Researching Cancer with the minion: Methylation and Structural Variation

Transcription

1 Agilent Webinar November 2017 Researching Cancer with the minion: Methylation and Structural Variation Winston Timp Department of Biomedical Engineering Johns Hopkins University

2 Nanopore: Single Molecule Sequencing Oxford Nanopore Technologies, CsgG biological pore No theoretical upper limit to sequencing read length, practical limit only in delivering DNA to the pore intact Palm sized sequencer Predicted sequencing output 3-6Gb ATCGATCGATAGTAT TAGATACGACTAGC GATCAG Oxford Nanopore Google Hangout March 2016 Deamer et al 2016, Nature Biotech Disclosure: Timp has two patents (US 2011/ A1; US2012/ A1) licensed to ONT 2

3 Nanopore Sequencing Workflow Four steps to generating usable data with nanopore sequencing Base-calling : the process of converting raw signal into nucleotide sequences Nanopolish : uses alignment and current signal to improve base-calls 3

4 Sequencing Operation Oxford Nanopore Technologies Protein nanopores on a synthetic polymer Multiple base-pairs at a time ( k-mers ) Characteristic current signature is converted to nucleotide sequences 4

5 Nanopore Library Prep Library prep is very similar to methods for short-read sequencing For DNA shearing we used Covaris gtubes or Diagenode Megaruptor After end-repair and A-tailing, leader adapter with motor protein is ligated MinION arrays 512 channels (with 4 pores possible per channel) (shown bottom left from running software); dark green pores are sequencing, light green available, other colors inactive. 5

6 Problems with Nanopore basecalling Multiple bases influence the current passing through the pore. Through simulation with Brownian Dynamics, we calculated the contribution from triplets of DNA in a solid-state nanopore - 64 current levels. Not all of these different currents are distinguishable Comer and Aksimentiev J. of Phys Chem C 116(5) (2012) 6

7 Hidden Markov Model Markov chains represent a series of states occurring in sequence, with defined transition probabilities. In a hidden Markov model the state is only observed indirectly. As an example, consider inferring the weather (without going outside) by observing whether people coming in from outside are wearing gloves or not 7

8 Prior Information for Decoding With no prior information, a given current value may not be called correctly (333pA would be called as GGG) If we know the previous triplet, the next triplet is well defined, leaving only four possibilities, resulting in the correct call of TCG Timp, et al. Biophys J. 30, (2012) 8

9 Nanopore HMM basecalling By using a sequence of observables and maximizing the total joint probability given below, we find the sequence of states. This is done using the Viterbi algorithm which grows, finding the most likely path for each step, saving the probabilities, to avoid recalculation. 1 st generation basecallers from Oxford used a HMM for basecalling similar to the one detailed in our Biophysical paper Transistion probablility matrix for oxford seems to allow for a 0, 1 (most common), 2, or 5 (reset) move. We think that Oxford trained its basecalling model on unmethylated lambda d k Õ t P(I(t) k t ) T (t-1)(t) 9

10 Basecalling shifting to RNN Recently (~6 months) there has been a shift to neural network based basecalling A recurrent neural network is still one with memory, that has a dependence on past computations Specifically two layers of Bidirectional Long Short Term Memory (BLSTM) These still require the same training data to learn what current distributions correspond to which k-mers and the results are still k-mer based, as multiple bases still influence the current. Oxford Nanopore 10

11 Assemblies SPAdes (Illumina only) Canu (Nanopore only) For Research Use Only. Not for use in diagnostic procedures. Long reads really help in getting complete assemblies using twice as many Illumina reads, the N50 is still ~0.3 Mb for Illumina only, while it s full length chromosome from nanopore only 11

12 Pilon Correction Though ONT error rate is improving, still not sufficient from raw basecalled for SNPs Application of Illumina data via pilon allows for correction Multiple rounds of correction needed initially Illumina reads won t align correctly to error prone scaffold Next we are trying to apply nanopolish which returns to the electrical signal to correct nanopore only assemblies 12

13 Nanopolish tools Consensus Calling github.com/jts/nanopolish Reference-based SNP Calling Methylation Detection Read Phasing TAGAAGATATCATGTATAGTACGAT TAGAAGATATCATG TAGCAGATATCATGTATATTACGAT CATGTATATTACGAT 13

14 Nanopolish SNP Calling and Genotyping input: pairs of haplotypes ACTACGATCGAC ACTACGATCGAC ACTACGATCGAC ACTACCATCGAC ACTACCATCGAC ACTACCATCGAC hidden Markov model /1 output: genotype call 14

15 Human Genotyping Results Genotype accuracy at all sites: 99.2% Genotype accuracy at variable sites: 94.8% Jain et al biorxiv

16 Structural Variation Defined as an abnormality in large region (50b-3mb) of a chromosome Pervasive in cancer 50% of pancreatic ductal adenocarcinoma (PDAC) have SVs Common in tumor suppressor genes such as CDKN2A and SMAD4 Short-read sequencing has difficulty resolving SVs, but nanopore sequencing long reads can stretch across them. High coverage needed to detect, but yield from nanopore sequencing still relatively low per flowcell (~3-5Gb) Baker, Monya Nature Methods 16

