Data integration in proteomic and biomarker

discovery pipelines: a case study using the

ProteINSIDE webservice.

M Bonnet, M. Reichstadt, J. Bazile, M.P Ellies

UMR Herbivores – Clermont-Ferrand - France

2

Why proteomics ?

Monti et al. 2019, J proteomics

Technological improvements aided the proteome depiction to have a pivotal role in the study of biological systems

3

MW

Gel-based proteomics

pH4 7

Proteomics is the large-scale study of proteinsin cells, tissues, organs, fluids

Proteinextraction

2-DEImage

Analysis

MS or MS/MS analysis

ProteinIdentification

Protein A

• Identification of proteins in a medium range of pH and molecular weight• Time-consuming and expensive (one mass spectrometry analysis = 1 spot)

Ruiz et al., 2016 ; Zargar et al., 2016

4

Gel-free proteomics

Proteinextraction

Proteinconcentration

Label free, chemical/

isotopical labelling

LC-MS/MS analysis

ProteinIdentification

Identification of the most abundant proteins => a depletion step can beincluded for the very highly abundant proteins

Ruiz et al., 2016 ; Zargar et al., 2016

Protein A

5

Proteomics : data analysis and mining

Gel-based or Gel-free proteomics provide lists of protein IDs with theirabundances to be analysed => this talk focus on the data integration of

protein IDs and abudances

6

Descriptive and differentialstatistics

Data analysis and data mining of lists of

protein IDs and abundances

7

http://prabig-prostar.univ-lyon1.fr/Articles/fdrtuto.pdf

Descriptive statistics : a very useful tutorial

8

Descriptive statistics : PCA, clustering, heatmap

• Observation of the data to assess data distribution, overall variation,

variability within each treatment group, outliers…

• Comparison of means and variabilities from those means

• Are the data normally distributed ? => to be transformed for parametrics

approaches or analyze by non-parametric methods.

• Correlation plots to compare treatment groups

9Karimpour-Fard et al., 2015

The major classification methods

Methods What does it do? Sample size/data characteristics

Principal Component Analysis

Separates features into groups based on commonality and reports the weight of each component’s contribution to the separation

Unlimited sample size, data normally distributed

Independent Component Analysis

Separates features into groups by eliminating correlation and reports the weight of each component’s contribution to the separation

Unlimited sample size; data non-normally distributed

Random Forest

Separates features into groups based on commonality; identifies important predictors

Performs well on small sample size ;resistant to over-fitting

Partial Least Square

Separates features into groups based on maximal covariation; reports the contribution of each variable

Unlimited sample size; sensitive to outliers

Support Vector Machine

Uses a user-specified kernel function to quantify the similarity between any pair of instances and create a classifier

Performs well on small sample size and resistant to over-fitting

10

The major clustering methods

Methods What does it do? Sample size/data characteristics

K-meansclustering

Separates features into clusters of similar expression patterns

performs best with a limited dataset, i.e., ~20 to 300 features

Hierarchicalclustering

Clusters treatment groups, features, or samples into a dendrogram

Performs best with limited dataset, i.e., ~20 to 300 features or samples

See the publication by Karimpour-Fard et al., 2015 for more information on the strengths, weaknesses and mathematical mechanisms of each

classification and clustering methods

Karimpour-Fard et al., 2015

11

Differential statistics : concepts of p-values (1/2)

• This is an answer to 2 questions : Whether the putative discovery is

true or false, by relying on the null hypothesis ? What is the

probability that a given false discovery is included in the set of

selected discovery

The null hypothesis and the null distribution refer to the “standard” behavior”

that is the non-differential abundance

Putative discovery : any

protein quantified in the

experiment

True discovery : a protein that

is differentially abundant

between the biological

conditions

False discovery :

a protein that is not differentially

abundant between the biological

conditions

Selected discovery : a protein

that has passed some user-

defined statistical threshold

Adapted from Burger, http://prabig-prostar.univ-lyon1.fr/Articles/fdrtuto.pdf

12

• This is an answer to 2 questions : Whether the putative discovery is

true or false, by relying on the null hypothesis ? What is the

probability that a given false discovery is included in the set of

selected discovery

• To answer => statistical tests quantify the similarity between the

putative discovery and the null hypothesis

• a p-value of 0.05 means that there is 5% chance of getting the observed

result, if the null hypothesis (no difference among groups) were true.

