8/7/2012. Experimental Design & Intro to NGS Data Analysis. Examples. Agenda. Shoe Example. Breast Cancer Example. Rat Example (Experimental Design)

Transcription

1 Experimental Design & Intro to NGS Data Analysis Ryan Peters Field Application Specialist Partek, Incorporated Agenda Experimental Design Examples ANOVA What assays are possible? NGS Analytical Process Alignment of NGS Data Challenges of NGS Analysis Partek Flow Demonstration 2 Examples Shoe Example Breast Cancer Example Rat Example (Experimental Design) Tips on setting up your next experiment 1

2 The Role of Experimental Design The goal of statistics is to find signals in a sea of noise The goal of experimental design is to reduce that noise so true biological signals can be found with as small a sample size as possible Partek Shoe Example Question: Do shoes affect height? Hypothesis: Yes, shoes affect height. Assay: Measure the height 10 people with & without shoes. (Change only one variable.) Sample Size: 10 Partek (5 male, 5 female) Analysis: Use a two sample t-test to see if there is a difference between the mean of two groups: with shoes and without shoes. t-test A simple t-test does not have the power to correctly identify this pattern, because it assumes multiple samples from the same individual are independent when they are not. p= 0.51 Fold-change = 1.02 Conclusion - No statistically significant difference in height due to shoes. 2

3 Paired t-test The paired t-test provides substantially more statistical power by removing person-to-person differences from the noise. p(shoes)=1e-5 p(person)=2e-9 Introducing Gender Once person is known, gender is already known; thus the p-value for Shoe remains unchanged. We get the estimate of gender effect for free! Add Gender (3-way ANOVA) p(shoes)=1e-5 p(gender)=.04 p(person)=2e-9 It appears (p=.04) that men (at Partek) are significantly taller than women 3

4 Explore Gender/Shoe Interaction Do shoes have the same effect on men & women? p(shoes)=1e-8 p(gender)=.04 p(person)=2e-12 p(shoe*gender) =7e-5 Wow! Shoes affect women s height more than men s! Also note that p-values for shoe effect are even smaller because we explained more noise. Breast Cancer Example Example of Large Batch Effect Example Data, GEO Experiment GSE848 Control (E2) Plus Drug Treatment of Breast Cancer Cells 5 Treatments x 3 Time Points x 2 replicates Biological replicates were processed in 2 batches Control Estrogen (E2) E2 + ICI E2 + Raloxifene E2 + Tomoxifen 0 hr 2 8 hr hr Fortunately, treatments were perfectly balanced across processing batches. 4

5 As Seen Using PCA As Seen Using Hierarchical Clustering What is Analysis of Variance? Analysis (Source: m-w.com) Etymology: New Latin, from Greek, from analyein to break up separation of a whole into its component parts 17.49% 1.15% 17.40% 58.36% Treatment Time 1.64% Analysis of Variance ANOVA a technique that partitions the variance in data into separate components or factors 5

6 Good News! Balanced Experimental Design The treatments were perfectly balanced with the batches, so batch can be included as a blocking factor in ANOVA, and the batch effect (noise) can be removed from the data. In terms of p-values for this gene, the difference is dramatic. With a simple 2-way ANOVA, this gene was #228 on the gene list and would not pass multiple test correction for significance. With a 3-way ANOVA including batch, it was #2 on the gene list. Factor 2-way ANOVA 3-way ANOVA Treatment E-07 Time Treatment*Time E-05 #2 Most Significant Gene Monday Median A =8.5 Median B =9.7 Tuesday Tue vs. Mon more than 2-fold difference ANOVA Partitions Variability Total variance is partitioned into variability due to influencing factors and the rest is assumed to be due to random error (noise). R 2 =81% for 2-way ANOVA R 2 =99% when Batch included 6

7 Batch Effect Remover Before Batch Removal After Batch Removal 19 Batch Effect Remover For visualization purposes only! Factors you would normally add for ANOVA How do we account for batch without Partek Batch Remover? 20 Building Blocks of Experimental Design No Randomization Completely Randomized Subjects randomly assigned to treatment groups Randomized Block Subjects randomly assigned to treatment groups within similar blocks (e.g. gender, litters) Requires a priori knowledge of differences between the blocks 7

8 Simplest Design: Not Randomized 8 Male Rats 4 Treated 4 Control Stripe coated rats are faster or more alert. Completely Randomized 8 Male Rats 4 Treated 4 Control A Better Approach Randomized Block Design First divide into blocks, then randomly assign to treatment groups 8

