1. Building and using protein interaction networks: industry perspective
Systems Biology for Drug Discovery
Andrej Bugrim
GeneGo, Inc.

2. Topics
Annotation process and collecting network content for idustrial-type applications
Biological and disease ontologies – how to improve and use them in functional analysis
Tools: utilizing network data in pharmaceutical R&D

3. Multi-level understanding of human biology
Causative
relations
Mechanistic
Level of
phenotype
Level of
Cell process/
network
Level of
protein

4. Disease-centered knowledge base in MetaMiner (Oncology example)
GG annotation team
Causative disease associations:
DNA, RNA, protein levels
Disease group
Protein-protein; Protein-DNA;
protein-RNA interactions
Network group
Biomarkers
Specialty group
Ligand-receptor
interactions: drugs,
leads, hits
Chemistry group
Compare
BC-perturbed cell processes
Causative BC models
General BC schema
Other cancers chosen by Consortium

5. Content

6. Three interactions domains in MetaCore
1,600 drugs w/targets
4,100 endogenous metabolites
>21,000 ligand-receptor interactions
850 GPCRs and other membrane receptors
110 Nuclear hormone receptors
Ligands: metabolites, peptides, xenoboitics
Membrane receptors
Signal transduction:
G proteins,
Secondary messengers
Kinases
Phosphotases
172K manually curated physical signaling interactions
538 canonical maps
42, step canonical signal transduction pathways
924 Human transcription factors
6,000 target genes
Transcription factors
Only MC has all 3 levels needed for functional recinstruction: ligand-receptor (primary signaling) level; signal transduction level; effector level (core metabolism).
DONE
11,300 metabolic reactions
116 Fine metabolic maps
Core effect: metabolic pathways
Metabolites
4,100 endogenous metabolites

7. MetaBase Content Overview
Database
Chemical compounds 580,000
Drugs 8,590
Chemical Reactions 35,600
Metabolic networks 251
Network
Proteins + genes 13,402
Transcription factors 924
Chemical compounds 26,000
Drugs 2,740
Endogenous compounds 4,100
Proteins linked to drugs 2,711
Reactions 5,330
Small molecule ligands for human receptors 3,510
blockers for ion channels 629
Pubmed journals 3,100
Pubmed articles 81,400
Total amount of interactions 177,000
Content
GeneGo regulatory networks 120
GeneGo disease networks
Maps
Regulatory maps 325
Metabolic maps 116
Traditional metabolic maps (EC) 97
Diseases 4,920
DONE (VK, )
Will work on it, provide histograms

8. MetaBase content by type
Database
Genes
(human:
38,700)
Total:137,500
Chemical
compounds
580,000
Human
proteins
14,570
Metabolic
reactions
35,600
DONE

9. Network interactions Manually curated interactions (172,787)
Signalling interactions;
137,297; 79%
Metabolic reactions; 35,490;
21%
Y2H "Interactome"; 2,370; 1%
Logical relations; 1,934; 1%
Protein-protein; 87,675;
51%
Small molecule-protein;
42,383; 26%
With MicroRNA; 1,620; 1%
With virus protiens; 335; 0%
Chip-Chip; 980; 1%
DONE
Counted collapsed interactions only, uncollapsed are over 650,000!
All interactions taken from articles indexed in Pubmed
Pubmed journals 3,100
Pubmed articles 81,400

10. Type of interactions in network
Effects
activation
inhibition
unspecified
Direct interaction
Indirect interaction
Mechanism
phosphorylation
influence on expression
dephosphorylation
unspecified
other type of covalent modification
binding
transport
cleavage
transcription regulation
transformation
catalysis
competition
1. We split all protein-protein and protein-compound (chem.) interactions on two main types: direct and indirect (functional)
2. Within both types of interactions, we separate distinct mechanisms which are defined by experimental evidence for interactions. There are 10 mechanisms for direct interactions and two mechanisms for indirect interactions.
All interactions, direct and indirect, are split on directed and undirected ones and are marked with “effects”
3.1 Directed interactions are characterized by causative relationships between network objects (proteins, genes, compounds, complexes etc.).
3.2 Undirected interactions are marked with effect-mechanism

