Introduction

Ovarian cancer is the 8th most commonly occurring cancer in women and the 18th most common cancer in overall. Moreover, this kind of cancer take 5th place as the cause of death in women1. A major challenge in treating ovarian cancer is the time of diagnosis, and the possibility of therapeutic strategy, especially if around 66% ovarian cancer women present with stage III or IV disease at initial diagnosis.

Primary optimal cytoreduction is the first – line treatment for patients (if candidates can be treated surgically). In the following stage, patients should receive chemotherapy included 6 cycles of platinum-based combination chemotherapy, with platinum and taxane. In the case of patients with cytologically high-grade disease who are not good candidates for surgery, first 3–4 cycles of neoadjuvant chemotherapy (NACT) may be given and next surgery caused reducing tumor mass can be done. At the and additional chemotherapy was used2,3.

In clinical practice, the biggest problem is the recurrent ovarian cancer observed among patients with high-grade ovarian cancer, even after the first session of therapy. These patients can be subdivided into platinum-sensitive or platinum-resistance groups. If the time to recurrent is more than 6 months after the end of the previous platinum-based treatment (platinum free-interval PFI), patients are defined as platinum-sensitive and considered eligible for further platinum-based chemotherapy with or without anti-angiogenic therapy or polymerase inhibitors4. If the time to recurrent is from 1 to 6 months, patients are classified as platinum-resistance, and deemed not eligible for platinum-based chemotherapy5.

Currently, there is no optimal treatment strategy for platinum-resistance patients. Therefore, it is reasonable to search for a biomarker in patients with advanced ovarian cancer that will allow the selection of a group of platinum-resistance patients before starting treatment in order to save time and implement therapy without platinum-based chemotherapy. For this purpose, new fast and reliable diagnostic methods are under development enabling to find platinum-resistance markers. Fourier Transform Inferred spectroscopy (FTIR) is one of such diagnostic tools. This method allows developing the chemical characterization of samples, as well as obtaining chemical fingerprints6. It is known that each substance has a different chemical composition or a different ratio between respective compounds7. The same situation can be found in serum, tissues, where the same functional groups were visible, but different amounts of presented functional groups were noticed8,9. Consequently, using FTIR it is possible, for instance, to differentiate between healthy and un-healthy patients10,11,12. However, until now nobody used FTIR to find differences between platinum-resistance and platinum-sensitive patients. Therefore, the aim of this unprecedented study is to determine chemical differences between platinum-sensitive and platinum-resistance women suffering from ovarian cancer using FTIR spectroscopy in combination with multivariate analyses and machine learning methods. For this purpose, the serum from these two groups of women was collected and studied. Following the FTIR spectroscopy analysis, the data was analyzed by means of PCA, HCA and 4 differences machine learning methods. To show the most significant differences (specific wavenumbers), C5.0 decision trees were calculated.

Results and discussion

It is a lot of causes of platinum resistance. The most significant include genetic, epigenetic immunological and environmental factors, often involving more than one resistance mechanism13. Consequently, testing whether a cancer is platinum-resistance using molecular methods is a long and very expensive process. Therefore, in this study we showed for the first time, a cheap and fast method - FTIR spectroscopy in combination with multivariate and machine learning analyses, which allows distinguishing chemical changes caused by platinum-resistance ovarian cancer in serum in comparison with serum collected from women suffering from platinum-sensitive ovarian cancer. The obtained FTIR spectra (Fig. 1a) of serum collected from two groups of women suffering from ovarian cancer allowed to observe peaks corresponding to phosphorate groups from DNA, phospholipids, amides vibrations from proteins, and C-H vibrations from lipids. Stretching symmetric and asymmetric PO2-vibrations were placed at around 1090 cm-1and 1260 cm-1, respectively14. Proteins vibrations visible at amides III, II, I and amide A were positioned at around 1320 cm-1, 1540 cm-1, 1640 cm-1 and 3290 cm-1 wavenumbers, respectively15,16,17. Finally, the lipids vibrations of C=O and CH2, and CH3 functional groups were visible around 1750 cm-1 and in the range between 2800 cm-1and 3000 cm18,19. Moreover, CH3 as well as C=O vibrations from COO- building proteins and lipids fraction were located around 1455 cm-1and 1410 cm-1 (Fig. 1a)19. In Table 1a position of peaks marked in Fig. 1a with vibrations description was placed.

