9:00 a.m.–9:10 a.m.

Welcome — Seventh Computational Approaches for Cancer Workshop (CAFCW21)

Eric Stahlberg, PhD, Frederick National Laboratory for Cancer Research

9:10 a.m.–10:00 a.m.

Featured Speaker: Graham Johnson, Allen Institute

Simularium Viewer: An Online Tool for Democratizing the Analysis of Spatiotemporal Biological Models

Introduced by Eric Stahlberg, PhD, Frederick National Laboratory for Cancer Research

10:00 a.m.–10:30 a.m.

CAFCW21 Morning Break

10:30 a.m.–10:45 a.m.

An Integrated Simulation-HPC-Learning Approach to Create Cancer Patient Templates for Digital Twins

Presenter: Heber Lima da Rocha, PhD, Indiana University

Abstract: Building from recent COVID-19 modeling, we developed a multiscale agent-based model of melanoma micrometastases and immune response in lung tissue. The cellular components of the tumor microenvironment are formed by healthy epithelial cells from the lung, isolated melanoma cells, and immune cells: macrophages, dendritic cells, CD8+, and CD4+ T cells. The model also includes immune cell trafficking to and from the lymphatic system to drive an expanding immune response. Melanoma cells proliferate uncontrolled, causing mechanical stress in the region of the cell cluster. This mechanical factor leads to the death of healthy cells in the region, stimulating the activation of antigen-presenting cells (APCs). Apoptotic melanoma cell death can also activate APCs. Macrophages ingest dead cell debris and produce proinflammatory cytokines, recruiting more APCs to the region. Dendritic cells are activated when in contact with dying cells (lung and melanoma) and 'receive' their antigen signature. Activated dendritic cells migrate to the lymph node and recruit CD8+ and CD4+ T cells that can induce cancer cell death.

10:45 a.m.–11:15 a.m.

Oncolomics: Digital Twins & Digital Triplets in Cancer Care

Presenter: Asoke Talukder

Authors: Asoke Talukder, PhD, National Institute of Technology Karnataka, Surathkal; Roland Haas, PhD, QSO-Technologies

Abstract: To address the complex challenges in cancer care, we integrated two digital twins, namely, (A) Digital twin of oncologists' mind, and (B) Digital twin of the patients' physical state. The oncologist's twin is realized through the semantic integration of (1) NCI Thesaurus (NCIt), (2) Gene Ontology (GO), and (3) Disease-Gene Network (DisGeNET). These knowledge bases are mostly curated for researchers by expert researchers from hundreds of person-years of specialized knowledge. We earlier constructed the Diseasomics knowledge graph that combined Disease Ontology (DO), DisGeNET, SNOMED CT, PharmGKB, and disease Symptoms association network constructed through NLP, and other AI techniques from PubMed, Wikipedia, and curated knowledge for universal healthcare. In Oncolomics, we used the digital triplet to discover missing links (link prediction) or the unknown knowledge in NCIt and enhance the accuracy of the knowledge graph.

11:15 a.m.–11:45 p.m.

Topological Interpretation of Deep-Learning Models

Presenter: Adam Spannaus

Authors: Adam Spannaus, PhD, Oak Ridge National Laboratory; Shang Gao, PhD, Oak Ridge National Laboratory; Noah Schaefferkoetter, Oak Ridge National Laboratory; Lynne Penberthy, MD, National Cancer Institute; Jennifer Doherty, PhD, Utah Cancer Registry; Eric Durbin, DrPH, Kentucky Cancer Registry; Steven Schwartz, PhD, Seattle-Puget Sound Registry; Antoinette Stroup, PhD, New Jersey State Cancer Registry; Xiao-Cheng Wu, MD, Louisiana Tumor Registry; Charles Wiggins, PhD, New Mexico Tumor Registry; Georgia Tourassi, PhD, Oak Ridge National Laboratory

Abstract: Developing trust in the predictions made from an AI-based algorithm is a tantamount concern, especially in systems such as threat detection or medical diagnosis where outcomes may have tragic consequences. This work presents a topologically informed methodology for inferring prominent features in a deep-learning classification model trained on clinical text. Creating a graph of the model's prediction space, we cluster the inputs into the vertices of the graph and extract subgraphs demonstrating high-predictive accuracy for a given label. These nodes contain a wealth of information about features that the deep-learning model has recognized as important. We demonstrate fidelity of these inferred features to the original model by appending these terms to incorrectly classified documents and showing an increase in classification accuracy on these augmented documents. This work demonstrates that we may gain actionable insights for subsequent models by improving erroneous classifications and providing a path forward for continuous improvement.

