Trends in Cancer
Opinion
Integrative Approaches to Cancer Immunotherapy Gregory L. Szeto1,2,6,* and Stacey D. Finley
3,4,5,6,
*
Cancer immunotherapy aims to arm patients with cancer-fighting immunity. Many new cancer-specific immunotherapeutic drugs have gained approval in the past several years, demonstrating immunotherapy’s efficacy and promise as an anticancer modality. Despite these successes, several outstanding questions remain for cancer immunotherapy, including how to make immunotherapy more efficacious in a broader range of cancer types and patients, and how to predict which patients will respond or not respond to therapy. We present a case for integrative systems approaches that will answer these questions. This involves applying mechanistic and statistical modeling, establishing consistent and widely adopted experimental tools to generate systems-level data, and creating sustained mechanisms of support. If implemented, these approaches will lead to major advances in cancer treatment.
Highlights Cancer immunotherapeutic drugs have shown great success in treating cancer patients, but many outstanding questions remain. There is a lack of mechanistic understanding of the effects of immunotherapy that can be improved using systems approaches. Systems approaches can provide datadriven guidance on sampling strategies for predictive models and rational design of immunotherapies. There is a need for sustained mechanisms to support integrative and inclusive research in this area.
State of Systems Approaches to Immunotherapy Immunotherapy has revolutionized both the research and treatment of cancer. The approach centers on using a patient’s immune system to eradicate cancer and prevent recurrence. Cancer immunotherapy is not a new concept: in the 1890s, Dr. William B. Coley reported shrinkage and durable remission in patients with malignant sarcomas after injecting them repeatedly with bacteria and bacterial lysates (termed ‘Coley’s toxins’) derived from erysipelas, an acute bacterial skin infection [1]. These early treatments included one patient who experienced remission for 26 years until death by heart attack. Widespread progress of cancer immunotherapy for multiple indications was significantly slower throughout the 1900s, due in part to the complex immunosuppressive mechanisms at play in most cancers. Excitingly, however, the field’s impact and maturity has increased exponentially in the past decade, and significant strides have been made by targeting key immunosuppressive molecules. There are now a diverse set of immunotherapies available for many cancers, with some achieving status as first-line treatments. This progress has captured the imagination of clinicians, researchers, and the general public. Clinical responses and deep, durable remissions have been obtained in nearly every type of cancer. In recognition of the field’s accomplishments, the 2018 Nobel Prize in Physiology or Medicine was awarded to Drs James P. Allison and Tasuku Honjo for their work leading to immune checkpoint inhibitors (ICI). ICI and adoptive T cell therapy (ACT) are the two classes of immunotherapy most widely tested and clinically approved. Both approaches focus on enabling specific antitumor responses by T cells. ICI use antibodies to block immunosuppressive molecules. These molecules are natural feedback mechanisms that limit activation of T cells by multiple mechanisms, such as attenuation of co-stimulation or recruiting phosphatases to inhibit T cell receptor signaling [2–4]. Checkpoint molecules become abnormally upregulated in cancer, resulting in T cell inhibition (detailed reviews can be found in [5,6]). The FDA has approved antibodies targeting programmed cell death-1 (PD-1), programmed death ligand-1 (PD-L1), and cytotoxic T lymphocyte-associated antigen-4 (CTLA-4) in diverse solid tumors. ACT is a broad class of ‘living drug’, mostly referring to either T cells or antigen 400
Trends in Cancer, July 2019, Vol. 5, No. 7 © 2019 Elsevier Inc. All rights reserved.
https://doi.org/10.1016/j.trecan.2019.05.010
1
Department of Chemical, Biochemical, and Environmental Engineering, University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, MD 21250, USA 2 Marlene and Stewart Greenebaum Comprehensive Cancer Center, University of Maryland, Baltimore, MD 21201, USA 3 Department of Biomedical Engineering, University of Southern California, 1042 Downey Way, DRB 140, Los Angeles, CA 90089, USA 4 Mork Family Department of Chemical Engineering and Materials Science, University of Southern California, 925 Bloom Walk, HED 216, Los Angeles, CA 90089, USA 5 Department of Biological Sciences, University of Southern California, 3616 Trousdale Parkway, AHF 107, Los Angeles, CA 90089, USA 6 These authors contributed equally to this work
*Correspondence:
[email protected] (G.L. Szeto) and
[email protected] (S.D. Finley).