if 100 statistical tests are performed and for all of them the null

hypothesis is actually true, it is expected that 5 of them will be

significant at the p < 0.05 level, by chance. In this case, five

statistically significant results are obtained, all of them being false

positives (type I error)

Differential statistics : concept of p-values (2/2)

A p-value is applied to one protein (individual test)

13

Differential statistics : available methods

t-tests :

Students (the first version),

Welch (numerous variations)

Limma

ANOVA, mixt models...

Likelihood-ratio tests

compare two models provided the simpler model is a special case of the more

complex model (i.e., “nested"). LRTs can be presented as a difference in the log-

likelihoods (recall that log(A/B) = logA – logB) and this is often handy as they can

be expressed in terms of deviance.

Non parametric tests

do not assume anything about the underlying distribution and have great statistical

power, which means they are likely to find a true significant effect : Mann-Whitney,

Kruskal-Wallis, Log-rank..tests

14

Some tools for statistical proteomics

P Packages R 'mixOmics‘ to analyse and integrate several types of

“omics data” http://mixomics.org/a-prop os/publications/

Several software packages for analyzing proteomic data include

statistical tools : MaxQuant (Perseus), Skyline, Progenesis

Packages R specifically developed for the analysis of proteomic data:

• MSstats (Choietal.,2014,Bioinformatics,30:2524-2526)

• DAPAR & ProStar (Wieczoreketal.,2017,Bioinformatics,33,13513)

• SafeQuant (Ahrné, unpublished, https://github.com/eahrne/SafeQuant/)

• GiaPronto (Weineretal.,2017,Mol.Cell.Prot.doi:10.1074/mcp.TIR117.000438)

• MCQ (PAPPSO platform, Paris, unpublished)

• Bazile et al., 2019

https://github.com/jane-bzl/Differential_abundance_and_correlation_to_one_factor)

15

DOI : 10.5281/zenodo.2539329

https://github.com/jane-bzl/Differential_abundance_and_correlation_to_one_factor

Used in the publication Pathways and biomarkers of marbling and carcass fat

deposition in bovine revealed by a combination of gel-based and gel-free

proteomic analyses. Bazile et al. 2019, meat science

Management of proteomic data to identify differentially

abundant proteins according to one discriminant factor

and to correlate their abundance with the value of this factor

Jeanne BAZILE, Ioana MOLNAR, Brigitte PICARD, Muriel BONNET

An available R script by Bazile and al.

16

group id adiposity Protein1 Protein2 Protein3 Protein4

group1 1 6.3 230795072 72392559 -1308293 54238

group1 6 3.7 267670438 49956496 -1099355 23984

group1 3 4.1 1736366 57744002 -1188752 3574699

group1 4 5.1 72392559 65141951 -1234186 368571

group1 5 6.8 203517615 64672505 -1551479 65781

group2 2 3.0 275028504 64089037 -1125215 5348673

group2 7 2.1 2043358 41658946 -1132811 687253

group2 8 2.2 64089037 50876533 -1049944 68357

group2 9 2.3 231769965 37805488 -1134805 573521

group2 10 2.2 305818430 39682550 -1086689 68967

Groups made accordingto the biological

conditions

Identification of the samples (number,

letters or both)

Quantitative values related to the

biological conditions

Values of proteins abundances

R script by J. Bazile : the input table

17

Shapiro test reports the normal (or not) distribution of an

abundance (TRUE=normal distribution)

p-values from Student-t-test for

normally or Kruskal-Wallis test for non-normally

distributed (TRUE= significant pval)

p-values from Student-t-test whatever the distribution of

the data (TRUE= significant pval)

pval.shapiro normalitypval.testTorKW

signif.testTorKW

pval.Ttest signif.Ttestpval.KW

signif.KW

Protein1 0.089 TRUE 0.801 FALSE 0.801 FALSE 0.465 FALSE

Protein2 0.403 TRUE 0.040 TRUE 0.040 TRUE 0.047 TRUE

Protein3 0.018 FALSE 0.047 TRUE 0.061 TRUE 0.047 TRUE

Protein4 0.000 FALSE 0.175 FALSE 0.675 FALSE 0.175 FALSE

R script by J. Bazile : the output table

• Results saved in your directory as an Excel file (.csv)• No correction for multiple tests for the selected discoveries because the size of

the set is below than 150 proteins in most of our proteomic experiments

p-values from Kruskal-Wallis -

test whatever the distribution of the

data (TRUE= significant pval)

