A Novel Spectroscopically Determined Pharmacodynamic Biomarker for Skin Toxicity in Cancer Patients Treated with Targeted Agents

In recent years, targeted therapy has become a standard treatment against various types of cancer such as melanoma, lung cancer, and breast cancer. Many of these therapies target the intracellular MAPK pathway, also known as the “RAS–RAF–MEK–ERK pathway,” which is abnormally activated in more than 7% of all human cancers (1). In particular, the BRAF gene coding for the serine/threonine-protein kinase BRAF is mutated in 60% to 70% of the melanomas (2). The MAPK pathway is involved in many cellular processes such as proliferation, differentiation, migration, and apoptosis (3). The use of tyrosine kinase inhibitors (TKI), for example, the inhibitors of the MEK, the EGFR, and the BRAF gene, are privileged strategies to regulate this pathway. Many TKI (4, 5) have recently received regulatory agency approval for the treatment of patients. Vemurafenib (GSK2118436), a BRAF inhibitor commercially available since 2011, has significantly improved the overall and progression-free survival of BRAF^V600E mutation-positive melanoma patients (6–8). In the case of EGFR mutation-positive non–small cell lung cancer, studies have demonstrated a significant progression-free survival benefit in patients treated with erlotinib (EGFR inhibitor), in comparison with platinum-based chemotherapy when given as first line treatment to European and Asian patients (9, 10). Combination of these treatments is possible and contributes to improve the overall survival in patients with melanoma (11).

Although the efficiency of TKI against cancer is established, it is well-known that it induces strong cutaneous side effects. Lacouture and colleagues have shown that 92% to 95% of patients treated with vemurafenib showed dermatologic adverse events (12). Eighty-seven percent of patients treated with trametinib (a MEK inhibitor) experienced cutaneous toxicity (13). The underlying mechanisms of these adverse effects on the skin have been studied by many groups (14–16) but remains not completely understood.

Raman spectroscopy is a noninvasive technique that can provide information about the molecular composition of cells and tissues (17, 18). This nondestructive vibrational spectroscopy instrumentation is based on inelastic scattering of light caused by the interaction of light and molecular vibrations. It does not require reagents or the use of contrast enhancing agents. This technique is under investigation for many applications in biology. One of the main topics is the diagnosis of cancer through the identification of the biochemical differences between healthy and cancer tissue (19–22). Thanks to easy access, the skin is one of the major targeted tissues for in vivo Raman-based studies (23, 24).

In this study, we have evaluated the ability of Raman spectroscopy to detect cutaneous toxicity in patients under TKI treatment, namely with MEK, EGFR, or BRAF inhibitors, and so to use Raman signature of the skin as a new pharmacodynamic biomarker. We also report the specific methodology that we have developed to perform and analyze this dermatological application of Raman spectroscopy.

This study was approved by the French Agency for the Safety of Health Products under ID RCB number 2010-A01051-37 and reference B101285-30. Patients were recruited from the dermatology unit and the pulmonology unit of the Department of Medical Oncology of Gustave Roussy (Villejuif, France). From January 2015 to March 2015, all patients, older than 18 years old, who were going to be treated with trametinib (MEK inhibitors), erlotinib, afatinib (EGFR inhibitors) dabrafenib, or vemurafenib (BRAF inhibitors) were invited to volunteer to this study. The volunteers signed an informed consent (SkinTarget Protocol CSET no. 2010/1664) permitting the research team to acquire in vivo skin Raman spectra, to perform skin biopsies, to have a dermatologist consultation, to take blood samples, and to take standardized pictures of the affected skin. Nine patients were included in this study, three undergoing treatment with a MEK inhibitor, three patients treated with an EGFR inhibitor, and three patients treated with a BRAF inhibitor. Patients were between 49 and 68 years old.

For each patient, data acquisition was performed at three time points, apart from one patient due to a no-show. During each session, a complete dataset was collected the same day, comprising in vivo Raman measurements of the skin, a dermatologic report, a skin biopsy, a blood sample, and standardized pictures of the face, the hands, the nails, and the feet. The Raman measurements and the skin biopsies were done at a skin location that appeared healthy from a dermatologic point of view. This location was chosen in order to avoid potential interference of the Raman measurement by a local inflammatory reaction. The first data collection session was performed before the start of the treatment. The second and the last session were planned about 2 and 4 weeks after initiation of the treatment. This measurements schedule was based on the expected time to occurrence of skin toxicity observed with these agents and coincided with scheduled routine visits of the patients to the hospital.