11. Distribution of interactions by mechanism
DONE

12. Network objects
Total number of nodes: 40,229
DONE

13. Proteins: distribution by tissue & localization
DONE

14. Molecular functions in Database
DONE

15. Endogenous compounds (4,100 total)
3,070 endogenous compounds involved in metabolic reactions: 6,819 reactions with endogenous compounds only
751 endogenous ligand for 498 receptors with 2,455 interactions
4000 (98%) of endogenous compounds in network
15,962 network interactions with endogenous metabolites
3,600 compounds with structures and brutto-formulas (other 700 are “generic”: contain acyl-, alkyl- and other variable groups)
DONE
Note: 4,000 endogenous metabolic compounds: this is 4x more then Biocrates, the leader in human metabolomics has!!!

16. Network and pathway statistics in GeneGO
>40,000 nodes;
~177,000 edges;
Average node degree: 3,77;
241 million shortest pathways;
Average shortest pathway length: ;
42, step canonical signal transduction pathways;
200 canonical metabolic pathways- major metabolic fluxes like glycolysis or TCA;
72,000 pathways on metabolic maps: pathways analogous to KEGG (KEGG has 42,500)
DONE
Enzyme1
Enzyme2
reaction1
reaction2
metabolite

17. Pathways in regulatory network
Start: TMR
(transmembrane receptor)
TF (Transcription Factor) a a b
End: Target genes

18. Ontologies

19. Knowledge base (ontologies)
By year:
2007
2006
2005
2004
2003
By genre:
Drama
Action
Romance
Horror
Foreign
By director:
Lynch
Tarantino
Leone
Stone
Antonioni
By actor:
Pitt
Nicholson
Depp
Redford
Damon
How do you compare “action” movies vs. Tarantino movies vs movies?
These are incomparable as these are different categories
Molecular pthwy
Cellular process
Disease
Metabolic process
Mixed ontologies

20. Multiple ontologies in MetaDiscovery Platform: multi-dimensional knowledge base on human biology

21. GeneGo metabolic maps vs. KEGG human maps
Human genes in GG and KEGG
DONE, VK

22. Enrichment in GO and GeneGo processes
GeneGo process networks
Resolution: interactions between proteins
Connections between all proteins in folder
Clear signaling path, effect within process
Resolution: list of proteins
No connections between proteins
No sgnaling/effect within process
4 samples from 4 patiens
Disease/norm from same patients
Affy U133A arrays

23. 231 613 446 268 345 1642 1297 Inflammation Genes from GO process
“Inflammatory response”
231
Genes from GO-processes
“Inflammatory response”
“Immune response”
613
Genes from GO-process
“Immune response”
446
Not in networks 79
Not in networks
199
In networks
152
In networks
247
Not in networks
268
In networks
345
Синие блоки – исходные статистические данные
Зелёные блоки – объекты GО-процессов, которые привязали к созданным нетворкам
Красные блоки – объекты GО-процессов, которые пока не привязали к созданным нетворкам
Желтые блоки – объекты, добавленные по смыслу представленных в нетворках процессов, согласно литературным данным
Genes in 15 process networks
1642
Genes added to networks
1297

24. 21,264 unique articles, indexed in PubMed
Diseases
4,881 Diseases, based on MeSH
38,709 Human genes total
Human genes linked to diseases – 6,318
Diseases linked to genes – 1,630
Human genes not linked to diseases – 32,391
Diseases with no gene links – 3,251
DONE
6,318 genes are linked to 1,630 diseases
21,264 unique articles, indexed in PubMed

25. Disease tree – Neoplasms by Site

26. Folders created at GeneGo based on reviews
Drug toxicity tree
38 Drug-induced pathological processes
Folders from MeSH
Folders created at GeneGo based on reviews
Мы создали иерархическую структуру патологий – дерево токсичности. Эта работа выполенена, в основном, на основе анализа статей. Исходно, дерево болезней MeSH содержало folder Drug toxicity and subfolders Drug Erutions with its subfolders (with the exception of Sweet's Syndrome).
Нет отдельного обзора, где перечислены все эти виды токсичности. Они встречаются в различных статьях и обзорах.
К настоящему момнту мы собрали 38 видов органоспецифичных патологий, вызванных драгами.

