The development of cancer immunotherapies and of immune checkpoint inhibitors (ICIs), in particular, has revolutionized the treatment landscape of various solid malignancies and significantly improved the long-term survival outcomes of patients with aggressive tumors. However, innate and acquired resistance to immunotherapy limits the benefit of ICIs to a small portion of patients. As such, many patients are unnecessarily subjected to potentially severe treatment-related toxicities without any clear clinical benefit from the treatment. The use of ICIs in patients that are unlikely to respond to the treatment also drastically increases treatment costs, needlessly imposing a significant financial burden on national healthcare systems.
The urgent need for predictive biomarkers
Given the high resistance rates to these otherwise impressive anticancer agents, the elucidation of resistance mechanisms and the identification of predictive biomarkers have come to the forefront of cancer research. The expression of programmed cell death ligand-1 (PD-L1) and other immune checkpoint molecules on the surface of tumor cells is currently the most reliable means of predicting response to ICIs. Hence, histological analysis for PD-L1 expression using biopsy specimens is becoming standard clinical practice. Yet, PD-L1 expression is extremely heterogeneous, and the expression status of individual immune checkpoint molecules is not sufficient to capture the complexity of the tumor microenvironment.1
The tumor mutational burden (TMB) is another emerging factor predicting response to ICIs, with a high TMB being associated with treatment response. However, TMB and PD-L1 expression fail to accurately predict response to ICIs due to their spatial intratumoral heterogeneity, which undermines their value as biomarkers. Thus, the identification of robust predictive biomarkers integrating information from multiple components of the tumor microenvironment is an unmet clinical need.2
Digital pathology: a new approach for personalized cancer therapy?
Novel methods of quantitative image analysis are rapidly emerging. For instance, the recent development of radiomics and pathomics offers a new integrative approach to comprehensively analyze digitized histopathological images, providing critical spatial and temporal information that traditional histology methods simply cannot offer.3 This unique insight into the tissue’s spatial and temporal architecture allows for personalized cancer treatment, as it sheds light on the interaction between the tumor’s molecular drivers and the different components of the tumor microenvironment.
Using digital pathology to accurately predict response to ICIs
Although at an early stage of development and clinical implementation, analysis of digitized histopathological images (also known as pathomics) has proved successful in unpuzzling several key tumor features from biopsy specimens; these features include PD-L1 expression on tumor cells, CD3 (lymphocyte marker) and CD8 (cytotoxic T cell marker) expression in the tumor microenvironment—all these tumor features have been associated with response to ICIs and, therefore, have a high predictive value.3
Most importantly, the combination of pathomics approaches with unsupervised feature analysis approaches, such as deep learning, can correlate tissue architecture, molecular biomarkers, gene expression profiles, and vasculature characteristics.4 Such a comprehensive characterization of the tumor and its microenvironment is vital for capturing intratumoral heterogeneity and accurately predicting response to ICIs, tumor recurrence, and survival outcomes.
Digital scoring of PD-L1 and other immunological markers can help pathologists overcome many of the manual scoring barriers by virtue of standardized metrics for biomarker assessment at the single-cell level. The ability to apply these methods on whole-tissue sections overcomes the problem of intra-tumoral heterogeneity in PD-L1 expression.1
Mounting evidence underpins the prognostic value of tumor-infiltrating lymphocytes (TILs), especially in patients with immunogenic tumors. Traditional TIL scoring involves the semi-quantitative analysis of standard hematoxylin and eosin (H&E)-stained sections. Heavily infiltrated tumors (commonly referred to as “hot” tumors) are more likely to regress upon immunotherapy than cold tumors. Nevertheless, several factors hinder the wide clinical implementation of TILs to predict response to ICIs: (1) visual scoring of TILs is highly subjective; (2) the lack of standardized scoring guidelines; and (3) the low reproducibility of TIL analysis.4
The implementation of digital pathology approaches can overcome many—if not all—of the shortcomings of traditional TIL scoring. A recent study by Acs and colleagues has shown that in contrast to manual TIL scoring, automated TIL scoring on H&E slides independently predicted disease-specific overall survival in 621 melanoma patients. The authors concluded that their algorithm holds promise for the identification of patients likely to respond to ICIs and those that are not.5
Similarly, Saltz et al.6 developed a deep learning model correlating spatial and molecular features of TILs using deep learning on pathology H&E-stained images. The model identified tumor type-specific and molecular subtype-specific TIL patterns, which correlated with patient survival. Again, the model can be adapted to predict response to treatment with ICIs.
Furthermore, Johannet et al.7 developed a machine learning algorithm integrating information from histology specimens and clinical data to predict response to ICIs. The algorithm accurately predicted response to ICI in a cohort of patients with advanced melanoma, providing evidence of the ability of digital pathology approaches, combined with machine learning, to predict response to immunotherapy.
In addition to predicting response to ICIs, digital spatial profiling of tissue specimens can also be used to identify novel biomarkers for response to immunotherapy. In a recent study, Toki et al.8 used digital spatial profiling of immunohistochemistry specimens from 60 ICI-treated melanoma patients and identified 11 immune markers associated with prolonged patient survival.
