Computational Approaches for Cancer Workshop 2022

Cancer is a complex problem that causes morbidity and mortality and poses immense challenges to humankind. Advancements in medical science and the use of simulations are helping scientists and researchers to understand the disease better. Next generation sequencing and molecular simulation techniques have created new opportunities for computational methods to be applied in biological systems. Biological data are converted to computable formats and computer algorithms are used to analyze biological data to provide greater insight. The methods being compute-intensive, high-performance computers are required to perform the analysis. Collaboration is the key to success for cancer research, and shared expertise and resources will provide more effective solutions. International Centre of Excellence on Computational and Biomedical Sciences (ICECBS), a global team science consortium, leverages the best of computer science to address key questions of biomedical sciences. ICECBS emphasizes whole-person health as opposed to reduction-based research, which mostly focuses on a single interventions impact on one or at the most a few physiological systems as separate processes. The consortium brings together multidisciplinary experts such as clinicians, data scientists, computational scientists, and quantum computing developers, for translational research and precision medicine for cancer. A medicinal plant dataset is used to train and test the models developed for classification. Natural Products Repository of the National Cancer Institute (NCI) is an important source of data on plants and marine organisms. Bioactivity Informatics of Indian Medicinal Plants (BIMP) maintains a database of medicinal plants for predicted biological targets, which links to structure and functions at a molecular level. Together with other similar datasets, these two resources for an expanding database could prove to be extremely valuable in identifying new cures based on natural products. Plant based compounds have been known to provide a scaffold for many drugs. The compute-intensive nature of cancer research requires High Performance Computing (HPC) and Artificial Intelligence (AI) solutions. The Accelerating Therapeutics for Opportunities in Medicine (ATOM) is a consortium brought together by the U.S. Department of Energy and NCI to develop an AI driven drug discovery platform. CANcer Distributed Learning Environment (CANDLE) is an effort to develop a broad deep learning infrastructure for use in cancer research. Quantum computing, an emerging technology that uses the laws of quantum mechanics to produce exponentially higher performance for certain types of calculations, offers the possibility of major breakthroughs in computational biology. Quantum computing is a natural fit for cancer research due to the computationally intense and complex nature of the problem. In the current study, we have used Quantum Support Vector Machine (QSVM) as a powerful classification algorithm that can classify objects in n- dimensions by finding a suitable hyperplane. We have performed SVM classification using quantum models executed on a Qiskit platform. Quantum computing can be used for screening known therapeutics relevant plant extracts to screen for new drugs and phytochemicals and identify molecular targets for the development of cancer therapeutics. All these consortia and databases will use quantum computing in a collaborative effort to address the cancer problem. Keywords—Quantum Computing, QSVM, Computational biology, Cancer research, medicinal plants *Corresponding author: Anil Srivastava (anil.srivastava@ ohsl.us) References: [1] Mukherjee, Siddhartha. The Emperor of all Maladies: a Biography of Cancer. Simon and Schuster, 2010. [2] Ansorge, Wilhelm J. "Next-generation DNA Sequencing Techniques." New Biotechnology 25, no. 4 (2009): 195-203. [3] Cacabelos, Ramón. "The Incorporation of Pharmacogenomics to Drug Development in Neuropsychiatric Disorders." Novel Approaches in Drug Designing & Development 1, no. 4 (2017): 60 63. [4] https://ohsl.us/icecbs [5] http://www.scfbio-iitd.res.in/plants_scfbio [6] https://qiskit.org [7] https://dtp.cancer.gov/organization/npb/introduction.htm [8] https://atomscience.org [9]https://datascience.cancer.gov/collaborations/joint-design- advanced-computing/candle

Akshay Seetharam

1. Nalanda University, India 2. Arizona State University, USA 3. Indian Institute of Technology (IIT Delhi), India 4. Open Health System Laboratory (OHSL), USA 5. Chalmers University, Sweden 6. National Cancer Institute, National Institutes of Health (NIH), USA 7. All India Institute of Medical Sciences (AIIMS), India 8. Poznan Supercomputing and Networking Center, Poland 9. Centre for Development of Advanced Computing (C-DAC), India 10. SP Pune University, India

