Abstract Background The response to neoadjuvant chemotherapy (NAC) differs substantially among individual patients with non‐small cell lung cancer (NSCLC). Major pathological response (MPR) is a histomorphological read‐out used to assess treatment response and prognosis in patients NSCLC after NAC. Although spatial metabolomics is a promising tool for evaluating metabolic phenotypes, it has not yet been utilized to assess therapy responses in patients with NSCLC. We evaluated the potential application of spatial metabolomics in cancer tissues to assess the response to NAC, using a metabolic classifier that utilizes mass spectrometry imaging combined with machine learning. Methods Resected NSCLC tissue specimens obtained after NAC (n = 88) were subjected to high‐resolution mass spectrometry, and these data were used to develop an approach for assessing the response to NAC in patients with NSCLC. The specificities of the generated tumor cell and stroma classifiers were validated by applying this approach to a cohort of biologically matched chemotherapy‐naïve patients with NSCLC (n = 85). Results The developed tumor cell metabolic classifier stratified patients into different prognostic groups with 81.6% accuracy, whereas the stroma metabolic classifier displayed 78.4% accuracy. By contrast, the accuracies of MPR and TNM staging for stratification were 62.5% and 54.1%, respectively. The combination of metabolic and MPR classifiers showed slightly lower accuracy than either individual metabolic classifier. In multivariate analysis, metabolic classifiers were the only independent prognostic factors identified (tumor: P = 0.001, hazards ratio [HR] = 3.823, 95% confidence interval [CI] = 1.716–8.514; stroma: P = 0.049, HR = 2.180, 95% CI = 1.004–4.737), whereas MPR (P = 0.804; HR = 0.913; 95% CI = 0.445–1.874) and TNM staging (P = 0.078; HR = 1.223; 95% CI = 0.977–1.550) were not independent prognostic factors. Using Kaplan‐Meier survival analyses, both tumor and stroma metabolic classifiers were able to further stratify patients as NAC responders (P < 0.001) and non‐responders (P < 0.001). Conclusions Our findings indicate that the metabolic constitutions of both tumor cells and the stroma are valuable additions to the classical histomorphology‐based assessment of tumor response. Keywords: cancer metabolism, machine learning, mass spectrometry imaging, metabolic classifier, Non‐small cell lung cancer, prognosis, spatial metabolomics, treatment response __________________________________________________________________ Abbreviations Chemo‐naïve chemotherapy‐naïve CI confidence interval CPA cyclic phosphatidic acid EGFR epidermal growth factor receptor FFPE formalin‐fixed paraffin‐embedded H&E hematoxylin and eosin HR hazards ratio IQR interquartile range LUAD lung adenocarcinoma LUSC lung squamous cell carcinoma LysoPA lysophosphatidic acid LysoPC lysophosphatidylcholine LysoPE lysophosphatidylethanolamine LysoPI lysophosphatidylinositol MALDI‐FT‐ICR‐MSI matrix‐assisted laser desorption/ionization fourier‐transform ion cyclotron resonance mass spectrometry imaging MC metabolic classifier MPR major pathological response NAC neoadjuvant therapy NSCLC non‐small cell lung cancer OCFAs odd‐chain fatty acids OS overall survival PA phosphatidic acid PC phosphatidylcholine PE phosphatidylethanolamine PGP phosphatidylglycerophosphate PI phosphoinositol RF random forest SM sphingomyelin SPACiAL spatial correlation image analysis TG triglyceride TNM tumor‐node‐metastasis 1. BACKGROUND Neoadjuvant chemotherapy (NAC), with or without radiotherapy, followed by surgical resection, improves survival in patients with locally advanced non‐small cell lung cancer (NSCLC) compared with surgery alone, particularly among patients with complete pathological response or major pathological response (MPR), which is classically defined as a residual tumor burden of <10% [[46]1, [47]2, [48]3, [49]4]. NAC has become a vital strategy for reducing tumor size and facilitating surgical resection. NAC also allows for intermediate evaluations of treatment response and prevents the development of micrometastases [[50]5, [51]6]. Along with the recent successes reported for targeted and immune checkpoint therapies in advanced inoperable NSCLC, recent studies indicate that the adjuvant use of these regimens is also beneficial [[52]7, [53]8]. However, evidence supporting the therapeutic efficacy of these regimens remains scarce in the neoadjuvant setting, although preliminary outcomes reported for immune checkpoint inhibitors [[54]9] and epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors have been promising [[55]10]. The development of more accurate biomarkers for patient stratification or use as surrogate endpoints would likely result in the efficient identification of patients with resectable or potentially resectable NSCLC. Accurate patient stratification is becoming increasingly important. The pathological tumor‐node‐metastasis (pTNM) classification is the most important and routinely applied prognosis prediction tool for malignant disease. MPR has been associated with long‐term overall survival (OS) among patients with NSCLC who undergo NAC [[56]4, [57]11, [58]12]. MPR represents an estimate of residual tumor cell quantity, used to evaluate tumor regression following NAC. Generally, MPR is used to evaluate the therapeutic response and is often used as a surrogate endpoint in clinical studies of resected NSCLC following preoperative therapies [[59]13, [60]14, [61]15, [62]16]. Until recently, 10% of the baseline tumor has been used as the cutoff value for defining residual tumor, which has been associated with significant prognostic value [[63]4, [64]17‐[65]^19]. However, we [[66]20] and others [[67]16, [68]21] have shown that although a 10% cutoff is appropriate for lung squamous cell carcinoma (LUSC), the cutoff for lung adenocarcinoma (LUAD) should be greater than 50% [[69]16, [70]20, [71]21]. The response to NAC presents substantial variability, ranging from complete to subtotal residual tumor regression, and using a cutoff value of 10%, MPR is only achieved in 17%–33% of patients, depending on the therapeutic strategy [[72]13, [73]22, [74]23]. Due to clinicopathologic heterogeneity and different biological behaviors within the tumor, patients with NSCLC at similar pathological stages may have notable prognostic differences [[75]24]. Metabolomics is recognized as a crucial scientific field, offering a promising avenue for identifying diagnostic and prognostic biomarkers for use in clinical practice and may serve as a powerful method for screening potential biomarkers in LUAD [[76]25, [77]26]. Spatial metabolomics is an emerging domain of omics research. It could be used to investigate tumor heterogeneity [[78]27], evaluate surgical resection margins [[79]28], and contribute to tumor classification [[80]29]. Spatial metabolomics and its enabling technology—matrix‐assisted laser desorption/ionization mass spectrometry imaging (MALDI‐MSI)—localize hundreds to thousands of different metabolites directly from biological tissue sections with cellular spatial resolution [[81]30, [82]31, [83]32, [84]33]. The comprehensive analysis of metabolic heterogeneity and the use of MALDI‐MSI have improved our understanding of tumor metabolism [[85]34]. Previous studies using MALDI‐MSI to analyze human specimens, including tumor tissue and body fluids, identified several biomarkers associated with lung cancer and clinical outcomes [[86]35, [87]36, [88]37, [89]38]. Our previous study found that the high‐mass‐resolution matrix‐assisted laser desorption/ionization Fourier‐transform ion cyclotron resonance mass spectrometry imaging (MALDI‐FT‐ICR‐MSI) was suitable for deciphering therapeutic effects and allowed for the assessment of metabolic changes that occur during the treatment of idiopathic pulmonary fibrosis [[90]39]. The metabolic compositions of both tumor cells and stroma were able to provide rich molecular information and may contribute to estimating prognosis in patients with NSCLC following NAC. Spatial metabolomics enables immunophenotype‐guided in situ metabolomics, facilitating the automated identification of histological and functional features in intact tissue sections and the comprehensive analyses of metabolic constitutions of tumor cells and the stroma [[91]27]. To date, no published studies have examined the ability of spatial metabolomics and metabolite identification to characterize treatment response and differentiate patients into non‐responders and responders in NSCLC. Therefore, the purpose of the present study was to investigate whether spatial metabolomics could be applied for the evaluation of NAC treatment response and prognosis in NSCLC. 