Real-time glioblastoma tumor microenvironment assessment by SpiderMass for improved patient management

Summary

Glioblastoma is a highly heterogeneous and infiltrative form of brain cancer associated with a poor outcome and limited therapeutic effectiveness. The extent of the surgery is related to survival. Reaching an accurate diagnosis and prognosis assessment by the time of the initial surgery is therefore paramount in the management of glioblastoma. To this end, we are studying the performance of SpiderMass, an ambient ionization mass spectrometry technology that can be used in vivo without invasiveness, coupled to our recently established artificial intelligence pipeline. We demonstrate that we can both stratify isocitrate dehydrogenase (IDH)-wild-type glioblastoma patients into molecular sub-groups and achieve an accurate diagnosis with over 90% accuracy after cross-validation. Interestingly, the developed method offers the same accuracy for prognosis. In addition, we are testing the potential of an immunoscoring strategy based on SpiderMass fingerprints, showing the association between prognosis and immune cell infiltration, to predict patient outcome.

Graphical abstract

Highlights

• •SpiderMass combined to AI provides a diagnosis of glioblastoma with over 90% accuracy
• •99 lipid markers are identified in relation to diagnosis, prognosis, and immune cells
• •An immunoscore based on SpiderMass data is being developed
• •Immunoscore applied to SpiderMass MS imaging data provides access to patient prognosis

Zirem, Ledoux et al. demonstrate that SpiderMass combined with AI enables accurate molecular diagnosis and prognosis of glioblastoma. The evidence of specific lipid biomarkers corroborating protein biomarkers offers potential therapeutic targets. This study provides immunoscoring that shows promise in predicting patient outcomes based on the infiltration of immune cell populations.

Introduction

With an incidence of 308,102 new cases and 251,329 deaths worldwide in 2020,¹ brain and central nervous system (CNS) tumors are among the deadliest cancers. Gliomas are the most commonly occurring tumors of the CNS, which account for almost 80% of all malignant primary tumors of the brain.² Gliomas can originate from different cell types and their classification represents a challenge. However, precise classification of gliomas is paramount to estimate the patient prognosis and foresee the best possible care and personalized treatment options.³ Since 2016 the World Health Organization (WHO) recommendation is to integrate histological features and molecular alterations for the classification of tumors of the CNS.⁴ Yet, the classification is based on the fifth edition of the WHO Classification of Tumors of the CNS,⁵ which considers further the importance of the molecular features while keeping the histology. The first criterion of classification is the isocitrate dehydrogenase (IDH) mutation status, which separates the glioma into the IDH-wild-type versus the IDH1- or IDH2-mutant groups. IDH-mutant status is not only important for the classification but it is also associated with a better prognosis.⁶ The tumors classified as IDH wild type are further categorized into glioblastoma (GBM) and astrocytoma/not elsewhere classifiable. Within the adult-type diffuse gliomas, GBM, histone-mutant gliomas, several rare entities, are the most prevalent and aggressive forms with an overall survival (OS) of about 15 months and a median of 14–24 months (as opposed to lower-grade gliomas, which have been possible to control for decades) and is classified as a high-grade malignant glioma (grade 4). While GBM shares some astrocytic histological features with astrocytoma, it also shows its own histological features including a high cellular density, a marker nuclear atypia, a high mitotic activity, and necrosis and/or microvascular proliferation, or at least one of the following specific molecular characteristics, i.e., EGFR amplification, TERT promoter mutation, and a concurrent gain of chromosome 7 and loss of chromosome 10 (+7/−10).

GBM is a highly heterogeneous disease, making it difficult to develop effective treatments.^7,8 So far, management of IDH-wild-type GBM involves a multimodal approach with maximal safe surgical resection surgery as a first-line approach followed by temozolomide chemoradiotherapy and maintenance (known as the Stupp protocol⁹). Treatment of GBM clearly meets different challenges. Indeed, novel immunotherapies, vaccination, or immune checkpoint inhibition, have shown disappointing results against GBM. One of the hurdles of GBM treatment is associated with the blood-brain barrier (BBB), which limits the passage of drugs into the brain, thus preventing the administration of many chemotherapies. Another difficulty is the heterogeneity of the tumor and signaling pathway associated, which make targeting therapeutic goals more complicated¹⁰ and also from the heterogeneity of the tumor, with the tumor microenvironment (TME) promoting the resistance and decreasing the efficiency of the patient immunological response. Indeed, the development of brain tumors disrupts the homeostasis of the CNS. The vasculature, extracellular matrix, and BBB are hijacked and integrated into the processes of tumor development.¹¹ It is known that the TME is immunosuppressive and that infiltration of different subpopulations of immune cells are correlated to both the response to therapy¹² and patient OS.¹³

On the other hand, the extent of resection (EOR) has been demonstrated to correlate with OS such that higher OS is observed for gross total resection (GTR) over subtotal resection or biopsy.¹⁴ In particular, increasing the EOR to >95% while sparing tissue to preserve functional integrity may improve surgical outcomes and OS.³ The importance of the EOR is thus recognized and has been incorporated into the European guidelines for the management of patients with GBM.¹⁵ However, GTR remains difficult due to the diffuse nature of GBM, which makes the accurate delineation of the tumor margins an issue. Different tools have been developed to help with finding the resection margins.¹⁶ Neuronavigation by magnetic resonance imaging with addition of 5-Ala has become a standard procedure to visualize margins and was shown to enable better EOR. Awake surgery has also been shown to limit the damages to functional tissues while increasing the EOR. Despite margins being more distinguishable for GBM, it is still infiltrating, hence awake surgery has proven its usefulness. More recently, Raman spectroscopy combined with machine learning (ML) has been developed for intraoperative assessment of glioma. Raman was shown to discriminate tumor from non-tumor tissue¹⁷ but also to distinguish between IDH-1 and -2 mutants from IDH wild type. This opens the possibility of a tailored EOR based on the aggressivity of the glioma.¹⁸ A handheld Raman probe is also available and was shown to achieve a sensitivity of 93% to classify normal brain from dense cancer.¹⁹ Therefore, both a real-time molecular analysis of the brain tissues during the surgery to achieve tailored resection and a better understanding of the microenvironment could become essential to improve the management of GBM patients.²⁰ Local therapies are increasingly seen as a unique option to deliver a higher dose of therapeutics while limiting the adverse effects of systemic administration. In this context, there is a need for stratifying patients based on their molecular features at the time of the surgery and correlate this stratification with patient survival to provide the best therapeutic option and start the treatment as soon as possible.

GBM has been the subject of numerous omics studies using mass spectrometry (MS)-based approaches and specifically using either shot gun proteomics or mass spectrometry imaging (MSI). Different MSI modalities, notably MALDI and DESI, have been used to follow known markers (e.g., IDH-mutant-associated metabolite 2-hydroxyglutarate),²¹ and identify potential biomarkers and therapeutic targets for the disease.^22,23 However, more recently MS has also emerged as novel technology for intraoperative assessment thanks to the development of novel ambient ionization mass spectrometry (AIMS) techniques.²⁴ Intra-operative AIMS is achieved from freshly resected tissues^25,26 or in vivo to advise neurosurgeons on the extent of tumor resection, with real-time feedback.^27,28,29 SpiderMass is one of these AIMS technologies that provides a minimally invasive solution for in vivo surgical applications.³⁰ Very interestingly, the technology is micro-invasive thanks to the high absorption of endogenous water present in the tissues at 2.94 μm (most intense absorption band of water) leading to a low penetration depth of the laser beam. Thus, there is only a white trace of dehydration left by the system with a few μm depth of tissues removal. It is also a very efficient process that works equally for all types of tissues because the water content of tissues is >70%.^27,30 In addition, the technology is based on MS and thus has the potential to become an essential tool for precise tissue characterization during excision surgery because of the high specificity of the molecular profiles, where the different markers are separated according to their mass (and charge). The MS spectra generated by the SpiderMass are indeed used to build a classification model that associates the MS molecular fingerprint obtained with a specific cell type (such as cancer). This association is achieved thanks to the use of a tissue cohort that has been annotated by a pathologist post-surgery. The classification model after validation can then be implemented in the instrument and is interrogated in real time during the surgery to provide an instant feedback, so far using a code, to the physician. The MS spectra are recorded in the millisecond range as well as the interrogation of the classification with a result that can be provided within less than a second. Moreover, we recently demonstrated that, if SpiderMass can be used to analyze fresh or fresh frozen tissues, it is also suitable for retrospective analysis in the pathology lab from formalin-fixed and paraffin-embedded (FFPE) samples.³¹ Interestingly, previous studies conducted intraoperatively both ex vivo and in vivo at the veterinary operating room on dog patients with sarcoma have also shown that the chemical background is not higher in vivo than ex vivo.²⁷ In cancer diagnosis-based MS, supervised ML algorithms play a critical role in building accurate classification models based on labeled datasets.³² Choosing and fine-tuning the appropriate classification model can have a significant impact on the accuracy of the results.³³ Cross-analysis by non-supervised ML strategies as well as interpretability of the results are also paramount and must be carefully considered.