11:45 a.m.–12:30 p.m.

Panel: Workforce Development and Inclusion

Moderator: Sally Ellingson, PhD, University of Kentucky

Panelists:

Belinda Akpa, PhD, Oak Ridge National Laboratory

Michelle Berny-Lang, PhD, National Cancer Institute Center for Strategic Scientific Initiatives

Caiden Lukan, Butler University

Nastaran Zahir, PhD, Cancer Training Branch Center for Cancer Training, National Cancer Institute

12:30 p.m.–2:00 p.m.

Lunch (on your own) and Virtual Poster Session

URL: https://cafcw21.virtualpostersession.org/

Posters

Using Keywords to Improve Classification Robustness on Cancer Pathology Reports, Andrew Blanchard, PhD, Oak Ridge National Laboratory

Non-invasive Classification of Isocitrate Dehydrogenase Mutation Status of Gliomas from Multi-Modal MRI Using a 3D Convolutional Neural Network: A Retrospective Multi-Institutional Analysis, Satrajit Chakrabarty, Washington University in St. Louis

Ensemble Learning for Deploying Robust TextCNN Models for Automatic Information Extraction from Cancer Pathology Reports, Kevin De Angeli, Oak Ridge National Laboratory

Identifying Sources of Mismatches in an Abstaining Classifier for Text Reports on Cancer Pathology, Sayera Dhaubhadel, Los Alamos National Laboratory

Applying Deep Learning Methods to Nuclei Detection in Whole Slide Images in LabCAS, Zachary Johnson, Alphan Altinok, Massachusetts Institute of Technology (MIT)

Voxel-Based Analysis to Investigate the Impact of Regional Dose Distribution on Radiation-Induced Lymphopenia After Radiotherapy for Lung and Liver Cancer Patients., Yejin Kim, Korea Advanced Institute of Science and Technology (KAIST)

Use of LogBB Data to Model and Predict Small Molecule Blood Brain Barrier Penetrance, Caiden Lukan, Accelerating Therapeutics for Opportunities in Medicine (ATOM) Consortium

Whole-Body Blood Flow Simulations to Investigate Lymphocyte Depletion Caused by Radiotherapy, Lucas McCullum, Harvard Medical School

OncoMiner: An updated pipeline for analyzing genomic sequencing data from changing technologies, Jonathon Mohl, PhD, University of Texas at El Paso

Medical Cybernetics for Telemonitoring and Telemedicine Supporting Primary Prevention and Treatment of Cancer, Zsolt Ori, MD, Ori Diagnostic Instruments

Cancer Compass: Biomarkers for Early Detection of Cancer, Asoke Talukder, PhD, SRIT India

Differentially Expressed Genes in NMIBC vs. MIBC Along with Their Targeted Networks and Pathways, Akhil Wali, National Institutes of Health (NIH), National Cancer Institute

Functional effect prediction and frequency-based analysis of genetic sequence variants in patients with cancer, Bofei Wang, The University of Texas at El Paso

Using Machine Learning Models to Predict Cancer Outcomes, Rachel Wofford, Oregon State University

Emerging gain-of-function mutations: their computational interpretation and multi-omic characterization, S. Stephen Yi, PhD, University of Texas at Austin

Dynamics-Adapted Radiotherapy Dose (DARD) for head and neck cancer radiotherapy dose personalization, Mohammad Zahid, PhD, Moffitt Cancer Center

2:00 p.m.–2:15 p.m.

From Regulatory Networks to Microenvironments: Multiscale Multicellular Cancer Modeling and Simulation with CompuCell3D

T.J. Sego, PhD, Indiana University

Abstract: Modeling the interplay of subcellular, intercellular and environmental factors is critical to producing meaningful, predictive simulations of cancer progression in vivo and testing biological hypotheses with virtual tissues. CompuCell3D is an open-source, cross-platform simulation environment that provides multiscale, multicellular biological modeling and simulation capabilities for a broad range of applications. New modeling, simulation and computing capabilities make CompuCell3D more relevant than ever to testing biological hypotheses in a multicellular context while considering processes that span scales from the subcellular to organismal scales. This talk presents new capabilities of CompuCell3D as relevant to modeling cancer onset and progression. The survey of new features includes capabilities like interactive Boolean network modeling, to computing leveraging high-performance computing environments. New features are demonstrated with examples related to proliferation, immune response and tissue recovery. An overview of an advanced application also demonstrates how CompuCell3D readily supports concurrent, collaborative development of complex modeling and simulation projects.

