Multi-resolution network modelling of T-cells for precision medicine of

multiple sclerosis

Mika Gustafsson, Assoc Professor

Translational bioinformatics

Linköping University

Motivation

• Most medications starts late, are ineffective for most patients and costly

• Patients who appear to have the same disease could be sub-types of the same disease, with different mechanisms

• Altered function or expression of thousands of genes in different cells changing as disease progresses

Aim is to develop network based methods for identification of predictive markers of complex diseases, so treatment can be early and individualised, ideally before disease becomes symptomatic

Multiple sclerosis (MS) and T-cells as models

• Complex disease with 25% concordance of monozygotic twins

• >100 SNPs exists and T-cell differentiation is the mosly enriched

• Relapse-remitting MS have recurrent non-symptomatic phases, so patients can be followed with and without symptoms

• T-cells can be activated into a ’disease alike state’, and short-time dynamics analysed (Hellberg, Cell Reports 2016)

Integrative disease models needed, e.g. networks

Many undiscovered disease genes have small individual effects

Interactome is

a hairball

Interlinked omics -> coarse-graining needed

Gustafsson et al, Genome Med 2014

GWAS co-localize in PPI disease modules

• GWAS is static and limited to genetics

• Gene expresssion profiling of a disease relevant cell type can increase sensitivity and includes epi-genetic effects

How do you validate modules containing hundreds of genes?

• Genomic concordance, i.e. overlap across omics

• GWAS was used for validating toplogical disease modules in DREAM 2016 by Marbach

• Clinical studies and functional animal studies of combinations of module gene products as biomarkers

Hellberg et al, Cell Reports, 2016

Bruhn et al, Science Translational Medicine, 2014

Gustafsson et al, Genome Med, 2014

nine ongoing studies on disease modules

https://gitlab.com/Gustafsson-lab/MODifieR Performed in collaboration with Z Lubovac, Skövde (modules)

• Diamond

• DiME

• CHD

• MCODE

• DiffCoex

• MODA

• Module Discoverer

• CliqueSum

• CliqueSum + Correlations

• WGCNA

Preliminary results

Different methods performs best on different data-sets

0

0,5

1

1,5

2

2,5

3

3,5

4

4,5

Multiple Scelerosis (Sawcer Nature Genetic 2011)

0

0,5

1

1,5

2

2,5

3

3,5

4

4,5

Rheuatoid arthritis CD4 (Okada2014 European)

- lo

g(P

)

Modules from the same methods most similar

PC

1 (

95

% o

f va

rian

ce)

Preliminary results using four methods on 24 diseases:

https://gitlab.com/Gustafsson-lab/MODifieR

Conensus module outperforms all indivdual methods

Summary: Modules

• Modules concept is vaguely defined

• Strong method bias

• Assessment needed for development of new methods and integrating new data typs

Gene regulatory networks (GRNs) provide directionality

• Modules are time-independent, undirected and does not provide mechanism

• GRNs give possibility to go upstream leading to early disease markers [1]

• Data compression: ≈7% of protein-coding genes are transcription factors (TFs), some TFs are master regulators

• TF targets can be predicted from DNA sequence and the functional targets can be inferred by time-series and perturbation experiments

[1] Gustafsson et al, Science Translational Medicine, 2015

Gustafsson et al., A N Y A S, 2009

Gustafsson et al, CMGRN, 2008

Gustafsson et al, IEEE Comput biol/Bioinfo., 2005

Hubs TFs in T-cell diff were GWAS

19 Gustafsson et al, Science Translational Medicine, 2015

• TFs degree might be easier to infer than indivdual interactions • mRNA is often used as a proxy of TF activity relied on average healthy GRNs, i.e. were not indvidualised

GRNs for analyzing multiple sclerosis (MS)

• MS has >100 reported SNPs that are annotated as eQTLs, which are in T-cell genes, but surprisingly those target genes are not differentially expressed

• Instead those genes exhibited the highest variance in relapsing patients

• We hypotheisize that personalised control for the upstream TF activity will resolve this