Trends in Cancer
presenting cells. T cells are the most commonly tested and FDA-approved cell type, with recent approvals of two chimeric antigen receptor (CAR) T cell products: Yescarta for diffuse large B cell lymphoma and Kymriah for B cell acute lymphoblastic leukemia (ALL) and non-Hodgkin lymphoma. CAR T cells are generated by transduction of a genetically engineered receptor into autologous patient T cells. This receptor has the extracellular domain of a tumor-specific antibody fused to intracellular domains that deliver powerful activating signals to T cells. Tumor infiltrating lymphocytes (TILs) are another type of T cell being investigated for ACT. These lymphocytes are tumor-specific T cells expanded from fragments of a patient’s tumor, purified, then infused. The FDA has also approved Provenge, an antigen presenting cell-based therapy for prostate cancer. Provenge is prepared by treating patient’s immune cells with an antigen–cytokine fusion to differentiate dendritic cells, which are then reinfused to recharge antitumor T cells in the body. Other populations in various stages of preclinical and clinical development include effector cells, such as natural killer (NK) or natural killer T (NKT) cells, and antigen presenting cells such as B cells (Roche/SQZ Biotech [7]). There have been many exciting advances as noted earlier, but significant hurdles remain for cancer immunotherapy. Major barriers facing the field are: (i) determining how to increase the number of patients who benefit from immunotherapy, and (ii) enhancing the quality of responses (see Outstanding Questions). ICI antibodies typically benefit a minority of patients: an estimated 43.63% of all US cancer patients were eligible for ICI drugs in 2018, and response rate (including complete and partial responses) was less than 13% [8]. Specifically, a meta-analysis of anti-PD-1/PD-L1 in solid tumors reports 2.19% complete response and 18.9% partial response in all patients receiving immunotherapy [9]. CAR T cells have seen major successes in certain blood cancers (lymphoma and leukemia), with clinical benefit in N80% of patients, but CAR T cells do not achieve efficacy in other blood cancers and solid tumors. Immune-related adverse effects (irAEs) remain a serious concern: both acute and chronic irAEs have been reported in patients treated with ICI and CAR T cells. The severity of irAEs range from acute and mild (i.e., diarrhea) to chronic (i.e., type I diabetes and other autoimmune disorders and syndromes) and potentially fatal (i.e., acute cytokine storms and myocarditis). For example, prevalence of irAEs in blood cancer patients treated with CAR T cells was 55.3% for cytokine release syndrome and 37.2% for neurotoxicity [10]. The rate of high-grade irAEs for ICI in melanoma have been reported at 55% for combination (anti-PD-1 and anti-CTLA-4), 27% for anti-PD-1 (nivolumab), and 16% for antiCTLA-4 (ipilimumab) [11]. Fatal irAEs for ICI remain rare (b1%) but important to understand [12]. Deaths in CAR T cell trials can be more prevalent, such as uncontrolled cytokine release syndrome in three out of 12 patients receiving CD19-targeted CAR T cells [13]. Other hurdles for cancer immunotherapy include: predicting responder/nonresponder status, identifying criteria to safely stop immunotherapy treatment, and matching patients to the correct immunotherapy with the goals of both improving clinical outcome and reducing incidence of adverse effects. We posit that systems approaches to immunotherapy hold unparalleled, underutilized promise in solving these issues by addressing the outstanding questions. Systems-level approaches have been applied in many contexts related to cancer [14], such as cell signaling cascades, metabolic networks, and cell–cell interactions leading to tumor growth. These approaches investigate how the ‘parts’ of a system give rise to the behavior of the ‘whole’ using computational modeling combined with experimental methods [15]. Systems approaches are inherently data driven, as quantitative data is needed to build robust, predictive computational models. Additionally, there is feedback from modeling to experiments [16], where the models generate new hypotheses that can be tested experimentally. The models can also be used to design new experiments needed to obtain greater insight into the system [17]. An advantage of systems approaches is that they allow controlled exploration of the roles of multiple cell types, molecular species, and biochemical reactions. The focus on how individual
Trends in Cancer, July 2019, Vol. 5, No. 7
401
Trends in Cancer
components of biological systems contribute to system function and behavior facilitates a deeper understanding of complex biological processes, and provides opportunities to develop new hypotheses and interventions [18,19]. Thus, systems approaches are particularly well suited to questions in cancer immunotherapy. Systems approaches to immunotherapy necessarily include both computational and experimental tools, which complement each other. Experimental data tells you the endpoint and, with some providence and careful design [17], some of the dynamics of processes at intermediate time points. In contrast, a computational model predicts the intermediate dynamics, which are of great importance in the context of immunotherapy, as both the body’s immune system and the tumor itself are constantly changing and adapting. Experimental measurements can provide a quantitative result or outcome, while a computational model can give the potential mechanism, or a set of possible mechanisms, that produces the result [20]. When combined, robust quantitative data and computational modeling and analysis can achieve the goal of providing detailed insight [19]. These approaches, enabled by mechanisms that provide sustained support, will drive forward immunotherapeutic strategies (Figure 1, Key Figure).
There Is a Lack of Mechanistic Understanding of the Effects of Immunotherapy That Can Be Improved Using Systems Approaches Several issues related to the effects of immunotherapy could greatly benefit from systems-level approaches. For example, there is a need to stratify patients, predicting those that will respond and those that will not respond to treatment. As discussed previously, immunotherapy has been successful in eliminating and controlling tumor growth for some patients. However, many patients do not respond to immune-based treatment strategies, even exhibiting strong resistance to the treatment [21]. There is currently little insight as to what distinguishes patients that do respond from those that do not. It is known, however, that there are complex and dynamic interactions between the immune system and the tumor. Therefore, systems-level models that account for these interactions, in tumor-specific and patient-specific ways, can aid in stratifying patients before deciding whether to initiate or continue with treatment. This requires incorporating characteristics about the tumor itself, the patient’s immune system, and the proposed immunotherapeutic strategy. Systems-level mathematical modeling and computational analyses are useful in addressing the question of patient stratification in the context of immunotherapy, predicting the efficacy of immune-targeting drugs, and many other pressing questions [18,22–24] (see Outstanding Questions). As always, the complexity of the model should match the questions at hand. In this regard, there are two broad classes of modeling: mechanistic and statistical. Mechanistic Modeling Ordinary differential equations (ODEs) or agent-based tools can provide detailed insight in order to answer questions, such as why the tumor continues to grow after treatment and which interactions contribute to response or resistance (see Outstanding Questions). There are examples of mechanistic systems-level models providing clinically relevant predictions regarding traditional cancer treatment. Such successes spur an intentional focus on applying the same approaches to answer questions in immunotherapy [24]. ODE-based mechanistic models provide a detailed framework to simulate dynamic interactions between biological entities. These entities, for example, could be intracellular molecular species or cells that comprise the tumor or the immune system. The ODEs predict how each entity evolves over time (i.e., the change in a species’ concentration or the change in number of cells). Such models are deterministic, predicting the same final state every time the model is run, if the same set of initial conditions are used. In 402
Trends in Cancer, July 2019, Vol. 5, No. 7
Trends in Cancer
Key Figure
The Keys to Success for Systems Approaches to Cancer Immunotherapy
Trends in Cancer
Figure 1. The development of immunotherapeutic drugs for cancer has led to durable responses in numerous patients. However, many questions remain. Combining mechanistic and statistical modeling, consistent and widely adopted experimental tools to generate systems-level data, and sustained mechanisms to support systems immunotherapy will produce answers to these questions and enable major advances in cancer treatment.
