Predicting the Efficacy and Outcome of I-O Therapies: Between Avatars, Ancers and Algorithms
In the era of precision medicine, there are two critical challenges the immuno-oncology field is facing: evaluating the efficacy of novel immuno-therapies (in comparison to current standard of care checkpoint blockers) and predicting the responses or (innate or acquired) resistance for each individual patient. There is an urgent need for predictive preclinical models to drive rational immunotherapeutic drug development, treatment combinations, and to minimize failures in clinical trials due to improper patient selection.
While some patients experience potent and durable clinical responses to the breakthrough immunotherapies, response rates are highly variable, and the treatments are often associated with side effects that differ from those of cytotoxic chemotherapy.
Two approaches (discussed herein) can be taken to improve I-O drug development processes:
the identification of predictive biomarkers of response (as well as intermediate end points that can provide early signals of efficacy or resistance to treatment) or
predictive modeling of responses with 'next-generation' humanized platforms (avatars vs ancers).
1. Algorithm-derived tumor signature instead of an unique biomarker for I-O therapy
While eg. driver genomic alterations involved in cancer initiation and progression can be successfully pursued as rational predictive biomarker for targeted therapies, there is none single factor sufficient to achieve accurate outcome prediction for immuno-therapies. Eg. PD-L1 does offer guidance as to who will respond better, however the survival benefit is seen in all subgroups of PD-L1 expression (see the side bar). And clinically useful biomarkers to predict the response to anti-CTLA4 treatment remain an unmet need.
Figure 1. Objective responses across tumour PD-L1 expression levels (CheckMate 012; Hellmann et al., 2016)
Multiple factors can affect the effectiveness of I-O drugs, including the degree of cytotoxic T cell infiltration, mutation or neo-antigen load, PD-L1 level, antigen presentation defects, interferon signaling, mismatch repair deficiency, tumor aneuploidy. Instead of a single actionable alteration, transcriptome signatures may better serve as reliable biomarkers in immuno-oncology. Comprehensive review on next-generation I-O biomarkers by Cesano and Warren, 2018.
1.1. PD-1/PD-L1 ICB introduce a new class of predictive biomarker assays: complementary diagnostics
Despite technical pitfalls that make clinical application challenging (PD-L1 negativity is unreliable and results may vary depending on antibody, assay, or tissue sample), two PDL1 immunohistochemistry (IHC) tests are currently approved by the US FDA for guiding treatment decisions. Currently, the only FDA-approved companion diagnostic (predictive biomarker assays linked to a specific drug considered an important treatment decision tool) is PD-L1 IHC 22C3 pharmDx, which is used to select patients for treatment with pembrolizumab, a PD-1 inhibitor marketed as Keytruda. According to a recent Blueprint Working Group analysis Hirsch et al., 2017) comparing IHC tests and cell scoring methods for PD-L1 expression, more data is needed before an alternative assay can be used to read specific therapy-related PD-L1 cutoffs. For nivolumab and atezolizumab, the assays PD-L1 IHC 22C3 pharmDx and Ventana PD-L1 (SP142) have status as complementary diagnostics, which means that there are no requirements for testing included in the labeling for these drugs.
1.2. Tumor mutation burden: from a biomarker to an indication for I-O therapies
At the October 2017 WCLC, researchers presented data from CheckMate-032 Phase I/II open-label trial comparing nivolumab monotherapy with nivolumab plus ipilimumab combination therapy in patients with advanced small-cell lung cancer. The data showed that a high TMB predicted better outcomes, regardless of the treatment arm, compared with medium of low TMB. In addition, patients with a high TMB who received combination therapy had significantly higher response rates and one-year overall survival than those who received monotherapy. These findings provided strong support for the clinical utility of TMB as a biomarker for nivolumab therapy, both alone and in combination with ipilimumab.
Tumour mutation load can be determined by whole-exome sequencing, but widespread access to this method is limited in a clinical setting because of its high cost and bioinformatics requirements. It has been shown that the mutation load can be estimated with similar accuracy using targeted sequencing panels - the Foundation One test (Foundation Medicine) is a validated (approved by the FDA in December 2017) targeted sequencing approach to characterize mutations in 324 genes known to be mutated in solid tumours for clinical management purposes, including the selection of appropriate FDA-approved treatments in certain cancer types.
