Role of DNA methylation in the genomic rearrangements in P. tetraurelia

PhD Defense DELEVOYE Guillaume

09/06/2022

Supervisor : Dr MEYER Eric

Jury members : Dr DUHARCOURT Sandra, Dr CHEN Chunlong, Dr DURET Laurent

Qu'est-ce que l'ADN ?

Où se trouve l'ADN ?

A quoi ressemble l'ADN ?

ATCGATGCGGATTCGATCATGCTAGCTGATCGATCGGAAGCTTGACTAGTCGATCGATCGATCGATCGATCGATCTTCTATATATATGCGCGTAGCTAGCTAGCTAGCTATATATGCATAGAGAGCTCGATCGCGCTATCTCCTCTGATCGATCGATCGGGATCGATCGGATCGATGCATTAGGATCGATCGGT

....

A quoi ça sert ?

Entre autres, c'est un livre de recettes de protéines

Certaines recettes... S'autorépliquent

La protéine qui correspond à un bout d'ADN peut parfois copier-coller, ou couper-coller ce bout d'ADN ailleurs

P. tetraurelia: Genomic architecture

Unicellular eucaryote with 3 nuclei:

• 2xMIC nuclei (2n)
  • Germline nucleus
  • Contains: TE & IES
    • No transcription outside meiosis
• 1xMAC nucleus (up to 800n)
  • Somatic nucleus
  • Amplified and "fixed" version of the MIC
  • Free from TE and IES
  • Transcriptionnally active

Profiling the IESs

• Non-coding
• Excised after sexual processes
• Remnant of TE ?
• Most recent IES are very short (< 27bp), and are the majority of IESs

 

45.000 Unique sequences

How are IESs recognized ?

(1/2)

• 100% TA-Bounded
• Weak consensus TAYAG
  • Degenerated TC1-Mariner TE insertion site
  • Not sufficient
• Periodic size distribution

 

30% of IESs only are small-ncRNA dependant (shown by DICER-like2-3 silencing)

 

• What about the majority remaining ?

 

 

 

 

How are IES recognized ? (2/2)

The sc-RNA pathway

The DNA methylation hypothesis

6mA likely to be abundant in Paramecium:

 

• Suspected 2.5% in the MAC of P. aurelia by Cummings et Al (1975)

  • Also documented 6mA in the MIC

  •  

• Detected methylation by SMRT in the MAC in our lab (unpublished data)
   
• Detection by SMRT in the MAC by Sandra Duharcourt's lab (unpublished)
   
• Detection in Oxytrichia by L. Landweber et Al (2019)
  
   
• no 5mC, 4mC in the MAC a priori

 

2) Crosstalks in the new forming MAC

1) Constant pattern in the MIC

Transcient ?

Methylase candidates

Candidates:

• 7 "DPPW motif IV" genes
  • NM4, NM9, NM10
• Another family: MT1a, MT1b, MT2

2) PacBio sequencing with short inserts | 6mA ++

Also at hand:

• Vegetative WT cells (HTVEG)
• Vegetative cells with a silencing of  a control gene (MAB)
• Cells that undergoe autogamy (HT2 - 2 hour , HT6 - 6 hour)

 

1) Grouped silencings by sequence homology

PacBio

TL;DR overview

PacBio SMRT principle

Methylation analysis needs local coverage> 25X

• IPD are captured
• Compared
  • either to WGA
  • or Trained model (ML - "in-sillico control") - SImulation from context

$$log(ipdRatio)= log(\frac{MeanIPD_{experience}}{Model})$$

 

Reads up to 80 kbp

• We use the isn-sillico model
• Nobody does

PacBio SMRT principle (2)

• Strategy 1: Long inserts (long reads)
  • Ideal for assembly of long repeated sequences
  • Poor resolution for DNA methylation analysis
     
• Strategy 2 : Short inserts (long reads)
  • Much higher resolution for DNA methylation analysis

 