addition to ODE models, other types of deterministic modeling approaches are useful for understanding cancer immunotherapy. These include partial differential equation models, which account for variations in space, and discrete deterministic models, where the values of the entities being modeled are distinct as opposed to continuous. ODEs account for the detailed mechanisms by which these changes occur, such as specific biochemical reactions. Given the level of detail, ODE-based models can be used to predict the dynamics of direct tumor–immune interactions [25] or the effects of targeted biochemical changes on immune cell signaling and activation [26]. Similarly, agent-based models provide detailed mechanistic insight. Here, the focus is on modeling individual agents (i.e., cells or molecules) in a discrete state space [27], as compared to ODEbased models, which fully specify the dynamics of a process continuously over the time domain. Trends in Cancer, July 2019, Vol. 5, No. 7
403
Trends in Cancer
Additionally, agent-based models are probabilistic, accounting for randomness in the system [27]. Agent-based modeling simulates the behavior of a population of cells arising from a complex network of interactions that follows a set of stereotypical rules [28]. The rules, which guide the individual cell’s decisions and influence the collective behavior of the population of cells, are based on the principles of thermodynamics, mechanics, and the physical sciences [29]. Agent-based modeling can be applied to predict the effects of changing interactions between species, for example, by specifying that the ability of an immune cell to promote tumor cell killing is diminished in certain tumor microenvironments [28]. Mathematical models provide detailed insight regarding how to enhance immune cell activation and identify potential therapeutic biomarkers; as an example, Arulraj and Barik recently developed an ODE model to investigate the mechanism used by PD-1 to inhibit T cell receptor signaling [30]. PD-1 is an immune checkpoint molecule that is the focus of ongoing preclinical and clinical studies. In fact, there are multiple FDA-approved ICI antibodies targeting PD-1, which are effective in only a subset of patients and produce serious adverse effects. Therefore, it is important to better understand how PD-1 regulates T cell activation. The model by Arulraj and Barik predicts two modes of regulation mediated by PD-1: directly dephosphorylating T cell receptor signaling molecules and indirectly regulating the T cell receptor by inhibiting the key kinase, LCK, leading to downregulation of LCK-activated signaling molecules. Collectively, their simulations highlight that T cell inhibition by PD-1 is mediated via LCK, a novel insight that is useful for controlling PD-1’s effect. The role of LCK in regulating T cell function and CAR activation has also been modeled using mechanistic ODE models [31,32]. Again, these models provide novel hypotheses, for example, how to enhance LCK activity or speed up overall activation of the engineered receptor. In another example, Gong et al. used agent-based modeling to predict the tumor response to inhibition of PD-1 and its ligand, PDL-1 [33]. The model predicted the spatial distribution of PD-1 in tumor tissue and how the spatial heterogeneity influences treatment outcome. Such a modeling approach is useful for identifying potential tissue-based biomarkers for immunotherapy. Statistical Modeling Data-driven statistical models are well suited to predict the drug sensitivity of cancer cell lines or what a patient’s disease course will be. The development and optimization of experimental tools that characterize biological scales ranging from whole tumor tissues to single cells enable these statistical, data-driven approaches to systems immunotherapy [34,35]. One method of constructing statistical models is machine learning (ML), which has gained great prominence in recent years. ML predicts response and outcome of immune-based treatment strategies. For example, Ding and coworkers used deep learning of genome-scale omics data (including gene expression, copy number, mutation status, and drug sensitivity) to identify features that predict drug effectiveness in cancer cell lines [36]. Their approach revealed features that could predict both effective drugs and sensitive cell lines with high specificity and sensitivity. The features were validated using an independent data set, indicating the predictive capability of the ML approach and potential applicability to the clinical setting. In another example, Charoentong et al. focused on the interactions between tumor and immune cells [37]. By mining cancer data in The Cancer Genome Atlas (TCGA)i, the authors revealed a clear association between tumor genotypes and the phenotype of immune cells and tumor escape mechanisms. Using ML on the genomic data, the authors also developed an ‘immunophenoscore’, a set of immune-related genes that can be used to classify patients that will respond to ICIs targeting CTLA-4 and PD-1. In related work, ML was used to develop a radiomic signature (based on computational medical imaging) of solid tumors that accurately predicts patient response to ICI [38]. Sun et al. identified eight 404
Trends in Cancer, July 2019, Vol. 5, No. 7
Trends in Cancer
radiomic gene expression signatures of CD8+ T cells that provide a noninvasive biomarker to characterize the tumor immunophenotype and clinical response to ICIs anti-PD-1 and anti-PD-L1 [38]. These are just a few examples from a growing body of work that applies ML to predict response and outcome of immune-based treatment strategies. There is a need to connect mechanistic and statistical models. The potential applications and utility of ML and the development of more computationally efficient deep learning algorithms are met with great excitement. However, we assert that ML is not the right answer all the time. In its current state, ML alone will not elucidate cancer immune networks, as there is little insight provided about the underlying biological mechanisms that yield the model predictions. We need approaches that go beyond the prediction of whether a tumor will grow or not given a specific drug treatment. Particularly in the case of cancer immunotherapy, we must focus on interpreting ML models to understand how and why certain features are used to make accurate predictions. Even in the example presented previously, Charoentong and coworkers admit that they can only speculate which mechanisms underlie the patient-specific response to ICI antibodies predicted by the immunophenoscore. Thus, while mechanistic and statistical models are each useful in specific contexts, they can be even more powerful when paired together.