Hence, SpiderMass has technology reeall, a potential to help with the management of patients with solid tumors. Here, we have explored this potential for patients with GBM. More specifically, we have identified two main aspects for which the technology would be useful. The first is to better delineate the surgical margin to move toward precision surgery. The second is to provide a stratification of GBM and be able to get a prognosis to help the oncologist with refining the therapy accordingly. Ultimately, if in the future some therapies could be administrated intraoperatively, we could offer all this information available by the time of the surgery. To perform SpiderMass intraoperatively, we must rely on a strong classification model. To address this, we have developed a more comprehensive artificial intelligence (AI) pipeline to provide accurate and reliable classification models and confident biomarker identification. The application of the pipeline to GBM demonstrates an accurate diagnosis and applicability to prognosis with more than 90% accuracy to distinguish patients with longer survival from those with shorter survival. We also show that specific lipid markers are associated with the prognosis of patients and are more specific to the molecular stratification than the tumor localization or histological class. In addition, we study the importance of the immune cell infiltration within the tissues and demonstrate that the ratio of macrophage subpopulation can be used to predict patient outcome. Ultimately, in this article we demonstrate that SpiderMass could be used to get a sub-stratification of patients with GBM as well as the prognosis in real time to afford improved management of the patient therapy.

Results

GBM IDH-wild-type lipidomic status assessment

Optimal classification model and AI processing pipeline

To undertake surgery, SpiderMass is combined to ML. Different ML algorithms are available, ranging from traditional ML to deep learning and the choice of the modeling will influence the accuracy of the built models. Traditionally, molecular diagnosis-based SpiderMass has relied on linear discriminative analysis (LDA) modeling using the Abstract Model Builder – AMX software, which has some limitations such as using only LDA classification, missing model evaluation and cross-validation, providing no explanation for predictions, and lacking statistical tests. In addition, ML models have been criticized for being opaque and providing limited insight into how or why they arrive at a particular outcome.^34,35 To address these limitations, we have developed a more comprehensive AI pipeline (Figure 1A) to provide reliable classification models and confident biomarker identification. Thus, we compared the performances of 24 different classifiers. This was performed from a retrospective cohort of FFPE tissues (Figure S1 and Table S1). All histologic regions (50 tumor, 42 necrotic, and 18 benign tissues) were analyzed using SpiderMass in both negative (138, 102, and 42 MS spectra, respectively) and positive ion mode (121, 88, and 39 MS spectra, respectively) (Table S2). Classification models were constructed to discriminate between tumor, necrotic, and benign tissue for which class assignment has been provided by pathologist and hence considered "the ground truth." In negative and positive ion mode, the optimal classification was obtained using the RidgeClassifier based on the highest accuracy on the 20% validation set, with 93% and 92% accuracy, respectively (Figures S2A and S2D). The corresponding accuracy of the training set was 100% for both (Figures S2B and S2E). After obtaining the classification report and confusion matrix for the optimal classifier, its performance was evaluated by using both accuracy and F1 scores, the latter considering both precision and recall and providing a more balanced measure of performance across classes. However, to ensure robustness and generalization of the data, a 20-fold cross-validation was performed. This showed that the actual accuracy of the model was 88% with a standard deviation (SD) of 0.03 and 87% with a SD of 0.02 in negative and positive ion mode, respectively. The model achieved F1 scores of 85%, 87%, and 89% for the benign, necrotic, and tumor tissue in negative ion mode, whereas the corresponding F1 scores in positive ion mode were 84%, 86%, and 88% (Figures S2C and S2F). The lower final accuracy obtained after 20-fold cross-validation could come from the small number of samples in the retrospective cohort, making the result less reliable. To address this issue, we added a second dataset (D2) based on a prospective cohort of 31 fresh-frozen tissues (Figure S1; Table S1) to get an extended dataset (DE). The added dataset D2 consists of 30 tumor regions and 7 necrosis regions, corresponding to 111 and 20 MS spectra in negative ion mode and 110 and 16 MS spectra in positive ion mode (Table S2). Despite some spectral differences (Figure S3) between the spectra of the FFPE and fresh tissues, the used of DE improved both the accuracy and F1 scores after cross-validation. In negative ion mode, the accuracy increased and stabilized at 92% after 20-fold cross-validation (Figure 1B). In addition, the SD decreased from 0.03 to 0.02. In positive ion mode, there was a slight improvement in performance (Table S3). Importantly, the accuracy remained stable even after cross-validation, allowing an actual accuracy of 88% with an SD of only 0.02.

Figure 1.

AI pipeline for optimal lipid classification model and discovery of molecular class-associated lipid markers

(A) Overall pipeline for molecular diagnosis. The continuous arrows are mandatory to obtain the final results (arrow ending with a dot) while a facultative step (dotted arrow) can be added to the pipeline, such as the spatial distribution of confident biomarkers thanks to WALDI-MSI.

(B) Classification report and confusion matrix of RidgeClassifier with the training set and after 20-fold cross-validation, obtained for the classification model using DE in negative ion mode.

(C) Table with the predicted label by SpiderMass blind diagnosis compared with the pathologist true label of two FFPE and seven fresh frozen (FF) unknown GBM tissue.

(D) Corresponding boxplot of 10 confident biomarkers specific for necrosis, benign, and tumor tissue in both ion mode with their tentative annotation.

ns, not significant: p ≤ 0.05; ∗p ≤ 0.01; ∗∗p ≤ 0.001; ∗∗∗p ≤ 0.0001.

Related to Figures S1–S6 and Tables S1–S6.

Automatic blind prediction of histological class from the built classification

To further test the performance of the statistical model in both ion modes, blind predictions were made on an independent test set consisting of two FFPE tissues and seven fresh frozen tissues. These nine tissues were not used to construct or evaluate the previously built model. The regions were subjected to SpiderMass analysis in a blind fashion and evaluated against the classification model trained on DE. In negative mode, all the analyzed tissues were correctly attributed to their respective region after comparing the SpiderMass result with the pathology examination. In positive mode, 2 out of 19 regions were misclassified, resulting in 90% of correct assignments (Figure 1C). All necrotic regions were correctly predicted, while two tumor regions were classified as necrotic regions. These two classification errors can be attributed to the fact that the blind analysis point resides at the border between the tumor and necrotic zones.

Lipid markers associated to the different molecular classes

We were then interested to obtain the identification of the lipids associated to the different histological classes to better understand the biological pathways responsible for the discrimination and address the issue of most models, which remain black box. The LIME³⁶ algorithm included in the developed processing pipeline provides interpretable explanations with potential biomarker identification by calculating the weight of each m/z feature to classify each tissue type. The top 120 positively and negatively contributing m/z features and their contribution are shown in Figures S4A–S4C for negative and positive ion modes. For example, in negative ion mode, the m/z feature that contributes most positively to the classification of benign tissue is m/z 766.55. The m/z 751.55 contributes positively to classify necrotic tissue but negatively to tumor tissue, while the opposite is true for the m/z 794.55. To identify the potential lipid biomarkers independently of classification, ion abundance variations between the different regions of interest (ROIs) were investigated using multivariate statistical analysis on the same data. In negative ion mode, peak picking, considering a signal with signal under noise >10, yielded 509 peaks out of a total of 5,000 m/z features. Hierarchical clustering was used to construct a heatmap and allowed us to identify three ion clusters, each of which was more overexpressed in a specific tissue type (Figure S4B). After heatmap visualization, a Kruskal-Wallis test was performed to determine whether each ion was statistically significant. Of these 509 peaks, 154 were found to be significant (p ≤ 0.05). After filtering out unresolved peaks and isotopes, 57 potential biomarkers remained. The same strategy was used in positive ion mode (Figure S4C–S4D), resulting in 100 significant peaks and a total of 37 m/z features were retained as potential biomarkers.

