Mutated Processes Predict Immune Checkpoint Inhibitor Therapy Benefit in Metastatic Melanoma

DOI: https://doi.org/10.21203/rs.3.rs-1269101/v1

Abstract

Immune Checkpoint Inhibitor (ICI) therapy has revolutionized treatment for advanced melanoma; however, only a subset of patients benefit from this treatment. Despite considerable efforts, the Tumor Mutation Burden (TMB) is the only U.S. Food and Drug Administration (FDA) approved biomarker for ICI treatment in melanoma. However, the mechanisms underlying TMB association with prolonged ICI survival are not entirely understood and may depend on numerous confounding factors. To identify more interpretable ICI response biomarkers based on tumor mutations, we trained classifiers using mutations within distinct biological processes. We evaluated a variety of feature selection and classification methods, and identified key mutated biological processes that provide improved predictive capability compared to the TMB. The top mutated processes identified are leukocyte and T-cell proliferation regulation; importantly, these processes demonstrate stable predictive performance across different data cohorts of melanoma patients treated with ICI. This study provides new methodology to construct biologically interpretable genomic predictors for ICI response with substantially improved predictive performance over the TMB.

Introduction

Melanoma is a highly aggressive disease and the deadliest form of skin cancer. Deaths from melanoma account for approximately 60% of skin cancer mortality1,2. Prognosis greatly depends on the stage at which the cancer is discovered. Whereas almost all patients diagnosed with localized melanoma survive for at least five years, less than a third of patients diagnosed with distant metastasized melanoma survive over the same period3. The majority of patients with metastatic melanoma do not benefit from surgery, chemotherapy and radiation alone4,5. Targeted therapies such as BRAF and MEK inhibitors have dramatically improved prognosis of patients with metastatic melanoma that harbor specific mutations6,7. However, only a subset of the patients can benefit from these treatments, and the majority of those develop resistance over time7,8. In recent years, Immune Checkpoint Inhibitor (ICI) therapy has been approved for patients with advanced disease, demonstrating durable remission in up to half of the patients 5,7,9.

The first antibody developed for clinical ICI treatment targets the cytotoxic T-lymphocyte antigen 4 (CTLA-4). CTLA-4 is a T-cell surface protein which binds to B7-1 and B7-2 expressed by antigen-presenting cells (APC)10, resulting in suppression of immune response by the T-cells. Ipilimumab, a human monoclonal antibody targeting CTLA-4, was the first ICI agent to demonstrate increased progression free survival (PFS) and overall survival (OS) compared to more traditional cancer treatment methods10–12. Subsequently, clinical targeting of the programmed cell death receptor 1 (PD-1), which binds to its ligand receptor PD-L1 to elicit tumor immune escape, has markedly improved the treatment of melanoma and demonstrated durable responses in other types of cancer. Several potential new ICI antibodies are currently being explored, such as those targeting the regulatory surface glycoprotein TIM-313. While 40-60% of patients with advanced melanoma experience benefit from ICI, a substantial fraction of patients do not benefit from this treatment, which can incur severe autoimmune adverse events11,12,14,15. Therefore, it is critical to uncover tumor characteristics that predict response to ICI.

