Scikit-ribo reveals precise codon-level translational control by dissecting ribosome pausing and codon elongation

October 26, 2016Biological Data Science

Han Fang

The hidden treasure in genomics

DNA RNA ProteinTranscription Translation

RNA Polymerase Ribosome

Replication

The hidden treasure in genomics

DNA RNA ProteinTranscription Translation

RNA Polymerase Ribosome

Replication

Ingolia et al. Science. (2009) Ingolia. Nat Rev Genet. (2014)

What is ribosome profiling (Riboseq)?

What is ribosome profiling (Riboseq)?

Less efficient translation Normal translation efficieny (TE) More efficient translation

log2(TE) < 0 log2(TE) = 0 log2(TE) > 0

T E = RP F rpkmRNA rpkm

Ingolia et al. Science. (2009) Ingolia. Nat Rev Genet. (2014)

What is ribosome profiling (Riboseq)?

Less efficient translation Normal translation efficieny (TE) More efficient translation

log2(TE) < 0 log2(TE) = 0 log2(TE) > 0

T E = RP F rpkmRNA rpkm

Less efficient translation Normal translation efficieny (TE) More efficient translation

log2(TE) < 0 log2(TE) = 0 log2(TE) > 0

T E = RP F rpkmRNA rpkm

Less efficient translation Normal translation efficieny (TE) More efficient translation

log2(TE) < 0 log2(TE) = 0 log2(TE) > 0

T E = RP F rpkmRNA rpkm

Ingolia et al. Science. (2009) Ingolia. Nat Rev Genet. (2014)

Calculate translational efficiency (TE)

Less efficient translation Normal translation efficieny (TE) More efficient translation

log2(TE) < 0 log2(TE) = 0 log2(TE) > 0

T E = RP F rpkmRNA rpkm

TE =

Riboseq rpkm

RNAseq rpkm

log2(TE) < 0

log2(TE) = 0

log2(TE) > 0

1

TE =

Riboseq rpkm

RNAseq rpkm

log2(TE) < 0

log2(TE) = 0

log2(TE) > 0

1

TE =

Riboseq rpkm

RNAseq rpkm

log2(TE) < 0

log2(TE) = 0

log2(TE) > 0

1

TE =

Riboseq rpkm

RNAseq rpkm

log2(TE) < 0

log2(TE) = 0

log2(TE) > 0

1

Hypothesis: TE distribution could be skewed by ribosome pausing events.

Ribosome footprints without bias Ribosome footprints with pausing

Simulated S. cerevisiae data

Null distribution

With ribosomepausing sites

Simulated S. cerevisiae data - TE distribution are negatively-skewed by ribosome pausing events

Null distribution

Randomly imputed ribosome pausing sites to 20% of the genes

Ribosome pausing sites (peaks) finding by negative binomial mixture model

Yes

SIMPLE BUSINESS PROCESS

Given a gene

Likelihood ratio test

Bimodal distribution, i.e.

potential pausing sites

Unimodal distribution, i.e.

no pausing

Bonferroni adjusted p-value

< 0.05

Posterior probability >

cutoff

Not a pausing site

No

Peak/Pausing sitesYes

No

Yifei HuangP (Xi|⇡i, µi, ki, ri) =Y

j

⇡iNB(Xij |µi, ri) + (1� ⇡i)NB(Xij |kiµi, ri),

for gene i at position j, where k � 5

H0 : ⇡ = 1

H1 : ⇡ 6= 1

1

P (Xi|⇡i, µi, ki, ri) =Y

j

⇡iNB(Xij |µi, ri) + (1� ⇡i)NB(Xij |kiµi, ri),

for gene i at position j, where k � 5

H0 : ⇡ = 1

H1 : ⇡ 6= 1

1

Yes

SIMPLE BUSINESS PROCESS

Given a gene

Likelihood ratio test

Bimodal distribution, i.e.

potential pausing sites

Unimodal distribution, i.e.

no pausing

Bonferroni adjusted p-value

< 0.05

Posterior probability >

cutoff

Not a pausing site

No

Peak/Pausing sitesYes

No

# genes # genes (rpkm > 100) # genes with pausing

# ribosome pausing sites identified

6664 1252 94 180

Yifei HuangP (Xi|⇡i, µi, ki, ri) =Y

j

⇡iNB(Xij |µi, ri) + (1� ⇡i)NB(Xij |kiµi, ri),

for gene i at position j, where k � 5

H0 : ⇡ = 1

H1 : ⇡ 6= 1

1

P (Xi|⇡i, µi, ki, ri) =Y

j

⇡iNB(Xij |µi, ri) + (1� ⇡i)NB(Xij |kiµi, ri),

for gene i at position j, where k � 5

H0 : ⇡ = 1

H1 : ⇡ 6= 1

1

Ribosome pausing sites (peaks) finding by negative binomial mixture model

mRNA with stronger secondary structure tend to have ribosome pausing events

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

0.0

0.5

1.0

1.5

No Yes

Seco

ndar

y st

ruct

ure

scor

e

factor(Pausing)NoYes

p-value< 2*10-16

PARS scores obtained from Kertesz et al. Nature (2010)