Collaboration with J Ernerudh, LiU (MS), S Baranzini, USA (MS) and R Lahesmaa, Turku (high-throughput biology)

42 TFs enriched for target MS-genes

• TFs were surprisingly not among target genes -> no genetic feedbacks

Individualised GRN for each MS patients

Model details

• Targeted DNA sequencing of 1800 MS associated eQTLs in 48 patients/controls

-> 48 different prior TF-target networks

• 96 gene microarrays from actve and non-active T-cells,

𝑋𝑖 = 𝑚𝑅𝑁𝐴𝑖 𝐴𝐶𝑇𝐼𝑉𝐸 −𝑚𝑅𝑁𝐴𝑖[REST]

𝑆𝑖𝑗𝑘 = 𝛾𝑖𝑗 𝑖𝑓 𝑝𝑎𝑡𝑖𝑒𝑛𝑡 𝑘 ℎ𝑎𝑠 𝑆𝑁𝑃 (𝑖, 𝑗)

0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝛾𝑖𝑗 = 𝑠𝑖𝑔𝑛(min𝑆𝑖𝑗𝑋 − 𝑆𝑋 2)

𝐴𝑗𝑘 − 𝑋𝑗𝑘 < 𝜂𝑗

𝑆𝑖𝑗 < 𝜆𝑗 Sparse network

L1 Deviation between mRNA of TF and its activity

min𝐴𝑋 − 𝑆 ∗ 𝐴 2

Removing putative edges and allow for deviation between TF activity and mRNA expression lowers cross-valdiation error

Some TFs are hubs in patients

Some TFs that predicted diff activated

Combining topology and diff activation

Summary: patients specific GRNs

• Longitudinal data opens the door of indivdual networks

• Prel data suggests: Hub upstream TFs can be used to cluster MS patients

• Validation experiments are ongoing, i.e. ATAC-seq and western-blots

• Generally applicable principle using paired longitudinal data at two disease associated states

• Ongoing SSF project: Use in vitro differentiation time-scale of ~24h and gruadual disease recovery during 9 months pregnancy to fit a nonlinear dynamic model

Realistic ODE modelling including feedbacks is needed for predicting effect of drugs • Early hub TFs are potential bottlenecks whose dysregulation could be

early markers for disease [1]

• We aim to identify their upstream early activity regulation, which could give earlier markers and predict drug responses

• Requires detailed ODE modelling of a core set of TFs, their upstream signaling paths and their thousands of target genes

• We therefore collect high-throughput proteomics, phospho-proteomics and transcript variants

[1] Gustafsson et al, Science Translational Medicine, 2015

Mechanistic modelling of the effect of TFs and their upstream regulators in health is possible using LASSIM

Magnusson, et al,PLoS Comput Biol, 2017

gitlab.com/Gustafsson-lab/lassim

Proteomics, transcriptomics and phospho-proteomics of six time-points of Th1 differentiation, combined with inhibition data and priori knowledge modelling to model MS

treatment in vitro

Protein

RNA-Seq

0.172229 0.173644 0.319013 0.00834591 0.046203005

-0.309512274 -0.647908085 -0.338634499 -1.226836197 -0.325007143

0.179318 0.219942 0.315562 0.079328 0.335925053

0.175871924 -0.40422443 -0.319218089 0.049761601 0.934979233

0.194538 0.247914 0.235476 0.205252 0.065867855

-0.439616434 -0.509785792 0.320426146 0.333162732 0.254898801

-0.198453 -0.288001 -0.343605 -0.41238 -0.247270831

0.244075857 -0.282880796 -0.497596803 -0.507578947 0.214281966

0.163254 0.21547 0.218799 -0.0611065 0.390113246

0.958503673 0.485058215 -1.75667742 -1.11987114 0.580351144

0.962017 1.27194 2.60522 1.86439 1.194252387

4.891851562 4.988597738 0.117176488 2.836889826 -1.870714429 1.00619 1.2994 2.21176 3.58514 4.26919974