Numerous biomarkers have been proposed for prediction of ICI response, but most have not been validated for clinical use. Gene expression biomarkers include PDL-116, CD3817, TIM318 and CXCL919 expression, cytolytic activity20, as well as machine learning-derived signatures such as IPRES21, TIDE22, IMPRES23, Immonophenoscores24, and others25,26. However, recent meta-analysis evaluated the reproducibility of ICI biomarkers and found that only a subset of these maintained any predictive performance27. To date, gene expression signatures predicting ICI response have not been incorporated into clinical use, likely due to limited reproducibility and lack of benchmarking standards, among other factors28. Genomic biomarkers of ICI benefit have met more success in terms of clinical use. In 2017, FDA approved the first biomarker for anti-PD1 efficacy based on high levels of microsatellite instability (MSI-H)29. However, MSI-H is only found in a subset of gastrointestinal and endometrial tumors. In 2020, the high tumor mutation burden (TMB-H), quantifying the number of mutations in a tumor, has been approved by the FDA as a marker for anti-PD1 efficacy30. While TMB-H has been associated with ICI benefit across different cancer types, there are several challenges for its utility. For examples, TMB is tumor type specific; moreover, TMB-H status does not preclude tumor progression and low TMB does not preclude response6,31. In addition, the mechanism underlying the clinical utility of the TMB is unclear. Therefore, there is a need for additional genomic ICI response biomarkers with improved predictive performance that are more biologically interpretable. Recent studies have examined the mechanistic link between anti-PD1 response or resistance and mutated biological processes such as interferon signaling, MHC presentation and beta catenin32,33, prompting a need for process-level ICI response biomarkers.

Here, we use tumor mutation data in the context of biological processes to predict patient response to anti-PD1 treatment. We first investigate whether the mutation burden in genes that belong to different biological processes correlate with anti-PD1 benefit. We then apply feature selection methods to distinct processes to identify subsets of genes in which the mutational count predicts anti-PD1 response. This revealed sets of mutated genes in several biological processes with a comparable predictive ability of anti-PD1 response to the TMB. Employing non-linear classification methods further enhanced the predictive performance of classifiers based on mutated genes in specific biological processes. The advantage of these methods is that they can capture intricate relations between the mutated genes in a process and anti-PD1 responses, simultaneously weighing mutations that contribute to either response or resistance. Evaluating decision-tree algorithms and neural network architectures, we found that random forest maintains the most robust performance across different datasets, accurately predicting response and overall survival in independent datasets spanning over 500 melanoma patients in total. In particular, mutations in genes belonging to the leukocyte proliferation and T-cell regulation processes demonstrate consistently high predictive performances. This study provides a potential way forward for understanding ICI treatment responses and constructing biologically interpretable predictors of treatment benefit based on mutation data.

Results

To evaluate whether mutated genes within biological processes can predict ICI treatment responses in metastatic melanoma, we obtained training and validation mutation and clinical datasets from metastatic melanoma patients treated with anti-PD1. For all experiments, models were trained on the same designated training dataset, and evaluated using the same designated validation dataset (See methods). Throughout this work, we used Gene Ontology (GO)34,35 to aggregate genes into biological processes. We first investigated whether the mutation load in genes belonging to distinct biological processes can accurately predict ICI responses. For each GO biological process, we counted the number of mutations in that process per sample in the training datasets and used these values to predict anti-PD1 responses. These analyses revealed that the total mutation counts in distinct biological processes were only mildly predictive of response (Supp. Table 1). We surmised that only a subset of the mutated genes within specific biological process may be predictive of ICI responses. To identify subsets of genes within distinct biological processes in which the mutation count best predicts ICI response, we applied feature selection methods to mutations in each biological process.

We used the sum of mutations in selected subsets of genes within distinct biological processes to predict melanoma ICI responders vs. non-responders. The area under the receiver operating characteristic curve (ROC AUC) was used to evaluate the predictive capacity of mutations in subtests of genes belonging to each biological process. We first employed greedy forward feature selection that iteratively finds the best new feature to add to a set of selected features. In this process, the algorithm starts with an empty set, and then iterates over all genes in a biological process, to add the gene that best improves the predictive performance. When using the greedy forward selected genes within each biological process, several biological processes showed high predictive performance on the training dataset, (ROC AUC>0.75). However, none of these predictors maintained high performance in the validation dataset (that is, at least 90% of the training performance, Supp table 2). We reasoned that the greedy feature selection strategy impaired generalization by converging into local optimum. We therefore applied randomized forward feature selection, which sequentially selects features to add using a probabilistic function (see methods for details). In contrast to the greedy forward selector, four processes that performed well on the training dataset maintained high performance when applied to the validation dataset (Supp 2 and Figure 1A). These include RNA polymerase II transcription regulation, enzyme regulator activity, establishment of protein localization and regulatory regions of nucleic acid binding (Figure 1A). We next applied a genetic algorithm feature selection36–38. This method outperformed the forward selection algorithms, where selected subsets of mutated genes in 15 processes maintained high performance on the validation dataset (Figure 1A and Supp table 2). The best performing processes include immune response, leukocyte differentiation and cell motility (Figure 1A). Several genes that were frequently selected within these processes have important roles in melanoma progression and prognosis. These include CD44, shown to have an effect on tumor progression and subsequent poor prognosis39,40 and TNFSF14, a regulator of T-cell proliferation that is commonly expressed in melanomas41.