2:15 p.m.–2:45 p.m.

Probing Decision Boundaries in Cancer Data Using Noise Injection and Counterfactual Analysis

Presenter: Ashka Shah

Authors: Rajeev Jain, Argonne National Laboratory; Ashka Shah, University of Chicago; Jamaluddin Mohd-Yusof, PhD, Los Alamos National Laboratory; Justin Wozniak, PhD, Argonne National Laboratory; Thomas Brettin, Argonne National Laboratory; Fangfang Xia, PhD, Argonne National Laboratory; Rick Stevens, Argonne National Laboratory

Abstract: Advanced analyses and computations based on gene expressions are prone to errors as they depend on experimental design, chemical operations/measurements, and data analysis. The assembly and aggregation of such data for creating deep neural network models may further influence the accuracy of these analyses. For example, the CANDLE NT3 Benchmark uses a table of laboratory-obtained data mapping RNA expression data to a normal or tumor designation and is used to make predictions about given expression samples. In this work, we use the NT3 Benchmark to study the effects of injecting bad data at different rates to study the impacts on the resulting predictions. Our data manipulations include flipping classification labels (label noise) and introducing noise in gene expressions (feature noise). We present results for the performance of both the base NT3 Benchmark and NT3 with the addition of the abstention class in the presence of various types of injected noise. For higher noise levels, the ability of the base network to correctly predict the normal/tumor classification (as measured by the validation accuracy) degrades significantly. Use of the abstaining classifier allows the model to learn when the labels have become unreliable and abstain from providing a prediction in that case, while retaining accuracy.

2:45 p.m.–3:00 p.m.

Predicting Tumor Time to Recurrence from Free-Text Notes

Divya Nagaraj, Stanford University

Abstract: Electronic medical records contain a significant amount of unstructured patient information from free text, but crucial information can be difficult to find within lengthy notes. Thus, we develop an automated tool that can detect mentions of tumor recurrence and progression in clinical, radiology and pathology notes to infer time to recurrence and progression. This approach avoids the need for re-training models and is flexible enough to be applied to a variety of institutions and note types. We tested this approach with a cohort of pediatric and adult brain tumor patients with 27,137 clinical, radiology and pathology notes from Stanford University Hospital, creating cumulative patient trajectory graphs and fine-tuning models to predict time to recurrence. To assess initial accuracy, we compared the patient trajectories to their diagnosis from the ICD-10 codes associated with the notes and found high accuracy from the weak labeling pipeline.

3:00 p.m.–3:30 p.m.

Afternoon Break

3:30 p.m.–4:15 p.m.

Panel: Digital Twins for Cancer Care

Moderator: Emily Greenspan, PhD, National Cancer Institute

Panelists:

Matthew McCoy, PhD, Georgetown University

Paul Macklin, PhD, Indiana University

Leili Shahriyari, PhD, University of Massachusetts Amherst

Tanveer Syeda-Mahmood, PhD, IBM

4:15 p.m.–4:30 p.m.

Image-informed Mathematical Modeling to Predict Patient-Specific Treatment Response to Neoadjuvant Systemic Therapy in Triple Negative Breast Cancer

Chengyue Wu, PhD, Oden Institute for Computational Engineering and Sciences, University of Texas at Austin

Abstract: Patients with locally advanced, triple-negative breast cancer (TNBC) typically receive neoadjuvant therapy (NAT) to downstage the tumor and for improved surgical outcomes. A critical, unmet need is a method to accurately predict an individual patient’s response to NAT, thereby allowing for the opportunity to guide further interventions. In this work, we construct and apply a clinical-computational framework that integrates quantitative magnetic resonance imaging with physics-based, mathematical modeling to predict the response of TNBC early in the course of NAT. Preliminary results demonstrate the potential of the clinical-computational framework as a powerful tool for predicting response to NAT. Ongoing efforts include applying the approach to the whole patient cohort and performing the systematic model selection. Once validated, the approach could also assist in optimizing treatment plans on a patient-specific basis or guiding patient selection in trials for novel NAT regimens.

4:30 p.m.– 4:45 p.m.

