25
Build and Optimize a BioTech Portfolio: Target Selection and Disease Indications Pankaj Agarwal Computational Biologist BioInfi [email protected]

[email protected] Build and Optimize a BioTech Portfolio ... · Pankaj Agarwal Computational Biologist BioInfi [email protected]. Acknowledgments & Conflicts of Interest GSK Colleagues

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

  • Build and Optimize a BioTech Portfolio:Target Selection and Disease Indications

    Pankaj AgarwalComputational Biologist

    [email protected]

  • Acknowledgments & Conflicts of Interest

    GSK

    Colleagues @ GSK

    BioAge

    Bill & Melinda Gates Foundation

  • Observations

    1. Most (expensive) failures are due to lack of efficacy i.e. poor target selection

    2. Causes of clinical trial result of p>0.05 a. Lack of pharmacology

    b. Poor target selectionc. Poor selection of patients/indication/sub-population

  • Outline: Data-Driven Methods for1. Build Portfolio

    a. Target Identificationb. Assessing Data Sources for Target Identification using ML and DL

    2. Systematic Drug Repurposing (SyDR) a. Experience b. Methods: Genetics and Cmapc. Example: Optimize Portfolio

  • Strategy

    Starting Point

    Target Id: Practical View

    ValidationPlan

    ● Disease● Technology/Modality● Data Source

    ○ Mendelian○ GWAS○ CRISPR○ Paper

    ● Decision Maker? Budget? ● Criteria? ● Robust?● Competitive Advantage?

  • Assessments of Target Data Sources Using SoTA Methods:Deep Learning, Network Diffusion &Tensor Factorization

  • Target outcomes from Pharmaprojects and Features from Harmonizome

    https://pharmaintelligence.informa.com/ http://amp.pharm.mssm.edu/Harmonizome/

    Phase IIIOutcome

    Targ

    ets

    ParsingFilteringMappingSummarization

    Omic Features

    Systematic interrogation of diverse Omic data reveals interpretable, robust and generalizable transcriptomic features of clinically successful therapeutic targets. Rouillard AD, Hurle MR, Agarwal P. PLoS Comput Biol. (2018)

  • Systematic interrogation of diverse Omic data reveals interpretable, robust and generalizable transcriptomic features of clinically successful therapeutic targets. Rouillard AD, Hurle MR, Agarwal P. PLoS Comput Biol. (2018)

    Standard ML Classifier for Clinical Targets

    – Logistic Regression– Random Forest

  • SDAE discovers hidden target features from transcriptomicHidden variables explain important patterns of expression across normal human tissues

    input corruptedinputhidden, explanatory

    variablesreconstructed

    input

    EncoderNetwork

    DecoderNetwork

    mRNA expression in normal human tissues from theGenotype-Tissue Expression (GTEx) Project

    ~20,000 Genes

    30 T

    issu

    es

  • SDAEs may discover features of clinically successful targets

    + phase III successx phase III failure

  • Non-specific

    TissueSpecific

    Conventional Supervised Feature Selection

    PP

    V

    TPR

    ----- model: 0.810±0.032 AUPRC- - - random: 0.78 AUPRC

    GTEx SDAE outperformed conventional ML approach using 1000s of features

    brain,pituitary

    blood,spleennon-specific

    liver, kidney,small intestine,pancreas,bladder, skin

    muscle,adipose

    Deep Unsupervised Feature Learning

    PP

    V

    TPR

    ----- model: 0.833±0.018 AUPRC- - - random: 0.78 AUPRC

    SDAEs may discover features of clinically successful targets

  • Using DTINetNetwork Diffusion

    12

    DTInet from A network integration approach for drug-target interaction prediction and computational drug repositioning from heterogeneous information, Luo et al Nature Communication (2017)

    1. PPI:InWeb: 625K2. MGI3. HPN

    – 25K

    HSDN: 148K

    Integrating Biological Networks for Drug Target Prediction and Prioritization.Ji X, Freudenberg JM, Agarwal P.Methods Mol Biol. 2019;1903:203-218.

    https://www.ncbi.nlm.nih.gov/pubmed/?term=Ji%20X%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/?term=Freudenberg%20JM%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/?term=Agarwal%20P%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/30547444

  • Proper negative examples from rewiring bipartite graph reduce accuracyRepurposing is harder than target identification?

    13

    Integrating Biological Networks for Drug Target Prediction and Prioritization.Ji X, Freudenberg JM, Agarwal P.Methods Mol Biol. 2019;1903:203-218.

    Also see: Predicting clinically promising therapeutic hypotheses using tensor factorization. Jin Yao, Mark R. Hurle, Matthew R. Nelson and Pankaj Agarwal. BMC Bioinformatics (2019) 20:69.

    https://www.ncbi.nlm.nih.gov/pubmed/?term=Ji%20X%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/?term=Freudenberg%20JM%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/?term=Agarwal%20P%5bAuthor%5d&cauthor=true&cauthor_uid=30547444https://www.ncbi.nlm.nih.gov/pubmed/30547444https://doi.org/10.1186/s12859-019-2664-1https://doi.org/10.1186/s12859-019-2664-1https://doi.org/10.1186/s12859-019-2664-1

  • 10 years go by and you have a thriving clinical portfolio

    Can you expand the indications and help more patients?