Hong-Jun Yoon

Oak Ridge National Laboratory and University of Dayton

Motivation: The development and use of high- performance machine learning (ML) models for cancer is accelerated by streamlining the ability of researchers to share information. As models become larger and more complex, the ability to leverage the compute capability of exascale machines to optimize training and hyperparameters will be more beneficial to researchers without access to such resources. As illustrated in Figure 1, we propose herein checkpoint conventions and an associated library to ease the generation, distribution, and reuse of ML models for cancer science. For example, the National Cancer Institute (NCI) has deployed the Predictive Oncology Model and Data Clear- inghouse (MoDaC) [1] repository that contains both data and models, but requires additional metadata beyond that stored in standard checkpointing. The Innovative Methodologies and New Data for Predictive Oncology Model Evaluation (IMPROVE) project [2] among other goals, also tries to validate, understand and improve the latest state-of-art drug-response models. By integrating the required MoDaC metadata (and providing a template for other repositories), we enhance the ability of cancer researchers who develop neural networks to make their models more widely available. The CANDLE Checkpoint Module: The Cancer Deep Learning Environment (CANDLE) [3] is a collection of cancer mini-applications called “Benchmarks” and a workflow framework around the Benchmarks called “Supervisor.” The candle_lib package is a pip-installable library designed to standardize and streamline machine learning code development and deployment. Originally developed as part of the CANDLE Benchmarks suite, it is now a independently-installable library that provides various utilities including integration with the CANDLE Supervisor layer to automate running complex workflows on exascale machines. At the core of CANDLE software is the notion of CANDLE compliance, a simple API that can be added to user models to support standardized CANDLE hyperparameters, which control both network architecture and system-level functionality including checkpoints. The CANDLE Checkpoint module inside candle_lib automates several checkpoint functions useful to both stand-alone Benchmarks and Supervisor workflows. It provides callback interfaces for the popular deep learning frameworks, along with standardized methods for con- trolling the frequency and number of saved checkpoints, and automatically generates metadata to satisfy the re- quirements of the various repositories. Additionally, it avoids modifying checkpoints in place, and uses a hard- linking scheme to ensure data consistency if a run crashes during a checkpoint operation. The associated MoDaC utilities allow the user to inter- face with the repository directly. Summary: By including the candle_lib package, the IMPROVE project enables check-pointing in a straight- forward standard way. It supports multiple DNN frame- works and new frameworks can be added as required while providing a consistent interface. This approach enables easy comparison, validation, hyper-parameter optimization and other studies across all the community ML models. We believe that this is critical to advance and standardize rapidly growing research area of cancer drug discovery.

Rajeev Jain

Argonne National Laboratory, Lawrence Berkeley National Laboratory, National Cancer Institute.

Polygenic risk scores (PRSs) are weighted genetic scores that are important in assessing and predicting risk for various clinical disorders. Multiple methods are being used to generate PRSs from summary genome-wide association study (GWAS) statistics. These methods are developed using techniques like pruning and thresholding, Bayesian approaches, and penalized regression and have relative strengths and weaknesses. Preparing the input files, configuring the analysis programs, and performing the post-analysis quality control and plots is a cumbersome process. We created a high performance, parallel analysis pipeline that enables rapid generation of scores on DOE computational resources.

Methods: We developed a workflow that uses the input summary GWAS data and performs standard quality control to generate PRS with multiple software such as PLINK, PRSice-2, LDpred-2, lassosum, PRS-CSx and SBayesR. This workflow was configured to leverage the high-performance computational resources available in the ORNL KDI cluster and can be enabled to run on cloud resources.

Results: We used the pipeline to generate new PRS and validate the PRS scores generated from other large consortia, across multiple VA MVP related projects. One such project was used to help evaluate the ability of genome-wide PRS the prostate cancer risk compared to a recently developed PRS of 269 established prostate cancer risk variants and multi-ancestry weights [REF]. Genome-wide PRS approaches included LDpred2, PRS-CSx, and EB-PRS. The results showed that the PRS constructed using 269 variants had significantly larger AUCs in both African and European ancestry men, with African and European ancestry men in the top PRS decile having larger odds of prostate cancer. The use of the pipeline in this investigation suggested that genome-wide PRS may not improve the ability to distinguish prostate cancer compared to a genome-wide significant PRS. The analysis using the workflow method was executed in less than 5 hours using 20 nodes in the OLCF KDI cluster for all of the software included in the analysis. The workflow ran each of the PRS software in parallel.