Leveraging High-Performance Computing and Quantitative Imaging for Personalized Spatial-Temporal Forecasts of High-Grade Glioma Treatment Response

Presenter: David Hormuth

Authors: David Hormuth, PhD, University of Texas at Austin, Oden Institute; Maguy Farhat, MD, MD Anderson Cancer Center; Brandon Curl, MD Anderson Cancer Center; Thomas Yankeelov, PhD, University of Texas at Austin, Oden Institute; Caroline Chung, MD, MD Anderson Cancer Center

Abstract: Maximal safe resection followed by combination radiotherapy and chemotherapy is the standard treatment approach for patients with high-grade gliomas to target residual and infiltrative tumor. Response to therapy depends on the ability to target the tumor and on the treatment sensitivity influenced by factors including tumor physiology and phenotypic behavior. While adaptive radiotherapy is possible, identifying subregions of disease that are likely to progress during the course of therapy would allow for anticipatory adjustments in the radiotherapy treatment to target more aggressive tumor areas. Towards realizing the goal of timely, personalized treatment adaptions, we have developed a family of biologically based, mathematical models of tumor growth and response, which are initialized and calibrated using patient-specific multi-parametric magnetic resonance imaging (mpMRI) data. mpMRI enables non-invasive measurement of tumor morphology, vascularity, and cellularity. In this report, we leverage high-performance computing resources to calibrate a family of models in patients with high-grade gliomas.

4:45 p.m.–5:00 p.m.

GenomicSuperSignature: Interpretation of RNA-Seq Experiments Through Robust, Efficient Comparison to Public Databases

Presenter: Sean Davis

Authors: Sehyun Oh, PhD, City University of New York; Levi Waldron, PhD, City University of New York; Sean Davis, MD, PhD, University of Colorado Anschutz School of Medicine

Abstract: Millions of transcriptomic profiles have been deposited in public archives yet remain underused for the interpretation of new experiments. Existing methods for leveraging these public resources have focused on the reanalysis of existing data or analysis of new datasets independently. We present a novel approach to interpreting new transcriptomic datasets by near-instantaneous comparison to public archives without high-performance computing requirements. All necessary data and functions to apply our approach to existing or new data are included in our software available as part of the Bioconductor project. In brief, we performed Principal Component Analysis (PCA) on the results of 536 public genomics studies comprising 44,890 RNA sequencing profiles. Sufficiently similar loading vectors, when compared across studies, were aggregated to form Replicable Axes of Variation (RAV). We annotated RAVs with metadata of originating studies and by gene set enrichment analysis, forming a knowledge graph. Functionality to associate new datasets with RAVs, extract interpretable annotations, and provide intuitive visualization are implemented as the GenomicSuperSignature R/Bioconductor package.

5:00 p.m.–5:15 p.m.

Electronic Health Records (EHR) Significantly Under-Capture Patient Co-Morbidity

Presenter: Thomas Dilling

Authors: Thomas Dilling, MD; Rachel Howard, PhD; Lynn Ansley, Moffitt Cancer Center

Abstract: The concept of a digital twin in healthcare is predicated upon mimicking, as closely as possible, the clinical state of the patient. However, EHR implementations might suffer in data quality, as they are incumbent upon accurate/complete data entry (typically) into discrete data fields. Unfortunately, the well-documented problem of healthcare provider “click fatigue” might preclude full capture of data.

One of the most important predictors of cancer outcome is performance status, which also conceptually incorporates the patient’s comorbid illnesses. Does the EHR potentially underestimate patients’ disease burden? In this analysis, we compared ICD-10 capture of patient co-morbidities from the EHR’s Diagnoses with those from the billing system (coded by billers from each clinic note). We hypothesized that discrete comorbidity data from the billing system would be statistically larger in number and scope than those captured within the EHR.

We utilized a cohort of lung cancer patients treated with radiotherapy who were covered under a retrospective research protocol. Comorbidity information was separately pulled from the EHR (PowerChart, Cerner Medical Systems, Kansas City, MO) and the billing (B) system (Soarian, Cerner Medical Systems, Kansas City, MO). Cohorts were standardized, utilizing only patients with data available from both systems. ICD-9 codes, SNOMED codes and invalid ICD-10 codes were removed. Because chronic conditions were recapitulated at each patient visit in B, all duplicates were removed. By doing so, we generated a “maximal” comorbidity list for each patient from each system for purposes of comparison.

5:15 p.m.–5:30 p.m.

CAFCW21 Wrap-Up