9 Randomized Block Design 8 Male Rats 4 Treated 4 Control Technical Blocks in Microarray Experiments Litter is an example of a biological block Examples of Technical/Processing Blocks: RNA Isolation Batch Hybridization Batch Operator As well as (although less so) Wash and Stain Batch Reagent, Cocktail Batches Chip Lot In Summary Block what you can and randomize what you cannot. Box, Hunter, & Hunter (1978) Blocking ensures that the differences in treatment cannot possibly be due to the blocking factor Blocking completely eliminates noise due to blocks Randomization gives approximate balance across other variables unaccounted for 9

10 Analysis of Variance Also Known As: ANOVA ANCOVA Linear Model Mixed Linear Model Invented in 1900, 1908, 1923 Still remains the most commonly used statistical method to analyze clinical trials! Simple ANOVA: Student s t-test t and F Statistics Fun fact In equal variance t-test is mathematically equivalent to a 1-way ANOVA. Student/Gosset Fisher 10

11 Assumptions of ANOVA Data is Normally distributed (bell shaped) within different treatment groups Ensure data is log transformed Variance is equal within different treatment groups Design balanced experiments Samples groups are independent. Don t make the shoe mistake *Replicates Required to get p-value Random vs Fixed Effects If the experiment were to be performed again, would the same levels of the factor be used? Yes - Fixed effect (e.g. gender, dose, time, dye) No - Random effect (e.g. hyb batch, wash batch, litter, subject) Why do I have to worry about this? In general, treating a random effect as a fixed effect will produce an overoptimistic p-value, leading to a false discovery. What Factors Belong in the Model? Obviously, the factors of interest to the researcher e.g. strain, time, strain*time Any factor needed to account for dependence of samples (don t violate assumption of independence!) e.g. donor Any additional blocking factors for noise reduction e.g. batch 11

12 Partek Expression Philosophy Use PCA to aid in quality control & sample grouping Use ANOVA to detect significantly expressed genes. Fold change is interesting for ranking, but not a great primary filtering metric Incorporate as much phenotypic and experimental design information into the ANOVA model as possible. Measure the experimental technical components.* Make sense of gene lists through functional groups How NOT to Run/Ruin Your Next Experiment! Samples are frequently organized by treatment groups. Samples are then processed in batches corresponding to treatment groups. But please do NOT process your control samples on Monday, and then process your treated samples on Tuesday. You will confound these two variables. ANOVA is powerful but not magical. Summary Experimental Design & Analysis Understand how separating variables in your analysis is critical to your success Design balanced experiments. Let p-values rank your data, but don t be a slave to FDR. 12

13 What kinds of assays are possible? DNA-Seq Copy Number SNP Structural variants Whole genome sequencing Metagenomics Targeted/Amplicon Sequencing ChIP-Seq Transcription Factor binding sites Methylation sites Histone modifications RIP-Seq (RNA-binding proteins) RNA-Seq Transcriptome Differential Gene Expression Alternative Splicing SNP detection Indel detection Novel exons/genes mirna-seq identify regulatory (non-coding) RNAs NGS Analysis Phases Primary Analysis Secondary Analysis Data File (Reads + Quality) Tertiary Analysis Reads aligned to genome FASTQ, BAM, Control Software Bowtie/BIOSC OPE/BWA, etc. Data File (Reads + Quality) FASTQ Reads aligned to genome Modified from Strand Life Sciences 38 NGS Analytical Process Illumina HiSeq SOLiD Roche 454 Ion Torrent PacBio Sequencing Genome Alignment GAGGTTGCAGTTTG chr R ACTGCTCCGCCTCA chr F GAATAAAAAATCCA chr F CGTCCTTCACCCTCT chr R CCTTAAGGAAAGGA chr F CAGCTAGGGTTGCC chr R CTGCTGGTGCTGCG chr F QC & Exploratory Analysis Powerful Statistics Intuitive Visualizations Integrated Genomics Biological Interpretation Publication 39 13

14 Comprehensive Analysis of NGS Data DNA-Seq SmallRNA- Seq RNA-Seq Methylation Seq ChIP-Seq 40 Read Types for NGS Single End Reads Paired-end Reads Junction Reads Multiple Aligned reads Strand-specific reads Paired End & Single End Reads DNA Space Single End Paired End chr2 DNA Space chr5 Multiple aligned 42 14