2. MATERIALS AND METHODS 2.1. Patient samples and tissue microarrays This study included two retrospective single‐center patient cohorts (Figure [92]1), comprising cases diagnosed at the Institute of Pathology of the University of Bern (Bern, Switzerland) between 2000 and 2016. All eligible patients had a pathology‐confirmed diagnosis. The NAC cohort included consecutive patients who received at least one cycle of platinum‐based chemotherapy prior to resection [[93]20]. All applied drug combinations for this cohort are summarized in Table [94]1. The chemotherapy‐naïve cohort included consecutive patients who underwent primary resection for NSCLC without prior chemotherapy or radiotherapy and were histologically and biologically matched with the NAC cohort, as previously described [[95]40]. Biological matching was accomplished by including only locally advanced NSCLC (at least stage IIIA), ensured by mediastinal lymph node metastasis (pN2), which qualifies for a multi‐disciplinary treatment approach integrating neoadjuvant or adjuvant systemic therapy. We did not statistically compare the stage distribution and pN distribution between the neoadjuvant and chemotherapy‐naïve cohorts due to the downstaging that resulted by NAC and the study design inherent bias towards only locally advanced tumors in the chemotherapy‐naïve cohort, chosen for better biological comparability. Exclusion criteria included patients who died within 30 days after surgery and patients without materials for appropriate evaluations of tumor regression. For the NAC cohort, patients were excluded if chemotherapy was applied without neoadjuvant intent prior to resection. FIGURE 1. FIGURE 1 [96]Open in a new tab Study design for the development of a metabolic classifier and the assessment of the predictive abilities of clinicopathological features (MPR and TNM staging) and metabolic factors for stratifying patients with NSCLC. Separate metabolic classifiers were established using the metabolic features evaluated in tumors and the stroma, resulting in individual survival risk categories, which were evaluated by Kaplan‐Meier and Cox regression analyses and compared with standard clinicopathological evaluations using MPR and TNM staging. Abbreviations: NAC, neoadjuvant chemotherapy; Chemo, chemotherapy; MALDI MSI, matrix‐assisted laser desorption/ionization mass spectrometry imaging; MPR, major pathological response; TNM, tumor‐node‐metastasis; NSCLC, non‐small cell lung cancer; RF, random forest TABLE 1. Baseline clinicopathologic characteristics of the the two cohorts of patients with NSCLC, who either have received neoadjuvant therapy before resection (NAC) or were chemotherapy‐naïve Characteristics NAC (cases [%]) Chemotherapy‐naïve (cases [%]) P value Total 88 85 Sex 0.529 Male 59 (63.6) 53 (62.4) Female 29 (36.4) 32 (37.6) Smoking status 0.934 Non‐smoker 12 (13.6) 9 (10.6) Ex‐smoker 29 (33.0) 26 (30.6) Active smoker 37 (42.0) 33 (38.8) No record 10 (11.4) 17 (20.0) MPR NA Present 37 (43.5) NA Absent 51 (56.5) NA ypTNM stage [97]^* 0 6 (6.9) NA I 17 (19.5) NA II 20 (23.0) NA III 40 (44.8) 77 (90.6) 0.766 IV 5 (5.8) 8 (9.4) ypT stage [98]^* pT0 11 (12.5) NA pT1 26 (29.5) 17 (20.0) 0.259 pT2 22 (25.0) 28 (32.9) pT3 13 (15.9) 19 (22.4) pT4 16 (18.1) 21 (24.7) ypN stage [99]^* pN0 35 (39.8) NA pN1 20 (22.7) NA pN2 31 (35.2) 83 (97.6) 0.318 pN3 2 (2.3) 2 (2.4) ypM stage [100]^* 0.352 pM0 83 (94.3) 77 (90.6) pM1 5 (5.7) 8 (9.4) Subtype 0.714 LUSC 40 (45.5) 41 (48.2) LUAD 48 (54.5) 44 (51.8) Neoadjuvant Chemotherapy NA Cisplatin + docetaxel 50 (56.8) NA Carboplatin + paclitaxel 3 (3.4) NA Cisplatin + pemetrexed 12 (13.6) NA Cisplatin + gemcitabin 7 (8.0) NA Cisplatin + vinorelbin 9 (10.2) NA Others 7 (8.0) NA [101]Open in a new tab ^* pTNM/pT/pN/pM stage for the Chemotherapy‐naïve cohort. Abbreviations: NSCLC, non‐small cell lung cancer; NAC, neoadjuvant therapy; NA, not applicable; MPR, major pathological response; TNM, tumor‐node‐metastasis; T, component/category coding the extension of the primary tumor; N, component/category coding regional lymph node metastases; M, component/category coding distant metastases; p, pathological TNM staging; yp, pathological TNM staging after neoadjuvant therapy; LUSC, lung squamous cell carcinoma; LUAD, lung adenocarcinoma Clinicopathological features and follow‐up data were retrieved, as previously described [[102]20], and all data were thoroughly re‐evaluated to update pathological tumor stages according to the Union for International Cancer Control (UICC) and the American Joint Committee on Cancer (AJCC) 8th edition pTNM classification guidelines and harmonize regression grading among other scales [[103]41]. OS was defined as the time elapsed between treatment initiation and death of any cause. Routine clinical follow‐up was performed for all patients, and all available information regarding relapse and disease progression was retrieved from the clinical files. Response to neoadjuvant therapy was histomorphologically assessed by a pathologist specialized in pulmonary pathology (Sabina Berezowska) and a pulmonary pathology‐experienced student (Philipp Zens) for each case, and the histological data including MPR were previously reported [[104]20]. Residual tumor content was assessed by the histological evaluation of all slides containing the tumor bed, as previously described [[105]4]. MPR was defined as ≤10% residual tumor cells for LUSC or as ≤65% residual tumor cells for LUAD, as previously described [[106]16, [107]20]. Patients were classified as NAC responders (MPR present) and non‐responders (MPR absent). Metabolic analysis was performed on tissue microarrays (TMAs) constructed for each cohort, as previously reported [[108]42]. From each patient at least two tumor‐containing tissue cores were collected. Briefly, 0.6‐mm‐diameter tissue cores were annotated on FFPE tissue blocks/slides by a pathologist specialized in lung pathology (Sabina Berezowska) and transposed into an acceptor TMA block. The study was approved by the Cantonal Ethics Commission of the Canton of Bern (KEK 2017‐00830), which waived the requirement for a written informed consent from patients. 2.2. High‐mass‐resolution MALDI‐FT‐ICR‐MSI analysis MALDI‐MSI was performed as previously described [[109]32, [110]43]. NSCLC TMA blocks were cut into 4‐μm sections using a microtome (HM 355S, Microm; Thermo Fisher Scientific, Waltham, MA, USA) and mounted onto indium tin oxide‐coated conductive glass slides (Bruker Daltonik GmbH, Bremen, Germany). The slides were coated in 1:1 poly‐L‐lysine (Sigma‐Aldrich, Taufkirchen, Germany) and 0.1% Nonidet P‐40 (Sigma‐Aldrich) before tissue mounting. The tissue sections were incubated at 60°C for 1 h and deparaffinized in xylene (2 × 8 min), followed by drying at room temperature (22°C). Subsequently, the samples were covered with 10 mg/mL 9‐aminoacridine hydrochloride monohydrate matrix (Sigma‐Aldrich) in 70% methanol using a SunCollect sprayer (Sunchrom, Friedrichsdorf, Germany). The matrix application was performed in eight passes using ascending spray rates (flow rates: 10, 20, 30 μL/min for the first three layers, followed by 40 μL/min for the remaining five layers). MALDI‐FT‐ICR‐MSI was performed on a Bruker Solarix 7T FT‐ICR MS (Bruker Daltonik) in the negative ion mode, utilizing 100 laser shots per pixel at a frequency of 1000 Hz. Mass spectra were acquired over a mass range of m/z 75‐1000 Da with a 50‐μm spatial resolution. 