DataDriven Solutions for Clinical Prediction and Functional Discovery

1

• Data-Driven Solutions for Clinical Prediction and
  Functional Discovery
• CHI, Molecular Medicine Tri-Conference
• Emerging Company Profile
• April 19, 2005
• Roland Somogyi, Ph.D.
• Larry D. Greller, Ph.D.
• Biosystemix, Ltd.
• rsomogyi_at_biosystemix.com
• ldgreller_at_biosystemix.com
• www.biosystemix.com
• (613)-376-3126

2
Personalized medicine The future of therapeutic
discovery, practice and business

• Diseases are complex
• Genes and pathways lead to the same symptoms in
  different ways in different individuals
• We must target these specific causes, not the
  symptoms
• Some drugs are only effective in specific
  individuals
• Drug targets can be specific for genetic variants
  of disease
• Individual pathway activity fingerprints may
  determine efficacy
• Some drugs cause adverse effects in a very small
  subpopulation
• Toxicity due to genetic variants of drug
  metabolism
• Physiological and pathway background patterns may
  lead to unanticipated side effects

3
Biosystemix value to customersSuccesful
personalized medicine programs ultimately depend
on understanding the data and deriving meaningful
predictions
You must pass through here
Discoveries and models
Biomedical data
Integrative data mining and predictive modeling
Personalized medicine predictors
Experimental platforms
Biosystemix solutions
Therapeutic markers targets
Genomics proteomics
Signaling pathways networks
Clinical data
4
Biosystemix focuses on the opportunity for
therapeutic solutions, services and products

• The predictive models which integrate the
  knowledge of markers, patterns and pathways
  associated with disease and therapeutic outcome,
  will become vital
• to personalized therapeutic practice,
• target and drug discovery, validation and
  approval, and
• an economic engine for the biomedical industry.
• Biosystemix provides key technologies and
  experience in
• extracting complex patterns of key markers from
  genomic and clinical data,
• integrating predictive molecular profiles,
  functional knowledge and clinical outcomes into
  comprehensive predictive models, and
• generating personalized medicine marker, target
  and model IP in a large variety of disease and
  biomedical application areas.

5
An advance in personalized medicine /predictive
medicine
6
A personalized medicine scientific case study
Predicting clinical drug response in MS
(multiple sclerosis)
Clinical RNA expression profiling data
Personalized medicine outcome
Computational modeling
Gene A expression
Good responder to interferon b
Nonlinear combinatorial predictive models
Gene B expression
Poor responder to interferon b
Gene C expression
7
Predicting drug response before IFNb treatment in
MSTwo genes work better than one
15 samples are misclassified by FLIP alone
1d IBIS models
10 samples are misclassified by Caspase 10 alone
2d IBIS models
Good response predictive region
Poor response predictive region
Only 5 samples are misclassified by FLIP and
Caspase 10 together
Blue poor responseRed good response
8
Predicting drug response before IFNb treatment in
MSThree genes work better than two
3d IBIS model

• The yellow, orange and blue arrows point to
  samples that are incorrectly classified in the 2d
  models and correctly classified in the 3d models
• Note 3d models pass stringent statistical
  cross-validation criteria
• A B Views of 3d model predicting good and
  poor drug responders from the expression of 3
  genes
• B, C D All 3 possible 2d predictive models
  involving the same genes

2d IBIS models
Blue poor responseRed good response
9
Reference
S. Baranzini1, P. Mousavi2, J. Rio3, S.
Caillier1, A. Stillman1, P. Villoslada4, M.
Wyatt1, M. Comabella3, L. Greller5, R. Somogyi5,
X. Montalban3, J. Oksenberg1 Classification and
prediction of response to IFNß using gene
expression profiling the supervised computational
methods. (2004) PLoS Biol 3(1) e2
1Department of Neurology, School of Medicine,
University of California at San Francisco 2School
of Computing, Queens University, Kingston,
Ontario, Canada. 3Department of Neuroimmunology,
Hospital Vall dHebron, Barcelona,
Spain 4Department of Neurology, Clinica
Universitaria de Navarra, University of Navarra,
Spain, 5Biosystemix Ltd., Sydenham, Ontario,
Canada
10
What have we found?

