HPC, Machine Learning, and Big Data
Download as PPTX, PDF1 like207 views
Warren Kibbe discusses challenges and opportunities in cancer research and precision medicine. Cancer is a grand challenge that requires deep biological understanding, advances in scientific methods and technology, and leveraging large amounts of detailed data. While genomics sequencing has become cheaper, analyzing and making sense of the vast amounts of multi-omic, clinical and other data generated poses new challenges. Emerging technologies like cryo-EM and single cell techniques are providing new insights. Collaborative team science bringing together experimentalists and computational/data scientists is critical to make progress on problems like understanding RAS activation. Precision medicine aims to understand health and disease at the population, individual and clinical levels by leveraging diverse clinical, molecular and other data, but connecting
HPC, Machine Learning, and Big Data
- 1. Cancer Research: Computing and Data Warren A. Kibbe, Ph.D. Professor, Biostats & Bioinformatics Chief Data Officer, Duke Cancer Institute warren.kibbe@duke.edu @wakibbe #HECBDML
- 2. Take homes • Cancer is a grand challenge • Data generation is no longer the bottleneck in biomedical research – data management, analysis, reasoning are • Highlight two technologies enabling a much more dynamic view of biology • Two vignettes highlighting computational and big data challenges in biomedical research
- 3. Data access is pervasive
- 4. (10,000+ patient tumors and increasing) Courtesy of P. Kuhn (USC) 2006-2015: A Decade of Illuminating the Underlying Causes of Primary Untreated Tumors Omics Characterization Cancer is a grand challenge • Deep biological understanding • Advances in scientific methods • Advances in instrumentation • Advances in technology • Data and computation • Mathematical models Cancer Research and Care generate detailed data that is critical to create a learning health system for cancer Requires:
- 5. Genomics – the poster child
- 6. Genomics – the poster child Cost per genome Consumer genetic testing
- 7. Biology is so much more than DNA https://xkcd.com/1605/
- 9. Biological Scales Molecular to Systems Biology DNA Large molecule Sequence encodes identity of protein Humans: 23 pairs of DNA molecules over 4 billion nucleotides per pair Protein 109 10-12 -10-4 Complexes 30,000 proteins with variants proteins, cofactors, metabolites, 2nd msgrs 10-6 -103 Organelles, cells 10-3 -104 Tissues 1 -108 Organs 1 -109 10-9- 10-4 10-9 -10-8 10-8 -10-7 10-7 -10-5 10-6 -10-2 10-3 -1 Size scale (meters) Time scale (seconds) compartments structures function signaling, networks emergent properties © 2004 WAKibbe, Northwestern University
- 10. Cryo-EM • Able to get atomic resolution of flexible molecules, like membrane- bound proteins
- 11. Single cell techniques • Sequencing • Proteomics • Metabolomics • Microenvironment https://arxiv.org/abs/1704.01379 Growing ability to focus on dynamics!
- 12. Example Basic Science Problem
- 13. NCI RAS Initiative + NCI-DOE Joint Initiative https://science.energy.gov/~/media/ascr/ascac/pdf/meetings/201809/ASCR2018_streitz_nomovies.pdf Fred Streitz
- 14. Multi-modal experimental data, image reconstruction, analytics Adaptive spatial resolution Adaptive time stepping High-fidelity subgrid modeling Experiments on nanodisc CryoEM imaging X-ray/neutron scattering Protein structure databases Adaptive sampling molecular dynamics simulation codes Unsupervised deep feature learning Uncertainty quantification Mechanistic network models RAS activation experiments (FNLCR) Phase field model Coarse- grain MD Classical MD Machine learning guided dynamic validation Granular RAS membrane interaction simulations Atomic resolution RAS-RAF interaction RAS Activation Predictive simulation and analysis of RAS Phase Field model of lipid membrane Cancer Moonshot Pilot 2
- 15. Lipid content: RAS/HVR binding by SPR, alpha assays in nanodiscs, liposomes, imaging in GVUs, lipidomics, SANS (possibly with contrast variation) RAS/HVR mobility & dynamics: single particle tracking, FCS, CG simulations of farnesylated HVR and RAS on nanodiscs and membranes, use to constrain phase field coupling RAF-membrane affinity: SPR in liposomes, biophysical measurements, MD simulations to identify regions of interest that interact with membrane RAS/HVR-membrane binding: SPR in liposomes, biophysical measurements, SANS (with contrast variation), AA and free-energy calculations of RAS/HVR binding to constrain CG parameters, free energies to inform phase field HVR structure/dynamics: crystallization, CD, MD of HVR in multi-component lipid platform to inform mobility in phase field model RAS activity & structure: GTPase, GTP off-rate, crystallization, NMR, cryo-EM?, SANS, AA MD simulations constrain CG parameters RAS-RBD structure: crystallization, NMR, AA simulations to constrain CG parameters RBD-CRD and CRAF structure: crystallization, NMR, cryo-EM, CG simulations validated against AA simulations RAS-RBD binding: SPR, ITC, alpha assays in nanodiscs, TIRF, SANS (possibly with contrast variation), compare with AA simulations and constrain CG simualtions RAF activation: dimerization, phosphorylation state(s), long time-scale CG simulations and kinetic estimation, multi-scale simulations multi-scale simulations of RAS/RAF dynamics on membrane RAS/HVR multimeric state: BRET, step photobleaching, PALM, AA and CG MD of KRAS/HV R on nanodisc and multi-component lipid platform Experimental data to inform modelSimulations to build model Farnesyl dynamics: solid state NMR, AA and CG simulations of farnesyl in membranes and lipid bilayers informs phase field model Lipid domains: Confocal microscopy RAS/HVR localization in GUVs, Calibrate coarse-grained (CG) simulations with all- atom (AA) Simulations, Calculate free energies of domains Close collaboration of experimentalists and theorists to build predictive model
- 16. Team Science is critical Clinical Trials Biostatists Bioinformatics Clinical Care Clinical Research EHRs, Imaging, Lab Systems Data Science, ML, BD, HPC Analytics and Visualization Open Data enhances collaboration and team science!