Conclusions: The open-source pipeline is available at: https://github.com/exascale-genomics/PRS-dev

Alex Rodriguez

Argonne National Laboratory, Data Science and Learning

Introduction A previously validated radiomic score (RS) predicts the proportion of infiltrating CD8 cells in the tumor microenvironment and response to immunotherapy. We sought to determine whether this radiomic score is stable across multiple lesions measured at varying timepoints within the same patient. Methods The RS was calculated utilizing the DICOManager and Python-based Pyradiomics packages in a manner analagous to the previously described workflow utilizing the LifeX package. A subset of TCIA (The Cancer Imaging Archive) patients were used to compare the RS calculated via our workflow versus those calculated on the LifeX platform. A cohort of melanoma patients treated with PD-L1 inhibitor therapy were then analyzed to assess the within-patient stability of the RS across time. Each patient had up to 3 lesions segmented at 2 timepoints during or after their PD-L1 inhibitor therapy course. The association of radiomic score at timepoint 1 with tumor volume change was also assessed. The correlation ratio (Eta) was calculated first by grouping all lesions for each patient and then after splitting the timepoints into separate nominal groups for each patient. Results Ten patients from TCIA were assessed with RS scores and our pathway yielding a Pearson correlation coefficient of 0.65 (95% CI: 0.03-0.91). One hundred thirty-six lesions from 33 patients were evaluable at least 2 timepoints in the melanoma cohort and with an additional 56 lesions from 28 patients were evaluated at a single timepoint. For lesions with two evaluable time points, our calculated RS at timepoint-1 did not correlate with tumor response (change in tumor volume) (Pearson correlation coefficient 0.16; 95% CI: -0.01-0.32). However, the RS was highly correlated for each patient as evident by eta coefficient of 0.83 (p<0.01), and within lesions measured at the same timepoint (eta coefficient 0.77 for timepoint-1 and 0.89 for timepoint-2; p<0.01). Conclusion Difficulties remain in reproducibility of even well-validated radiomic scores across analysis pipelines. The RS score remains stable across multiple timepoints for a given patient.

Nizam Ahamed

Department of Radiation Oncology, School of Medicine, University of Pittsburgh, PA, USA