To uncover lipid biomarkers with high confidence, we combined the data from the supervised and unsupervised analyses. Thus, for an m/z feature to be considered as a confident biomarker, it should be significant (either overexpressed or underexpressed) in a given ROI and should also have a positive contribution if overexpressed and a negative contribution if underexpressed in the same ROI. A total of 41 confident biomarkers were identified, 27 in negative ion mode and 14 in positive ion mode, which are listed in Table S4 with their corresponding identification, obtained by MS/MS thanks to the SpiderMass analysis, and their relative abundance variation presented in boxplots (Figures 1D and S5). In negative ion mode, of the 27 confident biomarkers, 9 ions are overexpressed in the tumor tissue, the majority of which are phosphatidylserines (PSs) and phosphatidic acids (PAs). Seven ions are specific for benign tissue, of which six are highly abundant and one diglyceride (DG) 40:9 (m/z 661.55) is poorly abundant. Normal brain tissues are composed of different lipid species, but the absence of DGs appears to be a potential marker of health. Interestingly, all the highly abundant ions found in benign tissue are large phospholipids with fatty acids containing more than three unsaturated. In glioma cells, the high concentration of polyunsaturated fatty acids appears to inhibit cell division and tumor formation.³⁷ In necrotic regions, three ions are overexpressed at low masses and eight ions are underexpressed at high masses. More specifically, phosphatidylinositols (PIs) are found to be underexpressed in these tissues. In positive ion mode, among the 14 confident biomarkers, two are specific for benign tissue, m/z 842.65 and m/z 850.65. For necrotic tissue, three confident biomarkers are poorly represented, again PIs and phosphatidylcholines (PCs). For tumor regions, eight were found to be overexpressed, the majority of which are phosphatidylethanolamines (PEs). Interestingly, the glycerophospholipids highly expressed in tumor tissues are mostly saturated lipids such as PC 34:2 (m/z 758.65) and PS 38:2 (m/z 814.55).

Unsupervised and supervised molecular sub-stratification of GBM IDH wild type based on lipidomics

In a previous study using spatially resolved shot gun proteomics guided by MALDI-MSI, we have shown that grade 4 GBM could be stratified in three specific molecular sub-groups.²³ Interestingly, these were shown to be involving different pathways respectively related to (1) neuro-developmental genes, characteristic of neuronal/glial lineages and neural progenitor cells (red region), (2) immune status with macrophage infiltration (yellow region), and (3) antiviral immune response and viral infection (blue region). In addition, a correlation was observed between these groups and the OS of the patients, revealing that those patients with identified inflammation markers, macrophage infiltration, and antibodies exhibited a shorter OS. We thus firstly performed MALDI-MSI in positive ion mode from the remaining tissues of our prospective cohort and performed a global segmentation from all the tissues (Figure 2A). As previously demonstrated for peptides, we found three main molecular groups with a good correlation with the global segmentation from the protein MALDI-MSI. Indeed, it gives the same sub-stratification of IDH wild type based on lipidomic and proteomic MALDI-MSI data. There would be a high value to obtain such a stratification in vivo and in real time during the surgery to adapt the EOR according to the prognosis of the patient. MALDI-MSI cannot be used in vivo but SpiderMass can. Interestingly, MALDI and SpiderMass lipid profiles are very close because SpiderMass is based on WALDI-MS, which recapitulates a similar process to MALDI using endogenous water as a MALDI matrix.³⁸ Thanks to the AI processing pipeline developed, classification models were obtained to differentiate the three specific molecular sub-groups with a correct classification rate, after 20-fold cross-validation, of 94% and 88% in negative and positive ion mode, respectively (Figures 2B and 2C). The automatic blind prediction of the different sub-groups made it possible to classify well six of the eight unknown tissues. Next, 18 confident biomarkers were discovered in both modes (Figure 2D). Interestingly, some ions were in common with the lipid markers of the different histological zones found above. For example, PS 34:1 (m/z 760.55) specific to the tumor area is also specific of the proteomic blue region related to tumor growth. In addition, two ions m/z 878.65 (PE 46:4) and 902.65 (PS 44:0) were found specific to benign tissue but also here of the red region related to neuro-developmental genes. This enables multiple correlations between proteomic and lipid data, as well as the possibility of using SpiderMass as an in vivo tool for stratifying GBM tissue to adapt the surgery response according to the survival time of the patient.

Figure 2.

Unsupervised and supervised molecular sub-stratification of IDH wild type based on lipidomics

(A) Global segmentation based on lipidomic and proteomic data for nine FF GBM tissues.

(B) Classification report and confusion matrix of RidgeClassifier with the training set and after 20-fold cross-validation, obtained for the classification model that distinguish the three segmentation clusters.

(C) Boxplots of two lipid biomarkers obtained for each segmentation clusters.

ns, not significant: p ≤ 0.05; ∗p ≤ 0.01; ∗∗p ≤ 0.001; ∗∗∗p ≤ 0.0001.

Prediction of patient outcome from SpiderMass lipidomic data

SpiderMass classification of patient according to survival

On top of the validated diagnosis and stratification based on SpiderMass, we were also interested in investigating the potential to predict the patient outcome. A classification was built using the previously developed processing pipeline, according to a 15-month median of OS of GBM patients, separating the cohort into <15 and >15 month OS. We used the MS spectra recorded from the tissues independently from their localization to find the lipids, discriminative of the outcome rather than the histological annotation. Both in positive and negative ion mode, we could build a classification model with a correct classification rate of 79% and 87% after 20-fold cross-validation (Figure 3A). To improve the specificity of the outcome-based classification, we then looked to the patients with the more extreme survivals, i.e., <10 and >36 months. By narrowing our model to these cases, we obtained, in negative ion mode, a classification rate of 100% for the training set, 94% for the training set of the 20% validation and 93% after the 20-fold cross-validation (Figure 3B). In positive ion mode, the classification model reaches 87% of correct classification rate after cross-validation (Figure 3B).

Figure 3.

Obtainment of a classification model to distinguish patients with a short or long survival time and the discovery of biomarkers for each survival

(A and B) Accuracies obtained for the classification model made with RidgeClassifier, with the training set, after 20% validation and after 20-fold cross-validation in negative (orange) and positive (red) ion mode to distinguish patients with a survival time less than or greater than 15 months and after narrowing our model to most extreme cases (<10 months and >36 months).

(C and D) Six potential biomarkers, their corresponding boxplot and their chemical structure, found specific prognosis <10 and >36 months in both ion mode.

ns, not significant: p ≤ 0.05; ∗p ≤ 0.01; ∗∗p ≤ 0.001; ∗∗∗p ≤ 0.0001.

Related to Table S7.

