Envisioning Computational Innovations for Cancer Challenges (ECICC) Community: MicroLab June 11, 2019

Welcome! ECICC Community

AGENDA Welcome and Introduction: Today’s Goals What is a MicroLab? How to Participate MicroLab Origins ECICC Scoping Meeting Overview Review 4 ECICC Cancer Challenge Areas Virtual Breakout Groups Select Your Breakout Next Steps Breakout Group Discussions

Today’s Goals Introduce the lean-in cancer challenge areas developed at the ECICC Scoping Meeting in March Expand and engage the community: Share your ideas! Deepen connections and inspire multi- disciplinary collaboration among cancer, data, and computational scientists to create transformative impact in cancer research that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

What is a MicroLab? (µLab): a 60-90 minute, highly interactive virtual event Purpose: Facilitate stimulating scientific discussions in smaller, more intimate virtual breakout groups  Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

How to participate in a MicroLab (µLab) Plenary session is recorded - available in the Hub later Questions via Chat Breakouts for deep dive into the topics Breakouts not recorded, but with scribes

Renaming yourself Hover over your name Click Rename Go to ‘Participants’ along the bottom. Hover over your name Click Rename  Rename yourself with one of the following: Synthetic – Your name ML – Your name Twin – Your name Adapt – Your name Hit Rename Your breakout room will open soon.

MicroLab Origins Emily Greenspan, PhD National Cancer Institute Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

Cancer Challenges and Advanced Computing MicroLab: Origins NCI-DOE Collaboration: Joint Design of Advanced Computing Solutions for Cancer (JDACS4C) Envisioning Computational Innovations for Cancer Challenges Scoping Meeting March 6-7, 2019, Livermore Valley Open Campus, Lawrence Livermore National Lab DOE-NCI partnership to advance exascale development through cancer research CANCER COMPUTING Vision for cancer challenge Vision for computing innovation Build a community Join the online community! https://nciphub.org/groups/cicc

Envisioning Computational Innovations for Cancer Challenges (ECICC): Scoping Meeting Meeting Goals Identify lean-in cancer challenge areas that push the limits of current cancer research computational practices and compel development of innovative computational technologies Build a community, multi-disciplinary engagement, and collaboration among cancer, data, and computational scientists to create transformative impact (e.g. CSBC, PSON, ITCR) Demonstrate how to break down silos and work across domains, disciplines, organizations Define the types of cultural shifts in cancer research that could be possible with high-performance computing (HPC) Meeting Outcomes Writing groups identified 9 challenge areas (go.hub.ki/ecicc2019) Meeting report summarizes these 9 challenges into four areas: Synthetic Data (PII and other data at scale) Hypothesis Generation Using Machine Learning Digital Twin Adaptive Treatments Identified culture shifts included incentivizing data sharing, collaboration in areas of less expertise, moving from descriptive to predictive analytics, optimizing experimental design with a data first approach

Team Leads: 4 Lean-In Cancer Challenge Areas Synthetic Data Nick Anderson, University of California, Davis Bill Richards, BWH/Harvard University Hypothesis Generation Using ML Jeremy Goecks, Oregon Health & Science University Amber Simpson, Memorial Sloan Kettering Cancer Center Digital Twin Tina Hernandez-Boussard, Stanford University Paul Macklin, Indiana University Tanveer Syeda-Mahmood, IBM Ilya Shmulevich, Institute for Systems Biology Adaptive Treatments John McPherson, University of California, Davis Rick Stevens, Argonne National Lab that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Presentation of Cancer Challenge Areas that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Cancer Challenge Area: Synthetic Data Nick Anderson, UC Davis Bill Richards, BWH/Harvard Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

Access to Large-scale Clinical Data Sets There is a scarcity of sharable high quality, high volume clinically-derived data sets available for research, education or validation purposes Existing data sets are either deidentified - and require specific attestations for use, or are provided through interfaces that enforce deidentification and prevent access to low-level data Making large-scale, information-rich deidentified clinical data sets available for sharing will continue to be limited by compliance and regulatory requirements, and require significant work Current methods of deidentification or access management face new challenges when applied to very large and uniquely integrative data sets, such as concerns such as whether such methods “over-sanitize” data or introduce bias when applied at scale Making broadly available large-scale data sets that include identifiable clinical data is currently challenging without changes to privacy law, or with development of new models of data access that supports oversight of secondary uses on sensitive data – and in health is not currently socially popular.