For this study, a commercially available gen2-SCA (Skin Composition Analyzer) from RiverD International B.V. was used (Fig. 1). This confocal Raman spectroscopy instrument has been optimized for in vivo skin measurements. It can measure the Raman spectrum in the Fingerprint (FP) region (400–1,800 cm⁻¹) using a 30 mW 785 nm continuous wave laser and the Raman spectrum in the high wavenumber (HWN) region (2,600–4,000 cm⁻¹) using a 20 mW 671 nm continuous wave laser. The laser light intensities comply with international laser safety standards (25). The laser is focused into the skin by a custom designed immersion microscope objective (100×, 1.2 numerical aperture). The depth resolution is better than 5 μm. Raman signal is collected using the same objective. An XY piezoelectric stage can translate the measurement window so as to access multiple skin locations. The objective can be Z-translated to focus the beam at different depths in the skin. The measurements were performed using the RiverICon software (RiverD International B.V.).

Before each Raman measurement session, the equipment was auto-calibrated and the power of the laser beams was checked. The measurements were performed on the lower forearm of the patients, far away from any skin lesion. Patients put their right arm on a fused silica measurement window through which the microscope objective focuses laser light in the skin. During each session, between five and nine different locations were acquired for both spectral regions (FP and HWN). At each location, Raman measurements with specific in-depth profile were performed to acquire several spectra from two skin layers: the stratum corneum (SC) and the viable epidermis (VE; ref. 26). For the FP spectral band, the Raman acquisitions were performed until 24-μm depth with a step size of 4 μm and two additional points at 32- and 40-μm of depths, which gives a total of nine spectra per in-depth profile. For the HWN spectral band, the Raman acquisitions were performed until 48 μm with a step size of 4 μm, which gives a total of 13 spectra per in-depth profile. To ensure spectra with high signal-to-noise ratio (SNR), the acquisition time was chosen in the range from 5 to 30 seconds in FP band and fixed to 1 second in HWN band. The maximum duration of the Raman data collection was 40 minutes. In total, 4,532 spectra were collected (1,854 in the FP region and 2,678 in the HWN region).

First of all, an analysis of the spectra quality on the FP dataset was performed. To do so, the signal-to-noise ratio (SNR) was calculated, with the Eq. 1, for each measured spectra in the FP band.

where S is the magnitude of the peak at 1,003 cm⁻¹, N is the total magnitude of the signal at 1,003 cm⁻¹ minus the charge-coupled device (CCD) offset, and G is the gain of the CCD detector. All the spectra with SNR < 10 were removed from the dataset (representing less than 10%). Moreover, a few outlier spectra were also removed from the dataset.

Caspers and colleagues (27) demonstrated that the water mass concentration profile and the natural moisturizing factor (NMF) profile are good indicators of the SC/VE interface. Water mass in-depth profile can also be used to determine the SC apparent thickness (28). The SkinTools software v2.0 (RiverD International B.V.) was used to extract water mass profile from the HWN Raman measurements and so to determine the SC/VE interface (29). Thus, each spectrum acquired in the FP spectral band could be attributed to the appropriate skin layer, meaning SC or VE. Then, the FP spectra were preprocessed by extended multiplicative signal correction (EMSC; ref. 30) to normalize each spectrum to a unique reference spectrum while removing the spectral variance due to fused silica and room light. Based on the dermatologic exam performed on the same day as the Raman acquisition, each recorded Raman spectrum was labeled as: “toxicity” meaning spectrum from patient suffering of skin toxicity or “no toxicity” meaning spectrum from patient not suffering of skin toxicity. For the two groups (i.e., “toxicity” and “no toxicity”), the mean spectrum for each skin layer investigated and each inhibitor was calculated.

Multivariable analysis was performed to classify the Raman signature of patient in the two groups. Multivariable analysis was based on partial least-square regression (PLS; ref. 31) coupled with linear discriminant analysis (LDA; ref. 32) in cross-validation by bootstrapping (33), coined her “PLS-DA.”

