All ions must serve: The role of various regimes of data acquisition in joint classifier for intraoperative mass spectrometry-based glial tumour identification  

Introduction

The expanding capabilities of ambient ionisation mass spectrometry make metabolic molecular profiling a promising approach for various life science and medical applications.^1,2 The simplicity and rapidity of sample collection, processing, and analysis are the key factors explaining the growing interest in ambient ionisation methods for diagnostics tasks inside and outside hospitals.^3–7 On the other hand, the lack of sample purification or separation makes the resulting mass spectra interpretation a challenging task. It arouses interest in the application of machine learning techniques for data analysis.^7–9 Indeed, molecular profiles obtained with ambient ionisation mass spectrometry is a multidimensional data which, however, represents only a part of the molecular composition of an analysed biological specimen. The different ionisation potential leads to competition for charge carriers during the ionisation process, while the limited dynamic range of the detector confines the simultaneous detection of all ions present in a wide range of concentration. Thus, the variety of the detected molecules (chemical space) became restricted, so only partial information on the underlying biological processes could be considered during the analysis.

Oncological surgery, especially neurosurgery, is one of the most demanding fields where new intraoperative tissue classification techniques are urgently needed.^10–13 The resection volume is the most important characteristic for the patient outcome since complete resection significantly reduces the likelihood of the tumour reoccurrence. In the meantime, excessive resection is not appropriate in neurosurgery, especially in cases where functional centres of the central nervous system are affected by the malignancy. Multiple studies demonstrate the potential of an ambient ionisation molecular profiling implementation as an additional decision-making technique for neurosurgery.^3,14,15 The classification of the resected tissue in these studies is based either on the lipid alterations¹⁶ accompanying the malignant transformation of cells or on the presence of specific water-soluble metabolites such as 2-hydroxyglutarate;^16–18 however, the detection of different classes of molecules requires the specific ionisation regimes: ion polarities and mass ranges. Further translation of such studies into clinical practice not only requires the translation of methods from high-resolution mass spectrometers, which are common in the scientific institutes, to low-resolution instruments that prevail in clinical practice,¹⁹ but will confront the problem of high natural variability of the molecular composition of malignant and, especially, healthy tissues. The realisation of research on cells or animal models is not suitable in the case of metabolic profiling as cell cultures do not represent the physiological environment of cells, while an animal model does not provide sufficient genetic intra- and intertumoural variabilities.²⁰

The development of classification models for intraoperative glial tumour tissue identification is suffering heavily from the limited availability of control samples, which could not be easily obtained due to medical and ethical reasons. In combination with the aforementioned limitations, it results in a “small sample size” problem which substantially reduces the accuracy of the classification models. As it was demonstrated earlier,²¹ the joint classifier that combines multiple weak learners fitted on the mass spectrometric data of various regimes boosts the classification accuracy significantly but parallel acquisition of multimodal data may slow down the analysis procedure substantially. Ensemble learning methods provide an opportunity to reduce the data set unbalancing impact on the classification model’s characteristics.²² Indeed, the concurrent consideration of ions detected in different regimes complements the chemical space with molecules of various classes, which is necessary to account for the natural molecular variability of intact and malignant tissues. In this work, we analyse the contribution of each regime of the data acquisition in attempt to evaluate their role in the joint classifier proposed and determine the requirements for the metrics for further mass spectrometry set-up optimisation.

Methods

Results

The initial data set was used to train classifiers to evaluate the contribution of various regimes. The different subsets were selected from all data for each of the following:

The results obtained on testing and validation data sets are shown in Table 1. For all the classifiers, the accuracy was found to be lower than for the joint classifier for all regimes.²¹ The accuracy obtained for the validation set was lower than for the corresponding training sets. However, the specificity and sensitivity of classifiers change variously through all the variations. The exclusion of positive polarity from the classification process significantly increases the false positive results, which means that many non-tumour samples have been incorrectly identified as tumour samples. In the case of the negative ion mode exclusion, the same effect was observed only for vote-aggregated classifies; however, the probability-based classifier trained solely on positive ion mode data did not show significant accuracy alteration but initially showed a poor false positive rate. Comparison of the impact of different resolutions on the classification demonstrates the tendency of false negative results for the probability-based classifier trained on high-resolution data in contrast to the tendency of false positive results for vote-based. In the case of low-resolution data, both aggregation types show some tendency to false positive tumour identification. The exclusion of any mass range from the data leads to a substantial loss of accuracy due to an increased rate of false positive identification. Such an effect is especially strong in the case of the analysis of narrow mass range data.