Commercially available digital pathology platforms with potential value in predicting immunotherapy response
Commercially available digital pathology platforms that could be used to predict response to ICIs include Immunoscore® (HalioDx), which enumerates peritumoral and intratumoral CD3+ and CD8+ T cells in tumor tissue sections. The prognostic value of Immunoscore® has been extensively studied, especially in colon cancer. However, the potential of Immunoscore® to predict response to ICIs remains to be investigated.9,10
A leading provider of computational pathology software Indica Labs provides two HALO analysis tools (Proximity Analysis and Tumor Infiltration Analysis) to analyze the spatial distribution of immune cells in the tumor microenvironment.
Similarly, the Denmark-based companyVisiopharm offers an artificial intelligence-powered multiplexed phenotyping tool that characterizes the spatial relationship of immune cells in the tumor microenvironment to evaluate immune responses.
The Cognition Master Professional Suite platform by VMscope offers automated scoring of different immunological markers (e.g., CD3, CD4, and CD8) and TILs on whole-tissue sections.
The combination of digital pathology approaches with artificial intelligence is expected to improve patient stratification in immuno-oncology and spare many patients from the unnecessary toxicities of ICIs. Yet, certain challenges need to be addressed before the full potential of these automated quantitative image analysis methods can be harnessed.
Importantly, quantitative image-derived biomarkers identified using digital pathology approaches should be thoroughly validated in multicenter prospective studies. Handling (i.e., analysis, curation, storage, sharing) of big data generated by digital pathology methods is a significant challenge that needs to be addressed. Additionally, new strategies are required to cope with the high computational power necessary to handle digital pathology data.
Another major concern limiting the clinical implementation of deep learning algorithms is the limited understanding of complex models and the interpretation of their findings—commonly known as the “black box” problem.
Despite these challenges, the clinical implementation of digital pathology approaches combined with artificial intelligence is expected to transform modern medicine, including cancer diagnosis, treatment, and prognosis.
In addition to the value of digital biopsies for accurate diagnosis and tumor classification, these approaches will likely form a major part of the evolving tool set for precision cancer immunotherapy. By analyzing digitized histopathological images of biopsy samples, pathologists and physicians will be able to accurately stratify patients and predict response to ICIs and other treatments.
1. V.H. K, K. S, J. R, K.D. M. Precision immunoprofiling by image analysis and artificial intelligence. Virchows Arch. 2018:511-522. http://www.embase.com/search/results?subaction=viewrecord&from=export&id=L625162633%0Ahttp://dx.doi.org/10.1007/s00428-018-2485-z.
2. Ung C, Kockx M, Waumans Y. Digital pathology in immuno-oncology – A roadmap for clinical development. Expert Rev Precis Med Drug Dev. 2017;2(1):9-19. doi:10.1080/23808993.2017.1281737
3. Banna GL, Olivier T, Rundo F, et al. The Promise of Digital Biopsy for the Prediction of Tumor Molecular Features and Clinical Outcomes Associated With Immunotherapy. Front Med. 2019;6(July):6-11. doi:10.3389/fmed.2019.00172
4. Acs B, Rantalainen M, Hartman J. Artificial intelligence as the next step towards precision pathology. J Intern Med. 2020;288(1):62-81. doi:10.1111/joim.13030
5. Acs B, Ahmed FS, Gupta S, et al. An open source automated tumor infiltrating lymphocyte algorithm for prognosis in melanoma. Nat Commun. 2019;10(1):1-7. doi:10.1038/s41467-019-13043-2
6. Saltz J, Gupta R, Hou L, et al. Spatial Organization and Molecular Correlation of Tumor-Infiltrating Lymphocytes Using Deep Learning on Pathology Images. Cell Rep. 2018;23(1):181-193.e7. doi:10.1016/j.celrep.2018.03.086
7. Johannet P, Coudray N, Donnelly DM, et al. Using machine learning to predict immunotherapy response in advanced melanoma. J Clin Oncol. 2020;38(15_suppl):10010-10010. doi:10.1200/jco.2020.38.15_suppl.10010
8. Toki MI, Merritt CR, Wong PF, et al. immunotherapy treated patients using Digital Spatial Profiling. 2020;25(18):5503-5512. doi:10.1158/1078-0432.CCR-19-0104.High-plex
9. Pagès F, Mlecnik B, Marliot F, et al. International validation of the consensus Immunoscore for the classification of colon cancer: a prognostic and accuracy study. Lancet. 2018;391(10135):2128-2139. doi:10.1016/S0140-6736(18)30789-X
10. Ascierto PA, Marincola FM, Fox BA, Galon J. No time to die: the consensus immunoscore for predicting survival and response to chemotherapy of locally advanced colon cancer patients in a multicenter international study. Oncoimmunology. 2020;9(1):1826132. doi:10.1080/2162402X.2020.1826132