An Ensemble SVM Model for the Accurate Prediction of Non-Canonical MicroRNA Targets - PowerPoint Presentation

Slide1

An Ensemble SVM Model for the Accurate Prediction of Non-Canonical MicroRNA Targets

Asish Ghoshal

1

,

Ananth

Grama

1

,

Saurabh

Bagchi

2

,

Somali

Chaterji

1

Purdue University, West Lafayette, IN

Slide2

MicroRNA (miRNA): “The genome’s guiding hand?”

miRNA are 22 nucleotide (nt) strings of RNA, base-pairing with messenger RNA (mRNA) to cause mRNA degradation or translational repression

Can be thought of as biology’s dark matter: small regulatory RNA that are abundant and encoded in the genomeDysregulation of miRNA may contribute to diverse diseases

Canonical (i.e., exact) matches involve the miRNA’s seed region (nt 2-7) and the 3’ untranslated region (UTR) of mRNA and were thought of as the only form of interactionRecent high-throughput experimental studies have indicated the high-preponderance of “non-canonical” miRNA targets

miRNA

Slide3

Background

microRNAmRNA

ArgonauteRISC (complex)

3

Slide4

We are attempting computational prediction of miRNA-mRNA interactions

Challenging because of:

Large number of features in miRNAs and mRNAsNoisy and incomplete data sets from experimental approachesWide variety of positive miRNA-mRNA interactions Prior computational approaches are deterministic and rely on perfect complementarity of the miRNA and mRNA nucleotides

“Avishkar”: Computational miRNA target prediction

Slide5

Background: CLIP Technology and Non-Canonical miRNA-mRNA Interactions

An experimental method to map the binding sites of RNA-binding proteins across the transcriptome. Proteins are crosslinked

to RNA using ultraviolet light, and an antibody is used to specifically isolate the RNA-binding protein of interest together with its RNA interaction partners, which are subjected to sequencing.5

Slide6

Our ContributionsAvishkar suite comprises of 4 different models: a global linear SVM model,

a global non-linear model, an ensemble linear model, an

ensemble non-linear model (radial basis function kernel)Feature computation and transformationSpatial features at two resolutionsSeed enrichment metricLinear classifier had a moderate amount of bias, so we switched to a non-linear, kernel-SVM classifier. This achieves a TPR of 76% and an FPR of 20%, with the AUC (ROC curve) for the ensemble non-linear model being 20% higher than for the simple linear model.

This is an improvement of over 150% in the TPR over the best competitive protocol.Since training kernel SVMs is computationally expensive, we provide a general-purpose and efficient implementation of Cascade SVM, on top of Apache Spark: https://bitbucket.org/cellsandmachines/kernelsvmspark

Parallel Support Vector Machines: The Cascade SVM: Hans Peter Graf, Eric Cosatto, Leon Bottou, Igor Durdanovic, Vladimir Vapnik

, NIPS 2005

Slide7

Problem StatementPredict if a miRNA targets an mRNA (segment)

miRNAs

mRNA segments

1 or 0 ?

7

Slide8

Problem StatementIdeally we would want some experimentally verified edge labels to train on.

miRNAs

mRNA segments

1 or 0 ?

1

0

1

Edges with ground truth labels

Edges whose labels have to be predicted

8

Slide9

Problem StatementWe have labels on vertices from CLIP-Seq experiments, i.e., which mRNA segments were targeted.We do not know which miRNA targeted a particular mRNA segment.

miRNAs

mRNA segments

1 or 0 ?

1

0

1

0

9

Slide10

Methods: Generating Ground Truth DataConsider only the most highly expressed miRNAs20 in humans, 10 in mouseAdd an edge if

the binding between a miRNA and mRNA segment is strong enough, i.e., ΔG is below a certain threshold or,There is at least a 6-mer seed match between the mRNA segment and miRNA.

miRNAs

mRNA segments

1

0

1

0

10

Slide11

Methods: Generating Ground Truth DataLabel the edge 1 (or 0) if it is incident on a mRNA segment that has label 1 (or 0).

miRNAs

mRNA segments

1

0

1

0

1

1

0

0

0

1

11

Slide12

Methods: Schematic of Features (Feature Engineering)

The alternating

blue

and green regions denote the thirteen consecutive windows around the miRNA target site (red). These are the windows where the average thermodynamic and sequence features are computed.

We hypothesize that the curves are different for the positive and negative samples.

The

miRNA seed region

is the heptametrical sequence at positions 2-7 from the 5’ miRNA end.

