Machine learning characterization of cancer patients-derived extracellular vesicles using vibrational spectroscopies: results from a pilot study.

Abstract

The early detection of cancer is a challenging problem in medicine. The blood sera of cancer patients are enriched with heterogeneous secretory lipid-bound extracellular vesicles (EVs), which present a complex repertoire of information and biomarkers, representing their cell of origin, that are being currently studied in the field of liquid biopsy and cancer screening. Vibrational spectroscopies provide non- invasive approaches for the assessment of structural and biophysical properties in complex biological samples.

Methods

In this pilot study, multiple Raman spectroscopy measurements were performed on the EVs extracted from the blood sera of n = 9 patients consisting of four different cancer subtypes (colorectal cancer, hepatocellular carcinoma, breast cancer and pancreatic cancer) and five healthy patients (controls). FTIR (Fourier Transform Infrared) spectroscopy measurements were performed as a complementary approach to Raman analysis, on two of the four cancer subtypes. The spectra were subjected to various machine learning classifiers with hyperparameter optimization to discriminate between healthy and cancer patients-derived EVs. The AdaBoost Random Forest Classifier, Decision Trees, and Support Vector Machines (SVM) distinguished the baseline corrected Raman spectra of cancer EVs from those of healthy controls (N = 18 spectra) with a classification accuracy of >90% when reduced to a spectral frequency range of 1800 − 1940 𝑐𝑚⁻¹ and subjected to a 50:50 training: testing split. FTIR classification accuracy on N = 14 spectra showed an 80% classification accuracy. Our findings demonstrate that basic machine learning algorithms are powerful applied intelligence tools to distinguish the complex vibrational spectra of cancer patient EVs from those of healthy patients. These experimental methods hold promise as valid and efficient liquid biopsy for artificial intelligence-assisted early cancer screening.

Appendix

• 1. EVs isolated from serum displayed exosomes and microvesicles characteristics .

Based on the minimal information statement for the study of EVs set by the ISEV (International Society for Extracellular Vesicles) [2], we characterized the isolated EVs both physically and phenotypically. By using Western blot analysis, we observed that these vesicles expressed selective markers of EVs (i.e., Alix and TSG101) (Fig. 5A). The highest expression levels of these markers were observed in fractions 3-5 (at iodixanol density of 1.107-1.13 g/ml). These fractions were subsequently pooled for further analyses. When assessed by NTA, isolated EVs displayed a mean diameter of 109 nm (range 59-145 nm) (Fig. 5B). In addition, TEM analyses showed that the isolated EVs were round-shaped vesicles with a mean diameter of 90 nm (Fig. 5C).

Fig. 5

Isolation and characterization of patient-derived EVs. (A) EVs were isolated as described under Materials and Methods using iodixanol (OptiPrep) gradient and ultracentrifugation. Proteins isolated from the different fractions were analyzed by Western blot for the expression of specific EVs markers. Note that the highest expression levels of EVs markers are located in fractions 3 to 5. These fractions were pooled for subsequent analyses. MW = Protein molecular weight marker. (B) NTA analysis of pooled fractions 3-5. (C) TEM micrograph of purified EVs (red arrowheads). Scale bar 100 nm

• 2. Renishaw Raman spectroscope identifies rich spots of EVs on air-dried CaF2 slides . Spectra acquisition consisted of scanning at multiple pink spots on the prepared sample slides (Fig. 6A). The spontaneous formation of pyramidal crystals on the slides are shown (Fig. 6B)

Fig. 6

EVs detection by spectroscopy. (A) EVs SPOT ON RAMAN CONFOCAL MICROSCOPE (50X Objective). The axes represent the microscope field of view sizes in X and Y directions corresponding to the 50x lens that was used for measurement. The inset bar gives the field of view scale corresponding to the 50x objective lens. The image shows the EVs spot focused on the air-dried CaF₂ slide through the confocal microscope of the inVia Renishaw Raman system, with the scale bar of 20 μm as indicated by the scale bar on the bottom. These pink spots are regions enriched with EVs and provide a method to infer where the optimal Raman spectra are acquired. The yellow patchy regions correspond to the slide with higher concentrations of the PBS buffer. (B). CRYSTAL FORMATION. Pyramidal crystals spontaneously self- organized in the EVs rich regions (pink spots)

• 3. Machine learning algorithms exhibit poorer performance in the classification of EVs on Raman spectra without baseline correction

. Various binary classification algorithms were trained on raw Raman spectra without baseline correction to assess their performance accuracy in distinguishing cancer samples from healthy controls (Figs. 5A-J). All results collectively confirm that without baseline correction, ML algorithms exhibit poor predictive performance.

The AdaBoost RF classifier is a meta-estimator and an iterative ensemble learner available on Sci- kit-learn (Python machine learning library). The AdaBoost RF classifier was assessed on 1021 data points from 19 raw Raman spectra without baseline correction from the two classes (Fig. 7). To assess the sensitivity/specificity of the ML predictions, receiver operating characteristic (ROC) curves are generated to show the diagnostic ability of the binary classifier with varied discrimination thresholds. As shown in Fig. 5, the ROC curve for the AdaBoost classifier’s performance on the raw Raman spectra (whole range) with a testing size of 0.5 (i.e., the ML is trained on 50% of the data and tested on 50%) is shown (Fig. 7A). A 0.5 test size ensures stringent training conditions for the classification assessment. A smaller test size of 0.2 and 0.3 always showed greater classification accuracy. 50 tree estimators and a learning rate of 1.0 were kept as the default hyperparameters. The classification accuracy was 77.78% with a mean-square error (MSE) of 0.222. As seen, the area under the curve (AUC) was 0.45 indicating a poor classification accuracy. The turquoise curve shows the relationship between the true positive rate (TPR) and false positive rate (FPR). The closer the turquoise curve comes closer to the red dashed curve at 45 degrees of the graph plane, the less accurate the classifier’s predictions, and lower AUC of the ROC curve. The ROC curve visually informs us the trade-off between the sensitivity (TPR) and the specificity (1-FPR) [26]. The f1 scores were 0.62 (i.e., a precision of 0.67 and recall of 0.57) for the cancer group and 0.29 (i.e., a precision of 0.25 and recall of 0.33) for the healthy groups. The F1 score of 1.00 indicates a perfect recall and precision. The F1-score is often used as a measure of statistical accuracy for binary classifiers in ML [25, 27].