Discriminative markers associated to the patient outcome

When searching for the biological explanation and potential biomarkers associated to the shortest and longest OS, some ions were found to be specific to one or other. Indeed, with both ion modes, 48 statistically confident prognosis biomarkers were found (Table S5). In summary, PCs and PIs are more present in patients with a longer OS, as PC 42:4 and PC 44:6 in positive ion mode (Figure 3D) and PI 38:4 and PC 42:6 in negative ion mode (Figure 3C). Moreover, ions m/z 850.75 (PE 44:5) and m/z 920.75 (PS 46:5) are more abundant in benign tissues and were found also to be associated with longer survival. It would appear that tumor tissue from a patient with long survival shares molecular markers with benign tissue. This could explain why some patients live longer. On the contrary, PSs, Pas, and PEs, such as PA 38:1 and PE 38:0, are more highly expressed in poor prognosis tumor tissue (Figures 3C and 3D). In particular, PS 40:6 was found to be associated with lower survival in both ion modes (m/z 834.55 and m/z 836.55) with a high significance and a high contribution. In addition, sphingolipids, specifically ceramides (Cer), were characteristic of a worsened prognosis in our study, since none were found for the longer survivals. It was also shown that the abundance of five Cer lipids, including Cer d42:2 (m/z 682.65) and CerP 34:0 (m/z 654.55), was significantly higher in patients with poor survival, regardless of the acyl chain length or the degree of unsaturation. Interestingly, our findings at the lipid level are closely related to those at the proteomic level. Indeed, the previously identified three prognosis markers²³ are closely linked to the prognosis lipid markers found here. Tumors from patients with longer survival showed increased expression levels of proteins RPS14 and PPP1R12A, whereas tumors from patients with shorter survival showed higher expression level of ANXA11. ANXA11 is a calcium-dependent phospholipid-binding protein,³⁷ that binds to negatively charged phospholipids in the presence of calcium ions.³⁹ The phospholipids identified as biomarkers of shorter survival, PSs, PAs, and PEs, are negatively charged, suggesting a specific lipid and protein pathway associated with a poorer prognosis. Targeting this pathway in the future may improve the survival of GBM patients. In addition, PPP1R12A is a subunit of myosin phosphatase, which regulates anti-tumor signaling pathways such as activating gene expression of tumor suppressors, such as Rb, Rap, and c-Myc.⁴⁰ On the other hand, there is a known ceramide-protein interaction between Cer and I2PP2A, leading to the degradation of c-Myc.⁴¹ This highlights a balance between the presence of PPP1R12A protein, which activates c-Myc and leads to a better prognosis, and Cer, which inhibit c-Myc and are associated to a shorter survival.

Immunoscoring of GBM by identification of SpiderMass signature of immune cells

SpiderMass classification of immune cells

Infiltration of various population of immune cells was found to be associated to the patient outcome in colon cancer^42,43,44 and is now recognized in many cancers.⁴⁵ Nowadays, the immunoscore is obtained from excised tissues or biopsies by immunohistochemistry (IHC) using antibodies specific to the different immune cell populations. However, the immunoscore is obtained post-surgery. Here, we explored the possibility to create an immunoscore based on SpiderMass data that could in the future be exploited in vivo. To this end, we analyzed with SpiderMass different populations of immune cells namely macrophages (M1-like and M2-like), CD4 and CD8 T lymphocytes, and NK cells (all regrouped as lymphocytes) versus cancer cells (NCH82 GBM cells). Immune cells were analyzed directly from culture well plates after cell sorting by flow cytometry to get the specific molecular profiles of each cell type. Indeed, since SpiderMass is not designed to offer single-cell analysis, as its spatial resolution is only 250 μm, hundreds of cells are analyzed at the same time from one analytical spot. To compensate for this more limited spatial resolution by comparison with IHC, we developed a solution to predict the ratio of one cell type from a SpiderMass MS spectrum. A Python library called LGBM (Light Gradient Boosting Machine)⁴⁶ was used to train an immunoscoring model using 107 MS spectra for each cell type for which a correct classification rate of 100% was obtained in both training and cross-validation (Figure 4A). Using our developed AI pipeline, 10 lipids markers of M1-like macrophages, 8 lipid markers of M2-like macrophages, 6 lipid markers of lymphocytes, and 3 lipid markers of NCH82 cancer cells were obtained (Figure 4B; Table S5). For example, the ions m/z 818.65 (GlcCer d18:1_22:0) and 819.55 (PG 18:1_22:6) were found specific to M1-like and M2-like macrophages, respectively. For the lymphocytes, PEs were reported as lipid markers: PE (20:4_16:0) (m/z 738.55) and PE 38:3 (m/z 768.55). Interestingly, two ions (m/z 844.65, m/z 848.65), specific to M1-like and M2-like macrophages, were also found specific to tumor tissue from patients with an OS less than 10 months.

Figure 4.

SpiderMass classification based on immune cell lipid profiles

(A) Trained LGBM model for cancer cells, lymphocytes, and M1-like and M2-like macrophages that were analyzed previously by SpiderMass technology after isolation.

(B) Boxplot of three biomarkers for GBM cancer cells, lymphocytes, and M1-like and M2-like macrophages.

(C) The overall pipeline for seeing the distribution of cancer and immune cells in GBM tissue analyzed by SpiderMass-MSI.

(D) Multiplex MALDI-IHC on three biomarkers in two GBM fresh frozen tissue sections. The display scale (arbitrary peak intensity units) is as follows (minimum intensity/full intensity threshold): 4/25 (Ki67), 2/6 (CD8α), and 4/25 (CD68).

ns, not significant: p ≤ 0.05; ∗p ≤ 0.01; ∗∗p ≤ 0.001; ∗∗∗p ≤ 0.0001.

Related to Table S8.

Correlation of immunoscores with patient prognosis

From a SpiderMass image, we were able to get the predicted distribution of the different immune cells across the tissue. Indeed, a second pipeline was dedicated to predict the probability of presence of each cell type based on SpiderMass images of fresh frozen GBM tissues from the prospective cohort. The results obtained provide estimated scores for each cell type, and ratio scores were computed to determine the relative score presence (RSP) of each cell type across the entire image. These scores were calculated by summing the scores provided for each cell type and dividing the sum by the total scores across all labels (Figures 4A–4C). A total of six fresh frozen tissues were analyzed, three from patients with a <10-month survival and three from patients with >36-month survival. The ratios provide insight into the distribution of the trained cell types across the image, allowing a comprehensive assessment of the cellular landscape in patients with shorter versus longer survival (Figures 5A and 5B). Interestingly, the predicted ratio of cancer cells is not significantly different according to the survival. However, the lymphocytes were predicted to have a higher abundance in tissue from patients with >36-month survival (Figure 5C). Indeed, the mean RSP of lymphocytes was above 40% and below 36% for longer versus shorter survival. The predicted ratio of macrophages also revealed an interesting association between immune cells and the patient outcome. Specifically, a higher proportion of M1-like macrophages (3.5%) was found in the tissues of patients with a better outcome compared with those with a worse outcome (2.8%) (Figure 5D). Conversely, M2-like macrophages had a mean RSP of 35.8% and 41.3% in patients with longer and shorter survival (Figure 5E). The ratio of M1-like to M2-like macrophages could serve as a prognosis marker. In fact, a proportion of M1-like macrophages below 7%, in comparison with M2-like macrophages, signifies a shorter OS. Conversely, a composition exceeding 10% of M1-like macrophages relative to M2-like macrophages is indicative of an extended OS (Figure 5F). Furthermore, a correlation appears to exist between the prognosis and the balance between the presence of lymphocytes and macrophages. Specifically, in cases of unfavorable prognosis, macrophages tend to outnumber lymphocytes (+22%), while the converse holds true for patients with a favorable prognosis (Figure 5G).

Figure 5.

Immunoscoring distribution from GBM FF tissues for cancer cells, lymphocytes, and M1-like and M2-like macrophages

(A and B) H&E scans and relative score of presence for cancer cells, lymphocytes, and M1-like and M2-like macrophages in three tissues of patients with good or a poor prognosis. For each cell, the ratio in percentage is indicated. In addition, a histogram was built to compare the percentage in each tissue.

(C–E) The mean percentage of lymphocytes and M1-like and M2-like macrophages in tissue with a prognosis of <10 months and >36 months.

(F and G) The mean ratio M1/M2 macrophages and lymphocytes (L)/macrophages (M) in tissue with a prognosis of <10 months and >36 months.

A 5-plex MALDI-IHC panel was constructed to corroborate previous results from the recently developed immunoscoring pipeline. MALDI-IHC relies on the use of antibodies conjugated with novel photocleavable mass-tags and enables highly multiplexed MALDI mass spectrometric imaging of targeted protein biomarkers.^47,48,49 For this purpose, the five biomarkers chosen were two controls, vimentin and a collagen cellular matrix marker; the Ki67 cancer-related biomarker (proliferation/cancer cells); and two biomarkers for immune cells CD8 (cytotoxic T cells) and CD68 (macrophages). The distribution of distinct cell populations (importantly immune cells), as achieved through the integration of SpiderMass and LGBM, was validated by MALDI-IHC (Figure 4D). Indeed, a higher concentration of macrophages were detected in tissue from a patient with a poor prognosis. In comparing immunoscores with the MALDI-IHC analysis, it is important to note that immunoscoring uses a probabilistic approach where each pixel indicates the likelihood of specific cell types being present, spanning a gradient from 0 to 1. In contrast, the MALDI-IHC analysis follows a binary strategy, assigning pixels a value of 0 (absence) or 1 (presence) for the considered cell types.