Systems Approaches Can Provide Data-Driven Guidance on Sampling Strategies for Predictive Models and Rational Design of Immunotherapies There are two challenges in cancer immunotherapy which require integrative systems approaches: (i) creating computational models to predict treatment efficacy in patients, and (ii) deriving comprehensive understanding of mechanisms of action to rationally inform dose scheduling and immunotherapy combinations (see Outstanding Questions). The overall goal of improving immunotherapy outcomes is shared by many research groups, but chosen experimental methods are paired to computational models that vary widely, and many risk oversimplifying the complexity of the problem. We assert that a harmonious solution can be found to improve immunotherapy outcomes in diverse cancers by balancing experimental and computational considerations. This effort will require close collaboration between immunologists and computational biologists to couple data-driven understanding of mechanisms to targeted use of experimental assays that capture multiple data types, developing algorithmic approaches for data integration, and multisite validation. Large-scale studies characterizing tumors have generated complex data structures of staggering size, spanning tumor type, anatomical location, and time. A large methodology gap currently exists: there is no consensus method or pipeline for integrating these factors and disparate data types into a cohesive framework. Data from TCGA recently provided the first multiomics study of the tumor immune landscape across many cancers [39]. Thorsson et al. presented an analysis of TCGA data providing the broadest pan-cancer assessment of immunity in cancer to date [40]. This study integrated six molecular assays (mRNA, microRNA, and exome sequencing; DNA methylation, copy number, and reverse phase protein arrays) with clinicopathologic data and outcomes for 33 cancer subtypes. Tumor gene expression profiles identified six distinct immune subtypes categorizing 30 nonblood cancers. Adding immune features {e.g., subtype scores, infiltrate data, immune gene signatures, T and B cell receptor diversity, neoantigen counts [insertion/deletion polymorphism (indel) and single nucleotide variant (SNV)], lymphocyte fraction, average cancer testis antigen expression} to tumor type and staging in elastic net regressionbased Cox proportional hazard models significantly increased predictive accuracy of overall survival. A high/low-risk binary cut-off demonstrated that immune features alone could classify patients with 30% increased probability of survival. Unique in this study was the combination of insights into predictive models and underlying mechanisms of action. The authors examined Trends in Cancer, July 2019, Vol. 5, No. 7
405
Trends in Cancer
the association between common driver mutations to immune subtype, and constructed consensus regulatory networks linking underlying genomic, epigenetic, and transcriptomic state to immune infiltration and survival outcomes. This study is both thought-provoking and robust, but three major caveats are: features were all analyzed at baseline, experimental data was gleaned exclusively from tumors, and multiomics analysis was not integrated across scales or provided as a package for analysis of validation cohorts. Significant work has focused on interrogating primary tumors to predict outcome and therapeutic efficacy. An important, related construct for evaluating immunotherapy’s potential is the paradigm of immunologically ‘hot’ or ‘cold’ tumors, defined as the extent of T cell infiltration in a tumor, and tumor cell-intrinsic features, such as tumor mutational burden (TMB), which together dictate whether an immune response can be mounted against a tumor [41,42]. For example, a solid tumor with high T cell infiltration and high TMB is classified as ‘hot’, while a tumor with little T cell infiltration and low TMB is ‘cold’. Multiple groups have attempted to provide simplified scoring systems inspired by pathological classification of tumors. The Immunoscore is one attempt at using this concept in practice by quantifying the extent of immune infiltration in a tumor [43]. The extent of immune infiltration as CD3+ and CD8+ T cells in the tumor and invasive margin is calculated as the mean of four percentiles (CD3 and CD8 in two regions) and converted into an Immunoscore which is used as a two or three level categorical variable. For three levels, Immunoscore was scored low (0–25% density), intermediate (25–70% density), and high (70– 100% density), while two levels scored all patients between 25–100% density as high. This proved useful in predicting outcomes, such as the time to recurrence in a large international cohort of colorectal cancer patients as well as smaller cohorts of kidney cancer [44]. Distinct from Immunoscore, work by Jiang et al. used computational methods to assess induction of T cell dysfunction and prevention of T cell infiltration into tumors using gene expression [45]. Notably, their metric outperformed PD-L1 protein staining and mutational load for predicting responses. A third recent scoring system [named IMPRES (immuno-predictive score)] identified nearly all responders and misclassified fewer than half of nonresponders in ICI-treated metastatic melanoma [46]. TMB has also been broadly studied, and predictive efficacy identified in diverse cancers for ICI. Overall, these scoring systems are all similar to TCGA findings in their singular focus on tumor samples, typically captured pretreatment. Notably, they provide no framework to evaluate response over time or to design rational immunotherapy dose schedules. The focus on primary tumor samples only also neglects the importance and potential of nontumor tissue samples in providing more predictive and mechanistic information. We posit that more robust models can be generated by optimizing sampling site and time. The utility of nontumor site data has been underscored in recent studies. In particular, biomarkers from blood have been heavily sought and are appealing due to ease of sampling and sample quantity. In one study, exosomal PD-L1 correlates with ICI response against PD-1 [47]. Similarly, sequencing of circulating tumor DNA from ‘liquid biopsies’ of blood have proved to be effective biomarkers of immunotherapy response and tumor burden [48–50]. These results also underscore the importance of time post-treatment when assessing biomarkers. Spitzer et al. closely examine immunotherapy response in multiple tissues using preclinical models [51]. They report numerous changes mirrored from tumor to systemic compartments, including blood and bone marrow. Key circulating CD4+ T cell subsets were identified as key to immunotherapy efficacy. Preclinically, an effector memory Th1 subset was identified in multiple tumor models and proven to be sufficient for tumor rejection by adoptive transfer. In a clinical cohort receiving anti-CTLA-4 and granulocyte–macrophage colony-stimulating factor (GM-CSF) immunotherapy, a different CD4+ T cell subset identified as PD-1 negative and CD127low was enriched in responders. This work demonstrates the emergence of a specific blood T cell population that can be used 406
