Background
Advanced melanoma is the most deadly skin cancer, with a total of 91,279 new cases, and 9320 deaths, expected in 2018 in the United States alone [
1]. While early-detected melanoma is mostly curable [
2,
3], advanced metastatic melanoma is life-risking. Over the past 10 years, increased biological understanding and access to innovative therapeutic modalities have transformed advanced melanoma into a new oncological model for treating solid cancers [
4]. In particular, immune checkpoint blockers (ICB) have shown a major success in the treatment of advanced melanoma [
5,
6]. The monoclonal antibody ipilimumab, blocking the cytotoxic T-lymphocyte antigen 4 (CTLA-4), was the first checkpoint blocker approved for the treatment of advanced melanoma, since it shows an objective response rate of 6–11% [
7,
8]. The approval was then followed by the one of pembrolizumab and nivolumab—two monoclonal antibody drugs which block the programmed cell death 1 (PD-1) receptor, and show response rates of 30–40% [
9,
10]. More recently, a highly toxic combination of ipilimumab and nivolumab was also approved for the treatment of advanced melanoma, with a resulting response rate of about 60% [
11,
12]. But in spite of the relatively high response rate of this treatment combination, PD-1 monotherapy, such as the one with pembrolizumab, still remains a pivotal treatment for patients with advanced melanoma, due to its relatively low toxicity and acceptable response rate. Moreover, results of the phase Ib KEYNOTE-001 trial show that a high proportion of patients with metastatic melanoma, who had achieved complete response on pembrolizumab, maintained their complete response for prolonged durations after treatment discontinuation [
13]. As ICBs become widely available, the ability to forecast duration of individual response can be critical. How to predict the patient’s response, and adjust treatment plans accordingly, is a big challenge in the current immunotherapy practice [
14].
Response rates would be improved and many treatment complications would be prevented if one could identify good responders already before therapy. Indeed, several biomarkers for response to pembrolizumab have been analyzed and the expression of programmed death-ligand 1 (PD-L1) on tumor and immune cells emerged as an acceptable response predictor [
15]. Yet, the significant fraction of PD-L1-negative patients who benefit from pembrolizumab suggests that PD-L1 cannot serve as a reliable response biomarker, on its own [
16]. In another endeavor, response scales were developed, based on several clinical factors, including localization of metastases, various blood measures, age, and gender. These scoring systems enable to stratify the patient cohort according to the overall response rate and the probability to survive a year from treatment initiation [
17,
18]. In other studies, certain immune signatures on the tumor tissue [
19,
20], and blood [
21] were associated with response, as well. However, the utility of these methodologies has yet to be validated [
21].
Acknowledging the urgent need of reliable response predictors, mathematical modelers have joined the efforts to develop tools for predicting personal response to immunotherapy [
22]. For example, Kogan et al. [
23] proposed a general algorithm for personalizing prostate cancer immunotherapy during the treatment for predicting future response. To this end the authors constructed personalized mathematical models and validated their prediction accuracy retrospectively, by accruing data from a clinical trial of prostate cancer vaccine. This was done using a new methodology of iterative real-time
in-
treatment evaluation of patient-specific parameters. Another algorithm for predicting response to cancer therapy is put forward in Elishmereni et al. [
24], attacking hormonal treatment of patients with prostate cancer. Here too, the authors developed personalized mathematical models, describing the dynamic pattern of Prostate Specific Antigen. By inputting the personal clinical PSA levels during the first months of treatment, the authors created personal models, and predicted correctly the time to biochemical failure under androgen deprivation therapy in 19 out of 21 (90%) patients with hormone-sensitive prostate cancer.
In the above described algorithms, prediction is made possible only by inputting personal clinical measurements collected during the first months of therapy. While this approach may still be of significant benefit in the design of clinical trials or in the clinics [
25,
26], most physicians would prefer to forecast the patient’s response to the drug before treatment onset.
This is the primary goal set in the present work: to develop an algorithm which could be of benefit in the current clinical practice. This will be achieved, first and foremost, by predicting the patient response to therapy before its administration, and secondly, by inputting data that are routinely collected in the clinics, e.g., describing disease progression by the sum of diameters (SOD), as prescribed by the Response Evaluation Criteria In Solid Tumors 1.1 (RECIST 1.1). Most importantly, our goal is to generate instructive output information for the physician’s decision-making process, e.g., aligning the prediction of disease progression with its effective confirmation by computed tomography (CT) or magnetic resonance imaging (MRI).