Fig. 1
figure 1

Average of FTIR spectra ± SD of serum collected from platinum-sensitive (black curve) and platinum-resistance (pink color) women suffering from ovarian cancer (a); PCA (b1, c1) and loading plots (b2, c2) in training datasets: 800–1800 cm−1 range (b) and 2700–3000 cm−1 (c).

Full size image

The higher absorbance level of all analyzed peaks was observed on the FTIR spectrum of serum collected for platinum-sensitive patients in comparison with serum collected from platinum-resistance women. Moreover, as could be observed from Table 1, significant shifts of both PO2− phospholipids and DNA, amide II vibrations toward higher wavenumber and shift of amide III toward lower wavenumber in platinum-resistance women in comparison with platinum-sensitive was noticed. Furthermore, in platinum-resistance women the absence of asymmetric stretching vibrations of the CH2 group from lipids was noticed. PCA analysis showed that observed changes in the FTIR spectra were so significant and noticeable in both selective ranges that it was possible to distinguish the two groups of studied samples (the serum taken from platinum-sensitive women and the serum from platinum-resistance patients), Figs. 1b1, c1. Loadings plots showed, that amides II and amide I in the FTIR range between 800 cm−1 and 1800 cm−1 (Fig. 1b2) were responsible for the distinction. While in the lipids range the differentiation could be noticeable in asymmetric stretching CH3 and symmetric stretching CH2vibrations (Fig. 1c2). It is well known, that copper transporter 2 (CTR-2) is also involved in regulating cellular platinum levels and that increasing of CTR-2 expression was correlated with platinum resistance20,21. Moreover, the mechanism of action of cisplatin is closely related to DNA22, where methylation23, histone modifications24 play important roles in cisplatin treatment of ovarian cancer. Consequently, changes in the DNA and proteins, which were observed as a shifts of both PO2−phospholipids and DNA, amide II vibrations can be responsible for platinum resistance. Moreover, ovarian cancer was developed in a unique tumor microenvironment that plays a key role in the natural history of the disease25. In the platinum resistance tumor microenvironment changes in the fibroblasts, adipocytes, mesothelial cells and five types of immune cells, and tumor infiltrating lymphocytes, as well as components of the extracellular matrix can be observed26,27. However, the number of variables is very large and a lot of time would be required to indicate which of the changes has the greatest impact on the phenomenon of platinum resistance using FTIR spectroscopy. Therefore, machine learning methods was used to which wavenumbers played the most significant role in the differentiation between platinum-resistance and platinum-sensitive ovarian cancer women.

Table 1 Average positions of the analyzed FTIR peaks of platinum-resistance and platinum-sensitive patients, differences between positions and band assignments, where “*” means shift higher than 4 cm−1 (value of spectral resolution).
Full size table

Furthermore, HCA analysis showed that the samples from platinum-resistance women were more homogeneity that serum collected from platinum-sensitive women (Fig. 2). In the case platinum-resistance patients in range between 800 cm−1 and 1800 cm−1, 20 measurements from 26 (77%) were connected with one homogeneity group (Fig. 2a). In the case of range from 2700 cm−1 to 3000 cm−1 19 samples from 26 (73%) created one similarity group (Fıg. 2b). In the case of platinum-sensitive women higher homogeneity was noticed for fingerprint FTIR range (Fig. 2a) than for lipids one (Fig. 2b). In the first range, 72% of samples created one similarity group, while in the second analyzed ranges – 69%.

Fig. 2
figure 2

HCA analysis of blood samples collected from platinum-sensitive (black color) and platinum-resistance (pink color) women suffering from ovarian cancer in the range 800 cm−1 – 1800 cm−1 (a) and 2700 cm−1 − 3000 cm−1 (b).

Full size image

Consequently, to evaluate the significance of observed changes in the FTIR spectra which allow marking chemical changes responsible for platinum-resistance occurrence, machine learning methods were employed.

Table 2 shows the results obtained for four datasets using ML algorithms: the original datasets for the 800 cm−1 − 1800 cm−1 and 2800 cm−1 − 3000 cm−1 ranges, as well as for the datasets after selection of relevant wavenumbers. As can be seen, the results of the quality of classification parameters (accuracy, sensitivity and specificity) were almost perfect. For accuracy, the value was ranging from 0.98 to 1.00 for 800–1800 cm−1 region and from 0.92 to 0.98 for 2800 cm−1 – 3000 cm−1. The sensitivity was from 0.88 to 1.00 for both analyzed FTIR ranges, while value of specificity was in almost all cases around 0.95.