Importantly, using all three feature selection methods, the biological processes with best performance on the training dataset performed significantly better on the validation dataset compared to processes that showed poor performance on the training dataset (Fig. 1b). We found positive correlation between the performances of selected subsets of mutated genes in different biological processes across the feature selection methods (Figure 1C). Overall, these results support the premise that subsets of mutated genes within specific biological processes maintain comparable predictive performance to that of the TMB.

Using selected subsets of mutated genes, none of the best-performing processes demonstrated a substantial improvement over TMB. We reasoned that accounting for complex interactions between mutated genes in biological processes may be critical for prediction of ICI response. We therefore applied non-linear classifiers to mutated genes within each biological process. First, we trained decision tree algorithms, including gradient boosting (GB) and random forest (RF) using mutations in all sequenced genes within a biological process. The top biological processes using both methods showed a strong predictive capability across the training and validation datasets (Figure 2A). In contrast to the sum of mutation classifiers, the top decision-trees predictors substantially exceeded TMB performance for the validation dataset (Figure 2A, Supp table 3). Interestingly, leukocyte proliferation regulation and T-cell proliferation regulation were among the top biological processes, both directly linked to ICI related immune responses; checkpoint inhibitor antibodies prevent T-cell inhibition and promotes the proliferation of effector T cells42, and their response to these treatments require their proliferation and presence in the tumor microenvironment43 (Figure 2B). We investigated the mutated genes in the leukocyte proliferation regulation process with the highest contribution to the RF prediction capacity. We found that mutations in beta catenin gene CTNNB1 had the highest contribution for prediction, in agreement with recent findings that activation of this gene is associated with a reduction in T-cell antitumor response44. In addition, among the top contributing genes in that process we found IL2, a gene with known antitumor activity by increasing T-cell proliferation and previously used clinically to treat cancers5,45, and CD137, another known target for antibody mediated immunotherapy target previously tested in clinical trials46 (Figure 2C). To further investigate non-linear predictors that may capture complex interactions between mutated genes within these processes, we evaluated two classes of neural network models using mutated genes within the top processes. Both the Forward Neural Network and Long Short-Term Memory Recurrent Neural Network models demonstrated high predictive capacity when applied to mutations within these biological processes (Figure 2D, Supp table 4).

To evaluate the potential clinical utility of these predictors, we examined their performance using an additional dataset where not all genes used for training are sequenced. This dataset21 comprises mutation and response data from 38 melanoma patients treated with anti-PD1, but included only 59-68% of the genes used to train the classifiers (Supp. Table 5). Remarkably, despite this, the process mutation decision tree classifiers maintained their high predictive performance for this dataset (Figure 3A-D, Supp Table 5). To test the robustness of this approach we evaluated these classifiers when retrained using different random seeds (see methods). This analysis revealed that the performance on both unseen datasets is maintained with the random forest classifiers and is consistently better compared to TMB (Figure 3E). Notably, random forest classifiers were the most robust when presented with missing features in the test dataset21 (Supp figure 1).