Of particular interest is the fact that the SpiderMass-MSI approach, based on immunoscoring, allows the differentiation of various subpopulations within the TME without necessitating techniques reliant on probe utilization. In conclusion, this innovative approach not only offers insights into the presence of immune cells within the TME but also presents the potential for a more rapid prognostication of survival time among GBM patients.

Discussion

This study represents a comprehensive investigation of the possibility to get an accurate diagnosis, sub-stratification, and prognosis of GBM using AIMS by SpiderMass combined with an AI pipeline. We explored various classification methods to discover the most appropriate one for our dataset. Given the lack of consensus in the community and the utilization of different algorithms (such as LDA⁵⁰ or Lasso⁵¹) for similar purposes, our goal was to identify the most suitable approach. A total of 24 classifiers were evaluated and an optimal model was obtained using the RidgeClassifier.⁵² It is a linear classifier that uses L2 regularization to avoid overfitting and demonstrate strong performances for MS data from tissues (high-dimensional datasets).⁵³ To address the issue of small sample size, which is responsible for the limited performances of the classification model, we investigated data extension by aggregating SpiderMass data from both fresh frozen tissues and FFPE tissues thanks to the ability of the technology to provide good quality data independently of the conservation of the sample. Data extension led to an increase in model accuracy of 92% and 88% after 20-fold cross-validation for tissue histological type classification in both ion modes. Blind predictions resulted in no misclassified spectra in negative ion mode and only two misclassified spectra in positive ion mode, with an average calculated sensitivity and specificity of 95% and 98%. By comparing the specificity and sensitivity obtained in our study for preoperative delimitation of brain tumors with other techniques in use,⁵⁴ SpiderMass technology shows an interesting potential by comparison with the other modalities such as fluorescence-guided surgery or even preoperative image guidance. SpiderMass also appears to be in the upper range of success compared with other MS-based modalities such as REIMS, while being far less invasive (Table S6).

In addition, to validate further our results and provide biological explainability, we included a secondary data processing pipeline. The LIME algorithm combined with unsupervised ML techniques and statistical tests was used to find confident lipid markers. The pipeline’s predictive explanations and statistical tests can increase the transparency and usefulness of ML models for biomedical applications. The combination of SpiderMass and ML pipeline represents a promising approach for advancing precision medicine for GBM. Interestingly, the study identified 41 reliable biomarkers that showed different phospholipid classes based on tissue type. Benign tissue was characterized by the presence of polyunsaturated fatty acids and the absence of DGs. Necrotic tissue showed the absence of PIs, while tumor tissue had a significant presence of PSs and PAs among the detected phospholipids.^55,56 These robust biomarkers warrant further investigation for their therapeutic potential.

Here, we have extended our previous work by showing that IHD-wild-type GBM could be stratified in three main molecular sub-groups with similar results based on lipidomic data as those obtained by spatially resolved proteomics. The accuracy of classification for these sub-groups was found to be 94% and 88% for negative and positive ion mode, respectively. Notably, within the set of 19 robust biomarkers uncovered, three biomarkers were already identified as specific to distinct histological tissue types. For instance, one biomarker exhibited specificity to the tumor region and also within the blue region linked to tumor growth. Two other biomarkers demonstrated specificity to both benign regions and the red region, which, is associated with neuro-developmental genes. This not only facilitates intricate correlations between proteomic and lipid data but also opens the door to employing SpiderMass as an in vivo tool for stratifying GBM tissue. This stratification can significantly contribute to tailoring surgical responses in alignment with a patient’s survival timeline.

Interestingly, our developed ML pipeline not only serves as a diagnosis tool but also for prognosis. The classification model developed in this study successfully distinguished tumors from patients with less or more than the median survival of 15 months. However, more specific results were obtained when comparing patients with more extreme survivals, i.e., <10 months versus >36 months with achieving accuracies of 93% and 87% in negative and positive ion mode, respectively. The presence of PCs and PIs was associated with a longer survival, while PSs, PAs, PEs, and Cer were more highly expressed in tumor tissue with a shorter survival. A recent study also found significantly more Cer in control tumor tissue compared with tissue treated with anti-PD-1 therapy, supporting our findings that Cer are markers of a poor prognosis.³⁷

A graphic depiction showcasing all identified biomarker ions from our study, alongside their corresponding groups of specificity, is presented in Figure 6. This visual representation facilitates an effortless assessment of whether a given ion exhibits specificity toward more than one group. Notably, a compelling correlation emerges between tumor methylation, tissue benignity, OS exceeding 36 months, and the segmentation clusters denoted in red and yellow. Indeed, a direct association between ions specific to tumor methylation and those specific to survival beyond 36 months (m/z 725.55 PA 38:3 and 906.65 PC 42:6), as well as benign tissue (m/z 904.65 PC 42:7 and 932.75), becomes apparent. Furthermore, ions exclusive to benign tissue also exhibit specificity toward survival exceeding 36 months (m/z 850.65 PE 44:5, 883.65 PI 38:5, and 920.75 PS 46:5), in addition to the red and yellow sub-stratification clusters (m/z 874.85, m/z 878.65 PE 46:4, m/z 881.75, and m/z 902.55 PS 44:0). In contrast, the blue sub-stratification cluster demonstrates a distinct association with tumoral tissue (m/z 760.55 PS 34:1) and survival periods of <10 months. Notably, ions specific with no-methylation also exhibit specificity toward tumor tissue (m/z 780.65 PE O-40:5 and 802.55 PC 38:7) and shorter survival durations (m/z 744.55 PE 36:1). Intriguingly, tumor tissue’s connection to short survival is underscored by the presence of three ions (m/z 749.55 PA 40:6, 778.55 PS 36:6, and 788.65 PE P-40:0). Considering that some ions were found to be specific to the methylation status, obtaining the methylation status by SpiderMass in a less time- and resource-consuming manner than currently achieved by genomics is something that could further be investigated in the future.

Figure 6.

Overview of all lipid markers identified using supervised and unsupervised discovery specific of the tumor methylation, tissue histology, sub-stratification of GBM tumor, and survival time of the patient

(A) Negative ion mode.

(B) Positive ion mode.

Related to Figure S6 and Tables S6–S8.

The third innovative aspect is the use of SpiderMass to predict the survival time of GBM patients based on immune cell infiltration. Here, we studied macrophages and lymphocytes as a first approach to assess the possibility of predicting the presence of immune cell infiltration based on SpiderMass data. The M1 phenotype of macrophages refers to pro-inflammatory and anti-tumor macrophages, while the M2 phenotype refers to anti-inflammatory and pro-tumor macrophages.⁵⁷ Creating an immunoscore for SpiderMass offers a way to translate the immunoscore in vivo and predict the patient outcomes already at the time of the surgery. This is offering a way for surgeons to tailor the therapy if a local intraoperative therapy (e.g., photodynamic therapy) is planned. So far, SpiderMass analysis, based on surgeon needs, was set to a spatial resolution of 400 μm, which is obtained by focusing the laser beam at the exit of the optical fiber. Thus, a recorded MS spectrum in an area of the tissue showing immune cell infiltration corresponds to a mix spectrum of different cell types. To address this problem, we have tested the performance of the LGBM algorithm to predict the ratio of different cells within the MS profile. However, this required training the LGBM algorithm to recognize the profile of the lymphocytes and macrophages. This was achieved by analyzing the different immune cells with SpiderMass to feed the algorithm. Remarkably, based on the prediction, we showed that the macrophage type (M1-like versus M2-like) were associated with the prognosis with different ratios of M1-like and M2-like according to the patient survival, patients with lower survival presenting more M2-like and less M1-like and vice versa for patients with longer survival. This is well known for cancer and it validates the approach. In addition, the ratio of M1 and M2 falls within the values previously calculated by histology and IHC.^58,59,60 Besides, the predictions were validated by MALDI-IHC using specific antibodies directed against immune cells. These results indicate that the immunoscore has the potential to become a prognosis prediction tool for GBM patients without the need for probes, allowing for personalized treatment for the patients based on their personal immunoscores. In addition, considering that the approach employed for the immunoscore enables to predict the percentages of presence of each cell type in the local environment analyzed, it can effectively be also translated for the definition of surgical margins. Indeed, for accurate margin assessment it is necessary to get the percentage of cancer cells versus normal cells. The same approach as we developed could be used to predict these percentage of these cell types. While none of the tissues in the present cohort were suitable for establishing this boundary, preliminary findings from an ongoing study from transgenic TgC3(1)-Tag mice models of breast cancer demonstrate its successful application for establishing the ratio of cancer/normal cells for improving the delineation of the tumor margins (Figure S6).