Step 1: Consensus

99% accuracy

Max

75% accuracy

Because our inserts are circular and shorts, we can make CCS of high accuracy despite a 15% error-rate

Step 2: Sorting

Deduced origin

MIC DNA

Alignment of consensus

TA

TA

TA

Step 3 : DNA methylation analysis

• Single-molecule
• Single-nucleotide resolution
• Independant yet pairable analysis on both strands

Some more details:

the random sampling approach

Deduced origin

MIC DNA

Alignment of consensus

Only a few remaining: ~ 10 to ~200 sequences

100% should carry a methylation pattern

• 1 out of 200 comes from the MIC
   
• 1/6 of MIC inserts will carry an IES
   
• 1/2 of IESs are just wrongly retained in the MAC
   
• 30% of the remnants are scanRNA dependant

Final categories

• MDS
  • = "MAC" for 99% of sequences
• MAC_IES
  • "true" MAC_IES (never seen retained)
  • other MAC_IES (sometimes retained)
• MIC
  • Other MIC specific (TE, repeated sequences)
• MAC (TA Junction)
  • Overlap a TA junction of excision
• rDNA
• mtDNA
• Other:
  • Contaminants
  • Alternative excision boundaries ("LOWID")
  • low identity consensus ("Trash")

Little bug to be corrected

30%

MDS and MAC_TA_Junction represent a vast majority

MIC specific and IES are as rare as expected

+ Applying the retention filter divides the number of "MAC_IES" approximately by 50%

Detecting m6A in details

Detecting m6A optimally

H0: ipdRatio of umethylated-Adenines

H1: ipdRatio of 6mA

S: Threshold on the pvalue

 --> Specificity

 --> Sensitivity

 

$$\alpha$$

$$1 - \beta $$

Adenine

6mA

ln(ipdRatio) ~ N(0,1)

log(ipdRatio)

Using E.coli as ground truth

E.coli:

• Feeds paramecium: contaminants+++
• Nearly 100% symetrically-methylated with m6A
  • GATC
  • EcoK
  • Few others
    • Depends a lot of the strain
• Outside of GATC and EcoK: very low levels of m6A

I investigated the PacBio's output on it's GATC & EcoK VS other sites

How log(ipdRatios) look in E.coli

Separability raises with coverage, which is expected

PacBio pvalues do have some biological meaning

PacBio pvalues do have a meaning...

• Just don't
• Obvious point n°1:
  • log(ipdRatio < 1) -> -inf

Normality assumptions under H0 are broken under high coverages

• The higher the coverage, the worse (here, >40x)
• Will cause bell-shaped pvalues under null
  • Hidden phenomena on the previous curve

Normality assumption matters

All Adenines' pvalues [E.coli] coverage > 40X

Long story short

PacBio's pvalues:

1. Are biologically relevant
   • We can build a reliable ad-hoc system with them
2. Are produced by a linear combinations of > 150 different coverages
   • Which forbids the usual statistical treatments
3. Are somehow broken on a coverage-dependant manner
   • Which forbids a simple fix for point 2
   • We can't use the classical statistical treatments directly on pvalues

Other PBio's scores for m6A

For n6mA, PacBio produces:

• A modification score --> Slowing of the polymerase
  • pvalue against H0 only (the one we presented earlier)
• An identification score --> Kinetic signature of a modification
  • loglikelihood between H0 and H1 (H1 = Other modifications or secondary peaks)
     

They are PHRED-transformed p-values of two different statistical tests, that rely on the mean of the IPDs

PHRED scores

The scores (Qv) are PHRED-transformed p-values

Typical covscore plot

Modification score / coverage

Using flat threshold on modification score = Hudge lack of power

Solution : A coverage-dependant threshold on the scores

From now on

 

"positive detection"

=

score > linear thershold

 

 

(only >25X considered)

Benchmarking

How good (or bad) is our method ?

$$Se = P(D^+|M) ~ 92\%$$

 

$$Sp = P(D^-|NM) ~ 99.8\%$$

Starting from sufficient coverages (~20X to ~30X), Se and Sp don't depend on the coverage anymore