• Combinatorial 3d models predicting IFNb response
  outcome in MS achieve high accuracy and
  statistical validation scores.
• These predictive models provide valuable
  diagnostic/prognostic answers in complex diseases
  for which no markers exist
• Next step is in-depth clinical validation
• Single genes and pairs do not achieve high
  predictive accuracy
• Finding the nonlinear and combinatorial patterns
  at the root of these models requires advanced
  data mining
• Conventional statistics not effective here

11
Gene function and pathway discovery through gene
network reverse engineering
12
Predicting the molecular mechanisms underlying
differential drug response Data-driven,
computational reverse engineering reconstructs
signaling pathways directly from clinical MS
gene expression data

• Red lines Gene interactions in good responders

Literature quote interferon-inducible stat2
stat1 heterodimer preferentially binds in vitro
to a consensus element found in the promoters of
a subset of interferon-stimulated genes

• Green lines Gene interactions in poor responders

Jak2 phosphorylates only Stat1 resulting in Stat1
homodimer formation and GAS (cis element)
activation of Interferon gamma induced genes
IFN gamma receptor heterodimers activate Jak2
SOS1 and Grb2 complex activates RAS/MAPK pathway
leading to FOS activation
13
What made it possible?

• Setting the stage with thorough experimental
  design
• Careful clinical study design and patient
  recruitment
• Sufficient number of high quality, clinical blood
  and RNA sample
• A solid foundation of precisions measurements
• Quantitiave, gene expression RT-PCR assays
• Reverse transcription polymerase chain reaction
• Combines stringent hybridization with
  amplification
• Only the best assays should be used for clinical
  applications
• Providing the edge with advanced computational
  analysis
• Nonlinear and combinatorial methods for pattern
  recognition
• Higher-dimensional predictive modeling and
  statistical validation
• In the words of by Kaminski and Achiron,
  highlighting the Baranzini study in PLoS Med
  2(2) e33.
• However, the importance of Baranzini and
  colleagues study lies not in its mechanistic
  insights, but in its clinical relevance. The
  careful design of the experiment, the use of
  reproducible real-time PCR instead of
  microarrays, the meticulous analysis, and the
  previous observations support the notion that
  PBMCs express clinically relevant gene expression
  signatures in MS and probably in other
  organ-confined diseases.

14
Data-driven predictive models provide
opportunities for better medical practice

• Step 1 Diagnosis of the disease
• Specific form of a disease is not apparent in
  superficial symptoms
• Higher-dimensional diagnostic models based on
  in-depth patient profiling
• Molecular and physiological fingerprints
  distinguish forms of a disease.
• Step 2 Prognosis of the outcome
• Complex prognostic models based on in-depth
  profiling data can enable reliable choices for
  timing of therapeutic interference
• Step 3 Therapeutic choice
• Therapeutic decision models based on detailed
  patient state information will significantly
  increase the probability of successsful
  treatment
• Step 4 Therapeutic discovery
• Data from personalized medicine studies will be
  used in the data-driven discovery of new disease
  mechanisms and pathways for individually-targeted
  intervention.

15
Biosystemix currently provides its expertise and
services to partners in predictive medicine and
genomics

• Immunogenomics
• S2K, Genome Canada / Genome Quebec-funded
  multi-center consortium
• Infectious diseases
• HIV
• SARS
• HTLV
• Transplant rejection
• Immune Tolerance Network, NIH/NIAID-funded
  multi-center consortium
• Autoimmune diseases
• Allergy
• Diabetes
• UCSF, Department of Neurology
• Predicting drug response in multiple sclerosis
• Cancer
• Queens University, Ontario Cancer Institute
• Predicting good and poor outcomes in Follicular
  Lymphoma
• Toxicogenomics
• University of Michigan
• Inference of pathways involved in toxicity from
  gene expression data