Randomly 90% of the complete dataset was selected to establish a model (training dataset), and the remaining 10% of the dataset (validation dataset) was used to test the model. To avoid any bias, the ratio in the number of spectra belonging to the two groups (i.e., “toxicity” and “no toxicity”) was preserved when splitting the dataset. PLS was performed on the mean-centered training dataset. The drug concentrations into patients' blood were used as the observable variable for the PLS analysis. The titration of the drugs is detailed in a specific paragraph and the results are reported in Supplementary Table S1. In the case of blood samples were missing, the corresponding spectra were removed from the PLS and subsequent LDA analysis. From the 10 first latent variables (LV) determined by PLS, LV selection was based on quantitative criterion. To be selected, the scores of the LV for the two groups (i.e., “toxicity” and “no toxicity”) had to be statistically significant (Student t test, Bonferroni adjusted P-value <0.05; ref. 34). To prevent any overfitting in the model, the maximum number of LVs selected was equal to 6. The scores on the selected LVs were used as input for the LDA analysis to determine the linear discriminant (LD). The LD established on the training dataset was applied to the mean-centered validation dataset. So LD score was calculated for each spectrum of the validation dataset. On the basis of the LD scores of the validation dataset, the discriminative power between the two groups (i.e., “toxicity” and “no toxicity”) was analyzed using receiver operating characteristic (ROC) curves (35). The area under the curve (AUC) of the ROC curve was calculated. Accuracy, sensitivity, and specificity were determined on the basis of the Youden index (36). Student t test was also performed to assess the statistical difference between the LD scores of the two groups in the validation dataset. At each iteration, ROC curve, AUC, accuracy, sensitivity, specificity, and P value were determined for the validation dataset. Moreover, LD scores of the validation were recorded. This individual cross-validation was repeated 5,000 times with different training and validation datasets. The number of iterations of bootstrapping was determined to include each spectrum of the complete dataset in the two sub-datasets (i.e., training and validation). The mean ROC curve, AUC, accuracy, sensitivity, and specificity were calculated. For each spectrum of the complete dataset, the mean LD score was calculated. The correlation between the Raman signature and the drug concentration was performed by calculating the coefficient of determination (R²) between the mean LD scores and the drug concentration.

This data process was applied to the two skin layers (SC and VE) and the three groups of patients (MEK, EGFR, and BRAF inhibitors). All data processing and data analysis were performed under MATLAB v2009b (MathWorks).

Four millimeters diameter punch biopsies were performed on clinically unaffected skin. The skin biopsies were formalin fixed and paraffin embedded. Microscopic slides with 3-μm-thick tissue sections were stained with hematoxylin, eosin, and Safran (HES) and periodic acid-Schiff (PAS). The slides were analyzed by three pathologists from two different hospitals (Erasmus MC and Gustave Roussy) in an independent way.

Blood samples were centrifuged (3,000 rpm, 2,200 × g for 10 minutes) and plasma was separated in aliquots and stored at −20°C prior to analysis.

Plasma samples of patients treated with dabrafenib (BRAF inhibitor), trametinib (MEK inhibitor), and erlotinib (EGFR inhibitor) were analyzed by ultra-performance liquid chromatography combined to tandem mass spectrometry (UPLC-MS/MS) on Acquity-Xevo TQ system with MassLynx 4.1 software (Waters). Dabrafenib, trametinib, and erlotinib were analyzed using respectively [²H₉]-dabrafenib, [¹³C₆]-trametinib, and bosutinib as internal standards (IS). Dabrafenib, trametinib, and erlotinib and their IS were separated on a Acquity UPLC BEH C18 chromatography column (2.1 mm × 50 mm I.D., 1.7 μm; Waters) using gradient elution with mobile phases of acetonitrile ± methanol, ammonium formiate, and ultrapure water at a flow rate of 0.7 mL/minute.

Vemurafenib (BRAF inhibitor) was analyzed by high-performance liquid chromatography combined to UltraViolet detection (HPLC-UV) on a HPLC-UV 1200-1290 system (Agilent) using sorafenib as IS. Vemurafenib and the IS were separated on a Kinetex PFP 4.6 × 100 mm, 2.6 μm (Phenomenex) using gradient elution with mobile phases of acetonitrile, KH₂PO₄ 20 mmol/L and ultrapure water at a flow rate of 1.0 mL/minute.