−4

0

4

0 100 200 300 400position

PAR

S sc

ores

−4 0 4PARS_score

0

500

1000

1500

2000

0 100 200 300 400position

ribos

ome

coun

ts

ribosome_pausing

FALSE

TRUE

mRNA with stronger secondary structure tend to have ribosome pausing events

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

0.0

0.5

1.0

1.5

No Yes

Seco

ndar

y st

ruct

ure

scor

e

factor(Pausing)NoYes

p-value< 2*10-16

PARS scores obtained from Kertesz et al. Nature (2010)

−4

0

4

0 100 200 300 400position

PAR

S sc

ores

−4 0 4PARS_score

0

500

1000

1500

2000

0 100 200 300 400position

ribos

ome

coun

ts

ribosome_pausing

FALSE

TRUE

mRNA with stronger secondary structure tend to have ribosome pausing events

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

0.0

0.5

1.0

1.5

No Yes

Seco

ndar

y st

ruct

ure

scor

e

factor(Pausing)NoYes

p-value< 2*10-16

PARS scores obtained from Kertesz et al. Nature (2010)

Simplifying the generalized linear model (GLM)

RNA count

Secondary structure

count

Dwell time(1/speed)

count

Translationefficiency (TE)

count

Simplifying the generalized linear model (GLM)

RNA count

Secondary structure

count

Dwell time(1/speed)

count

Translationefficiency (TE)

count

Simplifying the generalized linear model (GLM)

RNA count

Secondary structure

count

Dwell time(1/speed)

count

Translationefficiency (TE)

count

Simplifying the generalized linear model (GLM)

RNA count

Secondary structure

count

Dwell time(1/speed)

count

Translationefficiency (TE)

count

Joint inference of protein TE and codon dwell time using GLM, while accounting for secondary structure

GLM for joint inference of TE and codon dwell time:

Yij ⇠NB(mean = µij , dispersion = ↵), for gene i, position j

g(µ

ij) = �

0

+ x

im|{z}

mRNA

+ �

it|{z}

TE

+ �

kc|{z}

codon

+ �sxij

| {z }secondary structure

where g(.) is a log link function, µij = E(Yij),

x

im is mRNA abundance for gene i,

�

it is translational e�ciency for gene i,

�

kc is dwell time for codon k,

�sxij

is secondary structure e↵ect at position j for gene i.

1

Scikit-ribo perfectly reproduced relative codon dwell time from Weinberg et al

Scikit-ribo perfectly reproduced relative codon dwell time from Weinberg et al

CGG

CGA

CCG

Rarest codons

Significant correlation between tRNA abundance and codon elongation rates

GLM estimates vs. RPKM-based estimates reveals systematic bias in typical Riboseq analysis

GLM estimates vs. RPKM-based estimates reveals systematic bias in typical Riboseq analysis

The rpkm based approach overestimated TE of highly structured mRNA, while the rest of the mRNA were slightly under-estimated, as hypothesized.

Highly structured mRNA genes

Accurate TE estimation supported by proportional synthesis for heterodimeric complexes in S. cerevisiae.

Summary

Discussed:1) Introduced scikit-ribo for joint analysis of Riboseq and

RNAseq data.

2) Identified biases in Riboseq data due to ribosome pausing.

3) Corrected biases and revealed underlying biology

4) Joint inference of codon elongation rate and protein TE

5) Revealed precise translational control at codon levelAlign reads (eg. w/ STAR, HISAT)

Filter rRNA sequences(eg. w/ bowtie)

Identify & trim adapters (eg. w/ cutadapt)

Transcript quantification(eg. w/ cufflinks, salmon)

Joint analysisw/ scikit-ribo

Riboseq RNAseq

Ribosome A-site position predicition

A-site codonlocalization

Ribosome pausingsite calling

Translation efficiencyinference

Differential translation efficiency testing

https://github.com/hanfang/

Acknowledgments

Gholson Lyon Michael Schatz

Lyon LabMax DoerfelYiyang WuJonathan CrainJason O’Rawe

Schatz LabFritz SedlazeckTyler GarvinJames GurtowskiMaria NattestadSrividya Ramakrishnan

CSHL Stony Brook UniversityUCSF RutgersJHU

Yifei HuangAdam SiepelNoah Dukler

Rachel Green Jonathan WeissmanJoshua DunnDavid Weinberg

Premal Shah Rob Patro