15 Junction Reads Derived from transcripts, some RNA-Seq reads will read through splice junctions (single end or pairedend) They will not align well to genomic reference since the two ends are many nucleotides apart (separated by the intronic region) DNA Space 43 Next Gen File Formats Unaligned (FASTA, FASTQ, SCARF, QSEQ, SRA, RAW, TXT, others) Alignment Tools ELAND BFAST Bowtie TMAP BWA TopHat SOAP Etc. GAGGTTGCAGTTTG chr R ACTGCTCCGCCTCA chr F GAATAAAAAATCCA chr F CGTCCTTCACCCTCT chr R CCTTAAGGAAAGGA chr F CAGCTAGGGTTGCC chr R CTGCTGGTGCTGCG chr F Aligned: SAM, BAM, Vendor Specific Formats/Color Space Variant Call File (VCF, BCF) SNPs, indels 44 What to expect? Cluster Cloud Laptop File size depends on read length, read type 4GB single lane (~100 million reads) Bowtie w/ 8 cores = 20/25 minutes; reference genome - read length = 33bp (older) TopHat same file 1 day Read length x number of reads x 8 = file size (fasta, double for fastq) BAM file ~ 3-4x smaller than unaligned file 15

16 FASTQ Format(Unaligned GATTTGGGGTTCAAAGCAGTATCGATCAAATAGTAAATCCATTTGTTCAACTCACAGTTT +!''*((((***+))%%%++)(%%%%).1***-+*''))**55CCF>>>>>>CCCCCCC65 Line 1) begins with a '@' character and is followed by a sequence identifier and an optional description (like a FASTA title line). Line 2) is the raw sequence letters(acgt). Line 3) begins with a '+' character and is optionally followed by the same sequence identifier (and any description) again. Line 4) encodes the quality values for the sequence in Line 2, and must contain the same number of symbols as letters in the sequence Sanger format can encode a Phred quality score from 0 to 93 using ASCII 33 to What is Alignment? Read comes off a sequencing machine A T G G T C A Goal: Determine where on the genome that read belongs Method: Match sequence of read to sequence from a reference genome (reference G G C A T G G T C A T T C genome) (read) A T G G T C A Result: Genomic Location of read 47 Align junction reads Gene/Transcript Exon junction (Reference Genome) DNA Space G A T G C A C G G A T T G T C A T RNA Space A T G G T C A (Read) 1) Align to Genome gapped alignment time expensive -breaks up read in pieces (25mer) 2) Align to transcriptome lose genomic context! 48 16

17 SAM Format (Aligned reads) Sequence Alignment/MAP (SAM) format is TAB-delimited BAM is binary SAM M-match Position Header line I-Insertion position of mate D-Deletion sequence length quality (header) (Reference Sequence) CIGAR Read id Bitwise flag Quality score Reference genome Reference name of mate optional Explain flags 50 VCF Format (Variant Call Format) The Variant Call Format (VCF) is a TAB-delimited format with each data line consists of the following fields: Chromosome, Position, variant id, reference/alternative alleles, quality, information(read depth), event, sample Id (optional), format (optional) 51 17

18 Partek Flow Web based Application Cloud, Desktop, Server Chrome, Firefox, Safari Access from any terminal, smartphone Project centric Protocols Collaborate with others Current release 1.0 / 2.1 beta Alignment, QA/QC, GSA Export results to PGS Coming soon SGE, 52 Challenge: Data volume is a bottleneck Help, I m drowning in data! How do I handle all this data? Solution: Schedule Tasks Schedule & Queue tasks s you when tasks are complete Keep your hardware running 24/7/365 18

19 Challenge: The quality of the data will affect the alignment How do I determine data quality? Do I have outliers? Can I move forward with my analysis? Do I need to trim/filter my reads? 55 Solution: Pre & Post Alignment QA/QC Group and individual QA/QC for excluding outliers Quality score per read/position Look for drop in quality scores Make intelligent decisions for trimming/filtering adaptors, barcodes, low quality reads 56 Challenge: Alignment Different people, different parameters will result in different alignment. Which aligner to use? Some aligners have more than 50 different options. How do I know what to set? What options do I choose for RNA-Seq, ChIP-Seq, DNA-Seq, mirna-seq, MeDip-Seq? What options do I choose for the different read types? Junction reads? Paired-End reads? Multiple Aligned reads? 57 19

20 Solution: Multiple aligners with recommended defaults Vendor Specific default options Automatic Download of reference genomes Assay specific default options (RNA- Seq, ChIP-Seq, DNA- Seq) Advanced options also available through GUI Interface (no command line) 58 Challenge: How do I keep track of my samples? Which samples are Tumor? Control? Age? Sex/Gender? How am I ever going to keep track of this clinical information? 59 Solution: Advanced Sample Management Manage files associated with sample throughout life of project Keep track of reference genome Controlled vocabulary SNOMED List In-place editing of sample info 20

21 Plug-in for Torrent Suite Perform QA/QC within Torrent Suite and seamlessly upload data to Partek Flow for Comprehensive Data Analysis and Visualization Performs QA/QC within Torrent Suite Uploads data to Partek Flow Comprehensive Solution for RNA-Seq Alignment Mapping QC Statistics Visualization Integrated Genomics Biological Interpretation Acknowledgements Partek Flow Demonstration 63 21