In conclusion, this study demonstrates the potential of SpiderMass combined with AI predictions for accurate classification of tumors and tissue subtypes. The ML pipeline developed in this study provides accurate classification results, identifies potential biomarkers, and provides interpretability. The reliable biomarkers identified reveal the distinct phospholipid profiles associated with different tissue types. In addition, the lipid and proteomic findings suggest a specific lipid-protein pathway associated with both poor and good prognosis in GBM patients. The use of SpiderMass as a prognosis tool and the association between macrophage phenotypes and patient prognosis further contribute to the advancement of precision medicine for GBM and other heterogeneous cancers. Future investigations should focus on validating these findings in larger cohorts and exploring targeted therapies based on the identified pathways and biomarkers. Overall, the combination of SpiderMass and the ML pipeline holds great promise for possible application in the future for accurate diagnosis and prognosis as well as improved EOR to be achieved during the surgery directly from the patients and is even compatible with awake surgery considering the analysis is painless.

Limitations of the study

The diagnosis and prognosis aspects, rooted with Ridge model classification, yield highly favorable results, and the corresponding markers identified through both supervised and unsupervised approaches demonstrate robustness. From a molecular standpoint, this reliability is underpinned by the well-established presence of glycerophospholipids, primarily situated in the bilayer of cell membranes.

While the LGBM model performs admirably across various fingerprints of four cell types (NHC82, M1, M2, lymphocytes), it is crucial to acknowledge that the interpretability of scores on the SpiderMass image remains relative. To comprehensively capture the diverse landscape of the TME, the inclusion of additional cell types is warranted. To enhance the generalizability of the link between immunoscore and patient prognosis, there is a need to expand the cohort of tissues subjected to imaging analysis.

Moreover, considering the promising potential targets identified in this study, future investigations should explore these findings through drug screening and brainoid models. This proactive approach could shed light on the translational implications of the study and facilitate the development of targeted interventions.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-CD68	Miralys™ PC-MT Probe	Cat#AP1001124
Anti-CD8α	Miralys™ PC-MT Probe	Cat#AP100152
Anti-Ki67	Miralys™ PC-MT Probe	Cat#AP1001184
Anti-VIM	Miralys™ PC-MT Probe	Cat#AP1001122
Anti-Collagen-1A1	Miralys™ PC-MT Probe	Cat#AP100153
Antin-CD3ε	Miralys™ PC-MT Probe	Cat#AP1001181
Anti-CD4	Miralys™ PC-MT Probe	Cat#AP100173
Anti-CD3 PE (clone SK7)	Sony Biotechnology	Cat#2324030
Anti-CD4 FITC (clone A161A1)	Sony Biotechnology	Cat#2387030
Anti-CD8 Alexa Fluor 647 (clone HIT8a)	Sony Biotechnology	Cat#2104590
Anti-CD7 PE Cy5 (clone CD7-6B7)	Sony Biotechnology	Cat#2315550
Biological samples
Healthy and cancerous glioma tissues	University Hospital Center of Lille	https://www.chu-lille.fr
Blood samples for macrophages and lymphocytes	French blood agency (EFS)	https://dondesang.efs.sante.fr
Chemicals, peptides, and recombinant proteins
Hemalun	Merck	Cat#109249
Phloxin	VWR	Cat#10047229
Saffron	MM France	Cat#F/SAFRAN
Ethanol (EtOH)	Carlo Erba Reagents	Cat#4127012
Xylene	Carlo Erba Reagents	Cat#492301
Glycerol	Reidel de Haen	Cat#15523
Isopropanol (IPA)	Carlo Erba Reagents	Cat#415183
DMEM	Thermo Fischer Scientific	Cat#12-491-015
RPMI Medium	Thermo Fischer Scientific	Cat#31870025
DPBS, no calcium, no magnesium	Thermo Fischer Scientific	Cat#14190-094
Ficoll	Dominique Dutscher	Cat#17-5442-02
EDTA	Sigma Aldrich	Cat#03690
Leucine enkephalin	Waters	Cat#70002456
Phorbol-12-myristate-13-acetate (PMA)	Sigma Aldrich	Cat#P8139
Lypopolysachharide (LPS)	Invivogen	Cat#tlrl-3pelps
Macrophage colony stimulating factor (MCSF)	Ozyme	Cat#BLE574802
IFNγ	PeproTech	Cat#300-02
Il4	PeproTech	Cat#200-04
2,5-dihydroxybenzoic acid	Sigma Aldrich	Cat#149357
Methanol (MeOH)	Carlo Erba Reagents	Cat#4148831
Chloroform	Carlo Erba Reagents	Cat#508320
Trifluoroacetic (TFA)	Sigma Aldrich	Cat#T6508
Paraformaldehyde (PFA)	Alfa Aesar	Cat#43368
Acetone	VWR	Cat#20067.320
Acetic acid	Sigma Aldrich	Cat#320099
TRIS-HCl	Promega	Cat#H5125
TRIS-Base	Promega	Cat#H5135
NaCl	Fischer chemicals	Cat#S/3161/60
Octylβ-D-glucopyranoside (OG)	Sigma Aldrich	Cat#03757
Ammonium bicarbonate (NH4HCO3)	Sigma Millipore	Cat#149357
Deposited data
The raw data from SpiderMass analysis has been deposited to the Harvard dataverse site.	This paper	Harvard Dataverse: https://doi.org/10.7910/DVN/RUULD8
Experimental models: Cell lines
Human NCH82 stage IV glioma	Collaboration with Dr Regnier-Vigouroux	N/A
Software and algorithms
GraphPad Prism v9.5.1	GraphPad software	RRID: SCR_000306 https://www.graphpad.com
Abstract Model Builder (AMX) version 1.0.2053.0	Waters research center, Hungary	N/A
Python 3.11.5	Open source	RRID: SCR_008394 https://www.python.org
Jupyter Notebook 6.5.4	Open source	https://jupyter.org
FlexImaging 5.0	Bruker	https://www.bruker.com/fr/products-and-solutions/mass-spectrometry/maldi-tof/rapiflex-maldi-tissuetyper.html
SCiLS Lab 2022a Pro	Bruker and SCiLS GmbH	RRID: SCR_014426 http://scils.de/software/
QuPath 0.4.4	Bankhead et al.56	RRID: SCR_018257 https://qupath.github.io
MATLAB R2019a	MathWorks	RRID: SCR_001622 https://matlab.mathworks.com
MetFrag	Ruttkies et al.59	https://msbi.ipb-halle.de/MetFrag/
MassLynx 4.1	Waters	RRID: SCR_014271 éhttps://www.waters.com/waters/fr_FR/Logiciel-de-Spectrométrie-de-Masse-MassLynx/nav.htm?cid=513662&locale=fr_FR
The original code has been deposited on GitHub. The O-DAPIA script for one dimensional SpiderMass data were released and archived in Zenodo	This paper	https://zenodo.org/doi/10.5281/zenodo.10656851
The original code has been deposited on GitHub. The T-DAPIA script for SpiderMassMSI data were released and archived in Zenodo	This paper	https://zenodo.org/doi/10.5281/zenodo.10656830

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Isabelle Fournier (isabelle.fournier@univ-lille.fr).

Materials availability

This study did not generate new unique reagents.

Data and code availability

All SpiderMass raw data (MS spectra, MSMS spectra and MS imaging data) have been deposited at Harvard dataverse and are publicly available as of the date of publication. DOI is listed in the key resources table.