27. Gene-Disease connections in public domain and GeneGo
OMIM
Only genetic info (mutation, SNPs)
No expression
No protein activity, loc
GENE
MeSH
Only citation with
Diseases name. Low trust
Only
hierarchical structure
disease tree
Public domain does not have structured information
about disease connectivity(by clinical classification)
and causative relations withgenes and proteins
GeneGo
Сравнение публичных баз и GeneGo - связь генов и болезней
Hierarchical strusture
disease classification ,888 diseases
Genes associated with diseases ,429
Cited articles , 792

28. Content. Cancer maps and networks. Breast Cancer: general scheme
Далее на каждом слайде будут детально разобраны основные процессы и задействованные в развитии ВС сигнальные каскады.
На все нижеследующие нетворки промепплены описанные в литературе маркеры ВС - красные блямбы, чтобы детально отследить задействованость участников исследуемого процесса в ВС.

29. Angiogenesis in tumor growth
Для успешного роста опухоли крайне необходимо ее своевременное прорастание сосудами, поскольку в быстрорастущей опухоли центральные ее отделы испытывают недостаток кислорода – может наступать некроз тканей. В таких условиях главным образом активируется HIF-A фактор ответственный за реакцию на гипоксию – и активирующий ангиогенез через VEGF. Если опухоль растет быстрей чем сосуды, то наблюдается некротизирующий процесс и воспаление, через IL-1, 6, 18, что также через NF-kB приводит к синтезу VEGFs.
A key angiogenesis activator is VEGF-A, which can act alone or via its basic receptors VEGFR-1 and VEGFR-2 and neuropilins as corecetors. Other members of VEGF family are also involved in the process. One of transcriptional activators of VEGF-A is HIF-1 that is responsible to induction of angiogenesis under hypoxia. Moreover, some other growth factors (FGFs, EGF, PDGF, PlGF, TGF-beta) and MMPs are involved in regulation of angiogenesis. Angiopoietins 1 and 2 act on later stages on angiogenesis via TIE receptors. Integrins are required for an intercellular adhesion that occurs during angiogenesis.

30. Fine metabolic differences between rodents, human
Unique genes
Human
Mouse, Rat
Unique genes and orthologs catalyse one reaction
141 mouse genes
74 rat genes
There is no human orthologs
for Protein A
Unique genes catalyze unique reactions
9 mouse genes
2 rat genes
Orthologs catalyse different reactions
1 mouse gene
1 rat gene

31. Tools

32. Data analysis workflow in MetaDiscovery suit
PathwayEditor
MetaLink
MapEditor
Custom interactions data:
Y2H
Pull-down
Co-expression
annotation
Custom maps,
networks, pathways
ISIS DB
Molecular bio data
Structures
sdf, MOL
Metabolites
HTS, HCS
HTS, HCS
MetaCore/MetaDrug platform
Signature networks
Diseases
Drug response
P-value scoring
Ontologies:
GO processes
GeneGo processes
Canonical pathways
Metabolic networks
Toxicities
Cross-experiment comparison
Time series
Multi-patient cohorts
Multiple logical operations
Complete report
Network alignment
Multiple algorithms
Sub-network queries
This is the general schema for network analysis. High throughput data can be mapped directly on Metacore interaction tables and unique for the data file network will be built. Note that the network provides with highest possible resolution mapping of the data: at the level of individual proteins/genes and single-step physical protein-protein interactions and one step metabolic transformations.
The networks can be also built using third party tools and compared withing Metacore.
Upon generation, networks can be interpreted in the context of canonical pathways (in this case, Genego maps) and different process ontologies (for example, GO categories) and prioritized based on p-value and data saturation. Based on this analysis, new hypotheses can be generated for drug targets, siRNA perturbations and biomarkers
Med. chemistry:
Indications
- Toxicities
- Off-site effects
Modeling software:
CellDesigner
Virtual Cell
SBML, BioPax
Biology:
- Biomarkers
- Pathway-based targets

33. curated interactions from the literature
MetaCore™ Platform
Networks
Building Tools
Pathway editor
Statistics for pathways,
processes, networks
Visualization
Tools
Data:m-arrays, SAGE, proteomics,
siRNA, metabolites, custom interactions
Logical operations module
curated interactions from the literature
Self-explanatory. Composition of Metacore
Oracle Based Database

34. Signaling, regulation, metabolism, diseases
Pathways Integration
Interactive, static maps
550 maps
Signaling, regulation, metabolism, diseases
Backbone of formalized “state of art” in the field
Networks of protein interactions
Dynamic; built “on-the-fly”
Exploratory tool
Build new pathways for genes of interest

35. Choose direction and checkpoints within network building page
From – histamine
through – histamine H1 receptor
to – Actin
Выбираем направление сигнала:
From – histamine, through – histamine H1 receptor, to – actin.
Получили кратчайшие пути. Однако при этом не попали Rac1 и RhoA.