Table 2 FTIR mean values of the quality assessment parameters of the four obtained learning algorithms, using 4 datasets.
Full size table

Importance analysis of individual features (wavenumbers) (Fig. 3) obtained using the random forest and Boruta algorithms, for the categories of platinum-sensitive and platinum-resistant cases shows that about 5 relevant wavenumber ranges can be identified, which clearly distinguish both categories. These ranges were presented in Table 3.

Fig. 3
figure 3

Mean importance values of individual wavenumbers obtained with the feature selection algorithm during analysis of the FTIR dataset 800 cm−1 − 1800 cm−1 wavenumbers (a) and 2700 cm−1 − 3000 cm−1 wavenumbers (b).

Full size image

For the 800 cm−1 − 1800 cm−1 range, a strong difference in absorption can be seen in the 943 cm−1 − 951 cm−1 range as well as in the ranges corresponding to amide II, amide I and C=O lipids vibrations. While, in the second analyzed range, wavenumbers originating from symmetric stretching of CH2 from lipids can be a strong platinum-resistance marker (Fig. 3).

Table 3 Identified significant ranges of wavenumbers.
Full size table

Also, the C5.0 decision tree learning model generated from these data (Fig. 4a) indicated that amide I vibrations placed around 1631 cm-1 wavenumber was the most important that unambiguously distinguishes the two categories of patients. Similarly, for the 2800 cm-1− 3000 cm-1 range, a strong absorption difference can be seen in the 2972 cm-1− 3000 cm-1 range, which can be a strong platinum-resistance marker (Table 3). Also, the decision tree learning model generated from these data (Fig. 4b) indicated that the most important wavenumber was 2993 cm-1 allowing to distinguish positive and negative cases almost perfectly. As it was visible, lipids could also play an important role in the platinum resistance phenomenon. Particularly, machine learning method and PCA analyses showed that by using asymmetric stretching CH3 vibrations of lipids it was possible to differentiate serum collected from both analyzed groups of women suffering from ovarian cancer (Figs. 1c1, c2, 3b, 4, Table 3). Moreover, as we have shown, the accuracy of the presented results is high. However, taking into account, that until now, nobody used FTIR spectroscopy to determine factors responsible for platinum resistance, our results cannot be compared with the other ones. Consequently, to determine whether the observed chemical differences may be responsible for the phenomenon of platinum resistance, further research must be undertaken.

Fig. 4
figure 4

Decision tree for FTIR dataset 800 cm−1 − 1800 cm−1 wavenumbers (a) and 2700 cm−1 − 3000 cm−1 wavenumbers (b).