Synthetic Data: Overview Synthetic data sets: Statistically identically clinical derived data sets, considered not-relinkable to any patient, institution or community – determined to contain neither PHI or PII. The idea: Foster the development of high quality, statistically identical data sets and associated descriptive metadata and generation methods that can be broadly shared to support collaborative and reproducible analytic development and validation on large scale data Why: Expanding access to synthetic data sets generated in collaboration with a variety of clinical data domain expertise will democratize access to interdisciplinary experts, and accelerate options to access, test and share new methods and models. 1

Challenge Area: Synthetic Data Broad research questions: Methods for reproducibly generating new and unlinkable synthetic data sets, rules and appropriate descriptive metadata from “reference” clinical data (1:1) Methods for generating new and unlinkable synthetic datasets from clinical data “seeds” and permutation rules (1:many) Methods for comparing new algorithms tested solely on synthetic data sets on “reference” clinically identifiable data sets (and visa versa) (1:1) Methods for developing sets of uniquely distinguished and high-quality data sets for large “population simulation” studies (1:n-k) that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Synthetic Data Discussion Questions What are barriers to adoption and use of synthetic data sets? What justifies synthetic data use over clinically extracted or prospectively collected data sets? In what data areas is synthetic data generation most/least tractable? What serves as a gold standards for the content of synthetic data set – the ability to test against the “reference” data? Are there unexpected legal or other barriers to expecting “non-human subject” synthetic data sets to be shared? What is unique about clinically derived data that prevents other key interest groups (basic sciences, engineering, computing, law) from participating? What or whom might see funding value in these capabilities? What additional key components or challenges would you add? What is needed to move forward? that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Cancer Challenge Area: Hypothesis Generation Using Machine Learning Amber Simpson, Memorial Sloan Kettering Cancer Center Jeremy Goecks, Oregon Health & Science University Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

Hypothesis Generation Using Machine Learning: The Big Idea Patient Data Hypothesis Generation Using Machine Learning: The Big Idea ML has the potential to efficiently guide hypothesis generation and experimental design for cancer clinical trials. ML has the potential reveal new relationships in large, complex data sets to guide cancer research and novel targeted therapies. that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and + = Pre-clinical Data

Hypothesis Generation Using Machine Learning: Overview Patient Data Hypothesis Generation Using Machine Learning: Overview Multi-omic data are increasingly becoming available Data mining techniques could be used to find novel patterns in data to generate hypothesis for evaluation in clinical trials Transfer learning could be used to bridge the gap between pre-clinical and clinical data that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and + = Pre-clinical Data

Hypothesis Generation Using Machine Learning: Overview Machine learning is not really used clinically - how do we change this? Cancer data conundrum: researchers and clinicians are inundated with more information than they can handle, while, at the same time, there are sizeable gaps in the biological systems information that is required for research and clinical advance ML is key to addressing both hurdles: ML can provide guidance to experimentalists on what to study and the sequence in which to attack questions. ML is a critical bridge between large, complex data sets and mining actionable meaning from the data. ML has the unique ability to discover and use algorithms to cluster observations (data), and to do so iteratively with experimentation, in an active learning process. that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and How do we influence patient management???

Hypothesis Generation Using ML Discussion Questions What is your most pressing concern? (What keeps you up at night?) What do you know that we don’t? What are the cross-cutting themes you would like to explore in- depth (biological, clinical & computational)? What do we need to dive into greater depth? What existing funding sources are you aware of that can support research in this area? Anticipated funding? What additional key components or challenges would you add? Which specialty areas / collaborators should be involved? What introductions are needed? What is needed to move forward? that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Cancer Challenge Area: Digital Twin Tina Hernandez-Boussard, Stanford University Paul Macklin, Indiana University Tanveer Syeda-Mahmood, IBM Research Ilya Shmulevich, Institute for Systems Biology Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

Digital Twin: The big idea Treatment Exploration digital twin Digital Twin: The big idea Patient and oncologist discuss goals and preferences Clinicians build a "digital twin" Clinicians use HPC to simulate all treatment options on the virtual twin Patient and clinician explore risks, benefits, side effects They choose a plan and predict progress using their digital twin Patient Data Treatment Consultation that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

What would this enable? Model personalized benefit from different treatment modalities Incorporate genetic, environment, and social factors to predict individual trajectory Predict outcomes and side-effects throughout patients’ health trajectories Advance patient-valued care Virtual clinical trials Synthetic controls Expand generalizability Population Estimates that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and Not just looking at point in time but can predict where the patient will be with each treatment modality