In the core of our computational algorithm lies a mathematical mechanistic model for the interactive dynamics of the disease, the cellular immune arm and the drug. By inputting clinical and molecular measurements of the patient’s parameters before treatment, the algorithm enables to personalize the model and simulate it to predict the time to disease progression (TTP) of the individual patient under pembrolizumab. Such predictions are expected to assist the treating oncologists in planning the therapy program of the patient.
Discussion
Treatment with ICB has proven successful, as it produces a significant clinical benefit in a subset of patients. However, identification of the potentially responsive patients before treatment initiation still remains a challenge, and the availability of personal response predictors has been pointed out as an unmet clinical need [
44‐
47]. Here we showed that the personalization algorithm we developed can serve as a virtual response predictor in the clinic, along with clinical information about baseline tumor size, Breslow thickness, and the status of nodular melanoma. Taking into account the low life expectancy of untreated patients with advanced melanoma, and the involved side effects and high immunotherapy costs [
48], the ability to pre-select patients for these treatments can significantly improve the quality of life of the patients.
The personalization algorithm we developed enables predictions of the time to progression, as defined by RECIST 1.1. Nowadays, the first response assessment in the clinic takes place at around 3 months into the treatment. As many patients progress within these first 3 months [
49‐
51], the algorithm predicting the TTP before treatment can save several months of administration of an incompatible and expensive drug. For patients who benefit from the treatment, the algorithm provides information on the duration of the response.
Prediction of the type and duration of response is a unique addition of this study to the knowledge gained from previously developed biomarkers for ICB. Several markers in the tumor microenvironment and peripheral blood are associated with response to ICB in patients with malignant melanoma [
52]. However, there is no way to quantify the relationships between the biomarker levels and the expected response, as yet. For example, elevation of the baseline LDH level is associated with shorter overall survival (OS) of patients with malignant melanoma under anti-PD-1 treatments [
53]. However, the survival time of individual patients cannot be predicted by this marker. In our study, clinical disease progression was observed in all patients who had an elevated LDH level before treatment onset and more than 10% increase of the LDH level on the first CT scan (11 out of 29, 38%). In contrast, disease progression occurred in only 4 out of 18 patients who initially had elevated LDH levels, but less than 10% LDH change from baseline on the first CT scan. Therefore, the change from baseline of LDH levels can serve to predict disease progression within the first year of ICB initiation, but for many patients, the prediction does not considerably precede the detection of progression by imaging scans. Another study reports that an increase in tumor burden of less than 20% from baseline, during 3 months into treatment with pembrolizumab, is associated with longer OS of patients with advanced melanoma [
54]. However, we note the difficulty in using early increase in tumor load as a response predictor, as this increase can be detected only a while after the initiation of treatment, when patients may have already experienced disease progression. The ability to predict ICB treatment outcomes before treatment, by use of our suggested personalization algorithm, can be a significant contribution to the currently available methodologies for response evaluation.
Our results show that the Breslow thickness, the baseline tumor burden, and the status of nodular melanoma can serve as markers for TTP prediction under pembrolizumab, when integrated and processed by our personalization algorithm. We found that different values of Breslow thickness and status of nodular melanoma are associated with different rates of tumor growth. Breslow thickness has been known as a prognostic biomarker for melanoma [
55,
56], and here we show that it has a predictive power. Using the baseline tumor burden as a potential biomarker is supported by Joseph et al. [
57], who analyzed the relationships between baseline tumor burden and overall survival of 583 patients with advanced melanoma under pembrolizumab. In addition, the peripheral blood from patients with advanced melanoma has been analyzed, showing that response to pembrolizumab is associated with the ratio between the baseline tumor burden and the reinvigoration of effector CD8+ T cells [
42].
Using a small patient cohort (54 patients) for its training, our personalization algorithm yields moderately accurate predictions. We believe that by increasing the size of the training set we will significantly improve the performance of the regression and classification we employed for identification of the individual model parameters. Yet, considering the limited clinical information and the simple mathematical model implemented at the core of the algorithm, the results are encouraging.