All analytic methods were validated in terms of specificity, linearity range, precision, and accuracy according to EMA guidelines. Dabrafenib and trametinib concentrations were determined between 1 and 2,000 ng/mL with intraday and interday precision coefficients of variation (CV) below 2.4%. Vemurafenib concentration was determined between 1 and 50 μg/mL with intraday and interday precision CVs below 6.9%. Erlotinib concentration was measured between 1 and 1,000 ng/mL with an intraday and interday precision CVs below 8.3%.

The results are reported in Supplementary Table S1.

In this study, none of the nine patients included suffered from skin toxicity before the beginning of the treatment. After 1 month of treatment, all of them presented skin adverse events. The Supplementary Table S2 reports the various cutaneous adverse events for each patient at the two time points.

As expected (37, 38), patients of MEK and EGFR inhibitors groups were suffering from similar cutaneous adverse events. Five of 6 patients displayed folliculitis after 1 month of treatment. Figure 2 shows example of folliculitis on the face of patient treated with MEK inhibitors. Prurituses were detected in half of them. In the BRAF group, no folliculitis or pruritus was noticed contrary to patients treated by MEK or EGFR inhibitors. As expected (39, 40), two out of three patients of BRAF inhibitors groups were suffering from grade I keratosis. Two of 3 patients of BRAF inhibitors group were suffering from grade I xerosys. It has to be noticed that 6 of 9 patients of the cohort were displaying skin toxicity after 2 weeks of treatment.

The mean normalized Raman spectra of the two groups (i.e., toxicity and no toxicity) were calculated for the three inhibitors and the two skin layers investigated (Fig. 3). Because of the thickness of the SC compared with the VE, the number of spectra acquired in the SC for the FP band was 33% lower than in the VE (439 vs. 656). As expected, the shape of the Raman spectra in the SC and the VE was different. The SC spectra showed a peak at 880 cm⁻¹, which was not present in the VE spectra. This peak is attributed to the NMF, which is only present in the SC skin layer (27).

The differential spectra (toxicity minus no toxicity) did not show clear and consistent differences. In the case of MEK and EGFR inhibitors in the SC, the differential spectral displayed a small decrease in the peak at 1,650 cm⁻¹ attributed to keratin. For the same conditions, the peak at 1,442 cm⁻¹ related to NMF displayed a small increase. The Raman signature of the drug itself could not be observed in the Raman spectra of the skin patients even after 1 month of treatment. This might be due to the residual amount of drugs in cutaneous tissue, which was below the detection limit of the equipment and/or to the metabolization of the drug ending to different chemical products with different Raman signatures. Therefore, for each skin layer and inhibitor, multivariable analysis was used to investigate the presence of systematic differences between the two groups (i.e., toxicity vs. no toxicity).

Figure 4 shows the ROC curves to quantify the discriminative power between the two groups (i.e., toxicity vs. no toxicity) for the different analysis performed. Table 1 summarizes the discrimination performances for each ROC curve by indicating the AUC, the accuracy, the specificity, and the sensitivity.

For patients treated with MEK inhibitors, the AUC of the ROC curve was 90% when investigating the VE skin layer (Fig. 4A). Accuracy, sensitivity, and specificity were around 90% (Table 1), indicating a good discriminative power between the two groups. In comparison, the AUC of the ROC curve for the SC was much lower (70% vs. 90%). Although the discriminative power was lower in the SC than in the VE, the LD scores between the two groups (i.e., toxicity vs. no toxicity) were significantly statistically different with a P value lower than 0.01% (Supplementary Fig. S1A and S1D).

For patients treated with EGFR inhibitors, the AUCs of the ROC curve in both skin layers were higher than 95%. The AUC of the ROC curve was slightly lower in the SC than in the VE (96% vs. 97%). Accuracy, sensitivity, and specificity in the two skin layers were more than 95%. The differences in the LD scores between the two groups were statistically different with a P value lower than 0.01% (Supplementary Fig. S1B and S1E). It can be noticed that P value was lower in the VE than in the SC (1.51 × 10⁻³¹ vs. 4.3 × 10⁻¹⁷). It confirms that the discriminative power was slightly better in the VE than in the SC for patients treated with EGFR inhibitors.