In May 2017 FDA granted accelerated approval for ICB to a treatment for patients that have been identified as having a biomarker referred to as microsatellite instability-high (MSI-H) or mismatch repair deficient (dMMR). Mutations in genes for DNA mismatch repair pathways are associated with high tumour mutation load. This is the first time the agency has approved a cancer treatment based on a common biomarker rather than the location in the body where the tumor originated.
1.3. Computational frameworks and state of the art technologies
New genomic and proteomic technologies, combined with advanced bioinformatic tools, enable the simultaneous analysis of thousands of biological molecules. These cutting-edge techniques help enable the discovery of new tumor signatures, which are critical for making the leap to precision medicine and personalized I-O therapy. Mass cytometry, whole-exome sequencing, gene expression profiling, CyTOF (see global characterization of T cells in non-small-cell lung cancer by single-cell sequencing by Guo et al., 2018) and T-cell receptor clonality sequencing technology (extending on the information from the IHC-based Immunoscore® assay, HalioDx SAS, immunosequencing technology enables the profiling of T cell and B cell repertoires and is available as a commercial assay, ImmunoSeq, Adaptive Biotechnologies, although its clinical utility in this setting has not yet been established) are just a few of the novel technologies and high-throughput approaches being used for biomarker development. The resulting quantity and complexity of data requires unique analytical challenges and multi-disciplinary expertise for interpretation.
One gene signature, the Tumour Inflammation Signature (TIS), has been developed on the NanoString platform as a clinical-grade assay that provides both quantitative and qualitative information about the immune environment within a tumour, reporting on the presence of an immune infiltrate as well as the functional status of T cells. The TIS was developed on the NanoString nCounter® gene expression system (NanoString Technologies, Inc., Seattle, WA, USA) by Merck as a clinical-grade trial assay to predict immune response to pembrolizumab (Ayers et al., 2017). In HNSCC tumors, TIS has been shown to have greater sensitivity and improved negative predictive value relative to PD-L1 IHC to detect responders to pembrolizumab.
Figure 2. Actionable immune-based classification of cancer by The PanCancer IO 360™ assay (NanoString Technologies Inc.). Anticancer immunity in humans can be histologically segregated into three main phenotypes: the inflamed phenotype (also known as “hot”), the immune-excluded phenotype, and the immune-desert phenotype (the latter two considered “cold” tumours). The TIS gene expression profiling algorithm described by Ayers et al. (2017) is at the base of this decision tree. Additional mechanisms of peripheral immune suppression may exist, including other IC as well as negative regulatory cell subtypes. In the case of the non-inflamed phenotype, the next important question to be answered is whether there are defects in T cell trafficking or in appropriate T cell priming and activation (intrinsic to the tumor or specific to the host). The IO360 panel supports the development of signatures to potentially predict a patient response to a variety of immunotherapeutic interventions. Within the framework of the panel, the biology of the tumour can be matched with the mechanism of action of a particular drug. (Cesano and Warren, 2018)
To predict ICB response, researchers at Harvard (Jiang et al., 2018) developed a computational method called TIDE, which integrates the expression signatures of T cell dysfunction and T cell exclusion to model tumor immune evasion. The TIDE signatures, trained from treatment-naive tumor data, can predict ICB clinical response based on pre-treatment tumor profiles. TIDE predicted the outcome of melanoma patients treated with first-line anti-PD1 or anti-CTLA4 more accurately than other biomarkers such as PD-L1 level and mutation load. Furthermore, TIDE predicts regulators of ICB resistance whose inhibition might improve ICB response.
2. HIS preclinical models for evaluation of I-O therapies
In the recent years there has been much effort in improving the preclinical evaluation of onco-immunotherapies and increasing their translational power. Humanized mouse models (aka HIS, human immune system models) have been developed to study and modulate the interactions between immune components and tumors of human origin.
Technavio analysts forecasted (August, 2017) the global humanized mouse model market to grow at a CAGR of almost 10% from 2017-2021 and the market was projected to reach $118.15 million by 2021. Top vendors include genOway, Horizon Discovery, Taconic Biosciences and Jackson Laboratory.
Figure 3. Strategies to generate humanized PDXs. Sources of immune cells include tumour-infiltrating lymphocytes (TILs), peripheral blood mononuclear cells (PBMCs) or CD34‑positive haematopoietic stem cells (HSCs); HSCs may be purified from mobilized adult peripheral blood, bone marrow or umbilical cord blood. (Byrne et al., 2017)
Fully humanized systems use severely immunodeficient mouse strains such as NOG, NSG and BRG. See the very recent review by De La Rochere et al., 2018 for different immuno-deficient host mice used for humanization protocols. All these different strains show subtle differences to support the engraftment of functional human immune cells. In addition, these mouse strains have been improved by genetic modifications for the production of a variety of human cytokines (2nd generation of HIS models) that stimulate the differentiation of additional haematopoietic lineages. For example, strains such as NOG‑GM3, NSG‑SGM3 and MISTRG produce increased numbers of human myeloid and mast cells, regulatory T cells and NK cells.