17 Solution-phase Hybridization Capture Use of Agilent SureSelectXT Targeted Sequencing System in cancer research ~90 bps biotinylated RNA probes complementary to target sequence Biotin-streptavidin interaction to enrich for the targeted region Optimization for long-reads : > 2 kb 17

18 Targeted Capture Optimization Trial 1 Probe tiling, No empty spaces between probes Target region CDKN2A : 1.5 Mbp Low stringency to allow mismatches Result: 2.28 % on-target Trial 2 No tiling, average 400 bp space between probes Target regions CDKN2A : 1.5 Mbps SMAD4 : 850 Kbps High stringency to limit off-target capture Consideration of known SV breakpoints PDAC SVs from James Eshleman lab Result: 30 % on-target Collaboration with Josh Wang from Agilent 18

19 Targeted Sequencing Performance Control : NA12878 lymphoblast Sample : PDAC from Eshleman lab Illumina short-read targeted sequencing for comparison > 300-fold enrichment > 20X average coverage Agilent App Note: Total yield (reads) On-target On-target percentage Fold enrichment Coverage Illumina NA m 3.7m 85% 641X 113X Nanopore NA k 32k 30% 353X 27X Nanopore PDAC 56k 20k 26% 332X 20X 19

20 Nanopore Structural Variation Detection in Cancer Research NA12878 (ENCODE Human lymphoblast cell line) SVs detected with Sniffles (Schatz lab) chr9:21,038,354-21,038,506; 152 bps duplication Validated with PacBio data from Genome in a Bottle (Mt. Sinai School of Medicine) w/josh Wang (Agilent) & Jim Eshleman (JHMI) 20

21 Nanopore Structural Variation Detection in Cancer Research PDAC cell line (Eshleman): Novel, putative SVs detected from PDAC chr18: 51,198,535 51,199,143; 600 bps deletion Possibly allele-specific SV w/josh Wang (Agilent) & Jim Eshleman (JHMI) 21

22 Nanopore Structural Variation Detection in Cancer Research For Research Use Only. Not for use in diagnostic procedures. Large window of coverage; Absence in CDKN2A region chr9:21,950,000-22,436,000; 486 kbps SV Homozygous Deletion previously identified with Illumina data Sniffles did not detect this SV; No reads covering either breakpoint, unlucky coincidence of probes w/josh Wang (Agilent) & Jim Eshleman (JHMI) 22

23 Single Nucleotide Variation Detection in Cancer Research Phased SNV analysis is possible with coverage from targeted sequencing Illumina Pre-polish Post-polish Avg. Coverage Correct Total Precision 94% 60% 93% Sensitivity 32% 69% 26% Number of True SNVs: 3587(Eberle,et al. biorxiv, 2016) 23

24 Epigenetics: Historical The historical definition of epigenetics, by Waddington, is how genotype interacts with the environment to create phenotype. The analogy is the rolling ball in the image to the right the genotype has set up the hills, environment nudges the ball into one valley or another, then the ball is, to mix metaphor, lineage committed. Which valley the ball rolls into is not predetermined, but a stochastic behavior. Waddington, 1940; Raser et al. Science (2005) 24

25 Epigenetics: Modern For Research Use Only. Not for use in diagnostic procedures. Modern Definition of epigenetics involves heritable changes other than genetic sequence, e.g., positive feedback, high order structure, chromatin organization, histone modifications, DNA methylation. An analogy to a computer system: DNA Sequence = Hardware User input = Environment Systems Biology = Running programs Epigenetics = RAM 25

26 Single Read Methylation: Distribution We performed hybridization capture, then Illumina bisulfite sequencing on 6 paired colon cancer and normal samples. We then examined methylation patterns *within reads* and looked at the distribution in normal vs. cancer samples. Colors in the stacked bar graph represent different sequenced samples. Areas are clusters in regions which show significantly different methylation levels (ttest). Number of reads per pattern Number of reads per pattern 26

27 Single Read Methylation: Distribution We performed hybridization capture, then Illumina bisulfite sequencing on 6 paired colon cancer and normal samples. We then examined methylation patterns *within reads* and looked at the distribution in normal vs. cancer samples. Colors in the stacked bar graph represent different sequenced samples. Areas are clusters in regions which show significantly different methylation levels (ttest). Number of reads per pattern Number of reads per pattern 27

28 Single Read Methylation: Distribution We performed hybridization capture, then Illumina bisulfite sequencing on 6 paired colon cancer and normal samples. We then examined methylation patterns *within reads* and looked at the distribution in normal vs. cancer samples. Colors in the stacked bar graph represent different sequenced samples. Areas are clusters in regions which show significantly different methylation levels (ttest). Number of reads per pattern Number of reads per pattern 28

29 Nanopore: Methylation Cytosine methylation increases DNA stiffness, decreasing stretching force This may allow for filtering of methylated versus unmethylated DNA using a pore Mirsaidov, et al. Biophys J. (2009) 29