All original code has been deposited at GitHub (https://github.com/yanisZirem/Pipeline_SpiderMass1D and https://github.com/yanisZirem/Pipeline_SpiderMass2D) and was archived in Zenodo. The DOIs are available in the key resources table. If you have any questions or feedback, please contact yanis.zirem2016@gmail.com.

Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Experimental model and study participant details

Study design

According to the French Public Health Code and in application of the General Data Protection Regulations, all patients had been informed at the time of care that their standard clinical and biological data could be used for research purposes regarding the retrospective analysis of FFPE samples, and none had expressed his opposition. Regarding the prospective collection of samples, each patient’s informed consent for the collection and publication of clinical and biological data was obtained at the time of hospitalization prior to surgical intervention.

Patients

A retrospective cohort of 50 FFPE glioblastoma tissues was obtained from the Pathology department of Lille Hospital, France. A prospective cohort of 31 fresh frozen glioblastoma tissues were also included in this study. 31 patients with newly diagnosed glioblastoma were prospectively enrolled between September 2014 and November 2018 at Lille University Hospital, France (NCT02473484). All patients gave written informed consent before enrollment. The demographic and clinical data of the cohorts are reported in Table S1.

Cell line

Human NCH82 stage 4 glioblastoma cells were obtained from Dr Regnier-Vigouroux.

Method details

Histological staining and annotations

The paraffin blocks and the fresh-frozen tissues were sliced into 7 μm sections and placed on poly-lysine coated slides. This tissue slice consecutive to the SpiderMass-analyzed tissue was treated with hemalum solution for 1 min, then rinsed with tap water. The tissue section was dehydrated in 70% and 100% ethanol baths after being dyed in phloxine 0.1% solution for 10 s. The section was cleaned in xylene, cleaned twice in alcohol, dipped in saffron for 5 s then mounted with cover slips and the EUKITT slide mounting media. Using the Panoramic MIDI slide scanner, the stained slide was scanned to capture a digital image (3DHISTECH LTD. Budapest, Hungary) and the images were viewed and exported using QuPath 0.4.4.⁶¹ On tissues that had been stained, annotations were made.

Sample preparation

All the FFPE blocks were sectioned into 8 μm tissue sections using a microtome (Leica Biosystems, Wetzlar, Germany) at room temperature. A dewaxing step was performed through two 5 min washes in a xylene solution. Then, each dewaxed tissue section of the retrospective glioblastoma cohort was manually sprayed with a glycerol/isopropyl alcohol (IPA) (8:2, v/v) solution in two successive passes using a manual sprayer (Agilent). The syringe pump (74900 series Cole Parmer Instrument Company) was set to a 500 μL/min flow rate. As for fresh frozen tissues, a Leica CM1510S cryostat (Leica Microsystems, Nanterre, France) was used to cut 20 μm sections. Their analysis by the SpiderMass does not require any sample preparation.

SpiderMass analysis

The overall layout of the instrument setup has already been covered elsewhere.³⁰ In brief, the system is made up of three parts: the mass spectrometer itself, a laser system for remote micro-sampling of tissues and a transfer line allowing for the transfer of the micro-sampled material. The first component consists of a pulsed Nd:YAG laser (pulse duration: 5 ns, = 1064 nm, Quantel, Les Ulis, France) pumping a tunable wavelength OPO (Radiant version 1.0.1, OPOTEK Inc., Carlsbad, CA, USA). A handpiece with a 4 cm focusing lens is attached to the end of the biocompatible laser fiber, which is connected to the laser system output and has an inner diameter of 450 microns and a length of 1 m. In these studies, the laser intensity was set to 4 mJ/pulse for a fixed irradiation time of 10 s, resulting in a laser fluence of approximately 3 J/cm². The second component of the system is a 2 m transfer line made of Tygon ND 100-65 tubing (Akron, Ohio, USA, 2.4 mm inner diameter, 4 mm outer diameter). The transfer line is directly connected to the mass spectrometer (Xevo G2-XS, Waters, Manchester, United Kingdom) from which the conventional electrospray source was removed and replaced by a REIMS interface on one side and is attached to the laser handpiece on the other. A 200 μL/min infusion of isopropanol was administered before each acquisition. 200 μg/mL of Leucine enkephalin was added to the infusion to play the role of a lockmass. The sampling position was determined based on the histopathological annotations. The acquisition was composed of a burst of 10 laser shots resulting in an individual spectrum. Spectral acquisition was performed both in positive and negative ion mode in sensitivity mode with a scan time of 1 s. The mass range was set to m/z 50–2000.

SpiderMass MSI

The SpiderMass setup was described in the previous section. To perform imaging analysis, the Spider-Mass microprobe was coupled to a stiff robotic arm described in a previous work.⁶² The spatial step size was set to 250 μm for fresh frozen tissue to achieve oversampling. The mass-range was fixed between m/z 100–1500. The acquisition sequence was composed of 3 consecutive laser shots and 3 s between each step. The laser bursts and the spectrometer acquisition were automatically triggered through a MATLAB in-house user interface developed for the robotic WALDI-MSI.⁶² The data was acquired in negative and sensitivity ion mode.

Cell line and immune cell analysis

Human NCH82 stage 4 glioma cells were obtained from Dr Regnier-Vigouroux. These cells were cultured in DMEM. The medium was supplemented with 10% fetal bovine serum and 100 U/mL penicillin-streptomycin in a humidified air incubator at 37°C under an atmosphere of 5% CO₂. After 70% confluence, cells were washed two times with DPBS, dried under PSM during 10 min at RT than analyzed by SpiderMass directly into cell plate. First, to obtain primary macrophages, 50 mL of blood, from Etablissement Français du Sang (EFS), were diluted two times into PBS-EDTA than leukocytes were isolated thanks to 25 min centrifugation 2200 rpm with a Ficoll gradient. Leukocytes were washed three times with PBS-EDTA than resuspended into RPMI medium and incubated into a cell plate 1h30 at 37°C. Cell plate were washed three times with PBS. Macrophages were than grown 7 days with RPMI medium with 10% fetal bovine serum, 100 U/mL penicillin-streptomycin and MCSF. Next, primary macrophages were stimulated into two different conditions. M1-like macrophages were stimulated with 0.5 mg/mL of LPS and 20 ng/mL of IFN-γ during 48h in contrary to M2-like macrophages that were stimulated with 20 ng/mL of Il4 during 48h. Like NCH82 cells, macrophages were washed two times with DPBS, dried under PSM during 10 min at RT than analyzed by SpiderMass directly into cell plate. In addition, primary lymphocytes were isolated. For that, peripheral blood mononuclear cells (PBMCs) were isolated from whole blood samples using Ficoll density gradient centrifugation. Then, cells were labeled with mix antibodies (Sony): anti-CD3 PE (clone SK7), anti-CD4 FITC (clone A161A1), anti-CD8 APC (clone HIT8a) and anti-CD7 PE-Cy5 (clone CD7-6B7) for 20 min at 4°C in dark. After washing, CD3+-CD4⁺ cells, D3+-CD8⁺ cells and CD3⁻CD7⁺ cells were sorted using the BD FACS ARIA II SORP. One million of each population was transferred onto glass slides using a Cytospin centrifuge (Thermo Shandon Cytospin) and conserved at −80°C. The SpiderMass analysis were made directly onto the glass slide.

MALDI-MSI

Nine prospective tumors were analyzed using MALDI-MSI. For this purpose, 12 μm sections were cut using a Leica CM1510S cryostat (Leica Microsystems, Nanterre, France), and these sections were then placed on ITO-coated glass slides from LaserBioLabs (Valbonne, France). The application of the 2,5-dihydroxybenzoic acid (DHB) matrix was made thanks to a manual sprayer. The MALDI-MSI analyzes were carried out utilizing an Ultraflex II MALDI-TOF/TOF instrument (Bruker) operating in positive ion mode. The spatial resolution was adjusted at 70 μm, and the mass range was fixed at m/z 60–1000. Subsequently, the MALDI-MSI data was subjected to analysis employing SCiLS Lab software (SCiLS Lab 2022a PRo, SCiLS GmbH). The data was normalized using Total Ion Count (TIC) normalization. The segmentation of the images was then performed using the bisecting k-means algorithm, facilitating global and individual segmentation across the nine images. This comprehensive spatial segmentation enabled the identification of regions of interest, which were found to correspond with those obtained through proteomic data.