To further evaluate the potential clinical utility of these classifiers, we assessed their ability to predict overall survival in an independent dataset, the Memorial Sloan Kettering Cancer Center (MSKCC) data of patients treated with anti-PD147. This MSKCC dataset includes 321 melanoma patients treated with anti-PD1; in this dataset the mutation data are limited to 468 genes in the MSK-IMPACT targeted set. Nevertheless, the RF mutated process models trained previously were significantly predictive of survival in this dataset, and in particular, the leukocyte proliferation regulation process was significant and strongly predictive (Figure 4A, Supp Fig. 2). Using the predictors based on sum of mutations and the genetic algorithm feature selection, we found that higher number of mutations in the leukocyte differentiation process was predictive of ICI response (Fig. 1A). We found that the sum of mutations in selected genes in this process was also strongly predictive of overall survival in the MSKCC dataset (Figure 4B).

We then evaluated whether the leukocyte proliferation regulation RF classifier, which obtained the best performance over all datasets, may be applicable to other cancer types. To this end, we applied it to predict overall survival for other cancer types included in the MSKCC dataset. In addition to melanoma, three cancers (colon, bladder, and renal) showed positive association between the leukocyte proliferation regulation predictor and overall survival following anti-PD1 treatment (Figure 4C). When pooling samples from these four cancer types together, the leukocyte proliferation regulation predictor demonstrated significant overall survival predictive capability (Figure 4D).

Discussion

Understanding the mechanisms underlying response and resistance to ICI therapy is critical to improving treatment of melanoma as well as other types of cancer. Through different feature selection and classification methods, we have shown that analyzing tumor mutations in the context of biological processes enhances the predictive performance of ICI response compared to existing genomic predictors. Using feature selection methods, we identified subsets of genes within distinct biological processes in which the mutation burden presents an alternative biomarker to the genome-wide TMB. To further enhance the predictive performance, we trained nonlinear classifiers using mutated genes in distinct biological processes. We reasoned that nonlinear classification methods have the potential to capture complex associations between ICI responses and mutated genes within a process. We found that using a random forest method substantially improves the predictive capability of predictors trained using mutations in specific processes, demonstrating significantly better performance compared to the TMB. Among the processes that maintain the best performance are leukocyte and T-cell proliferation regulation, known to play an important role in immune infiltration and ICI treatment. The predictive performance of these process classifiers is consistent across multiple datasets, and remain stable across varying sequencing coverage.

We investigate different methods to predict treatment benefit using mutations in the context of biological processes, which demonstrate several notable improvements over the TMB. First, the models in this work require substantially fewer genes to be sequenced for prediction. For instance, the leukocyte proliferation regulation predictor require sequencing of 99 genes, and the T-cell proliferation regulation predictor require sequencing of 73 genes. Second, developing biomarkers based on distinct biological processes improves their interpretability, and allows investigation of the mechanisms underlining their clinical utility. In particular, we found that using non-linear classifiers substantially improves the predictive capability of mutated processes, by simultaneously accounting for mutations associated with either resistance or response to treatment. The methods implemented throughout this work may be applied to construct mutated process predictors of response to other treatments in different cancer types.

More generally, we found that somatic mutations within distinct immune and signaling processes have a strong predictive performance of ICI responses in melanoma. This finding suggests that interactions between tumor genetic alterations and the microenvironment underline, at least in part, ICI responses. This could be facilitated through altered antigen presentation, supported by several HLA mutations that are frequently selected in trees within the random forest classifier (Figure 2C). Alternatively, or in complement, it is possible that mutated signaling processes modulate immune infiltration in the tumor microenvironment, supported by the selection of mutations in multiple signaling genes such as beta catenin and protein kinase and phosphatase genes (Figures 1A,2C).

We additionally found that different processes were identified when using the mutation count classifiers than those identified with nonlinear classification methods. Interestingly, the leukocyte differentiation process was selected using the genetic algorithm feature selection, whereas the leukocyte proliferation regulation was selected using the decision tree algorithm. It is possible that while mutated leukocyte differentiation process is associated with ICI response, some of the mutated genes in the leukocyte proliferation regulation process may be associated with ICI resistance. Importantly, genes belonging to the leukocyte proliferation regulation process but not in the leukocyte differentiation process include several MHC class I complex genes (HLA-A,E,G,DRB1,DRB5 and DPB1), which are known to be associated with immune evasion and ICI resistance48,49.