18

Correlations computed either by the Pearson test (for normally distributed data) or

Kendall/Spearman test (for non normally distributed data)

adiposity Protein1 Protein2 Protein3 Protein4

adiposity1 0,06723532 0,82398759 -0,8845269

-0,15145677

Protein10,06723532 1 -0,0077549 -0,04223508

-0,04956437

Protein2 0,82398759 -0,0077549 1 -0,60280322 0,245925

Protein3 -0,8845269 -0,04223508 -0,60280322 1 0,16626722

Protein4 -0,15145677 -0,04956437 0,245925 0,16626722 1

R script by J. Bazile : correlation results

• A figure to review the proteins significantly correlated with one parameter• 2 tables produced and save to your directory :

one with the correlation values one with the p-value of the correlations

19

Descriptive and differentialstatistics

Data mining : when proteomics turns functional

Data analysis and data mining of lists of

protein IDs and abundances

20

Data mining : when proteomics turns functional

List of proteins

Gene Ontology analysis

=> GO

Prediction of protein secretion,

=> secretome

Protein-Protein

Interaction PPI,

=> interactomics

Interpretation of the data

FUNCTION

21

GO “annotations”

• An annotation is a statement linking a gene to some aspect of its function

=> GO term

• Each annotation is based on some evidence, recorded as part of an

annotation :

• Evidence code (type of evidence)

• Reference (published data, journal article)

22

Semantics and structure of GO annotations

The association of a gene product with a class of GO gives statements about the

•Molecular function, MF : molecular activities of a gene product

•Cellular Component CC: where the gene product is active

•Biological process BP : pathways and larger processes made up of the activities of

multiple gene products

GO annotations represent :• the normal, in vivo, biological role of gene products• Current knowledge that is Human understandable and machine computable

Ontology is a structured representation of biology, composed of

•classes

•relations

•definitions

For more information, http://geneontology.org/docs/go-annotations/

23

One of the main uses of the GO is to perform

enrichment analysis on gene/protein sets. For example,

given a set of genes/proteins that are up-regulated

under certain conditions, an enrichment analysis will

find which GO terms are over-represented (or

under-represented) using annotations for that gene set

GO enrichment analysis

The output of GO enrichment analysis is a GO term with a p-value most of the time corrected for multiple testing (Benjamini & Holsberg, Bonferroni…)

24

Tool Description

DAVIDIt provides a set of functional annotation tools to understand biological meaning, such as identification of enriched GO terms, clustering of redundant annotation terms and visualization of genes using BioCarta & KEGG pathway maps (https://david.ncifcrf.gov/).

EnrichrIt contains a large collection of gene set libraries available for analysis and download. It is a source of curated gene sets and a search engine that accumulates biological knowledge for further biological discoveries (http://amp.pharm.mssm.edu/Enrichr/).

KOBAS

A web server for functional gene set enrichment that can accept either gene lists or gene expression data as input. It generates enriched gene sets with the corresponding name, the p-value or a probability of enrichment and an enrichment score based on results of multiple methods (http://kobas.cbi.pku.edu.cn/).

PANTHERIt classifies proteins according to family (evolutionarily related proteins) and subfamily (proteins with the same function), molecular function, biological process and pathway (http://www.pantherdb.org/).

PathVisioIt provides a basic set of features for pathway drawing, analysis and visualization(https://www.pathvisio.org/).

WebGestaltA functional enrichment analysis web tool. It supports several organisms gene identifiers and three methods for enrichment analysis, including over-representation analysis, Gene Set Enrichment Analysis (GSEA), and Network Topology-based Analysis (NTA).