Trends in Cancer, July 2019, Vol. 5, No. 7
Trends in Cancer
to identify immunotherapy responders. The pipeline developed in this publication, Statistical Scaffoldii, is publicly available on GitHub for replication and validation studies and may facilitate evaluation of temporal dynamics in immunotherapy responses. However, the importance of timing in sampling is vastly understudied. Typically, samples are taken either pretreatment or at the time of observed clinical response. However, these may be the least useful, as they may represent the system at steady state. This is clear, given the change in response dependent on dose schedule in preclinical and clinical studies. Tzeng et al. demonstrate that a two-day delay between tumor ablation and immunotherapy drastically changes response from 0% to 85% in preclinical models [52]. Park et al. show that chemotherapy can ablate antitumor antibody responses in preclinical models of breast cancer [53], while Liu et al. demonstrate superiority of neoadjuvant (presurgical) compared to adjuvant immunotherapy in preclinical breast cancer models [54]. These results are mirrored in clinical studies showing neoadjuvant immunotherapy promotes survival and favorable clinical outcomes in a wide range of cancers including melanoma and glioblastoma [55–57]. Altogether, these studies argue that many sites and data types can provide robust biomarkers of immunotherapy efficacy and increase our understanding of their mechanisms of action. Overall, a quantitative statistical evaluation is needed to determine the relative efficacy of different assay types, sampling sites, and sampling times for predicting immunotherapy response. Comparative studies with parallel sampling of multiple sites longitudinally and of multiple assays are a necessity for accurate comparison of models. While these initial studies will be complex, the end goal should be a streamlined assay and data analysis pipeline to allow the broadest clinical use for multiple cancers and immunotherapy types.
There Is a Need for Sustained Mechanisms to Support Integrative and Inclusive Research in This Area The use of new high-throughput assays that generate large, complex volumes of data is now commonplace in immunology and cancer. These fields are searching for solutions to complex problems, such as predicting vaccine responses, personalizing therapies based on individual mutations, modeling immune system function, and computationally identifying neoepitopes. These problems have required dedicated interdisciplinary teams and will require a new generation of scientists trained to think intersectionally between immunology and computational biology. The requirements for this new breed of computational immunologists are daunting and ill-defined. Graduate training programs need to carefully consider what degree to incorporate curricular additions, and the field as a whole needs to converge upon a minimum standard for competency in training. Multiple organizations have recognized and responded to the need for cross-trained scientists with funding for research projects and training. The American Association of Immunologists (AAI)iii launched the Intersect Fellowship Program for Computational Scientists and Immunologistsiv in 2018. The goal of the Intersect Fellowships is ‘to improve understanding and communication between immunology researchers and computational scientists.’ These fellowships support up to one year of training for a computational science or basic bench science postdoctoral in the reciprocal discipline. The Society for Immunotherapy of Cancer (SITC)v has now held two rounds of Sparkathons: classes of early career investigators from diverse training are convened and funded for US$200 000 over 12 months to address major barriers in cancer immunotherapy. Two projects funded via this mechanism focused on tumor heterogeneity to stratify patient responses, and a clinical trial with data repository to help clinicians decide when to stop immunotherapy. The Cancer Research Institutevi awards grants of up to US$200 000 through their Technology Impact Awardvii. However, none of these mechanisms is sustained, with the longest being two Trends in Cancer, July 2019, Vol. 5, No. 7
407
Trends in Cancer
years in duration. The National Cancer Institute (NCI) has provided the most durable support for interdisciplinary research in cancer through the Physical Sciences-Oncology Network (PS-ON)viii and the Cancer Systems Biology Consortium (CSBC)ix, offering at least five years of funding. The National Institutes of Health (NIH)-funded PS-ON initiative supports Physical Sciences-Oncology Centers (PS-OCs) and Physical Sciences-Oncology Projects (PS-OPs), bringing together biologists and oncologists with researchers from disciplines ranging from chemistry to engineering to address outstanding questions in cancer research. Several of the PS-OCs and PS-ONs are pursuing research aimed at establishing a mathematical and systems-level understanding of the immune system in the context of cancer, including the PS-OCs at the Houston Methodist Research Institute, the University of Minnesota, and the H. Lee Moffitt Cancer Center and Research Institute; as well as PS-OPs at California Institute of Technology and Georgia Institute of Technology. Lastly, the CSBC supports the integration of experimental and computational tools to investigate cancer through center and project grants. Studying tumor-immune interactions using systems biology approaches is a focus of CSBC centers at the City of Hope, Harvard Medical School, the Memorial Sloan Kettering Cancer Center, and Stanford University. Collectively, these funding mechanisms enable scientific advances to be made in the context of immunotherapy. However, more sustained funding is needed, particularly to support trainees and early stage researchers.