Fig. 7

ML performance on raw raman spectra without baseline correction. (A). Adaboost random forest (RF) on whole range raw raman spectra. (B) Adaboost 5-fold cross-validation curve on raw raman spectra for whole-range. (C) Adaboost RF classification on raw raman spectra with reduced frrequency. (D) CV curve for adaboost on reduced raw raman spectra. CV accuracy score: 20.00 ± 24.49%. (E)ROC curve for decision trees on raw raman spectra no baseline. (F) CV curve for decision trees performance on raw raman spectra (at 1808.25 cm⁻¹). The CV accuracy score was determined as 60.00 ± 20.00%. With an increased sample size of patients and hence, increased training datasets the performance of these algorithms can be better optimized for the intended impact of the presented findings. (G) ROC curve for SVM classification on raw raman spectra (complete). (H) classification report for SVM on raw raman spectra (full range). CV accuracy score: 80.00 ± 24.49%. (I) ROC curve for SVM performance on raw raman spectra with reduced frequency range. (J) CV curve for SVM predictions on reduced spectra. CV accuracy score: 90.00 ± 20.00%

The poor performance of the RF classifier in Fig. 7 indicates the data must be filtered to a narrower spectral range or alternately, undergo a baseline correction, as indicated by the PCA plot of the reduced frequency space above. In binary classification, recall of the positive class is defined as sensitivity while the recall of the negative class is specificity. The precision is defined as the ratio tp/ (tp + fp) where tp is the number of true positives and fp the number of false positives. The recall is the ratio tp/(tp + fn) where fn is the number of false negatives. The recall denotes the ability of the classifier to find all the positive samples. The f1 score defines the weighted harmonic mean of the precision and recall, where an f1 score reaches its best value at 1 and worst score at 0. Here, the f1 score was found to be of 1.00 (i.e., precision and recall were 1.00) implying both, a 100% sensitivity and specificity.

Figure 7B displays the cross-validation learning curve corresponding to graph 7A. It shows that the AdaBoost classifier is not optimally tuned to predict the classes of the newly presented test datasets as indicated by the vast grey shaded region (indicates training uncertainty). The grey fill space on the plot denotes the standard deviation of the training performance by the classifier as the training size increases. The broad range of grey fill indicates a heavier computational training is required for the classification accuracy to be optimized. The cross-validation score, or also known as out-of-sample testing, indicates the likelihood of the RF classifier’s performance when new results are presented with the current amount of training it has undergone. It is a validation technique to generalize the performance of the RF classifier to an independent dataset. The five-fold CV accuracy score was found to be 70.00 ± 24.49%. The curve in turquoise corresponds to the CV score curve and the optimal training score of 1.00 is indicated by the dashed violet curve.

The RF classifier was assessed on the frequency reduced spectra (1800-1940 cm⁻¹ wavenumber region) of 18 samples (12 cancers and 6 healthy controls). With a 0.5 test size the classification performance was 50.00% with a MSE of 0.5 indicating poor performance of the classifier. The AUC is 0.67 indicating a poor sensitivity and specificity [25, 27]. An f1 score of 0.86 was observed for the cancer group and of 0.50 for the healthy group (Fig. 7C).

Decision trees are a supervised learning technique which use multiple algorithms to decide to split a node into two or more sub-nodes with tree-like diagrams to classify some target variable/data. The performance of the decision trees classifier is shown with a 0.5 test size on a randomly selected single frequency (at 1808.25 cm⁻¹). The classification accuracy was found to be 88.89%with a MSE of 0.111. The AUC is 0.94 indicating a high classification accuracy. An f1 score of 0.93 (i.e., precision of 1.00 and recall of 0.88) for cancer and 0.67 for healthy (i.e., precision of 0.50 and recall of 1.00) was observed. The entropy criterion (one of the learning parameters) was used for the tree classification. The results remained unchanged with baseline correction for the Decision Tree performance (Fig. 7E).

The classification predictions by the SVM algorithm with a linear kernel is shown on the full range raw Raman spectra. SVM is a supervised ML algorithm which finds the optimal hyperplane that maximizes the margin between the data classes using gradient descent learning. Classification accuracy of 80.00% and MSE: 0.2 were observed. An f1 score of 0.75 for the cancer class (i.e., precision and recall of 0.75) and of 0.83 for the healthy class (i.e., precision and recall of 0.83) were obtained corresponding to an AUC of 0.79 (Fig. 7G). The SVM classification accuracy on the raw spectra within the selected frequency range reported above and a 0.5 test size was found to be 88.89% with a MSE of 0.111. The f1 scores for the cancer and healthy group classification were 0.92 (precision of 0.86 and recall of 1.00) and 0.80 (precision of 1.00 and recall of 0.67), respectively (Fig. 7I).