1.012118145 2.888966616 6.61475359 7.815029445 7.069508923

#N /A #N /A #N /A #N /A #N /A

0.158502153 2.412843053 8.355836692 11.58024281 10.2193337

#N /A #N /A #N /A #N /A #N /A

-0.996670673 -1.042070673 0.296515886 -0.600143736 -0.564410533

-0.00247007 -0.0571366 -0.169218 -0.369788 -0.480510915

-0.496917424 -0.42916554 0.192262702 -1.120203906 -2.121066947

-0.0902545 0.790325 0.149592 0.657427 0.11948294

0.528703699 0.425547298 -0.75672533 2.246075755 3.002246612

-0.00820334 -0.0936787 -0.110092 0.891726 0.80128827

0.094182644 -0.316919898 -0.197078072 0.464574589 0.2445864

0.305004 0.396862 0.495147 0.725293 0.694410725

-0.541039098 -0.107932393 0.509085125 -1.27993021 -1.639710973

0.619702 0.699076 1.11085 0.71581 0.817489289

2.705314645 1.534350071 -2.004124457 -0.047850555 -2.516189326

0.124337 0.204385 0.198768 0.0792561 0.684639703

-0.176162156 -0.240920939 0.308033951 0.478875326 0.630974662

0.581939 0.402258 0.340685 0.440113 0.468735434

0.204136489 -0.038939407 -0.92280189 -1.92269937 -2.236200193

-0.154232 -0.122263 -0.182047 -0.0394336 0.036493394

-0.705026586 -0.035072744 0.204904687 0.231208718 0.18752275

0.204208 0.163731 0.266377 0.12595 -0.341823282

0.039170007 0.196453078 -0.266846099 0.019201448 -0.492807854

0.349388 0.344859 0.537454 0.610781 0.830761537

-0.695533511 -1.00911776 0.701890412 -0.304037673 0.326928842

0.157578 0.24092 0.392948 0.862859 0.090121252

-1.032478795 -0.00596674 2.071140267 2.413984179 1.04195881

0.119292 -0.020047 0.0327164 -0.201421 -0.137726232

0.961307671 0.326624465 -0.894313193 0.09567435 -1.62659809

0.139907 0.288115 0.199657 -0.0708241 -0.300249556

0.866566767 0.347641634 -0.179917778 -0.390892569 1.433634894

0.0847869 0.00893925 -0.0950481 0.270325 0.705097791

-0.090076511 0.11928289 0.304722309 0.190094046 -0.199039482

0.605098 0.54346 1.29067 2.33199 1.033689581

0.153866145 1.442158985 1.836850816 3.626554991 0.834842716

0.0847869 0.00893925 -0.0950481 0.270325 0.705097791

-0.090076511 0.11928289 0.304722309 0.190094046 -0.199039482

0.370984 0.385355 0.325066 0.0969046 0.117309532

0.326375862 0.115799864 0.524430124 0.58169903 -0.307391529

0.255345 0.234616 0.244797 0.783022 1.778730113

-0.231847812 -0.02372434 0.380498184 2.144834001 2.034715346

0.255345 0.234616 0.244797 0.783022 1.778730113

-0.231847812 -0.02372434 0.380498184 2.144834001 2.034715346

0.255345 0.234616 0.244797 0.783022 1.778730113

-0.231847812 -0.02372434 0.380498184 2.144834001 2.034715346

#N /A #N /A #N /A #N /A #N /A

-0.925668906 0.050723229 0.822341825 0.077661961 -0.132427925

0.311843 0.283361 0.343082 0.609991 0.160668786

-0.827150015 -0.812891686 1.484365893 -0.143805789 -0.877207719

0.263471 0.14661 0.77794 2.33932 1.322042021

-0.632272543 -0.50146355 2.160455684 4.116607379 3.416822379

0.920842 0.481933 0.388637 -0.129385 -0.214682741

-0.16621994 -0.45685769 0.023321858 -0.310151374 0.08464365

-0.0140578 0.0692668 -0.0564365 -0.262897 -0.03558365

-0.0371606 -0.353831787 -0.183740696 -0.550811259 0.252952433

0.107125 0.185436 0.263647 0.0444783 0.159309497