Outstanding Questions
Publishers play another key role in dissemination of findings from the field. They have begun to provide outlets for research outputs, but must do more. Currently, Immunohorizons explicitly welcomes articles for novel computational methods and tools. Immunity provides the ‘Resource’ article type and Cancer Research provides a similar ‘Resource Report’ format for computational methods and databases. eLife recently published an article with a focus on computational reproducibility [58] that may relieve a fundamental bottleneck for validation studies. There are also a range of journals with a long-standing commitment to publishing mathematical and computational tools and studies to interrogate complex biological processes, including PLOS Computational Biology, BMC Bioinformatics, and Molecular Systems Biology. Overall, the field as a whole needs to focus not only on the building of new software tools for complex data analysis, but also on elevating the importance of generating tools that can be used by any scientist readily. Requirements for new software and algorithms to be deposited in GitHub and Bioconductor can ensure a minimal level of documentation and versioning. Creation of interactive portals, such as ImmPortx and ImmGenxi, will further increase the likelihood of widespread adoption of specific analytical methods with consistent user experiences and increased reproducibility [59,60]. Many other databases are available for data curation and in silico testing or validation (many databases reviewed in [37]), such as the large public data repository Gene Expression Omnibus (GEO)xii for functional genomics [61], database of Genotypes and Phenotypes (dbGaP)xiii [62], the CODEX databasexiv [63], and open-access online-only publication Scientific Data. Altogether, these efforts will accelerate and enable validation and replication studies, which will determine the clinical usefulness of any given strategy for systems modeling of cancer immunotherapy.
How does time post-treatment change the predictive potential of biomarkers?
Concluding Remarks The body’s immune response to cancer cells is complex, involving cascades of biochemical reactions that occur in different types of cells and across different scales of time and space. Such complexity requires integrative approaches that incorporate our existing knowledge of cancer immunotherapies and extend that knowledge to establish reliable biomarkers and design effective treatment strategies. There is great potential for such approaches to build on and broaden the clinical successes that the field of cancer immunotherapy has already seen. Intentional focus on a few key areas will increase the chances of realizing this potential, including: (i) taking advantage of the vast amounts of in vitro and in vivo systems-level data that are being amassed, (ii) 408
Trends in Cancer, July 2019, Vol. 5, No. 7
How can we accurately predict treatment efficacy in patients? What criteria should be used to safely stop immunotherapy treatment? How can oncologists match patients to the correct immunotherapy? How can the incidence of adverse effects be reduced? Why do some patients’ tumors continue to grow, or grow faster, after treatment? What tumor-immune interactions contribute to response or resistance? How do local and systemic immunity change in primary versus metastatic tumors?
Which tissues, timepoints, and molecules should be monitored for robust predictive modeling of outcome? Can we more clearly define the mechanisms of action for immunotherapeutic drugs? Can modeling be used to rationally inform combination treatments and dosing schedule? Which algorithms can be used to optimize and cross-validate models across space, time, and research groups? What advances in ML techniques are needed generate mechanistic insight? How can statistical and mechanistic models be integrated to address clinical needs?
Trends in Cancer
combining data-driven statistical and detailed mechanistic modeling, (iii) establishing sustained mechanisms to ensure integrative approaches and continued support for researchers, and (iv) ensuring active, equal collaboration between experimental immunologists and computational scientists from study conception through to completion. Finally, we note the importance of considering ethical issues surrounding immunotherapy [64]. For example, there is uncertainty about what criteria should be used for selecting patients for clinical trials, protecting the vulnerable population of cancer patients from harmful side effects of immunotherapies, and setting the costs of these drugs. Moreover, there is a question as to whether it is appropriate to offer patients unapproved drugs shown to be effective in the research setting, particularly under the FDA’s Expanded Accessxv (sometimes termed ‘compassionate use’) designation in patients without comparable alternative therapies. By addressing research and ethical issues (see Outstanding Questions), we expect that the growing number of immunologists, cancer biologists, engineers, physicists, computer scientists, and others from a range of disciplines, coming together to study cancer immunotherapy, along with medical ethicists and policy makers considering how to implement these therapies, will make unprecedented advances in treating cancer. Acknowledgments The authors thank Dr. Neda Bagheri for critical comments that helped to refine this opinion article.
Resources i
www.cancer.gov/about-nci/organization/ccg/research/structural-genomics/tcga
ii
www.github.com/spitzerlab/Modeling_Effective_Cancer_Immunotherapy/
iii
www.aai.org/
iv
www.aai.org/Careers/Fellowships/Intersect
v
www.sitcancer.org/home
vi
www.cancerresearch.org/
vii
www.cancerresearch.org/scientists/fellowships-grants/impact-grants/technology-impact-award
viii
https://physics.cancer.gov/
ix
https://csbconsortium.org/
x
www.immport.org/home
xi
www.immgen.org/
xii
www.ncbi.nlm.nih.gov/geo/
xiii
www.ncbi.nlm.nih.gov/gap
xiv
http://codex.stemcells.cam.ac.uk/
xv
www.fda.gov/news-events/public-health-focus/expanded-access
References 1.