Testicular germ cell tumors (TGCT) being the most common solid malignancy in adolescent and young men, are second in terms of the average life years lost per person dying of cancer. Two major types of TGCTs are seminoma (SE) and non-seminoma (NSE). Management of patients with seminoma includes orchiectomy, platinum-based chemotherapy or radiation therapy. Despite a high patient survival rate, current treatments significantly decrease patients’ quality of life and lead to around 40 severe adverse long-term side effects. Platinum-based regimens have heterogeneous outcomes even within patients with the same TGCT histotype that frequently leads to under- and over-treatment. Moreover, about 20% of seminoma patients will experience relapse, and the reason for this phenomenon is unclear. Here we conducted a computational study of 64 pure seminomas (the most common subtype of TGCTs) available at The Cancer Genome Atlas data portal (TCGA). All 162 matching histopathological slides were evaluated by a pathologist for diagnosis confirmation. Consensus clustering approach of 64 pure seminoma samples based on transcriptomic data identified two distinct subtypes. Our analysis revealed that two seminoma subtypes differ in 1) pluripotency stage, 2) activity of double stranded DNA breaks repair mechanisms, 3) rates of loss of heterozygosity, DNA methylation and telomere elongation, 4) expression of lncRNA associated with cisplatin resistance. In comparison to more pluripotent seminoma subtype 1, seminoma subtype 2 shows signs of differentiation into non-seminoma TGCT and therefore may have higher resistance to platinum-based chemotherapy. Moreover, clinical data from TCGA demonstrated higher occurrence of lymphovascular invasion for subtype 2 (43%) in comparison to subtype 1 (25%). We conducted deconvolution of bulk RNA-seq data and also revealed significantly higher immune score for subtype 2, which means higher level of infiltration by immune cells. Despite of the high level of lymphocyte infiltration, five independent clinical trials evaluating the efficiency of immune checkpoint inhibitors in TGCTs treatment were shut down due to lacking clinical efficacy. We identified 20 significantly overexpressed genes in subtype 2 that are related to senescence-associated secretory phenotype (CCL3, CCL13, CSF2, CXCL8, etc.). This fact in combination with data on altered pathways in subtype 2 (downregulation of cell cycle and upregulation of inflammation) allows us to hypothesize that senescence of seminoma infiltrating lymphocytes can be one of the reasons for immunotherapy failure. Using publicly available single cell RNA-seq data of pure seminoma we found that most of senescence-associated genes are overexpressed in immune cells, which agrees with our hypothesis. Using all available histopathological slides of pure seminoma at TCGA we developed test version of deep learning decision making tool for identification of seminoma subtypes using only slide images. Highest achieved validation accuracy was 0.864. Overall our findings showed that seminoma subtype 2 reveals similarity with non-seminomatous TGCTs, which are more aggressive and require different treatment strategy. Therefore, patients with subtype 2 seminoma may require adjustments of a treatment protocol or development of alternative treatment approaches. We propose that consideration of identified subtypes might help to improve seminoma clinical management by administrating subtype-specific treatment.

Kirill Medvedev

University of Texas Southwestern Medical Center, Department of Biophysics

Objective: Multi-parametric MRI (mpMRI), consisting of T2-weighted (T2W), apparent diffusion coefficient (ADC) and dynamic contrast enhanced (DCE) sequences, is widely used for prostate cancer (PCa) screening. Prostate Imaging Reporting and Data System (PI-RADS) guidelines allow standardization of radiologist’s interpretation of PCa (score 1-5) indicating probabilities of PCa being clinically significant (csPCa: Gleason Grade Group (GGG)>1 on biopsy). PI-RADS=3 indicates equivocal presentation between csPCa and clinically insignificant PCa (ciPCa: GGG=1), limiting the scope of non-invasive clinical decisions. Bi-parametric MRI (bpMRI: T2W and ADC) is being explored as a rapid screening protocol for PCa, however, PI-RADS = 3 lesions require DCE, despite a diminished role as per recent PI-RADS (v2.1). DCE is cost intensive, increases scan time compared to bpMRI. In this study, we sought to compare whether radiomics or computationally derived features of PCa from bpMRI are comparable to pharmacokinetic (pK) parameters derived from DCE to distinguish csPCa and ciPCa within PI-RADS=3 lesions. Materials and Methods: We employed datasets from 2 institutions, D1 from PROSTATEx with ND1=20 patients (11 csPCa, 9 ciPCa), and D2 from University Hospitals Cleveland with ND2=25 patients (8 csPCa, 17 ciPCa). For D1, patients were selected on having at least 2 of 4 readers assigning PI-RADS=3 and for D2, PI-RADS=3 assigned by single reader. PCa regions of interest (ROI) came from a consensus of 4 readers for D1 and single reader ROIs were used in D2. Radiomic features (including 1st and 2nd order statistics, Gabor, Haralick, CoLlAGe) were extracted from the ROI’s of D1 and D2 on bpMRI. Top three most frequently selected radiomic features via 3-fold cross validation training were recorded. A set of 3 pharmacokinetic (pK) parameters including Ktrans, Kep, Ve, were extracted from DCE by fitting them to Tofts model. These features were then used to train classification models CbpMRI, CDCE and CmpMRI using bpMRI, pK and bpMRI+pK features, respectively. Classification performance was evaluated using area under receiver operating characteristics curve (AUC). Results: For D1, top ranked radiomics features that distinguished ciPCa from csPCa on bpMRI were gradient, Haralick from ADC and CoLlAGe from T2W. For D2, they were gradient, Gabor, and Haralick features from T2W. For D1, CbpMRI, CDCE and CmpMRI, resulted in AUCs of 0.76, 0.87, and 0.79, respectively. On D2, classifiers CbpMRI, CDCE and CmpMRI, resulted in AUCs of 0.95, 0.64 and 0.86 respectively. CbpMRI and CDCE show comparable performance distinguishing csPCa and ciPCa in PIRADS=3 lesions. However, features that differentiated ciPCa from csPCa on D1 had significantly different presentation on D2 (p<0.05). Classifiers trained on D1 when tested on D2 resulted in AUCs of 0.55, 0.61 and 0.33 in the same order, indicating poor generalizability. Conclusions: Our findings suggest that radiomics from bpMRI are comparable and may outperform pK parameters from DCE in identifying clinically significant prostate cancer among PIRADS=3 lesions. Future research should explore whether radiomics from bpMRI obviates the need for DCE-MRI in PCa significance identification with larger patient sets and investigate how batch effects arising from site and scanner specific variations may influence generalizability of classifiers.

