Large-scale public data reuse to model immunotherapy response and resistance

1Despite growing numbers of published immune checkpoint blockade (ICB) trials in different cancer types with available omics data and clinical outcomes, ICB response prediction remains an open question. Many published ICB response biomarkers had been trained and tested on limited cohorts and showed variable performance in different cohorts. Moreover, with the limited data size in each clinical study, it is challenging to comprehensively evaluate the complexity of ICB response and immune evasion mechanisms. To address these challenges, we present a data-driven approach integrating large-scale omics data and biomarkers on published ICB trials, non-immunotherapy tumor profiles, and CRISPR screens on a web platform TIDE (http://tide.dfci.harvard.edu).

2Previously, we developed TIDE as a transcriptome biomarker of ICB response by modeling tumor immune dysfunction and exclusion. The statistical model of TIDE was trained on clinical tumor profiles without ICB treatments since the immune evasion mechanisms in treatment-naïve tumors are also likely to influence patient response to immunotherapies. The TIDE model has been applied to evaluate T cell dysfunction and exclusion signatures across over 33K samples in 188 tumor cohorts from well-curated databases, including TCGA, METABRIC, and PRECOG, as well as our in-house collections. In the current work, we significantly expanded the scope of our previous work by incorporating many new datasets and function modules (Additional file 1: Table S1).

3We processed the omics data for 998 tumors from 12 published ICB clinical studies (listed in Additional file 2: Table S2), and eight published CRISPR screens that identified genes modulating lymphocyte-mediated cancer killing and immunotherapy response. The clinical study data from ICB naïve cohorts includes 33K samples in 188 tumor cohorts from well-curated databases, including TCGA, METABRIC, and PRECOG. We integrated these data on the TIDE web platform using the MySQL database. The web platform is based on the Django 3.0 framework. We provided three interactive modules for hypothesis generation, biomarker optimization, and patient stratification (Fig. 1).

4The first module of the TIDE web platform can help cancer biologists prioritize genes in their input gene set for mechanistic follow-up experiments (Fig. 1A). Typically, a genomic experiment, often conducted on model systems in limited sample size, will yield tens to hundreds of gene hits. The large-scale omics data and clinical cohorts collected in TIDE enable cancer biologists to focus on genes with the highest clinical relevance and consistent behavior in other similar experiments. Generally, for any gene sets, a cancer biologist can utilize this module to evaluate each gene for its expression associations with ICB response outcome, T cell dysfunction levels, T cell exclusion levels, and phenotypes in genetic screens in diverse cohorts. To probe a candidate gene further, the user can also use a single gene as query to evaluate how the expression, copy number, somatic mutation, and DNA methylation levels of this gene influence clinical outcome in all collected datasets. Therefore, the prioritization module, integrating many independent cohorts, can help identify genes with improved robustness and clinical relevance.

5To demonstrate an example of usage of the regulator prioritization module, we queried 696 druggable genes annotated by the OASIS database, to find potential therapeutic targets in synergy with ICB (Fig. 2). For example, AXL, a Tyro3/Axl/Mer family receptor tyrosine kinase, is among the top targets ranked by this module to render the tumor microenvironment resistant to ICB. High AXL expression is associated with T cell dysfunction phenotypes in all datasets enumerated (Fig. 2 left panel). Meanwhile, high expression of AXL is also associated with worse ICB outcome in bladder cancer and treatment-naïve melanoma treated with ICB (Fig. 2 second to left panel). Among the cell types promoting T cell exclusion, both myeloid-derived suppressor cell and cancer-associated fibroblast have very high AXL expression level (Fig. 2 right panel). Indeed, in a recent clinical trial NCT03184571, the combination of AXL inhibitor and anti-PD1 has shown promising efficacy among AXL-positive non-small cell lung cancer patients. Hence, this module can prioritize genes with the best potential for developing combination immunotherapies.