For patients treated with BRAF inhibitors, the AUC of the ROC curve in the VE was 71%. In comparison, the AUC of the ROC curve in the SC was lower (65% vs. 71%). Although the discriminative powers remained low, the LD scores between the two groups were statistically significantly different with a P value lower than 0.01% (Supplementary Fig. S1C and S1F). PLS data analysis implies that the variance of the dataset might be linearly related to a variable, in our case the drug concentration into patient's blood. Because of the low discriminative power of the PLS data analysis in the case of the BRAF inhibitors group (AUC = 65%), it was reasonable to consider that the previous hypothesis was not correct and was limiting the discriminative performance. To try to improve the discriminative power of our model, we have performed the same data process but PLS was replaced by principal component analysis (PCA), which is an unsupervised multivariable data analysis technique. In this case, the discriminative result was 12% better in the SC in comparison with the PLS-DA data process (77% vs. 65%).

We correlated the Raman signature of patients' skin with the titration of the drug into patients' blood (Fig. 5). As a confirmation of the previous results, the best correlations were obtained for the EGFR inhibitors. In this case, the coefficients of determination were around 80% for the two skin layers investigated. The coefficient of determination was slightly higher in the VE than in the SC (81.5% vs. 79.6%), confirming the previous discriminative performance results. For MEK inhibitors in the VE, the coefficient of determination was 72.7%. In the case of BRAF inhibitors, the coefficients of determination were below 55% in the two skin layers investigated (Fig. 5C and F).

Histopathologic analysis of the microscopic slides of skin biopsies did not reveal any significant tissue disarray after 1 month of treatment (Fig. 6). In some cases, small variations in the SC with hyperkeratosis were noticed. In two biopsies, focal slight epidermal hyperplasia and slight dysplasia were present without any correlation with the treatment. No inflammation was noticed.

The aim of this pilot study was to investigate the ability of Raman spectroscopy to detect skin toxicity in patients under treatment by TKI. Multivariable statistical analysis technique (PLS-DA) was tested to discriminate Raman signatures of skin from patient suffering of cutaneous toxicity and Raman signatures of skin from patient not suffering of cutaneous toxicity. The discriminative performances, evaluated by PLS-DA and cross-validation by bootstrapping, were found to be dependent on the treatment and the skin layers investigated. In general, the algorithm allowed better discriminative results for MEK and EGFR inhibitors than for patients treated with BRAF inhibitors. This might be related to the different mechanisms of these drugs on healthy tissue. It is well established that BRAF inhibitors activate the MAPK pathway in normal (wild-type BRAF) cells (41). This activation induces a dimerization of the RAF kinase inducing a phosphorylation of MEK and ERK kinase (42). In contrary, MEK and EGFR inhibitors inhibit the MAPK pathway in the wild-type cells (13, 38). Although, patients under MEK and EGFR inhibitors exhibit similar spectrum of skin adverse events with papulopustular rash, dry skin, perionyxis, and hair changes. Patients treated with BRAF inhibitors were suffering from follicular keratosis, palm and sole hyperkeratosis, eruptive papillomas, squamous cell carcinomas, or melanomas. We might speculate that Raman spectroscopy is more sensitive to chemical modifications of the skin induced by the inhibition of the MAPK pathway than its activation. It has to be noticed that even with a small number of patients, the discrimination results reached high level of accuracy, sensitivity, and specificity. A larger number of patients would contribute to enhance those differences. Moreover, Raman signatures were acquired at skin locations that were not-affected from a dermatologic point-of-view. Skin biopsies were also performed on healthy looking tissue and did not show modifications of the skin related to the treatment. Thus, Raman spectroscopy has access to information that is not detectable at the dermatologic and histologic levels. Similar results have been obtained by Le Naour and colleagues (43) when targeting liver steatosis by vibrational spectroscopy on tissue sections from histologically normal areas from an otherwise steatotic liver. Many studies have demonstrated the possibility of Raman spectroscopy to accurately discriminate cutaneous cancer tissue from normal tissue (44, 45).