- 17. NCI RAS Initiative + NCI-DOE Joint Initiative https://science.energy.gov/~/media/ascr/ascac/pdf/meetings/201809/ASCR2018_streitz_nomovies.pdf Fred Streitz
- 18. NCI RAS Initiative + NCI-DOE Joint Initiative https://science.energy.gov/~/media/ascr/ascac/pdf/meetings/201809/ASCR2018_streitz_nomovies.pdf Fred Streitz
- 19. NCI RAS Initiative + NCI-DOE Joint Initiative https://science.energy.gov/~/media/ascr/ascac/pdf/meetings/201809/ASCR2018_streitz_nomovies.pdf Fred Streitz
- 20. Example from Precision Medicine
- 21. Population vs Individual vs Clinic
- 22. Health vs Disease • What is ’normal’? • Systematic and measurement error • Biological heterogeneity • Population Health
- 24. Clinical, Lab, Molecular data http://blog.75health.com/what-components-constitute-an-electronic-health-record/
- 25. Access to data has changed-Epic
- 26. From Apple Health App
- 27. Healthcare • Evidence is not consistently accessible and structured • Outcomes are not connected to care • Patient trajectories are not calculated or accessible
- 28. Healthcare • More data is ‘digital first’ every day • Decision aids are needed • Good UX and responsive computing and analytics are critical for improving health outcomes
- 29. Understanding Cancer • Precision medicine will lead to fundamental understanding of the complex interplay between genetics, epigenetics, nutrition, environment and clinical presentation and direct effective, evidence-based prevention and treatment. Ramifications across many aspects of health care
Editor's Notes
- #15: Biochemical/biophysical properties of fully processed KRAS4b RAS on nanodiscs Lipid composition of RAS:membrane interaction RAS:RAF binding in the context of nanodisc membranes Structure of RAS on membranes Crystallography of KRAS, and KRAS:effector complexes NMR of KRAS bound to nanodiscs X-ray/neutron scattering of KRAS on nanodiscs Cryo-EM imaging of KRAS protein complexes (+effectors) Dynamics of RAS in membranes Supported bilayers in vitro Live cell imaging with single molecule tracking Adaptive spatial resolution (e.g., sub-grid modeling) Propagating both coarse-grained and classical (atomistic) MD information, we aim to maintain the highest fidelity possible at the point of interactions while capturing long distance effects Multiple time scales By judiciously switching between spatial scales we enable investigation of timescales that are orders of magnitude longer than possible with fine-scale simulation alone. Automated hypothesis generation and dynamic validation We will efficiently and accurately explore, e.g., possible interaction sequences by coupling Machine Learning techniques with large-scale predictive simulation. Extreme scale simulation Requried novel computational algorithms and techniques will be developed for use on Sierra-class architectures, and will be designed for exascale. Deep learning algorithms Powerful pattern recognition tools will accelerate our predictive simulation capability by giving rapidly identifying, e.g., the time or region where a sub-grid model is needed or by logically exploring an intractably large decision tree. Uncertainty quantification Application of our extensive capability will be tested in the new (highly uncertain) world of biology and healthcare, leading to new insights and the development of new methods Scalable statistical inference tools The continued convergence of data analytics and predictive simulation as we approach exascale will require statistical tools that scale far beyond what is current, requiring the development of new strategies.
- #23: We now have tools to let us both understand social determinants of health and build ‘early warning systems’ to identify people who are have acute medical issues