Discussion

The impact of various data acquisition regimes on the classification model is not unexpected, but the reasons for declined accuracy are directly related to the chemical and physical limitations of each experimental regime. The most expected difference is obviously interconnected with the polarity of the data.²⁷ Indeed, positive ions spectra are characterised by the presence of various types of phospholipids²⁸ such as phosphatidylcholines, phosphatidylethanolamines, and phosphatidylserines. Other lipids, such as phosphatidic acids and phosphatidylinositols, could be ionised solely in negative ion mode, as well as many of the metabolites, including specific oncometabolite 2-hydroxyglutarate^16–18 and nurometabolite N-acetylaspartate.^16,29 The lipid metabolism is considerably altered during malignancy resulting in significant changes in the lipid content of the cells. Thus, the classifier based on one polarity only will be characterised with decreased accuracy. For all single-polarity classifiers, the decreased specificity (i.e. the rate of false positive identifications) and increased sensitivity (i.e. the rate of false negative identifications) is observed, which means the aggregation of different polarities into a single classifier substantially helps to overcome overfitting caused by a small sample size problem.

High-resolution mass spectra could be transformed to the same dimension as the low-resolution data,¹⁹ but some details might be lost during the transformation. However, on the particular data set of glial tumour tissues analysed in this study, classifiers based on the high-resolution data show a significantly higher loss of accuracy during validation on the samples obtained from patients of the second cohort. The increased efficiency of the tumour tissue classification with the high-resolution-based models originated from the richer feature matrix as a high-resolution mass analyser provides peak-rich spectra, but the same fact negatively affects the classification results for validation data set as high-resolution-based classifiers are more sensitive to the natural molecular variability of the tissues.

Limited dynamic range of various mass analysers could also affect the classification efficiency. The wide mass range allows the accumulation of more ions in each mass spectrum, but some minor ions could be lost due to the limited capacity of ion traps.³⁰ The narrow mass range (from m/z 500 to 1000) is able to detect as many lipids as possible, which could be useful for the determination of the affected metabolic alterations in various lipid classes. However, it was found that the exclusion of the wide mass range from the training data set significantly reduces the classification accuracy. The model became substantially overfitted as the rate of false positive identifications increases significantly. The observed alterations between the lipid composition of the tumour and non-tumour tissues²⁸ affect relatively abundant lipids, thus providing changes in the intensities of the relatively high peaks in mass spectra which could be determined on a wide mass range data. The contribution of the metabolites, which are mostly small molecules,¹⁶ to classification accuracy, appears to be important as the classifiers trained on narrow mass range data only lose accuracy more noticeably compared to the wide mass range classifiers.

The obtained results demonstrate the non-linear contribution of each data acquisition regime to the effectiveness of an aggregated classifier regardless of the aggregation method implemented. As the underlying reasons for the observed results are interconnected with the biochemical background of the malignancy and the physical limitations of the method, it is not possible to evaluate the impact of any of the regimes directly and independently. Since the aggregation of multiple regimes improves the accuracy of the classification, it is required to collect a lot of data in every experiment, which means an increase in the time required for data acquisition for each sample. As it is opposed to the main idea of ambient ionisation mass spectrometry – rapidity of the analysis, it is crucial to develop the method to assess the impact of each regime. The required metrics of such an impact should be a numerical estimation of how the classification accuracy and other characteristics change if any of the given regimes are included or excluded from the consideration, which allows one to determine the optimal set of regimes required for the accurate and high-performance implementation of ambient ionisation molecular profiling in clinical practice. It is proposed that the Shapley value^31,32 calculation could be found as an appropriate metric; however, the current size of the data sets is supposed to be insufficient to make such calculations at this time, so the further collection of both tumour and control tissues is still required.