Where Will They Strike Next?microRNA targeting tactics

in the war on gene expression

Jeff Reid

Miller “Lab”

Baylor College of Medicine

Outline

• Introduction to miRNAs

• The “ask Bartel” model for targeting

• Our proposed model

• Discuss predictions made by our model– All positions on the miRNA are not equal– A given miRNA’s targets share function

• Have a quantitative model that does not suffer from the arbitrariness of ask Bartel

Plant microRNAs

• This talk is about plant miRNAs– Animal miRNAs different, more complicated– If you want to know more about them ask Tuan Tran!

• What is a microRNA (miRNA)?• ~21nt single-stranded non-coding RNAs• Processed from stem/loop precursors• Bind to mRNA in the cytoplasm• Regulate genes

– Often relevant to development

microRNA biogenesis (conventional wisdom)

1.miRNA gene is transcribed producing primary transcript

2.pri-miRNA processed by dicer…3. ..producing miRNA duplex4.duplex moves out of the nucleus5.helicase activity unzips duplex6.mature miRNA forms RNA-

induced silencing complex (RISC)7.RISC recognizes a target site8.Targeted mRNA is regulated

(mRNA cleavage or translational repression)

Figure from Bartel, D.P. (2004). Cell 116, 281-297.

Target “Acquisition”• How does the RISC identify target sites?• Based solely on mature miRNA sequence

– Consistent all with known examples– “just” string manipulation– With that in mind, consider a simple model…

• Targets have small “mismatch score” – M– Count non-WC pairs in miRNA/target duplex– Score is independent of position

A CG

CU

CC

CC CC

UUUU

U AA AAG

GG GGGG

U A G AC

target site

RISC

mRNA

M = 2

5’3’

5’3’

Complementarity Model*• Look for 21-mers (mRNA sequence) with M < 4

– Find targets…– mir172a1 [AP2]: At5g60120(1) At4g36920(2)

At2g28550(3) At5g67180(3)

At5g12900(3) – …turns out most targets of a given miRNA are in

genes which share a common function

• There are some ask Bartel elements to the model– M = 4 targets sharing function included case-by-case– Single bulges are sometimes allowed (mir162, mir163)

– Model specificity is problematic…*Rhoades, et. al. (2002) Cell 110, 513-520.

APETALA2 transcription factor

Selectivity and Specificity

• Selectivity (false negatives)– Bartel’s model finds “everything” for M < 5

• Putative targets from this model (most confirmed by experiment) define the target population

• Specificity (false positives)– Bartel’s model is problematic

•M < 5 includes many false positives•M < 4 and qualitative ask Bartel elements are

necessary for model specificity

• Our goal is to develop a quantitative model

Position Dependent Model

• Ask Bartel has been spectacularly successful• Build on existing model & make it quantitative• No a priori justification of position-independence

– assumed by the ask Bartel model

• Extend to a position-dependent mismatch model– Assign mismatch at position i weight i

• For ask Bartel modeli = 1

• Quantify target “strength” with binding probability– pt is the probability of finding the miRNA bound to

target site t in the mRNA population

• Now “mismatch score” is position-dependent

• Boltzmann factor gives binding probability

• Quantitative model built, but how to find i?

Boltzmann factors

L

itmit ii

tmE1

, )1(),,( m = miRNA* sequence

t = target site sequence = mismatch parameters

g

gmEgeβmZ ),,(),(

),(),,(

),,(

βmZ

eβtmp

tmE

t

t

A CG

CU

CC

CC CC

UUUU

U AA AAG

GG GGGG

U A G AC

RISC

mRNA1 2 3 4 5

5’3’

5’3’

Model Comparison• Follow DNA binding protein example*

– Consider a thought experiment….• Mix many copies of the genome and N copies of the protein

and count the number of examples of protein bound to site t

– ft = nt / N

• If the model works ft and pt must agree!

• Determine i by looking for this agreement

– Maximize the probability that the data (ft) could have

come from the model (pt)…

*Brown, C.T., and Callan, C.G. (2004). Proc. Natl. Acad. Sci. 101, 2404.