30 Nanopore: Methylation Differences between methylated and unmethylated cytosine have been detected using nanopores. Methylation state can be called with 90% accuracy. We have implemented a classifier for mc on using Oxford Nanopore signals. Laszlo, et al. PNAS (2013) Schreiber, et al. PNAS. (2013) 30

31 Generation of methylated Samples To generated methylated samples, we treat unmethylated DNA (lambda, dam-/dcm- E. Coli, PCR product) with M. SssI methyltransferase We confirm the CG specific methylation using Illumina bisulfite sequencing of the sample pictured right is methylation in different contexts E. Coli dataset treated with M. SssI (red) versus untreated (green) 31

32 Nanopore: Methylated Error Simpson, Workman, et al. Nature Methods (2017) We sequenced samples with either SssI or no treatment. Notably, mismatch error rate is higher on methylated samples than unmethylated, though indel rate seems mostly unchanged At CG locations, there is not a clear alternative basecalling, though error is higher; Ts are not substituted for methylated Cs 32

33 Emission Probabilities We measured distributions of current for k-mers from E. Coli M.SssI treated (methylated; green) and untreated (unmethylated; red) samples on two different sets of pores - R7.3 and R9 flowcells. Boxplots of AGGTCG and TCGAGT k- mers which both contain CGs show significant differences in current in some cases (AGGTCG R7.3) and little to none in others (TCGAGT R7.3) R9 current distribution seem wider in both cases, but gives better discrimination in TCGAGT. Simpson, Workman, et al. Nature Methods (2017) 33

34 Distance of methylation effect We looked at the difference in current levels dependent on the position of the methylated base plotted are the current differences for R7.3(blue) and R9 pores(orange). Signal seems again stronger but more variable for R9 pores than R7.3 Methylation can either reduce current or increase it. Some positions are more sensitive to methylation than others. Simpson, Workman, Nature Methods (2017) 34

35 Nanopore: nanopolish methyltrain Multiple bases influence the current passing through the pore. Current basecallers use a neuralnetwork based methodology to call bases. We currently use a HMM based classifier to call methylation With nanopolish we can call the probability: Where S m is the probability methylated for a given observable D and S r the probability unmethylated We then take the log of this likelihood ratio. d Õ k P(I(t) k t t ) T (t-1)(t) 35

36 Simpson, Workman, Nature Methods (2017) NA12878 Methylation R7.3 R9 For Research Use Only. Not for use in diagnostic procedures. NA12878 (lymphoblast) gdna: Illumina WGBS on X-axis (24X coverage) (SRA: GSM ) vs. R7.3 (0.02X) or R9 (0.13X) nanopore sequencing. Correlation of 0.83 (R7.3) and 0.84 (R9) most gene promoters unmethylated 36

37 Nanopolish Methylation Methylation comparison between bisulfite and nanopore calling In whole genome Jain et al biorxiv

38 Simpson, Workman, Nature Methods (2017) Binned Methylation vs. Transcription Start Sites On human genomic samples: Binning methylation levels vs. distance to TSS sites, compared to bisulfite data (NA12878). We also generated completely methylated (M.SssI treated; ~95% meth) and unmethylated used to generate the ROC curve (right) R7.3 91% accurate at 68% of sites R9 94% accurate at 77% of sites 38

39 Cancer-Normal Comparison Reduced representation method:12.5mb of the genome (3.5-6kb size selection) We sequenced this fraction on nanopore and bisulfite Illumina seq Long reads measure phased methylation For Research Use Only. Not for use in diagnostic procedures. Simpson, Workman, Nature Methods (2017) 39

40 Haplotype-Phased Methylation nanopolish has experimental support for phasing methylation patterns this haplotype is highly methylated 40

41 Haplotype-Phased Methylation nanopolish has experimental support for phasing methylation patterns this haplotype isn t 41

42 Summary Nanopore technology is full of potential for sequencing, but always choose the right tool for the right job. Often multiple approaches with complementary data yield the best results. Multiple bases affect the electrical signal from nanopores; rather than a problem, this can be an advantage, as each base is interrogated multiple times. Using a hidden Markov model and Viterbi algorithm, we can decode the electrical signal to a DNA base sequence. Long-read sequencing is great at detection of structural variants in cancer samples used in clinical research when coupled with hybridization capture this can be a powerful approach. Modifications to the primary DNA sequence (e.g. cytosine methylation) can be detected directly using nanopores, as compared to the gold standard of chemical modification (bisulfite treatment). 42

43 Acknowledgments Timp Lab Johns Hopkins University Winston Timp, PhD Rachael Workman, MS Stephanie Hao, MS Yunfan Fan Isac Lee Amy Vandiver, MD, PhD Eshleman Lab Johns Hopkins School of Medicine James Eshleman, MD, PhD Alexis Norris, PhD Simner Lab Johns Hopkins School of Medicine Patricia (Trish) Simner, PhD Belita Opene Annie Antar Yehudit Bergman Ontario Institute for Cancer Research Jared Simpson, PhD P.C. Zuzarte, PhD Matei David, PhD L. J. Dursi, PhD Agilent Technologies Josh Wang, PhD Jonathan Levine, PhD 1R01HG A1 43

44 Other mod detection 44