MALDI-IHC

Multiplex imaging was conducted on two fresh-frozen GBM tissue samples that were previously analyzed using SpiderMass-MSI. One tissue sample had a survival rate of less than 10 months, while the other had an OS of more than 36 months. The MALDI-IHC analysis was made on an adjacent tissue section from the same tumor analyzed by SpiderMass. The tissue preparation and imaging protocol utilized was the recommended one for AmberGen (Billerica, MA) Miralys probes. Initially, the tissues were vacuum-dried for 10 min and then fixed with 1% PFA for 30 min. Subsequently, the tissues underwent a series of baths: one bath in PBS for 10 min, two baths in acetone for 3 min each, and one bath in Carnoy solution for 3 min. Following this, the tissues were rehydrated through two baths in 100% ethanol for 2 min each, succeeded by three consecutive baths in 95% EtOH, 70% EtOH, and 50% EtOH, each for 3 min. A 10-min TBS bath (50 mM Tris, pH 7.5, 200 mM NaCl) preceded antigen retrieval, which occurred in 20 mM Tris buffer at pH 9 for 30 min at 95°C. After a 10-min TBS wash, the tissues were treated with a tissue blocking buffer (2% each of normal mouse and rabbit serum and 5% BSA in TBS-OG [TBS with 0.05% w/v Octyl β-D-glucopyranoside]) for 1 h. Following this, the tissues were incubated at 4°C overnight in the same blocking buffer, which contained CD68, CD8α, Ki67, Vimentin, and collagen probes at a concentration of 2.5 μg/mL. Each slide was individually washed with three 5-min TBS baths and three 2-min baths in 50 mM NH₄HCO₃, all conducted in darkness. The tissues were then vacuum-dried for 1 h and 30 min at room temperature before subjecting them to a 365 nm UV exposure (Miralys Light Box from AmberGen, Inc., Billerica, MA) for 10 min to cleave the probes. DHB matrix at a concentration of 20 mg/mL in MeOH:TFA 0.1% (70:30, v/v) was sprayed onto the tissues using the HTX sprayer M5 from HTX technologies, LLC (Chappel Hill, NC). The two slides were subjected to MALDI-MSI analysis using a rapifleX MALDI-TOF-MS instrument (Bruker Daltonics, Germany) in reflector mode, positive ion mode, with a laser spot size of 20 μm and continuous raster scanning of 20 μm. Each pixel underwent 500 laser shots, and a TIC normalization was employed for multiplex image processing. The resulting images were analyzed using flexImaging (Bruker Daltonics, Billerica, MA).

Data analysis

Preprocessing and data importation

Before obtaining the matrix data for each dataset, several preprocessing steps were applied. First, the data was binned to 0.1 Da to reduce the number of data points. Then, all the spectra were aligned based on the lockmass (m/z 554.2 and 556.2 in negative and positive ion mode respectively) and the mass range was set between 600 and 1100 m/z. Finally, the total ion count (TIC) normalization was applied to remove any intensity variations between spectra. All final datasets contain 5000 m/z data points. For the data importation, there are two methods to obtain the matrix data from Waters RAW files. The first method involved a conversion of the raw files to mzML files using the MSConvert (Proteowizard)⁶³ and an importation of the mzML files to Python using the pyopenMS library. The second method, mainly used, included the importation of the raw files into “Abstract Model Builder” - AMX (version 1.0.2053.0, Waters research center, Hungary) and the extraction of the matrix data as csv files. The pandas library is used to import the csv files into Python.

Optimal classification model, cross-validation and blind prediction

The Lazy Predict library (https://lazypredict.readthedocs.io/en/latest/) was used to build multiple models from the scikit-learn library by training and testing a range of 24 classifiers. The random state was always kept at 1. Subsequently, the optimal model was reconstructed individually using the scikit-learn library, which enabled tuning of its parameters for optimization and evaluation of its accuracy. To further evaluate the model’s performance, 20-fold cross-validation was performed using KFold and cross_val_score functions, and the classification report was generated using the classification_report function. Additionally, the ConfusionMatrixDisplay function from the matplotlib library was used to display the confusion matrix. The optimal model was then saved and loaded for blind prediction using the joblib library, with the joblib.dump and joblib.load functions.

Prediction explanations

The Local Interpretable Model-agnostic Explanations (LIME) algorithm³⁶ was employed to explain the classification model. This algorithm calculates feature contributions that can be positive or negative. The ELI5 library (https://eli5.readthedocs.io/en/latest/overview.html) was utilized to generate a LIME table containing the weight of feature contributions using the explain_prediction function.

Significant features

A peak picking algorithm, the find_peaks_cwt function from the scipy library was applied to remove instrument noise. Then a clustering heatmap was generated based on the peak list. The heatmap used hierarchical clustering between m/z variables with euclidean distance, displaying the ions that were more or less abundant in each target clustered together. For this purpose, the clustermap function from the seaborn library was utilized. Next, a non-parametric statistical test Kruskal-Wallis with Bonferroni correction, employing the stats.kruskal function from the Scipy library, was used to evaluate the significance of each feature. Only significant features with a p value of equal or less than 0.05 were retained. Finally, a step of filtering was added to only keep the mono-isotopic peaks. The corresponding boxplots were then generated from the seaborn library.

Immunoscores

The immunoscoring model was trained using the LGBM (Light Gradient Boosting Machine)⁴⁶ Python library, a gradient boosting framework developed by Microsoft. This model leveraged cell spectra within the m/z range of 600–1100 in negative ion mode. The spectra were categorized into different cell types: macrophages (M1 & M2) with 107 spectra, cancer cell line (NCH82) with 107 spectra and lymphocytes (NK, CD8, CD4) with 107 spectra. Due to the lack of spectra for each type of lymphocyte and to prevent creating an inconsistent (unbalanced classes) model, the lymphocytes were all regrouped. To predict the cell types on SpiderMass images, the predict_proba function of the LGBM model was utilized. This function provided probability estimates for each cell type, allowing for a more nuanced understanding of the likelihood of each cell type’s presence in the local environment. Furthermore, ratio scores were computed to estimate the probability of presence over the whole sample, which is interpreted as a relative presence of each cell type across the entire image. These scores were calculated by summing the scores provided for each cell type and dividing it by the sum of the total scores across all labels. The ratios provided insights into the distribution of the trained cell types throughout the image, offering a comprehensive assessment of the cellular landscape in the analyzed sample.

MALDI-IHC quantification

The Python PIL library was employed to bring in the IHC image in PNG format. Subsequently, all the images underwent a grayscale conversion and were transformed into NumPy arrays with the help of the NumPy library. An algorithm was designed and applied to compute the count of black pixels in the image using the NumPy array, where each channel ranges from 0 (representing black) to 255 (representing white). The quantification of each IHC marker was determined by subtracting the count of black pixels from the total image pixels. To visualize the image and generate a bar plot, the matplotlib.pyplot library was utilized.

MS/MS analysis

SpiderMass technology facilitated the MS/MS investigation using the Xevo G2-XS instrument from Waters. MS/MS spectra were recorded after the isolation of the precursor ion and they were subjected to collision-induced dissociation (CID) in the transfer cell with a transfer collision energy ranging between 30 and 40 V, depending on the selected precursor ion. The lipid annotations were performed manually through fragmentation spectra guidelines and compared to LipidMaps database, Alex123, MetFrag database⁶⁴ and literature.

Quantification and statistical analysis

Three datasets were employed for training, cross-validation, and testing of all classification models, including those for tissue type, tissue region, prognosis, and immunoscore. Evaluation involved a 20% validation split and 20-fold cross-validation, with a classification report providing metrics such as accuracy, recall, precision, and F1 score. To assess the statistical significance of biomarkers, a non-parametric Kruskal-Wallis test was employed. Bonferroni corrections were applied to adjust p values for multiple comparisons. Values are presented as medians and visualized through scatter boxplots.

Published: March 28, 2024