2.
3.
4.
5. 6. 7.
8.
Coley, W.B. (1991) The treatment of malignant tumors by repeated inoculations of erysipelas. With a report of ten original cases. 1893. Clin. Orthop. Relat. Res. 262, 3–11 Chemnitz, J.M. et al. (2014) SHP-1 and SHP-2 associate with immunoreceptor tyrosine-based switch motif of programmed death 1 upon primary human T cell stimulation, but only receptor ligation prevents T cell activation. J. Immunol. 173, 945–954 Parry, R.V. et al. (2005) CTLA-4 and PD-1 receptors inhibit T-cell activation by distinct mechanisms. Mol. Cell. Biol. 25, 9543–9553 Hui, E. et al. (2017) T cell costimulatory receptor CD28 is a primary target for PD-1–mediated inhibition. Science 355, 1428–1433 Wei, S.C. et al. (2018) Fundamental mechanisms of immune checkpoint blockade therapy. Cancer Discov. 8, 1069–1086 Baumeister, S.H. et al. (2016) Coinhibitory pathways in immunotherapy for cancer. Annu. Rev. Immunol. 34, 539–573 Szeto, G.L. et al. (2015) Microfluidic squeezing for intracellular antigen loading in polyclonal B-cells as cellular vaccines. Sci. Rep. 5, 10276 Haslam, A. and Prasad, V. (2019) Estimation of the percentage of US patients with cancer who are eligible for and respond to
9.
10.
11.
12.
13.
14. 15.
checkpoint inhibitor immunotherapy drugs. JAMA Netw. Open 2, e192535 Carretero-González, A. et al. (2018) Analysis of response rate with ANTI PD1/PD-L1 monoclonal antibodies in advanced solid tumors: a meta-analysis of randomized clinical trials. Oncotarget 9, 8706–8715 Grigor, E.J.M. et al. (2019) Risks and benefits of chimeric antigen receptor T-cell (CAR-T) therapy in cancer: a systematic review and meta-analysis. Transfus. Med. Rev. 33, 98–110 Larkin, J. et al. (2015) Combined nivolumab and ipilimumab or monotherapy in untreated melanoma. N. Engl. J. Med. 373, 23–34 De Velasco, G. et al. (2017) Comprehensive meta-analysis of key immune-related adverse events from CTLA-4 and PD-1/PD-L1 inhibitors in cancer patients. Cancer Immunol. Res. 5, 312–318 Frey, N.V. et al. (2014) Refractory cytokine release syndrome in recipients of chimeric antigen receptor (CAR) T cells. Blood 124, 2296 Werner, H.M.J. et al. (2014) Cancer systems biology: a peek into the future of patient care? Nat. Rev. Clin. Oncol. 11, 167–176 Ma’ayan, A. (2017) Complex systems biology. J. R. Soc. Interface 14, 20170391
Trends in Cancer, July 2019, Vol. 5, No. 7
409
Trends in Cancer
16. Ideker, T. et al. (2001) A new approach to decoding life: systems biology. Annu. Rev. Genomics Hum. Genet. 2, 343–372 17. van Riel, N.A.W. (2006) Dynamic modelling and analysis of biochemical networks: mechanism-based models and modelbased experiments. Brief. Bioinform. 7, 364–374 18. Bhinder, B. and Elemento, O. (2017) Towards a better cancer precision medicine: systems biology meets immunotherapy. Curr. Opin. Syst. Biol. 2, 67–73 19. Brodland, G.W. (2015) How computational models can help unlock biological systems. Semin. Cell Dev. Biol. 47–48, 62–73 20. Neves, S.R. and Iyengar, R. (2002) Modeling of signaling networks. BioEssays 24, 1110–1117 21. Zaidi, N. and Jaffee, E.M. (2019) Immunotherapy transforms cancer treatment. J. Clin. Invest. 129, 46–47 22. Agur, Z. et al. (2016) Employing dynamical computational models for personalizing cancer immunotherapy. Expert. Opin. Biol. Ther. 16, 1373–1385 23. Kidd, B.A. et al. (2015) Integrative network modeling approaches to personalized cancer medicine. Per. Med. 12, 245–257 24. Konstorum, A. et al. (2017) Addressing current challenges in cancer immunotherapy with mathematical and computational modelling. J. R. Soc. Interface 14, 20170150 25. Talkington, A. et al. (2018) Ordinary differential equation models for adoptive immunotherapy. Bull. Math. Biol. 80, 1059–1083 26. Rohrs, J.A. et al. (2019) Understanding the dynamics of T-cell activation in health and disease through the lens of computational modeling. JCO Clin. Cancer Inform. 3, 1–8 27. An, G. et al. (2009) Agent-based models in translational systems biology. Wiley Interdiscip. Rev. Syst. Biol. Med. 1, 159–171 28. Norton, K.-A. et al. (2019) Multiscale agent-based and hybrid modeling of the tumor immune microenvironment. Processes (Basel) 7, 37 29. Soheilypour, M. and Mofrad, M.R.K. (2018) Agent-based modeling in molecular systems biology. BioEssays 40, 1800020 30. Arulraj, T. and Barik, D. (2018) Mathematical modeling identifies Lck as a potential mediator for PD-1 induced inhibition of early TCR signaling. PLoS One 13, e0206232 31. Rohrs, J.A. et al. (2016) Predictive model of lymphocyte-specific protein tyrosine kinase (LCK) autoregulation. Cell. Mol. Bioeng. 