Full size image

As we have shown, four machine learning methods based on different approaches were used in the study. The SVM model is based on determining the appropriate hyperplane between two categories of points (patients) based on the values of the attributes that describe them, in our case the absorption values of wavenumbers of a given wavenumbers. SVM works by maximizing the margin, or distance between the closest data points and this hyperplane, which increases the generalizability of the model. The algorithm is particularly effective for linear problems, but it can also be extended to nonlinear problems through the use of a so-called “kernel” (kernel) that maps the data to a higher dimension. The kNN model, in turn, is based on calculating the distance between the k nearest neighbors of the case under the study. This distance is calculated using Euclid’s measure and a different number of neighbors is used, in our case the number is 3. The new point (patient) is assigned to the class to which most of its k nearest neighbors belong. KNN does not require a training step because all knowledge is stored in the dataset. However, it can be computationally burdensome for large datasets because it requires comparing the new point with all existing data. The decision tree model, in turn, is based on the application of a measure of the entropy of information of individual attributes in the context of the category, and based on it the calculation of the so-called Information Gain. The attribute giving the highest information gain is placed in the next node of the tree. It works by creating a tree structure in which each node represents a question or feature, and each branch represents the result of a given question. The decision-making process involves moving from the root of the tree (the beginning) to the leaves (the final decisions) based on the values of the features in the data set, see Fig. 4. Decision trees are intuitive and easy to interpret. A somewhat similar model is random forest, which is a commonly-used machine learning algorithm, that combines the output of multiple decision trees to reach a single result. This is a type of collaborative learning bootstrap aggregating (bagging). Each tree in the forest is trained on a random subset of the data and a random subset of the features, which helps reduce the risk of overfitting that individual decision trees face. The outcome of Random Forest classification or regression is determined by aggregating the results from individual trees - in classification usually by majority voting, and in regression by averaging. The model is very effective and often used because it offers high accuracy and robustness to noise in the data. From the results of the experiments presented in Table 2, it is clear that all the methods used make it possible to achieve near-perfect recognition of platinum-resistant and platinum-sensitive cases. For the most part, the results of the accuracy, sensitivity or specificity parameters are at the level of 0.97-1.00. In some applications, worse results were obtained, especially can be seen in the application of the SVM model, which may be due to the fact that it is more difficult to determine the ideal hyperplane between the tested points. It is also possible to identify certain ranges of wavenumbers that are absorbed in a significant way. This makes it possible to reduce the attribute (wavenumber) space from 2076 to 151 for the 800 cm−1 − 1800 cm−1 range and from 624 to 55 wavenumbers for the 2800 cm−1 − 3000 cm−1 range (Tables 2 and 3; Fig. 3). At the same time, the classification quality parameters improve somewhat. For example, the RF and kNN models for the 800–1800 range achieve ideal classification, i.e. all evaluation parameters reach a value of 1 (Table 2). Classification performance is also shown by the structure of decision trees constructed using both ranges (Fig. 4). It can be seen that one significant wavenumber from the found ranges is enough to build ideal decision trees. It is these wavenumbers that obtained the highest Information Gain index value based on the entropy of information we obtain by using individual features (wavenumbers). Using only absorption value of these wavenumbers allows us to identify 36 platinum-resistant (PR) cases, distinguishing them from 26 platinum-sensitive (PS) cases, see Fig. 4. Such models as the most human-understandable and simple in construction can be recommended for use in the problem under study.

In our previously study, where we used FT-Raman spectroscopy as a complementary method to FTIR, we showed that not only absorption (FTIR), but also scattering technique (FT-Raman) can be used in determination platinum-resistance ovarian cancer28. Results obtained from PCA for FTIR and FT-Raman data showed clear differentiation between serum collected from platinum-sensitive and platinum- resistance ovarian cancer women. However, HCA analysis showed that higher similarity within the group was visible from FT-Raman data analysis. Decission tree results also showed differences between FTIR and FT-Raman measurement. Using FTIR, wavenumbers peaks at 1631 cm−1 (amide I) and 2993 cm−1 (CH lipids vibrations) were proposed as marker of platinum-resistance, while using FT-Raman – 1224 cm−1 (amide III) and 2713 cm−1 (CH lipids vibrations). As was visible, wavenumbers were different, but still amides and lipids vibration were proposed as markers of platinum-resistance phenomena. Consequently, presented in this paper results confirmed our previously study, which is especially important when spectroscopic methods were used first time in platinum-resistance phenomena in ovarian cancer study.

Conclusions

The FTIR spectroscopy in combination with multivariate and machine learning analysis showed to be a perspective tool for platinum resistance identification in women suffering from ovarian cancer. The most significant differences between platinum-resistance and platinum-sensitive women were observed in amide II, amide I, PO2- symmetric vibrations, as well as asymmetric stretching CH3vibrations from lipids. These results were confirmed by PCA analysis and the C5.0 decision tree learning model. PCA analysis showed a possibility to differentiate the serum collected from platinum-resistance and platinum-sensitive women, where significant for this distinction were PO2-, amides II and amide I, as well as asymmetric stretching CH3 and symmetric stretching CH2vibrations. The C5.0 decision tree showed that FTIR peaks at 1631 cm-1 and 2993 cm-1 can be a strong platinum-resistance marker. These results were correlated with PCA and suggested that changes in these molecules could be responsible for the phenomenon of platinum resistance. Finally, accuracy of FTIR method in differentiation between platinum-resistance and platinum-sensitive women suffering from ovarian cancer was above 92%.

Materials and methods

Participants

All methods were performed in accordance with the relevant guidelines and regulations. Moreover, written informed consent was obtained from all participants included in the study, which was approved by the Bioethics Committee of the Regional Medical Chamber in Rzeszow on 24th November 2016 (Resolution No. 90/B/2016).

