Researching Cancer with the minION: Methylation and ... · Researching Cancer with the minION: Methylation and Structural Variation ... to neural network based basecalling ... Collaboration

Researching Cancer with the minION:

Methylation and Structural Variation

Winston Timp

Department of Biomedical Engineering

Johns Hopkins University

Agilent Webinar

November 2017

For Research Use Only. Not for use in diagnostic procedures.

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

Deamer et al 2016, Nature Biotech

Oxford Nanopore Google Hangout March 2016

ATCGATCGATAGTAT

TAGATACGACTAGC

GATCAG

Disclosure: Timp has two patents (US 2011/0226623 A1; US2012/0040343 A1) licensed to ONT


3

• 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

Nanopore Sequencing Workflow


4

• Protein nanopores on a synthetic polymer

• Multiple base-pairs at a time (“k-mers”)

• Characteristic current signature is converted to nucleotide sequences

Sequencing Operation

Oxford Nanopore Technologies


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.For Research Use Only. Not for use in diagnostic procedures.

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) 3376-3393 (2012)For Research Use Only. Not for use in diagnostic procedures.

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 notFor Research Use Only. Not for use in diagnostic procedures.

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, 349-353 (2012)For Research Use Only. Not for use in diagnostic procedures.

9

• 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.

• 1st 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

dk P(I(t) | ktt

Õ )´T(t-1)(t )

Nanopore HMM basecalling


10

Basecalling shifting to RNN

Oxford Nanopore

• 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.


11

Assemblies

SPAdes (Illumina only) Canu (Nanopore only)

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 onlyFor Research Use Only. Not for

use in diagnostic procedures.

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


13

Nanopolish tools

Consensus Calling Reference-based SNP Calling

Methylation Detection Read Phasing

…

TAGAAGATATCATGTATAGTACGATTAGAAGATATCATGTAGCAGATATCATGTATATTACGAT

CATGTATATTACGAT

github.com/jts/nanopolish


http://github.com/jts/nanopolish

14

Nanopolish SNP Calling and Genotyping

-332-300-350

…

hidden Markov

model

…ACTACGATCGAC……ACTACGATCGAC…

…ACTACCATCGAC……ACTACCATCGAC…

…ACTACGATCGAC……ACTACCATCGAC…

input: pairs of

haplotypes

0/1

output:

genotype call


15

Human Genotyping Results

Genotype accuracy at all sites: 99.2%

Genotype accuracy at variable sites: 94.8%

Jain et al bioRxiv 128835For Research Use Only. Not for use in diagnostic procedures.

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. 2012. Nature MethodsFor Research Use Only. Not for use in diagnostic procedures.

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


18

Targeted Capture Optimization

Collaboration with Josh Wang from Agilent

• 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


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:

https://goo.gl/8V2Fei

Total yield

(reads)On-target

On-target

percentageFold enrichment Coverage

Illumina NA12878 4.4m 3.7m 85% 641X 113X

Nanopore

NA12878107k 32k 30% 353X 27X

Nanopore PDAC 56k 20k 26% 332X 20X


https://goo.gl/8V2Fei

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)For Research Use Only. Not for use in diagnostic procedures.

21

• 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)For Research Use Only. Not for use in diagnostic procedures.


Research

22

• 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)For Research Use Only. Not for



Research

23

Single Nucleotide Variation Detection in Cancer Research

Number of True SNVs: 3587(Eberle,et al. bioRxiv, 2016)

Illumina Pre-polish Post-polish

Avg.

Coverage 113 27 27

Correct 1133 2485 947

Total 1211 4138 1017

Precision 94% 60% 93%

Sensitivity 32% 69% 26%

Phased SNV analysis is possible with

coverage from targeted sequencing


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)


25

Epigenetics: Modern

• 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

For Research Use Only. Not for


26

Single Read Methylation: Distribution

Number of reads

per pattern

Number of reads

per pattern

• 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 (t-

test).


27


Number of reads

per pattern

Number of reads

per pattern











samples.




test).


28


Number of reads

per pattern

Number of reads

per pattern











samples.




test).


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)For Research Use Only. Not for use in diagnostic procedures.

30

Nanopore: Methylation

Schreiber, et al. PNAS. (2013)Laszlo, et al. PNAS (2013)

• 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.


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)


32

Nanopore: Methylated Error•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

Simpson, Workman, et al. Nature Methods (2017)


33

Emission ProbabilitiesS

impson

, Work

man, e

t al. N

atu

re M

eth

od

s (2

01

7)

• 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.


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)


35

Nanopore: nanopolish methyltrain

• Multiple bases influence the current

passing through the pore.

• Current basecallers use a neural-

network based methodology to call

bases.

• We currently use a HMM based

classifier to call methylation

• With nanopolish we can call the

probability:

• Where Sm is the probability

methylated for a given observable D

and Sr the probability unmethylated

• We then take the log of this likelihood

ratio.dk P(I(t) | kt

tÕ )´T(t-1)(t )


36

NA12878 Methylation

• NA12878 (lymphoblast) gDNA: Illumina WGBS on X-axis (24X coverage) (SRA:

GSM1002650) 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

Sim

pson

, W

ork

man, N

atu

re M

eth

od

s (

201

7)

R7.3 R9

For Research Use

Only. Not for use in

diagnostic procedures.

37

Nanopolish Methylation

Jain et al bioRxiv 128835

Methylation comparison between bisulfite and nanopore calling

In whole genome


38

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

Sim

pson

, W

ork

man, N

atu

re M

eth

od

s (

201

7)


39

Cancer-Normal Comparison

Simpson, Workman, Nature Methods (2017)For Research Use Only. Not for


• 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

40

Haplotype-Phased Methylation

nanopolish has experimental support for phasing methylation patterns

this haplotype is highly methylated


41

Haplotype-Phased Methylation

nanopolish has experimental support for phasing methylation patterns

this haplotype isn’t


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).


43

Acknowledgments

• Timp Lab – Johns Hopkins

University

• Winston Timp, PhD

• Rachael Workman, MS

• Stephanie Hao, MS

• Yunfan Fan

• Isac Lee

• Amy Vandiver, MD, PhD

• Ontario Institute for Cancer

Research

• Jared Simpson, PhD

• P.C. Zuzarte, PhD

• Matei David, PhD

• L. J. Dursi, PhD

• Eshleman Lab – Johns

Hopkins School of Medicine

• James Eshleman, MD, PhD

• Alexis Norris, PhD

Agilent Technologies

Josh Wang, PhD

Jonathan Levine, PhD

1R01HG009190-01A1

• Simner Lab – Johns Hopkins

School of Medicine

• Patricia (Trish) Simner, PhD

• Belita Opene

• Annie Antar

• Yehudit Bergman


44

Other mod detection


Documents

Researching Cancer with the minION: Methylation and ... · Researching Cancer with the minION: Methylation and Structural Variation ... to neural network based basecalling ... Collaboration