AmigoInteractively search the Gene Ontology data for annotations, gene products, and terms using a powerful search syntax and filters (http://amigo.geneontology.org/amigo/search/annotation/)

Main GO “annotations” software tools

25

Database Description

ReactomeIt collects the so called “reactions”, where molecules (e.g., proteins and metabolites) form a network of biological interactions and are grouped into pathways (https://reactome.org/).

KEGGIt integrates molecular building blocks with the knowledge on molecular wiring diagrams of interaction, reaction and relation networks (https://www.genome.jp/kegg/).

GO ConsortiumIt is based on the Gene Ontology (GO) classification, which defines concepts used to describe gene function (http://geneontology.org/).

WIKIPATHWAYIt is an open, collaborative platform for capturing and disseminating models of biological pathways for data visualization and analysis (https://www.wikipathways.org/index.php/WikiPathways).

PathwayCommons

It provides a web-based interface that enables biologists to browse and search a collection of pathways from multiple sources (http://www.pathwaycommons.org/pc/).

Main GO “annotations” databases

26

Prediction of secreted protein

Adapted from Nickel, 2003. European

Journal of Biochemistry, 270 (10), 2109-2119

Prediction of

proteins

secreted thanks

to signal

peptide

Prediction of

proteins

secreted

thanks to

other

pathways1

2

3

4

27

Bioinformatics tools to predict secreted proteins (1/2)

Caccia et al., 2013, Biochim Biophys Acta. 2013 Nov;1834(11):2442-53

28

Bioinformatics tools to predict secreted proteins (2/2)

29

Protein-Protein Interaction concept

PPI are considered as physical contacts with molecular docking between proteins

that occur in a cell or in a living organism in vivo

The methods to determine PPI are based on two alternative approaches:

•binary :techniques that demonstrate the direct physical interaction in proteins

pairs, yeast two-hybrid (Y2H)

•co-complex :approaches that study the direct and indirect interactions among

groups of proteins, (tandem) affinity purification coupled to mass spectrometry

((T) AP-MS)

De Las Rivas et al., PLoS Comput Biol. 2010 Jun; 6(6): e1000807

Focus on specific pathways

Node

Edge

30

Database Description

PSICQUIC

Proteomics Standards Initiative Common QUery InterfaCe, is a Community effort to standardise the way to access and retrieve data from Molecular Interaction databases.http://www.ebi.ac.uk/Tools/webservices/psicquic/view/main.xhtml

IMEXIt contains data of major public interaction data providers such as IntAct, BioGRID, Uniprot (https://www.imexconsortium.org/) [68]

IntActIt is an open source PPI database and it provides also analysis tools. All interactions are derived from literature curation or direct user submissions (https://www.ebi.ac.uk/intact/) [64]

STRING

It collects and integrates data obtained by consolidating known and predicted protein-protein association data for a large number of organisms. The associations include direct (physical) interactions, as well as indirect (functional) interactions, as long as both are specific and biologically meaningful (https://string-db.org/) [62]

Pathguide

A compendium of databases for molecular interaction/species.Contains information about 702 biological pathway related resources and molecular interaction related resources.http://www.pathguide.org/

Main PPI databases

31

Examples of complete and curated

databases

32

Protein-Protein Interaction analysis + GO

Adapted from Miryala et al., 2018,Gene

PPI search

Analysis of network topology

GO related to subnetwork

Data mining using

ProteINSIDE online tool

www.proteinside.org

34

Why ?

Generation of large amount of genomic and proteomic data

How can we extract the most relevant biological information ?

Transcriptomics and proteomics largely used to•identify biomarkers of ruminant traits•understand mechanisms behind traits

HOW ?

Main biological

information

Biological Function

Secreted proteins

Protein Interactions

HOW TO ANALYSE AND SUMMARISE DATA ?

An all in one system to automatically query and analyse the results from many tools and databases, and for a list of proteins / genes from three ruminant

species (bovine, ovine caprine) and well-annotated species (Human, mice, rat)

37

datbase

Web interface

Workflow

ProteINSIDE uses many resources that have been chosen according to the probability of their maintenance. ProteINSIDE is constantly updated.