Aaron Ng

Case Western Reserve University, Picture Health Inc., University Hospitals Cleveland Medical Center, Emory University, Georgetown University

Introduction

There are significant variations in hepatocellular carcinoma (HCC) screening and diagnosis protocols across various institutions; and as a result, patients may receive a mix of imaging studies (US, CT, MRI) across their longitudinal screening which makes standardized reporting of the outcome challenging. Natural language processing (NLP) can be utilized to classify imaging reports following standard guidelines; however structured LI-RADS reporting for HCC-screening using MR have only limited data available due to recent introduction of the standard as well as high cost of the exam. If NLP algorithms can be systematically adapted between different domains (UR or MR) and institutes, diagnosis and screening reporting can be standardized which will accelerate information dissemination to medical practitioners, thus improving patient care and treatment planning.

Methodology

We experimented with transferring language models (LM) between radiology reports of ultrasound (US) and MR scans obtained from two different institutes (Inst1 and Inst2) and developed a HCC diagnosis extraction model for both MR and US screening studies. Transferred LM for MR implies that an LM was trained on US reports of Inst1 and then used to initiate the representation of common words with MR vocabulary of an LM which was further trained on MR reports from Inst2 (vice-versa for US LM). We collected 12116 abdominal US studies performed in Inst1 to screen for HCC without LI-RADS score (untemplated), and 1744 templated studies with LI-RADS score (~10% malignant cases with LI-RADS>2). We also collected 9470 untemplated MR studies conducted in Inst2 without LI-RADS reporting, and prepared 1087 LI-RADS annotated reports (~50% malignant cases) with the help of radiologists as benin cases are often not coded with LI-RADS in practice. We experimented with transferring two LM by training the model from scratch including vocabulary space; 1) word2vec (context-independent) 2) BERT (context-dependent). Three classifier pipelines were tested, i.e., word2vec+RandomForest, word2vec+1DCNN, and BERT+RandomForest with malignant vs. benign as target labels. Classification performance was used to benchmark quality of natively-trained LM and transferred LM.

Results

Fine-tuned LM performs better when paired with any of the selected classifiers for the more challenging task of classifying MR reports without template with highest overall weighted f1-score 0.90. Similarly, US-fine tuned language models (best weighted f1-score 0.95) perform better when paired with any of the classifiers (Random Forest or 1DCNN) for the more challenging task of classifying US reports without a template. Learnt LM space clearly demonstrates that semantically similar words (e.g., ‘isointensity’, ‘hypointensity’, ‘hyperintensity’, ‘intense’, ‘bright’, ‘hypointense’) in the original and fine-tuned language space are mapped close together even when they originated from two separate domains.

Discussion

The study reports a successful transfer of language models from one domain to a similar domain in radiology and compares it to not performing adaptation. Experimental results showed that fine-tuning of the word-embedding models with similar domain adaptation (US → MR and MR → US), even for multi-institutional reports, provides more opportunity for semantic knowledge preservation for down-steam HCC classification tasks compared to training the language model from scratch.

Amara Tariq <tariq.amara@mayo.edu>

Mayo Clinic, AZ