This study also has several potential limitations that are important to discuss. First, despite the improved predictive performance of random forest classifiers, RF and similar methods are more complex and often less interpretable for clinical use. Nevertheless, this is not the first study demonstrating that non-linear classification methods can significantly improve prediction of ICI benefit50. Incorporating clinical features to train random forest models may potentially further improve the performance obtained in this work, when data becomes available50. In addition, future developments may dissect the biological processes distinguished in this work to identify candidate targets to enhance treatment sensitivity. Second, similar to the TMB, the predictive models developed in this study account only for tumor factors and not for the tumor microenvironment. Third, it remains open to investigation whether the biological processes distinguished throughout this work for melanoma also determine ICI response in other types of cancer.

In conclusion, this study investigates mutated biological processes that predict ICI response by employing different machine learning method, and pinpoints specific processes that are highly predictive of ICI benefit in melanoma. If further investigated and validated using additional data cohorts, the predictors developed throughout this work may present a compelling alternative to the tumor mutation burden for predicting patient response to ICI therapy.

Methods

Datasets

For training, we used 144 melanoma patients’ samples from Liu et al51, including somatic mutations and anti-PD1 response information. For validation, we used 68 melanoma patients’ samples with somatic mutations and clinical data from Riaz et al52. To further test the models, we used 38 anti-PD1 treated melanoma patients’ samples from Hugo et al21. For all datasets, responders were defined as patients with complete or partial response. We additionally utilized targeted mutation data and overall survival data from the MSKCC cohort47, including melanoma, colorectal, bladder, renal, lung, esophagus, glioma and head and neck cancers.

Feature Selection for Biological Processes Mutation Load Predictors

We applied three feature selection methods to mutations in genes belonging to each biological process, to select a subset of genes that best predict ICI response. To this end, the predictive performance is defined to be the resulting ROC AUC when using the number of mutations in selected genes in a process as scores, and the ICI response as labels. The following feature selection methods were applied to the training dataset:


1. Greedy Forward Selector. The greedy forward selection algorithm iteratively select genes within a process that improves the predictive performance. The algorithm starts with an empty list of genes, and at each step, it adds to that list the gene (in a specific biological process) that results in the highest performance when added. For each biological process, we ran a maximum of 10 iterations, where the stopping criterial was when 10 iterations were completed, or when none of the genes in a process improved the performance when added.

2. Probabilistic Forward Selector. The probabilistic forward selector algorithm is similar to the greedy forward selector, except that the selection of the gene to add in each step is randomized over a set of possible genes. We defined a probability to add a gene that improves the performance when added to be 

3. Genetic Algorithm. The following steps of the Genetic Algorithm were applied to each biological process (a) Initialization of a population of size 20, where approximately 10% of the genes in the biological process were randomly selected for each instance in the initial population. (b) Evaluation of each instance in the population, where mutations in each gene set in the population were summed to predict ICI response. (c) The top half of the instances in the population, that is, those with the best predictive performance, were selected for reproduction, with randomly selected pairing. (d) Crossover was applied to the randomly selected pairs, until a population size of 20 was reached. 10 iterations of steps (b)–(d) were repeated, and the best solution was retained, corresponding to the sets of mutated genes that yielded the best performance predicting ICI response.


Decision Tree Predictors for Mutations Within Different Biological Processes

We trained decision trees to predict ICI response using the training dataset, where the classification scores obtained with these predictors were used to predict ICI response. The following algorithm were considered:


1. Random forest. Random Forest generates multiple decision trees from subsets of features of the data, which are ensembled into a single classifier, therefore reducing the risk of overfitting for large decision trees. We used RandomForestClassifier method from the sklearn.ensemble package, with 100 estimators, a max depth of 5 and a minimum sample split of 2. Other parameters were defined to default.