Figure 4. Genetic modification of HIS mice to improve HSC and PBMC engraftment, and to diminish xeno-GvHD. (De La Rochere et al., 2018)
One methodology for the generation of humanized mice involves the transplantation of total PBMC from healthy human donors or patients or, in particular applications, the infusion of tumour-infiltrating lymphocytes (TILs) - these procedures are known to cause severe graftversus- host disease (GvHD) beginning 2–5 weeks after injection, seriously limiting the useful investigative time window of these models and the translational value of these studies. The other methodology - involving injection of human CD34‑positive HSCs into the mice - leads to the generation of major histocompatibility complex (MHC)-restricted T cells and B cells, as well as to limited amounts of monocytes, macrophages, neutrophils and dendritic cells. This can be eluded in BLT model, where the transplanted human fetal liver and thymus provide a human thymic microenvironment that supports the development of human T cells and their selection on human MHC molecules, T cells with affinity for mouse MHC are not eliminated, with the consequence of a higher incidence of GvHD than seen in other CD34+ HSC-engrafted models. See the very recent review by De La Rochere et al., 2018 for comparison of PBL and CD34 humanization of HIS mice.
Webinar: Cutting-edge development of mouse models possessing reconstituted human immune system.
One of the first studies in this area showed that CAR-T cells designed to recognize mesothelin, an antigen highly expressed on mesothelioma cells, exerted potent anti-tumor effects on malignant mesothelioma of Hu-PBL-mice. HIS models allow the study of human antibody-dependent cellular cytotoxicity (ADCC). Bi-specific antibodies targeting T cells to a tumor antigen have been evaluated in humanized preclinical models of colon carcinoma (anti-EpCAM/CD3), lymphoma (anti-CD20/CD3), and ovarian carcinoma (anti-CD3/CLDN6 and anti-CD3/EpCAM).
2.1. Cancer avatars as the epitome of personalized medicine
Patient-derived xenograft (PDX) provide arguably the closest model to human cancer available without using humans themselves. Malaney et al., 2014 summarize available commercial services in relation to PDX, encompassing 15+ providers, as well as several public PDX initiatives (US NCI repository, US pediatric preclinical testing consortium, PRoXe open-source repository, Novartisu Institutes of Biomedical Research Encyclopedia, The EurOPDX Consortium).
Development of mouse avatars entails implantation of PDX in mice for subsequent use in drug efficacy studies. To make a model system, a piece of a human tumor is implanted and grown in several immunocompromised mice, then harvested, fragmented, and implanted again in a larger set of HIS mice, without any in vitro manipulation.
The ability of PDX models to predict clinical outcomes is being improved through mouse humanization strategies and the implementation of co‑clinical trials, within which patients and PDXs reciprocally inform therapeutic decisions. The generated data is further used to learn the algorithms for future predicitons. Apart from predicting the I-O therapy responses of a specific patient’s tumor, PDX-engrafted HIS mice may be used to interrogate primary and acquired resistance to ICB. PDX models can as well be used in preclinical drug testing.
Figure 5. PDX preclinical/co-clinical study designs. a | PDX models allow population-based studies to be carried out, which better mimic the inter-tumour heterogeneity that is seen in patients and are more predictive of clinical efficacy than conventional xenografts of immortalized cancer cell lines. PDX molecular characterization and correlation with therapeutic response also facilitates biomarker discovery, as well as the identification of primary (and acquired) resistance mechanisms. b | Co‑clinical avatar studies allow for simultaneous drug testing in mice and patients for real-time adaptive therapeutic decisions. c | In the ‘biofacsimile’ or ‘proxy’ study format, integrative systems-based bioinformatics analysis can be used to pinpoint the best-matched PDX for a given patient from a collection of molecularly profiled models. PDX associated information is then leveraged to instruct clinical treatment options and/or to derive prognostic indicators. (Byrne et al., 2017)
The two greatest barriers to widespread use of mouse avatars are the time (the growing process of PDX takes three to six months) and expense (the cost of developing individualized PDX mice stands at tens of thousands of dollars—for just a single patient’s tumor) required to breed and maintain mice engrafted with human tumor tissue. For those who can afford them, mouse avatars are commercially available from Champions Oncology (chairman of the company’s board of directors is David Sidransky, Johns Hopkins University).