What are the barriers? Analytics: HPC-driven healthcare informatics: Multiscale analysis and coupling of data Integration of dynamical models with AI HPC-driven healthcare informatics: Simulation and virtual evaluation of care pathways Assess novel therapies simulated data to guide clinical assertions, inform clinical guidelines and develop health policies Data Sandboxes and platforms: Assembly of coordinated data across scales per patient Development of data commons to support the analysis Multi-scale analysis of cancer that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

What help is needed? Analytics, simulation and modeling: Multimodal analytics AI research in causal reasoning and pathways Multiscale data and model standards Uncertainty quantification Data Commons: Unified data and model shares Data visualization Patient data ingestions Validation efforts Scale out and test Regulatory: Licensing, regulatory, and funding landscape Open source and open data where possible! Integration with socioeconomics that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Digital Twin Discussion Questions Do you think this vision is realistic and feasible in our time frames? Would you like add any additional perspective on this vision? What do you see as the main challenges in achieving this vision? What existing funding sources are there to support research in this area? Which specialty areas/collaborators should be involved? What do you see the role of academia versus industry in addressing this vision? In what ways can you help expand/realize this vision? What are the cross-cutting themes you would like to explore in- depth (biological, clinical & computational)? What do we need to dive into greater depth in the breakout? that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Cancer Challenge Area: Adaptive Treatments John McPherson, UC Davis Rick Stevens, Argonne National Lab Unlike webinars which are focused on disseminating information, the purpose of µLabs is to

Adaptive Treatments: The Big Idea Precision oncology involves analyzing a patient tumor and tailoring therapy to best treat this individual tumor. A significant portion of tumors do not respond as predicted. A large portion of patients experience relapse. This process is repeated with serial biopsies and analyses each time a therapy fails. Can we better predict the response of a tumor to a particular therapy and most importantly chart the expected course that the tumor will take to evade this therapy? Treatment then become proactive and less reactive. Can we generate adaptive therapies knowing the expected course that will change during the course of treatment to thwart tumor evolution? that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Adaptive Treatments: What would this enable? Predictive modeling of treatment response short-term and long-term. More precise monitoring for adverse outcomes. Proactive early intervention to attack at tumor before it gains a second foothold. Therapies that can change their mode of attack on the tumor(s) as they adapt to continue to escape treatment and proliferate. that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Adaptive Treatments: Challenge We have generated a large knowledgebase of tumor characteristics with respect to mutational landscape but: Each tumor and individual is unique and must be considered holistically incorporating as much data as possible – We must get past single gene targeting mind set. Longitudinal tumor profiling data sets are lagging – Much data are produced clinically but are not accessible or integrated. More data are needed on changing tumor microenvironments during treatment and progression. that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Adaptive Treatments: Barriers Datasets: Robust longitudinal mutational profiling with treatment courses and outcomes are needed. Deep mutational analysis to identify tumor subclones. Deep analyses of changes in tumor and metastases microenvironments over time. Analytics: Scalable ML approaches to systems biology Dynamic modeling of all processes impacted by somatic mutation. Prediction of potential emerging tumor drivers. Modeling of changes in tumor microenvironments. Culture changes: Data sharing has often hit ethical and legal barriers. Patient advocacy and involvement is essential. that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Adaptive Treatments: Discussion Questions What is the most pressing need to achieve these goals? What data sets are out there and which need to be generated? What cross-cutting themes should we explore in-depth (biological & computational)? What resources do we need to dive into greater depth? (in vitro models, in vivo sensors for microenvironments) What technologies are best suited for adaptive therapies? (nanodevices, nanotherapeutics, bacteria, viruses) What existing funding sources are you aware of that can support research in this area? Anticipated funding? Needed funding? What additional key components or challenges should be added? Which specialty areas and collaborations are needed? Who should we try and engage to participate? What is needed to move forward? that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Select your Breakout Go to ‘Participants’ along the bottom. Hover over your name Click Rename  Rename yourself with one of the following: Synthetic – Your name Hypoth – Your name Twin – Your name Adapt – Your name Hit Rename Your breakout room will open soon.

Next Steps Join the Community Hub Site! Accept the invitation, add your profile and get to know the community: https://nciphub.org/groups/cicc Look for notes from all the Breakout Groups and add your comments Participate in the NEXT MicroLab: Watch for announcement coming soon! that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and

Breakout Group Discussions that push the limits of current cancer research computational practices and compel development of innovative computational technologies; and