One of the major problems in medical biomathematics is its failure to propose algorithms that can be of aid in the medical practice. Specifically, the two significant hurdles to mathematical models of cancer growth becoming clinically useful, are that in most of the models the required input information cannot be extracted in a straightforward manner from data that are routinely collected in the clinics, and that in most cases, the output information is not instructive for the physician’s decision-making process. Wishing to overcome these shortcomings, we developed our algorithm and tested it using data that are routinely collected in the clinics, namely, the sum of diameters (SOD) or sum of the longest diameters (SLD), as prescribed by the RECIST 1.1. In our case, we could increase the physical and mechanistic realism of the description of tumor growth by asking the radiologists to measure, with little additional effort, more informative tumor size parameters than SOD. But the current standards in the field involve longitudinal measurement of SOD, and as our goal commands, we wish to adjust our tools to the reality in the field, rather than developing an idealized solution.
By the same token, our discretization policy, inevitably, entails loss of information. Treating oncologists do not evaluate the patient’s disease progression status continuously, but rather, every 2–4 months, using the costly imaging technology (CT/MRI). As stated above, our goal was to generate clinically relevant output. For that it would be sufficient to align the prediction of disease progression with the time of its effective substantiation by imaging. For this reason, the resolution of TTP predictions is as coarse as the planned CT/MRI scanning schedule. Still, it would be of a significant help to the doctor to know whether the patient is expected to progress early, or will have moderately long TTP, or a very long TTP, as evaluated by RECIST1.1. The discrete categories of TTP used in this study roughly correspond to these possibilities of response duration.
As one can note, most of the recruited patients are non-progressing (censored). Our approach is to use their longitudinal lesion sizes for model training and validation, so that they have the same weight as the progressing patients in the major part of the work. We then sorted the censored patients as a separate category, checking whether the model had not falsely predicted progression for them during the follow-up period. The alternative way for taking account of censored patients is to construct the survival curves, e.g., by the Kaplan–Meier method, and to use log-rank tests or Cox regression for analysis. The latter methodology would be more suitable if we wished to compare two different populations, and to compare between individuals over the whole patient group.
Model simplicity is a prerequisite for generating a beneficial algorithm, since it requires to evaluate only a small number of personal parameters. A more complex model would entail the evaluation of a relatively large number of clinical measurements in the patients for determining the personal models. It should be borne in mind, also, that our evaluation of disease progression was not required to be more sensitive than that of RECIST 1.1, which takes into account only significant changes in tumor load. Our simple model is well suited for the estimation of similarly rough changes in disease progression.
One of the limitations of the personalization algorithm developed here is that it uses the RECIST 1.1 criteria, which include the appearance of new lesions. This option was not evaluated in our algorithm and we determined disease progression only by the change in size of the target lesions between following imaging scans. Inspecting the clinical patient data, we noted that in about 50% of those in whom new lesions were detected, treatment by pembrolizumab was continued after detection, practically implying that often clinicians do not consider the new lesion criterion as progressive disease. This finding is in line with the recent understanding that formation of new lesions under immunotherapy does not necessarily indicate actual progressive disease [
58,
59]. Indeed, in the recently developed immune-related RECIST (irRECIST) criteria, pertinent to immunotherapy, appearance of new lesions is not a criterion for progressive disease [
54]. The indicated response is then “unconfirmed progressive disease”, and validation is required in another imaging scan, at least 4 weeks later. Adaptation of our algorithm to the irRECIST criteria will be made upon clinical validation of these criteria as part of the clinical follow-up routine.
Future recommendations for improving the predictive power of our personalization algorithm include training by a larger dataset, as well as validation of the algorithm by clinical data from an independent dataset. Following improvements in the prediction accuracy, our algorithm can be used as a tool in selecting personal treatment. In addition, our innovative methodology can be adapted to other available immunotherapies, including anti-CTLA-4, anti-PD-1 combination, or other immunotherapies when becoming clinically available. Taken together, this study demonstrates that using computational algorithms for predicting the response to immunotherapy in patients with metastatic melanoma is feasible in the clinical realm.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.