16
Biosystemix sees growing opportunities in
personalized medicine

• Growing market for diagnostic and prognostic
  products
• Marker sets, assay kits and hardware for more
  effective diagnostic/prognostic profiling
• Information products
• Computational models linking complex
  diagnostic/prognostic patterns to outcomes
• Web-based, personalized medicine tools for use by
  physicians and patients
• Product linkage
• A drug may only be effectively applied if linked
  to a prognostic test
• Patent and regulatory approval for product sets
  that are only effective in combination
• May be required in the future by regulatory
  agencies for specifically-targeted drugs
• Opportunity for extracting value from generic
  drugs
• Novel combinations of generic drugs to match
  individual patient need
• Combinations and predictive models generating
  these combinations constitute valuable IP
• Creating new markets
• Providing new tools and therapies where they are
  currently non-existent or unreliable

17
Acknowledgements

• Larry D. Greller, Ph.D.
• Biosystemix CSO, Co-Founding Director
• Parvin Mousavi, Ph.D.
• Assist. Prof. Queens University School of
  Computing
• Sergio Baranzini, Ph.D.
• Assist. Prof. Neurology University of California
  San Francisco

18
Linking genes and pathways to predict
therapeutic outcome in a complex disease
Poor response predictive region
Good response predictive region
Good response predictive region
FLIP
Caspase 2
Caspase 10
Good response predictive region
19
Supplementary Slides
20
A collaborative, predictive medicine study in MS

• Investigational Groups
• Sergio Baranzini, Jorge Oksenberg UCSF
• Xavier Montalban Hospital Vall dHebron
  (Barcelona, Spain)
• Parvin Mousavi, Larry Greller, Roland Somogyi
  Biosystemix, Ltd.
• Multiple Sclerosis
• Autoimmune, neuroinflammatory CNS disease
• Primary therapy interferon-beta (IFNb) treatment
  
• Study Design
• RNA isolated from peripheral blood mononuclear
  cells after IFNb treatment at 6 time points (0,
  3, 6, 9, 12, 18 and 24 months)
• 70 genes measured by kRTPCR
• 52 patients
• 33 good responders
• 19 bad responders

21
Scientific challenges in personalized medicine

• High-quality molecular and physiological
  profiling
• Study design to capture key components of medical
  outcomes
• Study design to assist better post-hoc discovery
  of outcome-predictive profiles
• Adequate samples for statistical support
• Data management and integration
• Making different assay types commensurable
• Standards for data integration
• Data-driven computational discovery and modeling
• Complex outcome-predictive patterns
• Predictive models for clinical decision support
• Mechanistic discovery for novel intervention
  strategies

22
The need for data-driven models for tuning
therapies to individual need
Diagnostic and prognostic profiling
Personalized medicine therapy
Computational modeling
Gene expression A
Therapeutic compound X
Complex predictive models
Compound cocktail Y
Protein abundance B
Clinical assay intensity C
Drug dose Z
23
Effective inference and modeling for personalized
medicine must deal with biological complexity

• Interaction networks
• Nonlinearity
• Combinatorics

Curse of dimensionality
N 10,000
N 1000
k4
Log10 (C(N,k))
N 100
N 10
k
e.g. 400 million million combinations from 10,000
genes
24
Personalized medicine The ultimate application
of systems biology
Biomedical validation
Systems Biology
Target marker discovery
RNA, protein, metabolite profiling
Computational analysis and Bioinformatics
Data mining
Predictive modeling
Genetic variation characterization
Laboratory validation
Clinical assaydata
Clinical testing
Drugs, diagnostics predictive models
Personalized medicine
25
Recipes for success

• More than a vision
• It will be difficult
• Personalized medicine and integrative biology is
  technologically challenging
• but its tractable.
• Many technological components are there they
  now need to work together
• The devil is in the details
• Thorough and integrative scientific study design
• High quality assay technology and execution
• Advanced computational data mining and predictive
  modeling
• It all depends on people and technologies working
  together
• Integration of biomedical, physical, and
  math/statistical/computational sciences
• Acceptance of new technologies by regulatory
  bodies and medical practioners
• Support of RD and commercialization by
  businesses community