2 RPMK1 RPKM2 Fold- change Gene nts Gene nts

15 Surprises of RNAseq data Positive correlation between expression level and transcript length. Also with RPKM!!! Correlation between gene expression and length raw_data RPKM Marioni MAQC Griffith Equal transcript distributions between samples do not always hold

16 Surprises of RNAseq data With RNAseq, there is a relationship between the chance that a gene is declared differentially expressed and its length RNAseq microarrays

17 RNAseq and sequencing depth * Amount of reads sequenced in a RNAseq experiment * More sequencing depth Rare genes detected Better estimation of expression * How does SD affects gene detection and differential expression? * How many reads do I have to generate to saturate the system?

18 Saturation in RNAseq MARIONI: Solexa 5 lanes Kidney vs liver 22 million reads MAQC: Solexa 7 lanes Brain vs UHR 45 million reads Griffith: Solexa 22 lanes 2 cancer lines 200 million reads

19 Saturation in RNAseq Saturation Curves and New Detection Rates (NDR) Saturation does not seem to be reached even in large datasets!!!

20 Saturation per transcript biotype MAQC (45 M) Important expression of Pseudogenes, processed_transcripts and lincrnas

21 Saturation per transcript biotype Griffith (200 M) Off-target RNA species increase at high sequencing depths

22 Sequencing depth affects dataset transcript distribution For differential expression comparing samples should have similar library sizes.

23 Expression levels increase with sequencing depth at different rates Few, abundant RNA species sneak into sequencing output!

24 Sequencing depth influences the length of detected transcripts * As more it is sequenced, small genes are easier detected. * Still, RNAseq is biased towards longer genes

25 RNAseq & differential expression * Robinson and Smith (2007, 2008, 2010): edger Exact test based on negative binomial distribution. * Marioni et al. (2008): Likelihood ratio test based on Poisson model. * Anders and Huber (2010): DESeq Exact test based on negative binomial distribution. * Srivastava and Chen (2010): Gpseq Likelihood ratio test for two-parameter generalized Poisson model. * Wang et al. (2010): DEGseq (MATR & MARS) MA-plots based methods, assuming normal distribution for M A. * Hardcastle and Kelly (2010): bayseq Empirical Bayesian method to compute posterior probabilities of models, based on Poisson or Negative Binomial data distribution.

26 RNAseq & differential expression * Robinson and Smith (2007, 2008, 2010): edger Exact test based on negative binomial distribution. * Marioni et al. (2008): Likelihood ratio test based on Poisson model. * Anders and Huber (2010): DESeq * Parametric assumptions Exact test based on negative binomial distribution. * Need of replicates * Srivastava and Chen (2010): Gpseq Likelihood ratio test for two-parameter generalized Poisson model. * Unstable with low expressed genes * Wang et al. (2010): DEGseq (MATR & MARS) MA-plots based methods, assuming normal distribution for M A. * Hardcastle and Kelly (2010): bayseq Empirical Bayesian method to compute posterior probabilities of models, based on Poisson or Negative Binomial data distribution.

27 NOISeq * No parametric assumptions. No need of replicates. * Statistics for each gene, exon, transcript, tag, etc.: M = log 2 (expression in condition 1 / expression in condition 2) D = expression in condition 1 expression in condition 2 * Noise distribution: M-D null distribution estimation.! NOISeq-real: uses available replicates! NOISeq-sim: simulates replicates from a multinomial distribution with probabilities derived from the counts in the samples

28 NOISeq Probability for a gene of being differentially expressed (deg): Computed by comparing M-D values of that gene against noise distribution A gene is declared as deg if this probability is higher than 0.8

29 Differential expression vs Sequencing depth * edger, DESeq and bayseq, d.e.g. depend on sequencing depth * FET and NOISeq are constant

30 Differential expression by biotype

31 Sequencing depth & characteristics of selected genes NOISeq is robust to the length of detected genes, the foldchange of differential expression and the mean expression level

32 False discoveries at high sequencing depth Parametric methods tend to identify more false positives (up to 70%) as more it is sequenced. NOISeq controls FDR

33 Normalization by length (RPKM) maintain sequencing depth biases d.e.g. length True positives False positives

34 More understanding of RNAseq data Sequencing depth affects the composition of the RNASeq dataset Short transcrips are in disadvantage Most parametric RNAseq d.e. methods tend to overdetection as large library increases.

35 More understanding of RNAseq data Sequencing depth affects the composition of the RNASeq dataset Short transcrips are in disadvantage Most parametric RNAseq d.e. methods tend to overdetection as large library increases. NOISeq takes a non-parametric approach that better adapts to the noise with large reads numbers. NOISeq is robust to sequencing depth biases.

36 More understanding of RNAseq data Sequencing depth affects the composition of the RNASeq dataset Short transcrips are in disadvantage Most parametric RNAseq d.e. methods tend to overdetection as large library increases. NOISeq takes a non-parametric approach that better adapts to the noise with large reads numbers. NOISeq is robust to sequencing depth biases. Identification of low expression genes with RNASeq is possible but differential expression assessment remains difficult.

37 And more to do real RNASeq More/faster/robuster computer resources are needed to do large RNASeq analyses: Dynamics of gene expression Mul6factorial experiments Popula6on transcriptomics..

38 Acknowledgements Sonia Tarazona Fernando Garcia Aaron Weimann Stefan Götz Samuel Martín David Jovaní Dpto. Bioinformatics CIPF Ximo Dopazo

1. Introduction Gene regulation Genomics and genome analyses

1. Introduction Gene regulation Genomics and genome analyses 2. Gene regulation tools and methods Regulatory sequences and motif discovery TF binding sites Databases 3. Technologies Microarrays Deep sequencing