  • Systematic Drug Repurposing Mandate: Systematically Identify New Indications for Active and Terminated Assets

    Primarily evaluated internal compounds, but used others as comparators

    Considerations

    1. What is the Goal? End Point: PoC, Approval, Reimbursed, Accolades?2. Decision-making and budget3. Alignment of disease/phenotype with company strategy4. Quality, IP, targets of assets 5. Validation strategy and critical path6. Evidence connecting asset/target to disease: Key computational focus

  • Prescriptions(Adverse Events, EHRs,

    Social Media)

    Inte

    rnal

    Dat

    abas

    es

    Targ

    et

    Iden

    tific

    atio

    n

    Genetic Targets

    Animal Models(Safety, Efficacy, Markers)Phenotypic Assays (Screens, Expression)

    Clinical Trials(Safety, Efficacy, Markers)

    Scie

    ntifi

    c Li

    tera

    ture

    Hypotheses Sources Hurle, .. Agarwal (2013) Clin Pharm Therap 93 (4), 335-341Fast

    Follower

    Novelty

    Launch

    Clinical

    Pre-Clinical

    http://www.nature.com/clpt/journal/v93/n4/full/clpt20131a.htmlhttp://www.nature.com/clpt/journal/v93/n4/full/clpt20131a.html

  • Repurposing from Genetics (GWAS)

    Genome Wide Association Studies

    (GWAS)

    Industry pipeline155 GWAS genes targeted by drugs/biologics actively pursued by

    the industry (2.5X expected)

    GWAS trait similar to drug

    indication?

    Increased confidence in the indication (63 targets)

    (ex., PPARG)

    Potential drug repositioning opportunities (92 targets)

    No

    Yes

    Functional Annotations

    ENCODE (Nature, Sep 2012)

    FANTOM5 (Nature, Mar, 2014)

    New Functional Annotations

    Jhamb D, Magid-Slav M, Hurle MR, Agarwal P. Pathway analysis of GWAS loci identifies novel drug targets and repurposing opportunities. Drug Discov Today. 2019.

    Sanseau P, Agarwal P, ... Cardon L, Mooser V. Nature Biotechnology 33,317-20 (2012)

    https://www.ncbi.nlm.nih.gov/pubmed/?term=Jhamb%20D%5BAuthor%5D&cauthor=true&cauthor_uid=30935985https://www.ncbi.nlm.nih.gov/pubmed/?term=Magid-Slav%20M%5BAuthor%5D&cauthor=true&cauthor_uid=30935985https://www.ncbi.nlm.nih.gov/pubmed/?term=Hurle%20MR%5BAuthor%5D&cauthor=true&cauthor_uid=30935985https://www.ncbi.nlm.nih.gov/pubmed/?term=Agarwal%20P%5BAuthor%5D&cauthor=true&cauthor_uid=30935985

  • External and internal expression dataConnectivity Map: Drug-Disease mapping

    ● Numerous tools on Lincsproject.org

    ● Lamb, …, Golub: Science. 2006 Sep 29;313(5795):1929-35. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease.

    ● Profiled entire preclinical and clinical pipeline (400+ compounds) in 6+ cellular systems

    Systematic evaluation of connectivity map for disease indications. Jie Cheng, Lun Yang, Vinod Kumar & Pankaj Agarwal. Genome Medicine 6, 95 (2014)

    http://www.genometry.com/open-innovation/screening.png?attredirects=0

  • Discovered using Connectivity MapsEH: EPHX2 repurposing for IBD

    Reisdorf, Xie,.., Agarwal (2019). Preclinical evaluation of EPHX2 inhibition as a novel treatment for inflammatory bowel disease. PLoS One.

  • EPHX2 experimental data

    Reisdorf, Xie,.., Agarwal (2019). Preclinical evaluation of EPHX2 inhibition as a novel treatment for inflammatory bowel disease. PLoS One.

    Mouse DSS-induced colitis model Ex-vivo human samples

  • SyDR Repurposing Examples

    1. EPHX2 for IBD & wound healing (published)2. BET for NASH & Liver Fibrosis (published)3. Immune: RA, Psoriasis, Atopic Dermatitis, UC, Crohn’s, IBD4. Neuro: ALS, MS5. Respiratory: Asthma, COPD6. Metabolic: HF7. Cancer

  • Model for Data-Driven “Public” Drug Discovery

    Reisdorf, Chhugani, Sanseau, Agarwal: Expert Opinions in Drug Discovery, 2017 May..

  • Learnings1. Fantastic Opportunity: we were making medicines!2. People: Diverse Team & Decision-Making3. Data & Computation Pedigree

    a. Be computationally humble, open, agnostic4. Faster experimental-computational cycle5. Always test multiple hypotheses

    a. Avoid experimental results dumpster diving -- Independent statisticianb. Do not to be hard on yourself -- Internal comparators

  • Conclusions1. Build and Optimize Portfolios through target selection and indication

    expansion 2. AI/ML promising, but easy to make mistakes: limited # of training examples,

    unbalanced sets, and domain of applicability3. Happy to brainstorm and collaborate

    a. Open to scientific & business ideasb. Please find me here or [email protected]

    mailto:[email protected]

  • Conclusions1. Build and Optimize Portfolios through target selection and indication

    expansion 2. AI/ML promising, but easy to make mistakes: limited # of training examples,

    unbalanced sets, and domain of applicability3. Happy to brainstorm and collaborate

    a. Open to scientific & business ideasb. Please find me here or [email protected]

    10,000+ disease ÷ 10 new treatments/year

    = 1,000 years

    mailto:[email protected]