Gathering of biological

information

STRUCTURE OF ProteINSIDE

GOannotation

Prediction of secreted

proteins

BlastP

PPI

38

Basic Analysis

- Paste identifiers- Automatic settings

Custom Analysis

- Paste identifiers- Select your own settings- Use preset buttons

1

2

4

Example of 133 proteins :- Glycolysis (33)- Respiratory chain (12)- Hormones (72)One protein in duplicate

3

To launch a new analysis

39

Pro

teIN

SID

EBiological

information

Ontologie de gène (GO)

Processus biologiques

Localisations cellulaires

Fonctions moléculaires

Protéines sécrétées

Interactions protéine-protéine

Example with 133 proteins :- Glycolysis- Respiratory chain- Hormones

ProteINSIDE automatically summarizes the available information in biological databases for a gene /

protein.

RESULTS of « ID resume » MODULE (1/2)

40

ProteINSIDE automatically summarizes the available information for a gene / protein in the NCBI and UniProtKB databases.

Identifiers

Function

Specificities

NCBI - Coordinators NR 2013 UniProtKB - Magrane et al. 2011

RESULTS of « ID resume » MODULE (2/2)

41

Pro

teIN

SID

EBiological

information

Gene Onotlogy (GO)

Biological Process

Cellular Component

Molecular Function

Protéines sécrétées

Interactions protéine-protéine

ProteINSIDE researches terms (called annotations) for proteins or genes under the "Gene Ontology (GO)"

Hierarchical dictionary of GO terms

Example with 133 proteins :- Glycolysis- Respiratory chain- Hormones

RESULTS of « GO » MODULE (1/2)

42

ProteINSIDE annotated genes / proteins according to the terms of the "Gene Ontology (GO)" and calculated the representativeness of a GO term in the dataset

Function Expected Identified (no IEA)

Identified (with IEA)

Glycolysis 33 15 32

Hormone activity 79 34 78

Validation with the sample dataset:

QuickGO - Binns et al. 2009GO.org - Ashburner et al. 2000

RESULTS of « GO » MODULE (2/2)

43

Pro

teIN

SID

EBiological

information

Gene Ontology (GO)

Biological Process

Cellular Component

Molecular Function

Secreted Proteins

Interactions protéine-protéineProteINSIDE predicts the secretion of a protein with two

complementary analyses: - Looking for a signal peptide on the amino acid sequence - Looking for GO terms related to secretion

Example of 133 proteins :- Glycolysis- Respiratory chain- Hormones

RESULTS of « SECRETED PROTEINS » MODULE (1/2)

44

ProteINSIDE predicts the secretion of a protein by using the algorithm SignalP and GO terms related to secretion

+

Expected Identified(only peptide prediction)

Identified(peptide + GO)

Identified(peptide + GO + IEA)

79 85 63 78

Validation with the sample dataset:

SignalP - Petersen et al. 2011

+

RESULTS of « SECRETED PROTEINS » MODULE (2/2)

45

Pro

teIN

SID

EBiological

Information

Gene Ontology (GO)

Biological Process

Cellular Component

Molecular Function

Secreted Proteins

Protein-Protein Interactions

Analysis of interaction networks (interactome) to identify the function and / or the central role of

proteins in biological mechanisms

Example of 133 proteins :- Glycolysis- Respiratory chain- Hormones

RESULTS of « PPI » MODULE (1/2)

46

ProteINSIDE creates protein networks to reveal biological

mechanisms

Between proteins within the dataset (36 interactions)

Between proteins within and external to the dataset

(616 interactions)

PsicQuic - Aranda et al. 2011 Cytoscape web - Lopes et al. 2010

RESULTS of « PPI » MODULE (2/2)

47

BioMyn

6 “synthetic” datasets of 1000 random proteins=> 1dataset /species

Database for Agricultural species (plants and animals, including bovine and ovine ; McCarthy et al., 2006)

Human database with GO and PPI analyses (Ramírez et al., 2012)

Multi-species database, intensively used by biologists (Huang

et al., 2009)

Bench test of ProteINSIDE relatively to others tools

Predictions of secreted

proteinsAlgorithms PrediSi (Hiller et al., 2004 ) and Phobius (Lukas et al., 2004)

48

BENCH TEST

Kaspirc et al., 2015, PlosOne

49

Wo

rkfl

ow

Biological Information

Gene Ontology (GO)

Secreted Proteins

Protein-Protein Interactions

ProteINSIDE offers great support to the analysis of the large amount of data generated by genomics and proteomics. ProteINSIDE is also a

unique tool that performs these analyses using ruminant IDs.