36. Non-significant bars become semi-transparent
False discovery rate filter i
Threshold
0.01
Apply
Non-significant bars become semi-transparent

37. DNA level data MetaDiscovery tools
Input data
Illumina, Affy SNP arrays
Epigenomics data
ChIP/Chip, CGH,
Resequencing
Taq assays
Analysis
Correlation with clinical data
Direct data analysis
Correlation with molecular data
MetaDiscovery tools
Pharamacogenomics
Disease biomarkers
Applications

38. New customization modules
MapEditor: custom maps synchronized with MC/MD database
Draw pathways maps from scratch
Transform gene lists into networks into pathway maps
Edit MetaCore’s canonical maps
View and score your maps within the context of canonical maps
Map experimental data on custom maps
MetaLink: overlaying custom interactions
Import custom interactions (Y2H, co-expression, pull-down, etc.)
Visualize using GeneGo network building algorithms
Score “unknown” proteins (high IP potential) based on relevance to “benchmark” networks built from MetaCore interactions
PathwayEditor: annotation technology transfer, at the database level
Custom annotation of interactions, compounds, diseases, metabolism in the framework of internal annotation system at GeneGo
Use the annotation forms, workflows and QC system developed at GeneGo
Novel objects are imported and integrated with pre-existing data in MetaCore

39. Additional Localizations can be added
Adding Localizations
Additional Localizations can be added

40. Your NEW map is now an interactive part of MetaCore
Users can visualize their experimental data on the new map

41. Mapping interaction sets on networks
Resulting Direct Interactions network
Pink interactions are from the uploaded links file
Mouse over an interaction to see the uploaded weight value
Blue interactions are in both the links file and the MetaCore database

42. Systems level analysis
A NEW paradigm!
Many, many products
MetaDrug
Novel compounds
Structure-based
modeling
QSARs models
activity
Dry lab
predictions
Pharmacophores
QSAR models
metabolism
toxicity
Present a structure
as a GROUP of proteins
Query KNOWLEDGE base:
Diseases
- Toxicities
Pathways
Processes
Networks
IND BIOLOGAL effects:
Indication
Toxicity
Off target effects: =
GROUPS OF PROTEINS!
Score based on analysis
Indication
Toxicity
Select best structures
3,400 disease related genes
500+ diseases w/genes
600 known drug targets
3,000 toxicity related genes
700 cellular processes
200,000 protein interactions
Report BY and FOR
end user = chemist =
1 protein at a time =
Systems level effects =
Systems level analysis

43. Integration with MDL DiscoveryGate
Search compound of interest or its metabolites in MDL databases
*(requires access to Discovery Gate)
Copyright GeneGo

44. Merge networks

45. Algorithms

46. Old and new ways to analyze data
Current way of analysis:
all significance calculations done before mapping onto network
Statistical procedures,
thresholds of fold, p-value either in MC or 3rd party tools
Full data
tables
Connect them on network by one way or another:
Too many choices, no clear way to choose
Sets of genes
New way of analysis:
significance calculations follow the mapping onto network
Statistical procedures in MC based on concurrent analysis of expression profiles and connectivity
Full data
tables
Apply to global network
Sets of network modules

47. Samples are analyzed in pathway’s expression space
Calculating distance between samples in the pathway’s gene expression space. Hypothetical pathway is shown, consisting of three proteins A, B and C. Proteins are the products of corresponding genes. (A) Samples are represented as points in 3-dimensional space of gene expression. Note grouping of samples. (B) Representation of the pathway: corresponding gene expression (fold change compared to control) is represented by arrows.
Sample 1
Sample 2
Sample 3
Sample 4
Gene 1 1 4 3 2
Gene 2 7 6
Gene 3 9 8
Gene 4 5