For EGFR and MEK inhibitors, the best discriminative results were obtained in VE skin layer. In both groups, the AUC of the ROC curves for VE skin layer were more than 90%. These results are supported by dermatologic knowledge indicating that EGFR is mainly expressed in undifferentiated keratinocytes in the basal and suprabasal layers of the epidermis (46). Moreover, constitutive activation of upstream kinase MEK1 perturbed the differentiation of human keratinocytes located in the basal and suprabasal layers of the epidermis (47). This study points out that VE is the skin layer primarily impacted biochemically by MEK and EGFR inhibitors. The best discriminative result was obtained for the EGFR inhibitors in the VE skin layer with an accuracy of 96%, a sensitivity of 96% and a specificity of 95%. Even with a small number of patients, the correlation of the skin Raman signature and the drugs concentration was over 81%.

Our results show that Raman signature of patients dermatologically and histologically normal patients' skin can be used as a pharmacodynamic biomarker for cutaneous adverse events toxicity due to TKI treatment. The opportunity to predict skin toxicity via Raman spectroscopy, a noninvasive and label-free optical instrumentation, would be a great contribution to improve the life quality of patients under TKI treatment. Indeed, it could open the possibility to modify the treatment before the detection of visible and quality-of-life impairing skin toxicity. Moreover, because meta-analysis studies have demonstrated that skin toxicity is correlated with the response to the anti-EGFR treatment in non–small lung cancer (48, 49), investigating the correlation between Raman spectroscopy of the skin and anti-EGFR treatment efficacy would be interesting in the perspective of monitoring the treatment efficacy in a noninvasive way. Further investigations are thus necessary.

This pilot study is the first one to investigate the skin toxicity induced by new anticancer-targeted therapies by the means of Raman spectroscopy. Our results show that Raman spectroscopy is able to discriminate spectra from patients suffering of skin toxicity and patients not suffering of skin toxicity. By correlating the skin patients Raman signature and the drugs concentration into patient's blood, we can conclude that Raman spectroscopy can be used as a pharmacodynamic biomarker for skin toxicity induced by TKI treatment.

D. Planchard is a consultant/advisory board member for Astrazeneca, Boehringer, Roche, BMS, MSD, Pfizer, and Novartis. S. Koljenović, V.N. Hegt, and G.J. Puppels have ownership interest (including patents) in RiverD International B.V. C. Robert is a consultant/advisory board member for Roche, BMS, MSD, Novartis, and Amgen. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A. Azan, P.J. Caspers, S. Koljenović, V. Noordhoek Hegt, A.M.M. Eggermont, C. Robert, G.J. Puppels, L.M. Mir

Development of methodology: A. Azan, P.J. Caspers, T.C. Bakker Schut, B. Besse, A. Seck, S. Koljenović, A. Paci, G.J. Puppels, L.M. Mir

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Azan, C. Boutros, C. Mateus, E. Routier, B. Besse, D. Planchard, A. Seck, N. Kamsu Kom, G. Tomasic, S. Koljenović, V. Noordhoek Hegt, M. Texier, A. Paci

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Azan, P.J. Caspers, T.C. Bakker Schut, D. Planchard, S. Koljenović, V. Noordhoek Hegt, E. Lanoy, A. Paci, C. Robert, G.J. Puppels, L.M. Mir

Writing, review, and/or revision of the manuscript: A. Azan, P.J. Caspers, T.C. Bakker Schut, C. Boutros, D. Planchard, S. Koljenović, V. Noordhoek Hegt, E. Lanoy, A.M.M. Eggermont, C. Robert, G.J. Puppels

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Azan, A. Seck, G. Tomasic, S. Koljenović

Study supervision: C. Robert, G.J. Puppels, L.M. Mir

The authors thank the patients involved in this study, Cathy Philippe, Arthur Tenenhaus and Jane Merlevede for fruitful advice on the data process, Martin Van der Wolf and Kevin Stouten from River D International for technical support on the equipment, Bruno Thuillier from the dermatology unit of Gustave Roussy for logistic support, and Marie Breton from UMR8203 and Aliette Ventéjoux from EA 4398–PRISMES for carefully reading this article.

This study was financially supported by Gustave Roussy.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.