2. Gradient Boosting. Gradient uses boosting to integrate relatively shallow decision trees and ensemble a set of weak learners into a single strong learner. We used GradientBoostingClassifier method from the sklearn.ensemble package, with 100 estimators, a max depth of 2, a learning rate of 0.1, and the deviance loss function. All other parameters defined to default.


For reproducibility, the random state was set 100 throughout this work, except for the robustness analysis.

When testing on datasets with missing values (where some of the genes were not sequenced) the decision tree classifiers were retrained on the training dataset with the original random seed, for the subset of genes present in the new data.

Neural Network Predictors for Mutations within different Biological Processes

We additionally trained two neural network architectures to predict ICI response, where the resulting classification scores were used for prediction. These include:

1. Feed Forward Neural Network, using one fully connected hidden layer with 5 hidden units and sigmoid activation.

2. Long Short-Term Memory (LSTM) recurrent neural networks, using one LSTM cell with five hidden units.

All neural networks were trained with tensorflow.keras, using Adam optimizer, with 100 epochs and a batch size of 27.

Robustness Analysis

To evaluate the robustness of different methods, we retrained the classifiers using the mutations within the selected processes and evaluated the performance of 50 retrained classifiers for each selected process.

Survival Analysis

Survival analysis was performed using the proportional hazards, using python lifelines.statistics package. Either the sum of mutations per process (genetic algorithm and forward feature selection) or the classification scores (decision trees and neural networks) were for prediction. We evaluated all results when controlling for age and sex as confounders and stratified for different cancer types in analyses aggregating patients with different cancer types.

Code availability

The code to implement and reproduce all analyses presented in this work is provided in:

 https://github.com/AuslanderLab/Mutated_pathway_ICI_prediction 

References

  1. Cancer Facts & Figures 2021. 72 (1930).
  2. Street, W. Cancer Facts & Figures 2017. 76 (1930).
  3. Melanoma Survival Rates | Melanoma Survival Statistics. https://www.cancer.org/cancer/melanoma-skin-cancer/detection-diagnosis-staging/survival-rates-for-melanoma-skin-cancer-by-stage.html.
  4. BHATIA, S., TYKODI, S. S. & THOMPSON, J. A. Treatment of Metastatic Melanoma: An Overview. Oncol. Williston Park N 23, 488–496 (2009).
  5. Domingues, B., Lopes, J. M., Soares, P. & Pópulo, H. Melanoma treatment in review. ImmunoTargets Ther. 7, 35–49 (2018).
  6. Jardim, D. L., Goodman, A., de Melo Gagliato, D. & Kurzrock, R. The Challenges of Tumor Mutational Burden as an Immunotherapy Biomarker. Cancer Cell 39, 154–173 (2021).
  7. Sharma, P. & Allison, J. P. The future of immune checkpoint therapy. Science 348, 56–61 (2015).
  8. Villanueva, J. et al. Acquired Resistance to BRAF Inhibitors Mediated by a RAF Kinase Switch in Melanoma Can Be Overcome by Cotargeting MEK and IGF-1R/PI3K. Cancer Cell 18, 683–695 (2010).
  9. Larkin, J. et al. Combined Nivolumab and Ipilimumab or Monotherapy in Untreated Melanoma. N. Engl. J. Med. 373, 23–34 (2015).
  10. Gide, T. N., Wilmott, J. S., Scolyer, R. A. & Long, G. V. Primary and Acquired Resistance to Immune Checkpoint Inhibitors in Metastatic Melanoma. Clin. Cancer Res. 24, 1260–1270 (2018).
  11. Hodi, F. S. et al. Improved survival with ipilimumab in patients with metastatic melanoma. N. Engl. J. Med. 363, 711–723 (2010).
  12. Robert, C. et al. Ipilimumab plus dacarbazine for previously untreated metastatic melanoma. N. Engl. J. Med. 364, 2517–2526 (2011).
  13. Friedlaender, A., Addeo, A. & Banna, G. New emerging targets in cancer immunotherapy: the role of TIM3. ESMO Open 4, e000497 (2019).
  14. Schadendorf, D. et al. Pooled Analysis of Long-Term Survival Data From Phase II and Phase III Trials of Ipilimumab in Unresectable or Metastatic Melanoma. J. Clin. Oncol. Off. J. Am. Soc. Clin. Oncol. 33, 1889–1894 (2015).
  15. Wolchok, J. D. et al. Nivolumab plus ipilimumab in advanced melanoma. N. Engl. J. Med. 369, 122–133 (2013).
  16. Gibney, G. T., Weiner, L. M. & Atkins, M. B. Predictive biomarkers for checkpoint inhibitor-based immunotherapy. Lancet Oncol. 17, e542–e551 (2016).
  17. Chen, L. et al. CD38-mediated immunosuppression as a mechanism of tumor cell escape from PD-1/PD-L1 blockade. Cancer Discov. 8, 1156–1175 (2018).
  18. Holderried, T. A. W. et al. Molecular and immune correlates of TIM-3 (HAVCR2) and galectin 9 (LGALS9) mRNA expression and DNA methylation in melanoma. Clin. Epigenetics 11, 161 (2019).
  19. House, I. G. et al. Macrophage-Derived CXCL9 and CXCL10 Are Required for Antitumor Immune Responses Following Immune Checkpoint Blockade. Clin. Cancer Res. Off. J. Am. Assoc. Cancer Res. 26, 487–504 (2020).
  20. Rooney, M. S., Shukla, S. A., Wu, C. J., Getz, G. & Hacohen, N. Molecular and genetic properties of tumors associated with local immune cytolytic activity. Cell 160, 48–61 (2015).
  21. Hugo, W. et al. Genomic and Transcriptomic Features of Response to Anti-PD-1 Therapy in Metastatic Melanoma. Cell 165, 35–44 (2016).
  22. Jiang, P. et al. Signatures of T cell dysfunction and exclusion predict cancer immunotherapy response. Nat. Med. 24, 1550–1558 (2018).
  23. Auslander, N. et al. Robust prediction of response to immune checkpoint blockade therapy in metastatic melanoma. Nat. Med. 24, 1545–1549 (2018).
  24. Charoentong, P. et al. Pan-cancer Immunogenomic Analyses Reveal Genotype-Immunophenotype Relationships and Predictors of Response to Checkpoint Blockade. Cell Rep. 18, 248–262 (2017).
  25. Pérez-Guijarro, E. et al. Multimodel preclinical platform predicts clinical response of melanoma to immunotherapy. Nat. Med. 26, 781–791 (2020).
  26. Du, K. et al. Pathway signatures derived from on-treatment tumor specimens predict response to anti-PD1 blockade in metastatic melanoma. Nat. Commun. 12, 6023 (2021).
  27. Litchfield, K. et al. Meta-analysis of tumor- and T cell-intrinsic mechanisms of sensitization to checkpoint inhibition. Cell 184, 596-614.e14 (2021).