Model Testing

• Probability of data arising from our position dependent mismatch model

• Obtain best match of model to data by maximizing the log probability

• Yields set of parameters i which maximizes the

probability of getting the data from our model

g gg

gtmpm ffP ),,(),,(

g ggg βmZβtmEβm )],(ln[),,(),,(

ffL

Optimization Cartoon

1

2

3

4

5

Parameter Controls Inputs

miRNAs

data

Binding Probabilities

miRNA sequence

UAGCA

measured fraction bound

f1 f2 f3 f4 f5 ... f24

0

• Maximize L to get i

f24p24

Optimization Cartoon

1 2 3 4 5

Parameter Controls Inputs

miRNAs

data

Binding Probabilities

miRNA sequence

0

f1 f2 f3 f4 f5 ... f24UAGCA

f24p24

• Maximize L to get i

measured fraction bound

Optimization Cartoon

1

2

3

4

5

Parameter Controls Inputs

miRNAs

data

Binding Probabilities

miRNA sequence

0

f1 f2 f3 f4 f5 ... f24UAGCA

f24p24

• Maximize L to get i

measured fraction bound

Model Testing

• Probability of data arising from our position dependent mismatch model

• Obtain best match of model to data by maximizing the log probability

• Yields set of parameters i which maximizes the

probability of getting the data from our model

g gg

gtmpm ffP ),,(),,(

g ggg βmZβtmEβm )],(ln[),,(),,(

ffL

Review

• Application of this procedure to miRNAs

• Optimize to get best agreement between

– position-dependent mismatch model: pg

– Ask Bartel complementarity model: fg

• Equal binding probability for each training target• Minimal binding to everything else (background)

– A contribution we made to the method– necessary to avoid overfitting

Multi-miRNA Optimization

• Given the amount of data we have • This method would fail on DNA binding proteins

• All miRNAs share the same machinery for target recognition (all form the RISC)– DNA binding protein recognition depends on

the each specific protein

• Solution to our problem– Simultaneously optimize for several miRNAs

Results - Parameters

• Multi-miRNA optimization of nine Arabidopsis miRNAs– 157b, 159b, 160b, 164a, 165b, 167b, 168a, 171, 172a1– A set of functionally diverse (21-mer) miRNAs

3’ 5’(i)

i

Position 14• Mismatch at position 14

– Has no effect on a target’s binding probability!

• Surprising and exciting because…• …this position is known to be special

– mir162a target• 1g01040 DEAD/DEAH box helicase

– Has a bulge at position 14

• This analysis did not include mir162a!• A provocative result…

14 151 213’

3’

5’

5’ target

mir162a

Results - Targets

• Training targets should have low energy– Found by ask Bartel model– Reside in genes which share majority function

• Targets in the background have high energy– Background targets with low energy are interesting

• We are particularly interested all the majority function targets for a given miRNA– Especially those which are not training targets

• Look at distributions of target energies– For each value of M

mir165b -- HD-Zip

majority functionnot training

targets!

training targetsmajority function

N(E)

N(E)

mir159b -- MYBN(E)

N(E)

Conclusions• Refined the qualitative complementarity model

– A quantitative model which is much less arbitrary• Whatever we get, we get – not “ask Miller”

– Majority function targets group together at low energy– Bartel finds most targets, our model finds all targets

• Appropriate experiments could falsify our model– How important is position 14?– Look at some specific ask Bartel targets

• Advanced technology of optimization– Resolution of the overfitting problem– Simultaneous optimization

Encoding of Networks

• Networks– miRNA families

• A single target mRNA can be regulated by different miRNAs• And a single miRNA can regulate many different mRNAs

– Apparently an overlapping and probably redundant regulatory network

• Encoding– All this regulation encoded in mere text!– How is this encoded in the sequence?– Why is it encoded in this way?

Acknowledgements

• Miller Lab Posse– Jon Miller– Tuan Tran– Will Salerno– Gerald Lim

• Curtis Callan (Princeton)• Keck Center for Computational and

Structural Biology • BCM Biochemistry Department