12

Slide13

Methods: Capturing Spatial Interaction using Smooth CurvesCompute thermodynamic interaction profile upstream and downstream of the target regionData is collected for fixed-size windows on both sides of the target region

Averaging is done within each windowCompute smooth curves to remove noise in the biological data setUsed B-spline interpolation for smoothing out the points

13

Slide14

Methods: Capturing Spatial Interaction using Smooth CurvesCompute interaction profiles for features

Binding energy ΔGAccessibility ΔΔGLocal AU contentCompute interaction profiles at two different resolutionsWindow size of 40 and using the entire miRNA: “site” curvesWindow size of 9 and only using the seed region of the miRNA:

“seed” curvesUse coefficients of B-spline basis functions as features for classifier14

Slide15

Methods: Capturing all Possible Non-Canonical MatchesEncode the alignment of the miRNA seed region with mRNA nucleotides using a string of 1s (matches), 2s (mismatches), 3s (gaps), and 4s (GU wobbles).

Compute the probability that the alignment string (e.g., 1111121) is not a random occurrence.Define enrichment score for the seed match pattern as 1 – probability.The “seed enrichment metric” captures, in a single numeric feature, the relative efficacy of various kinds of seed matches in a single, unified metric.This in turn converts the categorical feature of seed match or non-canonical match into a numerical feature (probability value) that can be better handled by ML methods.

15

Slide16

Methods: CLIP Data

Species# Positive examples (Seed, Seedless)

# Negative examples# mRNA

# miRNA

Positive target sites

3’ UTR

CDS

5’ UTR

HITS-CLIP (Mouse)

861,208 (6%,

94%)

35,608,333

4,059

119

56%

43%

1%

PAR-CLIP (Human)

141,109 (8%,92%)

2,659,748

1,211

35

57%

39%

4%

16

Slide17

Methods: Our ClassifierDistributed linear SVM using gradient descent (Apache Spark)9 Intel X86 nodes, 8 GB memory, 4 cores.17

Slide18

Methods: Validation Protocol used to Evaluate Avishkar18

Slide19

Results: ROC Curves19

Slide20

Results: Feature Importance20

Slide21

Methods: Improving Classification PerformanceLinear classifier suffers from high bias (large training error).Use more complex learning modelKernel SVMSVMs suffer from a widely recognized scalability problem in both memory use and compute time.

Kernel SVM computational cost: O(n3)Does not scale beyond a few thousand examples for feature vector of dimension ~ 150.21

Slide22

Methods: Distributed TrainingCascading SVM [Graf et al. NIPS 2005]Key ideas:Train on partitions of the whole data set and do this in parallel

Merge SVs from each partition in a hierarchical mannerFinal step is serial and it is hoped that the number of SVs is reasonably small at that stageImplemented on top of Apache Spark

22

1 2 3

Slide23

Distributed Training: Our ContributionsUse non-linear SVM for classification since we observe high bias when using linear SVM, leading to lower accuracyBut non-linear SVM does not scale well even with cascading SVM

Biological insight: miRNAs within an miRNA family share structural similaritiesTherefore, we create a separate non-linear classifier for each miRNA family, plus, within each family we speed up the optimization using the Cascading approachTraining for the model for each miRNA family can occur in parallel23

Slide24

Results: Misclassification Rate for Linear and Non-Linear Classifier

Five-fold cross-validation misclassification rate for different miRNA families.

Mean test error of the non-linear SVMs for each of the miRNA families is less than the corresponding linear

models.

Benefit of non-linear SVM is more pronounced for larger miRNA families: e.g., for let-7 and miR-320 families, benefits are 50% and 69.9% over linear model.

Insight

: Non-linear model can remove the prediction bias inherent in linear model.

24

Slide25

Results

ROC curves for the ensemble linear model and ensemble non-linear

model, obtained by varying the probability threshold of for the output of the SVM.

The

misclassification error, true positive rate, and false positive rate were computed using 5-fold stratified cross-validation for seedless sites in

the human data set

.

One

possible operating region

is with an FPR of 0.2, the TPR for the linear model is 0.469, while the TPR for the non-linear model is 0.756.

25

Slide26

Results26

Slide27

ResultsDoes the classification performance improve due to clustering by miRNA family or due to the use of a more complex model?Performance of linear classifier does not improve by clustering data.

27

Slide28

Take-Away PointsWe have developed “Avishkar”, a machine-learning, support vector machine-based model, to predict both

canonical and non-canonical miRNA-mRNA interactions Avishkar extracts thermodynamic and sequence features and does smoothening through curve fitting, in order to extract enriched features from CLIP dataWe use non-linear SVM to minimize bias and scale it up to the large biological data sizes through a biologically-driven parallelization strategyWe achieve the best-in-class precision (true positive rate), with an improvement of over 150%, over the best competitive protocol

The code runs on Apache Spark and is open sourced on bitbucket: https://bitbucket.org/cellsandmachines/avishkarWe also provide a general-purpose, scalable

implementations of kernel SVM using Apache Spark, which can be used to solve large-scale non-linear binary classification problems: https://bitbucket.org/cellsandmachines/kernelsvmspark

28

Slide29

Thanks!29

Slide30

EXTRA + Notes30

Slide31

31

Slide32

Background: RNA interference32