CONCLUSION

Target species:

- Bovine

- Sheep

- Goat

Model species:

- Human

- Mouse

- Rat

ProteINSIDE works using 6 species

ProteINSIDEDatabase

ProteINSIDEWeb interface

ProteINSIDE searches homologous proteins

Using Bovine IDs to search the homologous proteins in Humans

Q3Y5Z3 Q15848

Prediction of the Secretome of tissues

24 publications + 8 datasets (ruminants) proteins secreted : 1749 by bothmuscle & adipose tissue, 357 muscle-specific and 188 adipose-specific

Cross-talk between these tissues for their functioning and the construction of animal phenotype?

51

Descriptive and differentialstatistics

Data mining : when proteomics turns functional

Looking for biomarkercandidates

Data analysis and data mining of lists of

proteins ID and abundances

Proteomics to identify biomarkers

of marbling in bovineDiscovery

In groups of ruminantsdivergent

for a phenotype

QualificationIn ruminants with a range

of variability for a phenotype

ValidationNew

population

VerificationOn large cohorts

< 10 samples> 100 candidates

> 1000 samples< 10 candidates

2-DE or Shotgunproteomics

OR ?

Timeline of biomarker discovery

Trageted

proteomics

Immunological

tests

Step 1 Step 2 Step 3

How to select candidate biomarkers ?

Biomarker selection uses a variety of models to identify a subset of biomarkers that

best differentiate the control from test samples.

Recommended group sizes => from 10 samples and above. The models that are

generally used include:

• Logistic regression

• Linear discriminant analysis (LDA)

• Support vector machine (SVM)

• Random forest

• Other models that may be used

Biomarker Selection

https://www.tebu-bio.com/blog/2019/01/31/proteomics-biostats-get-the-most-out-of-your-array-data/

• MYDATAMODELS (12000 €/y)

https://www.mydatamodels.com/

• Free R package

ModVarSel (Ellies Oury et al., 2019) and one package in construction for predictive models (Molnar et al.)

Some « Biomarker Selection »

A new computational methodology that simultaneously selects the

best regression model and the most interesting covariates

Different regression models (including variable selection) :

1- parametric (multiple linear regression (MLR)),

𝑌 = 𝛽0 +

𝑗=1

𝑝

𝛽𝑗𝑋𝑗 + 𝜀

2- semiparametric (sliced inverse regression (SIR))

𝑌 = 𝑓(

𝑗=1

𝑝

𝛽𝑗𝑋𝑗) + 𝜀

and nonparametric (random forests (RF))

𝑌 = 𝑓 𝑋 + 𝜀

What is ModVarSel

Model choice using MSEtest

How to choose a model ?

How to select variables/proteins ?

• To install the current version of the R package modvarsel from github, use : https://github.com/chavent/modvarsel.

• DOI : https://zenodo.org/record/1445554#.W7YrYPY69PY

• More details are given here : https://chavent.github.io/modvarsel/modvarsel-intro.html

• Example of application in Ellies-Oury et al., 2019 Sci Reports

Supporting information

Descriptive and differentialstatistics

Data mining : when proteomics turns functional

Looking for biomarkercandidates

Take Home message

Data analysis and data mining of lists of

protein IDs and abundances

60

Before statistical analysis, try to answer to two simple questions: i) how many

features, and ii) how many samples.

Carefully consider the final goal of the proteomics analysis

• for classification purpose => multivariate supervised methods provide the best

results.

• to find some hidden patterns within features => unsupervised methods should be

chosen and batch effects must be carefully taken into account and avoided when

feasible.

Combine different pathways analysis tools to take advantages of their power of

analysis regarding your objectives, try ProteINSIDE and send me feedback

Take Home message

Challenges remain for the development of mathematical methods to select

biomarkers and predict a trait of interest from proteomics data.

www.proteinside.org

Thank you for your attention