48. Network signatures for compounds effects
Mestranol
Phenobarbital
Differentially respond to drug treatments on the pathways. (A) Direct interaction network assembled from the genes extracted from the pathways distinguishing between mestranol and Phenobarbital treatments. Relative change in expression between treated and untreated rats are mapped (log-ratios, averaged over 5 repeats). Blue circles mark down-regulated genes, red circles mark up-regulated genes. (B) The network assembled from the genes extracted from the pathways different between Phenobarbital and Mestranol. Note that both comparisons contain significant numbers of negatively correlated genes.
Tamoxifen
Phenobarbital

49. Finding topologically significant nodes
Not topologically significant B C
4 out 6 under nodes regulated by B are differentially expressed: more than random share = significant
Only 1 out of 6 nodes regulated by C is differentially expressed: could be due to random event
= not significant
In reality algorithm also considers nodes beyond first-degree neighbors
Differentially expressed genes
Non-differentially expressed genes

50. Why JAK1 is significant in this dataset?
Regulation via JAK1
Feedback loops
Uterus in labor vs. non labor Wayne State
JAK1 provides essential network conduit between PLAUR and many differentially expressed targets of STAT1
Topological significance helps to find important links in pathways that do not come up on HT screens

51. Regulation of lipid Metabolism
Topologically significant nodes revealed by the new algorithm
Differentially expressed genes identified by microarray and confirmed by proteomic screen

52. Putting it all together: network activity inference
Identifying causal relation between putative input and output signals
Tracking effects of molecular perturbation trough activation/inhibition cascades
Predicted input
Scoring intermediary nodes
Experimental data
Experimental data: terminate cascade
Predicted target
Experimental data: start cascade
Inferred activity

53. Work in progress
Finding Patterns of significance (based on one experiment):
Significant neighborhoods
Significant receptors (by underlying cascade)
Significant transcription factors (by upstream cascade)
Significant interaction types (by distribution of expression at terminals)
Finding common and different pathway modules (based on multiple samples:
Looking for “differential pathways” - modules that distinguish one group of samples from another
Finding common motifs in a group of pathway modules
Inferring patterns of network activity
Identifying causal relation between putative input and output signals
Tracking effects of molecular perturbation trough activation/inhibition cascades
Looking into mutual gene-process information and Bayesian inference of significance
If gene G occurs only in process P its up-/down-regulation is a significant evidence with respect to inferring P’s status
If gene G occurs in many other processes in addition to P its up-/down-regulation is not a significant evidence with respect of inferring P’s status

54. Future products

55. MetaMiner Consortiums for 2007
Oncology (breast cancer, 4 other cancers)
Metabolic diseases (diabetes II, obesity, metabolic syndrome)
CNS and neurodegenerative diseases
Immunological and autoimmune diseases

56. MetaMiner consortiums: Analytical platform for disease areas
Cancer relevant annotations, datatabases,
Active cpds analysis creening
Cancer consortium labs
HTS, HCS
MetaMiner (Oncology) platform
Compounds scoring:
Indications
- Toxicities
- Off-site effects
Drug targets:
Divergence hubs on networks;
“Druggability” testing
Pathways connectivity
Biomarkers:
Combination of different types
- Expression
- Secreted proteins
- Metabolites
Convergence hubs (core effectors)
Data parsing, normalization
Experimental data depository
Maps for disease, processes, drug action
Custom maps for projects
Data analysis
This is the general schema for network analysis. High throughput data can be mapped directly on Metacore interaction tables and unique for the data file network will be built. Note that the network provides with highest possible resolution mapping of the data: at the level of individual proteins/genes and single-step physical protein-protein interactions and one step metabolic transformations.
The networks can be also built using third party tools and compared withing Metacore.
Upon generation, networks can be interpreted in the context of canonical pathways (in this case, Genego maps) and different process ontologies (for example, GO categories) and prioritized based on p-value and data saturation. Based on this analysis, new hypotheses can be generated for drug targets, siRNA perturbations and biomarkers

57. MetaTox consortium. Functional descriptors
Mapping on descriptors
Enrichment by category
Pathways maps
Toxicity, process maps
Sub-networks, modules, nodes
Predictive models
Indexing & scoring by tox. category