0.377660534 0.126130242 -0.614060057 -1.18962268 0.382132678

-0.281417 -0.209216 -0.302301 -0.540069 -0.234579024

0.065620319 -0.593100002 -0.339814619 0.064291985 -0.862133544

0.309888 0.32506 0.486814 0.324265 0.341703242

-1.007686755 -0.870826288 0.920626313 -0.038809076 -0.412061811

0.582327 0.59437 1.64459 1.0735 1.354395111

-0.427892943 -0.810231303 0.063253191 -0.108661237 0.11722832

0.316134 0.176818 0.0960334 -0.216136 0.672668298

-0.181805392 -0.190156057 0.154821884 -0.639788838 -2.081126777

0.349388 0.344859 0.537454 0.610781 0.830761537

-0.695533511 -1.00911776 0.701890412 -0.304037673 0.326928842Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

0.0946169 0.0463398 0.0388869 0.398087 0.876554782

0.646895256 0.652197007 -0.540206529 0.734261688 0.667077338Protein

RNA-Seq

0.320818 0.29872 0.456471 0.729564 1.021647513

1.216582576 1.15510153 -1.163093131 0.845535735 -0.141648254Protein

RNA-Seq

0.0437414 0.1347 0.183243 -0.124816 0.163902382

0.900008328 0.592462137 -1.414031529 -0.311017452 0.860532067

0.439565 0.468622 0.55403 0.387486 0.350146193

-0.548503054 -0.403284227 0.508443382 0.251313244 0.041268818

Protein

RNA-Seq

0.397583 0.257947 -0.0127637 -0.260787 0.263467541

0.134505931 -0.251955551 -0.037141485 0.522493966 -0.259153782Protein

RNA-Seq

0.175557 0.154436 0.111639 0.321946 0.471640153

0.143812432 -0.107810047 -0.71599472 -0.061621369 -0.606891898Protein

RNA-Seq

0.291115 0.387453 0.341561 0.208296 0.124201869

-0.245812622 -0.061880372 -0.109441034 -0.957199827 0.473122018Protein

RNA-Seq

0.605098 0.54346 1.29067 2.33199 1.033689581

0.153866145 1.442158985 1.836850816 3.626554991 0.834842716Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

0.370984 0.385355 0.325066 0.0969046 0.117309532

0.326375862 0.115799864 0.524430124 0.58169903 -0.307391529Protein

RNA-Seq

0.309888 0.32506 0.486814 0.324265 0.341703242

-1.007686755 -0.870826288 0.920626313 -0.038809076 -0.412061811Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

Protein

RNA-Seq

0.311843 0.283361 0.343082 0.609991 0.160668786

-0.827150015 -0.812891686 1.484365893 -0.143805789 -0.877207719Protein

RNA-Seq

Protein

RNA-Seq

Collaboration with M Kim, Korea (mass-spec) and C Nestor, LiU (T-cell biology)

Preliminary results • 7000 proteins and 240,000 transcript variants detected, corresponing to about

15,000 unique genes that are differentially expressed at some time-point of Th1 differentiation

• Awaiting perturbation and phsopho-proteomics

Summary

• Modules are powerful integration tools that increase functional relevance, but still not well-defined and do not give mechanism

• Gene regulatory networks may identify upstream early master regulators • Nonlinear modelling of those factors for predicting drug effects • Simultaneous integration of those concepts should ideally add synergistically

2015- my group includs medical doctor, computer scientists, theo. physicist, immunologists, and engineers

Hard work performed by

• Tejaswi Venkata Badam

• Dirk de Weerd

• Simon Wu

• Julia Åkerström

• Rasmus Magnusson

• Andreas Tjärnberg

• Olof Rundquist

Thank you!