6The second module allows translational scientists to evaluate the accuracy of their biomarkers on many ICB cohorts in comparison with other published biomarkers (Fig. 1B). We implemented eight published ICB response biomarkers and applied them to our collection of published ICB trial samples. For a user-defined custom biomarker, which can be a gene set or weighted gene score vector, this module calculates the biomarker expression level in all ICB cohorts. The module displays the comparison between the custom biomarker and other published biomarkers based on their predictive power of response outcome and overall survival.

7To demonstrate an example usage of the biomarker evaluation module, we tested one biomarker containing seven genes with previously reported association with tumor immune evasion (Additional file 3: Table S3). These genes were weighted by their reported direction of mediating anticancer immune response. This example biomarker gave an area under the receiver operating characteristic curve (AUC) greater than 0.5 in 12 out of the 16 ICB sub-cohorts (Fig. 3), suggesting it to be a robust predictive biomarker. This signature also achieved significant associations with prolonging survival in two sub-cohorts (Fig. 4, two-sided Cox-PH p value < 0.05). In contrast, several recently published biomarkers trained on limited clinical cohorts have shown significant performance variations in other cohorts (Additional file 4: Figure S1), underscoring the importance of cross-cohort evaluation of biomarker robustness using all available cohorts.

8The third module of biomarker consensus aids oncologists in predicting whether a patient will respond to ICB therapy based on multiple biomarkers (Fig. 1C). Based on tumor pre-treatment expression profiles, oncologists could use this TIDE module and multiple published transcriptomic biomarkers (Additional file 4: Supplementary Methods) to predict patient response and potentially make informed treatment decisions. Notably, in the second and third TIDE modules, we only focused on evaluating transcriptomic biomarkers but not mutation biomarkers due to the following reasons. The results of tumor mutation analyses might be influenced by different experimental platforms (whole genome versus custom panel), sample types (FFPE versus fresh frozen), and computational mutation callers. Although tumor mutation burden (TMB) seems to be a consistent ICB response biomarker, the computation of TMB across different cohorts and platforms is still an open question.

9To demonstrate an example usage of the biomarker consensus module, we upload the pre-treatment expression matrix of a melanoma cohort treated with anti-PD1 therapy (Table 1). Patients with favorable predictions from multiple biomarkers are highly likely to be responders. For example, patient 2 tumor has a negative TIDE score, indicating a lack of tumor immune evasion phenotypes. In addition, patient 2 tumor has positive scores of interferon-gamma (IFNG) signature, macro-satellite instability (MSI), and PDL1 (CD274) levels, all of which are positive biomarkers of ICB response. With the support from multiple markers, an oncologist could be more confident that patient 2 will respond to anti-PD1, and indeed patient 2 is a responder in the original study. In contrast, this module also reported some patients who are unlikely to benefit from ICB (Table 1). For example, patient 10 tumor has high TIDE score and low IFNG, MSI, and PDL1 levels. Based on the predictions from multiple biomarkers, an oncologist might predict patient 10 as a non-responder and select an alternative therapy, and indeed patient 10 failed to benefit from anti-PD-1. TIDE also showed that patient 10 tumor has a significant enrichment of T cell exclusion signature due to high infiltration of myeloid-derived suppressor cell (MDSC) and cancer-associated fibroblast (CAF). Therefore, elimination of MDSC and CAF might be needed for patient 10 to respond to ICB. In summary, by presenting the predictions from multiple biomarkers in one integrated platform, the biomarker consensus module can potentially inform oncologists on treatment decisions.

10In conclusion, we present a TIDE web platform to infer gene functions in modulating tumor immunity and evaluate biomarkers to predict ICB clinical response. Our work underlines the value of data sharing of published trials and code sharing of published biomarkers. Notably, several published ICB clinical studies have not released their omics data or clinical data (Additional file 2: Table S2), and we hope their authors could release these data to bring invaluable resource to the whole research community. As the immunotherapy data becomes increasingly available, we foresee the TIDE web platform with increased value and benefit to the mechanism studies in cancer immunology and the biomarker discoveries in immune oncology.