In a study published last year Sidransky together with Hidalgo (Dana Farber Cancer Institute) et al. reported 87% association between the drug responses in a patient and their corresponding PDX (Izumchenko et al., 2017). This goes in line with several recent studies (eg. Gao et al., 2015) in which responses obtained in mice were highly consistent with responses in patients, testifying that PDXs can act as a powerful resource for large-scale genotype–response correlations and therapeutic studies in genetically defined tumour subsets. - the correlation data so far only refers to non-HIS settings.
Figure 6. Comparative quantitative data of response rates in PDXs versus human patients (all non-HIS model setting). (Byrne et al., 2017)
In order to reflect sensitivity to drugs, the xenografts need to be routinely analyzed using exome and RNA sequencing to confirm that they match the genetic profile of the original tumor. PDXs were shown to retain the idiosyncratic characteristics of different tumours from different patients and effectively recapitulate the intra-tumour and inter-tumour heterogeneity that typifies human cancer. Although some engraftment-associated selection has been documented, several studies demonstrated that PDX models seemingly preserve most of the genomic clonal architecture of the original patient sample and also seem to resemble patient counterparts at the transcriptomic, epigenomic and histological levels, as well as in terms of shared signalling pathways.
Notably, both stromal and immune components of PDX are replaced over time by murine analogues and the haematopoietic elements show important differences in their spatial distribution or may be missing overall, which affects the signal received from molecular profiling. Golub and Beroukhim (Broad Institute) et al. 2018, found that PDX undergo mouse-specific tumor evolution and some of the genetic changes (several CNAs recurrently observed in primary tumors gradually disappear in PDXs) are also associated with differences in how the PDXs respond to cancer drugs. Therefore, the application of specific algorithms might be required for signal correction to avoid or reduce artefacts and biases.
2.2. Ancers to replace avatars
To reduce experimental in vivo burden, alternative strategies could rely on step-wise approaches, in which large-scale pharmacogenomic screens are carried out using less laborious ex vivo formats followed by in vivo validation in selected, molecularly relevant PDX-HIS models. In this regard, it is noteworthy that patient-derived material from human solid tumours, such as colorectal, pancreas and prostate cancers, can be grown and nearly indefinitely expanded as 3D organoids amenable to drug screens in a semi-high-throughput manner.
Mitra Biotech engineered personalized tumor ecosystems that contextually conserve the tumour heterogeneity, and phenocopy the tumour microenvironment using tumour explants maintained in defined tumour grade-matched matrix support and autologous patient serum, together with autologous PBMCs. Importantly, the ecosystem preserves the immune system heterogeneity, including lymphoid as well as the myleoid compartment.
The functional response of tumour ecosystems to anticancer drugs, together with the corresponding clinical outcomes, was used to train a machine learning algorithm that allows prediction of clinical efficacy with 90% accuracy (see video below). The tumor ecosystem and algorithm, (CANScript technology) emerges as a powerful platform to study the effect of drugs that modulate the immune system and thereby require the complexity of native TME including autologous ligands and PBMC’s (Paris et al., platform poster) as well as for enabling personalized medicine - using the CANscript process they deliver actionable feedback to the physicians in 7 days from specimen receipt, enabling fast and informed treatment decisions.
Webinar: CANscript innovative, ex vivo TME culture platform for interrogating MOA, efficacy, and translatability of oncology compounds.
Molecular Response (development of human TME-aligned, ECM-based model development since 2003) similarly demonstrated that standard-of-care agents showed analogous responses in the 3D ex vivo and patient-matched in vivo models, validating the 3D-Tumor Growth Assay (3D-TGA) as a high-throughput screen for close-to-patient tumors using significantly reduced animal numbers (Onion et al., 2016). Their patient-aligned immuno-oncology model of TGA with functional immune readouts entails optimized isolation protocols and ex vivo 3D culture with tumor/cancer-associated fibroblast (TAF/CAF), and tumor infiltrating lymphocytes (TIL).
Apart from offering an ImmunoGraft platform mentioned above, Champions Oncology has devleoped ex vivo drug sensitivity platforms (utilizing Cypre Symphony™ and Versagel™ technologies) to compliment in vivo studies to enhamce early discovery research.