The 4 scenarii for Se and Sp

• We don't care about Se and Sp
   
• We care about the fact that eventual mis-estimations of them doesn't really change anything

Finding unbiased estimators for    and FDR

If p number of positive detections among N tests:

 

p = FP + TP

$$\pi$$

So,

Which means

And:

What it gives in Paramecium

Debiased levels of m6A in the MAC_TA inserts

Details in MAC HTVEG

• ~95% of the methylation locates in AT dinucleotides in the MAC

  • True in any condition
     

• 75% of the methylation in an AT dinucleotide is actually symetrically modified, independantly from being in the MAC or the MIC

Kept in MIC and MAC

All conditions

MAC_TA outside of AT sites

Detections outside AT sites are likely to be at least partly true positives

Present in all samples

Never erased

Methodological development to correct hemi-methylation detection (1/2)

Let FD1 and FD2 be resp:

• Fraction of hemi-methylated AT sites
• Fraction of sym-methylated AT sites
  
  PZ0, PZ1, PZ2:  unbiased estimators of non, hémi, symetrically methylated AT sites

Then:

With

Methodological development to correct hemi-methylation detection (2/2)

We can also find the number of hemi-methylated sites being detected as such, and the proportion of sites detected as hemi-methylated that are really hemi-methylated. This is possible because we now approximately know PZ0, PZ1 and PZ2, and P(D|Z) is easy to determine:

Then, P(Z|D) can be determined through Bayes theorem using P(D|Z), P(Z) and P(D) (which are all known)

P(Z=1|D=1)

is our case of interest

What it gives in Paramecium

Debiased Hemi-methylation in MAC_TA

n6mA in the MIC (preliminary)

Pure MIC sequences: Cannot yet be trusted (error in the pipeline)

Coverage >= 40X made us lose too many materials, should restart with >=20X

n6mA in the MAC_IES rarely retained ("TRUE MAC IES")

HTVEG

MT2

 

 

--> Some molecules carry all the detections, in sym-A*T

Very likely to be sequences comming from the MAC to my opinion

At first look, seems like the same in all samples

Conclusion

• Lots of sweat spent on methods : Now we start the cool things
• In the MAC
  • Quantification in the MAC is well-characterized in all the samples
  • Our 6 genes are implied in the bulk of the MAC m6A
  • MT and NM families: Functional analogs of DNMT1 ?
  • Lots of questions raised by the MAC methylation: TSS, IES Junctions, nucleosome positionning...
• In the MAC IES
  • Very preliminary analysis between TRUE mac IES and other MAC IES tends to show that all detections could come from accidental retention in the MAC, and the MIC could actually carry 0% m6A or very lower levels than 1%. Really to premature to be sure
  • HT2 and HT6 are still potentially full of surprises
• TE in the MIC: No idea yet
• It's just a question of time now ! Which we asked (prolongation LABEX memolife)

Sorry for the headaches !! Thanks for your time :)

2.1 Retroingineering

The capping of IPDs

 

• modelPrediction is the predicted IPD value by the model in a given context of nucleotides at this position

• globalIPD is the mean of all the IPD values of the read.

• localIPD represents all IPDs that have been mapped at a given position in the genome, including those from other sequences

Conclusion on the capping

• Isn't coded as advertized by PacBio
• The way it's implemented for AggSN is problematic and doesn't really make sense
• Paradoxally, it should be more relevant for our approach than for the default one
• We expect no methylation to be undetected due to the capping

 

Laura landwebehr 2020

Oxytrichia trifallax

p values (A)

p-values (A score >20)

Out GATC score < 20

Out GATC

ipdRatio score 20

ipdRatio idv20/score20

A outAT score 20 isQv20 (812 seq)

A outAT score20 idQv20 + Strong BH correction (176 seq)

ipdRatio out GATC before filtering BH vs after (qv20/idqv20)

In GATC