The study was conducted with patients who were treated at the University Clinical Hospital of Fryderyk Chopin in Rzeszow, in the years from 2017 to 2020. All patients were women diagnosed with ovarian cancer in stage III without other type of malignancy ever with age between 32 and 76 years old. A total of 21 women were in the study; 12 platinum-resistance and 9 – platinum-sensitive. More detailed description of studied groups was placed in work Kluz-Barłowska et al.28.

Materials

The samples were collected from venous blood, in the morning at the day of surgery, before administration of anesthetic drugs. The patients were on an empty stomach, however they were taking medications for chronic diseases (arterial hypertension). Preparation of the samples was described in work Kluz-Barłowska et al.28. In short: blood was centrifuged (3000 rpm, 15 min) and stored at -80 °C to the FTIR measurement.

FTIR measurements

The absorbance spectra of serum samples were measured by a Bruker Vertex 70v spectrometer using attenuated total reflection (ATR) technique with, where in ATR diamond crystal was applied. The measurements were collected from 400 cm−1to 4000 cm−1>with 32 scans and a 4 cm-1 of spectral resolution. A background spectrum of air was collected before samples measurement. While serum samples with volume of 4µL were placed on the CaF2 slide and after dried the samples was measured 3 times. Consequently, 63 spectra were obtained. However, 62 spectra were analyzed. After each measurement, ATR plate was cleaned using 70% of ethanol. The obtained spectra were analyzed using the OPUS 7.0 software (software (Bruker Optit GmbH 2012). For each spectrum, the baseline correction was performed using the rubberband method with the same 64 points for each spectrum, min-max. normalization, and smoothing using 25 points with a Savitzky–Golay filter.

Multivariate analyses: principal component analysis (PCA) and hierarchical component analysis (HCA)

Each FTIR spectrum was consisted of high number of data. Therefore, to determine the similarity between the serum collected from platinum-resistance and platinum-sensitive women, PCA was performed based on the selected spectral regions: between 800 cm−1 – 1800 cm−1 and between 2700 cm−1 – 3000 cm−1. Moreover, to determine similarities between samples within the same group HCA analysis with Euclidean distance and paired group (UPGMA) algorithm was performed for the same FTIR ranges as PCA. Both, PCA and HCA analyses were done using PAST software (developed by Oyvind Hammer, https://www.downloadcrew.com/article/34304/past). On the other hand, positions of the analyzed FTIR peaks visible in serum spectra collected from platinum-resistance and platinum-sensitive woman was calculated from the average spectra obtained from all spectra characterized for platinum-resistance and platinum-sensitive woman.

Machine learning (ML) methods

The spectroscopic data were analyzed using selected machine learning methods. The goal was to determine the efficiency of identifying two categories of cases. Four different ML methods were used: K Nearest Neighbors (kNN), Support Vector Machine (SVM), C5.0 decision trees and Random Forest (RF). Selected quality measures were used to evaluate the quality of the built models: accuracy, sensitivity, specificity, precision, F1 score, and the Matthew correlation coefficient (MCC). Next, the Boruta feature selection method was used for identifying all relevant wavenumbers, and for creation the subsets of relevant wavenumbers. In the case of classification models, i.e., the ones that predict the values of a qualitative variable, one of the primary ways to summarize is to assess the quality of the classification - to answer the simple question of whether the model correctly assigns cases to categories of the dependent variable. The smaller the discrepancy between the actual values and those predicted by the statistical model, the better. The data from spectroscopic experiments have been converted to information systems, tables consisting of columns and rows. The number of cases is 62, of which 26 are platinum-sensitive and 36 are platinum-resistance. Additionally, in order to determine the most important wavenumbers to distinguish two groups of patients, a process of selecting appropriate features will be carried out using a random forest algorithm. This method is perfect for this purpose, it allows one to assess the importance of features based on the impact of individual features on the performance of the model. Common measures of feature importance in Random Forest include Gini impurity, average impurity drop, or average accuracy drop. The features with higher scores are considered more important, while those with lower scores are less significant. Additionally, Boruta’s algorithm will be used to distinguish between important and irrelevant features. The selected significant wavenumbers were used to then create a decision tree. Based on the prepared datasets, ML algorithms built learning models tested on test cases using a leave-one-out cross validation (LOOCV) approach.