  28. Byron, S. A., Van Keuren-Jensen, K. R., Engelthaler, D. M., Carpten, J. D. & Craig, D. W. Translating RNA sequencing into clinical diagnostics: opportunities and challenges. Nat. Rev. Genet. 17, 257–271 (2016).
  29. Research, C. for D. E. and. FDA grants accelerated approval to pembrolizumab for first tissue/site agnostic indication. FDA (2019).
  30. Research, C. for D. E. and. FDA approves pembrolizumab for adults and children with TMB-H solid tumors. FDA (2020).
  31. Xuan, J., Yu, Y., Qing, T., Guo, L. & Shi, L. Next-generation sequencing in the clinic: Promises and challenges. Cancer Lett. 340, 284–295 (2013).
  32. Galluzzi, L., Spranger, S., Fuchs, E. & López-Soto, A. WNT Signaling in Cancer Immunosurveillance. Trends Cell Biol. 29, 44–65 (2019).
  33. Paschen, A., Melero, I. & Ribas, A. Central Role of the Antigen-Presentation and Interferon-γ Pathways in Resistance to Immune Checkpoint Blockade. Annu. Rev. Cancer Biol. <bvertical-align:super;>6</bvertical-align:super;>, null (2022).
  34. The Gene Ontology Consortium. Gene Ontology Consortium: going forward. Nucleic Acids Res. 43, D1049–D1056 (2015).
  35. The Gene Ontology Consortium. The Gene Ontology Resource: 20 years and still GOing strong. Nucleic Acids Res. 47, D330–D338 (2019).
  36. Tan, F., Fu, X., Zhang, Y. & Bourgeois, A. G. A genetic algorithm-based method for feature subset selection. Soft Comput. 12, 111–120 (2008).
  37. Wang, L., Wang, Y. & Chang, Q. Feature selection methods for big data bioinformatics: A survey from the search perspective. Methods 111, 21–31 (2016).
  38. Jagdhuber, R., Lang, M., Stenzl, A., Neuhaus, J. & Rahnenführer, J. Cost-Constrained feature selection in binary classification: adaptations for greedy forward selection and genetic algorithms. BMC Bioinformatics 21, 26 (2020).
  39. Wu, R.-L. et al. Hyaluronic acid-CD44 interactions promote BMP4/7-dependent Id1/3 expression in melanoma cells. Sci. Rep. 8, 14913 (2018).
  40. Dietrich, A., Tanczos, E., Vanscheidt, W., Schöpf, E. & Simon, J. C. High CD44 surface expression on primary tumours of malignant melanoma correlates with increased metastatic risk and reduced survival. Eur. J. Cancer Oxf. Engl. 1990 33, 926–930 (1997).
  41. Mortarini, R. et al. Constitutive expression and costimulatory function of LIGHT/TNFSF14 on human melanoma cells and melanoma-derived microvesicles. Cancer Res. 65, 3428–3436 (2005).
  42. Darvin, P., Toor, S. M., Sasidharan Nair, V. & Elkord, E. Immune checkpoint inhibitors: recent progress and potential biomarkers. Exp. Mol. Med. 50, 1–11 (2018).
  43. Jenkins, R. W., Barbie, D. A. & Flaherty, K. T. Mechanisms of resistance to immune checkpoint inhibitors. Br. J. Cancer 118, 9–16 (2018).
  44. Spranger, S., Bao, R. & Gajewski, T. F. Melanoma-intrinsic β-catenin signalling prevents anti-tumour immunity. Nature 523, 231–235 (2015).
  45. Agarwala, S. S. Current systemic therapy for metastatic melanoma. Expert Rev. Anticancer Ther. 9, 587–595 (2009).
  46. Yonezawa, A., Dutt, S., Chester, C., Kim, J. & Kohrt, H. E. Boosting Cancer Immunotherapy with Anti-CD137 Antibody Therapy. Clin. Cancer Res. 21, 3113–3120 (2015).
  47. Samstein, R. M. et al. Tumor mutational load predicts survival after immunotherapy across multiple cancer types. Nat. Genet. 51, 202–206 (2019).
  48. Dhatchinamoorthy, K., Colbert, J. D. & Rock, K. L. Cancer Immune Evasion Through Loss of MHC Class I Antigen Presentation. Front. Immunol. 12, 469 (2021).
  49. Lee, J. H. et al. Transcriptional downregulation of MHC class I and melanoma de- differentiation in resistance to PD-1 inhibition. Nat. Commun. 11, 1897 (2020).
  50. Chowell, D. et al. Improved prediction of immune checkpoint blockade efficacy across multiple cancer types. Nat. Biotechnol. 1–8 (2021) doi:10.1038/s41587-021-01070-8.
  51. Liu, D. et al. Integrative molecular and clinical modeling of clinical outcomes to PD1 blockade in patients with metastatic melanoma. Nat. Med. 25, 1916–1927 (2019).
  52. Riaz, N. et al. Tumor and Microenvironment Evolution during Immunotherapy with Nivolumab. Cell 171, 934-949.e16 (2017).