9, 351–367 32. Rohrs, J.A. et al. (2018) Computational model of chimeric antigen receptors explains site-specific phosphorylation kinetics. Biophys. J. 115, 1116–1129 33. Gong, C. et al. (2017) A computational multiscale agent-based model for simulating spatio-temporal tumour immune response to PD1 and PDL1 inhibition. J. R. Soc. Interface 14, 20170320 34. Kidd, B.A. et al. (2014) Unifying immunology with informatics and multiscale biology. Nat. Immunol. 15, 118–127 35. Newman, A.M. et al. (2015) Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods 12, 453–457 36. Ding, M.Q. et al. (2017) Precision oncology beyond targeted therapy: combining omics data with machine learning matches the majority of cancer cells to effective therapeutics. Mol. Cancer Res. 16, 269–278 37. Charoentong, P. et al. (2017) Pan-cancer immunogenomic analyses reveal genotype-immunophenotype relationships and predictors of response to checkpoint blockade. Cell Rep. 18, 248–262 38. Sun, R. et al. (2018) A radiomics approach to assess tumourinfiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. Lancet Oncol. 19, 1180–1191 39. Liu, J. et al. (2018) An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell 173, 400–416 40. Thorsson, V. et al. (2018) The immune landscape of cancer. Immunity 48, 812–830
410
Trends in Cancer, July 2019, Vol. 5, No. 7
41. Li, J. et al. (2018) Tumor cell-intrinsic factors underlie heterogeneity of immune cell infiltration and response to immunotherapy. Immunity 49, 178–193 42. Maleki Vareki, S. (2018) High and low mutational burden tumors versus immunologically hot and cold tumors and response to immune checkpoint inhibitors. J. Immunother. Cancer 6, 157 43. Mlecnik, B. et al. (2011) Histopathologic-based prognostic factors of colorectal cancers are associated with the state of the local immune reaction. J. Clin. Oncol. 29, 610–618 44. Pagès, F. et al. (2018) International validation of the consensus Immunoscore for the classification of colon cancer: a prognostic and accuracy study. Lancet 391, 2128–2139 45. Jiang, P. et al. (2018) Signatures of T cell dysfunction and exclusion predict cancer immunotherapy response. Nat. Med. 24, 1550–1558 46. Auslander, N. et al. (2018) Robust prediction of response to immune checkpoint blockade therapy in metastatic melanoma. Nat. Med. 24, 1545–1549 47. Chen, G. et al. (2018) Exosomal PD-L1 contributes to immunosuppression and is associated with anti-PD-1 response. Nature 560, 382–386 48. Gandara, D.R. et al. (2018) Blood-based tumor mutational burden as a predictor of clinical benefit in non-small-cell lung cancer patients treated with atezolizumab. Nat. Med. 24, 1441–1448 49. Xi, L. et al. (2016) Circulating tumor DNA as an early indicator of response to T-cell transfer immunotherapy in metastatic melanoma. Clin. Cancer Res. 22, 5480–5486 50. Diehl, F. et al. (2008) Circulating mutant DNA to assess tumor dynamics. Nat. Med. 14, 985–990 51. Spitzer, M.H. et al. (2017) Systemic immunity is required for effective cancer immunotherapy. Cell 168, 487–502 52. Tzeng, A. et al. (2016) Temporally programmed CD8α+ DC activation enhances combination cancer immunotherapy. Cell Rep. 17, 2503–2511 53. Park, S.G. et al. (2010) The therapeutic effect of anti-HER2/neu antibody depends on both innate and adaptive immunity. Cancer Cell 18, 160–170 54. Liu, J. et al. (2016) Improved efficacy of neoadjuvant compared to adjuvant immunotherapy to eradicate metastatic disease. Cancer Discov. 6, 1382–1399 55. Cloughesy, T.F. et al. (2019) Neoadjuvant anti-PD-1 immunotherapy promotes a survival benefit with intratumoral and systemic immune responses in recurrent glioblastoma. Nat. Med. 25, 477–486 56. Blank, C.U. et al. (2018) Neoadjuvant versus adjuvant ipilimumab plus nivolumab in macroscopic stage III melanoma. Nat. Med. 24, 1655–1661 57. Amaria, R.N. et al. (2018) Neoadjuvant immune checkpoint blockade in high-risk resectable melanoma. Nat. Med. 24, 1649–1654 58. Lewis, L.M. et al. (2018) Replication study: transcriptional amplification in tumor cells with elevated c-Myc. Elife 7, e30274 59. Bhattacharya, S. et al. (2018) ImmPort, toward repurposing of open access immunological assay data for translational and clinical research. Sci. Data 5, 180015 60. Heng, T.S.P. et al. (2008) The Immunological Genome Project: networks of gene expression in immune cells. Nat. Immunol. 9, 1091–1094 61. Clough, E. and Barrett, T. (2016) The Gene Expression Omnibus database. Methods Mol. Biol. 1418, 93–110 62. Tryka, K.A. et al. (2014) NCBI’s database of genotypes and phenotypes: DbGaP. Nucleic Acids Res. 42, D975–D979 63. Sánchez-Castillo, M. et al. (2015) CODEX: a next-generation sequencing experiment database for the haematopoietic and embryonic stem cell communities. Nucleic Acids Res. 43, D1117–D1123 64. Jecker, N.S. et al. (2017) From protection to entitlement: selecting research subjects for early phase clinical trials involving breakthrough therapies. J. Med. Ethics 43, 391–400