2.3. Immunophenotype‐guided MSI analysis, data processing, and pathway analysis The Spatial Correlation Image Analysis (SPACiAL) workflow was used for the immunophenotype‐guided MALDI‐MSI analysis of automatically annotated tumor and stroma regions in NSCLC tissues, as previously described [[111]27]. SPACiAL is a computational multimodal workflow that includes a series of image and MALDI data processing steps to combine molecular imaging data with multiplex immunofluorescence. The SPACiAL workflow includes MALDI and immunofluorescence data integration, multiple image co‐registration, image digitization, and data conversion. After MALDI imaging, the matrix was removed from the section surface by a 5 min incubation in 70% ethanol, and sections were subsequently stained with hematoxylin and eosin (H&E) in a HistoCore SPECTRA ST multistainer (Leica, Wetzlar, Germany). To remove the H&E stain, we incubated the sections in a Coplin jar containing 100% xylene at room temperature (22°C) for 12 h. The slides were then transferred to a second Coplin jar containing 100% xylene for a 2‐min incubation, to a third Coplin jar containing pure propanol for 2 min, to a fourth Coplin jar containing 100% ethanol for 2 min, and to a fifth Coplin jar containing 1% HCl in 100% ethanol for 5 min. The slides were then washed under running tap water for 5 min. The tissue sections were subjected to immunofluorescence after H&E removal and analyzed by double staining using pan‐cytokeratin (monoclonal mouse pan‐cytokeratin plus [AE1/AE3þ8/18] [1:75], cat#CM162, Biocare Medical, Pacheco, CA, USA) and vimentin antibodies (1:500, clone ab92547, Abcam, Berlin, Germany). Signal detection was conducted using fluorescence‐labeled secondary antibodies (anti‐rabbit IgG DyLight 633 [cat# 35563] and anti‐mouse IgG Alexa Fluor 750 antibody [cat# A‐21037], Thermo Fisher Scientific), and Hoechst 33342 was used for nuclear staining. Automated steps for the analyses and annotation of tumor and stroma regions were applied to mass spectrometry data using SPACiAL as follows: first, the epithelial marker pan‐cytokeratin (white) was used to stain tumor cells, and vimentin was used to stain stroma regions (red); second, single‐channel images of pan‐cytokeratin and vimentin were used to annotate and separate tumor and stroma using fluorescence imaging; third, the digitized and co‐registered fluorescence images were scaled to match the exact MALDI resolution and converted into numerical matrices comprised of values corresponding to the lightness values for each pixel; fourth, objective tissue annotations were assigned based on semantics and function. The entire workflow is applied to the same tissue section, allowing for the automatic integration of morphological and spatial metabolomics data for thousands of molecules. We established this method and have successfully applied it in previous works [[112]27, [113]44‐[114]^46]. Supplementary Figure [115]S1 displays representative immunofluorescence sections used during this process. Fluorescence stains were scanned at 20× magnification using an AxioScan.Z1 digital slide scanner (Mirax Desk, Carl Zeiss MicroImaging GmbH; Jena, Germany) and visualized using the software ZEN 2.3 blue edition (Zeiss; Oberkochen, Germany). All root‐mean‐square normalized mass spectra were exported from SCiLS Lab v. 2020 (Bruker Daltonics). Peaks in the mass range of m/z 75‐1000 Da were annotated by accurate mass matching with the Human Metabolome Database ([116]http://www.hmdb.ca/) [[117]47] and METASPACE ([118]http://annotate.metaspace2020.eu [[119]48]; ion mode: negative; adduct type: [M−H], [M−H−H[2]O], [M+Na−2H], [M+Cl], and [M+K−2H]; mass accuracy, ≤4 ppm). Pathway enrichment analysis was performed on tumor and stromal tissue using MetaboAnalyst 5.0 ([120]https://www.metaboanalyst.ca) [[121]49]. Pathway analysis algorithms, the hypergeometric test for overrepresentation analysis, and relative‐betweenness centrality were selected for pathway topology analysis. P value and impact score were calculated for each metabolic pathway, revealing substantial differences in enriched pathways between the tumor and the stroma. 2.4. Random forest classifiers The random forest classification, a robust machine learning algorithm, was performed for the classifier. Leave‐one‐out cross‐validation (R 4.0.2) was used to predict long‐ or short‐term survival for each patient. The cohort was separated into long‐term (≥35 months) and short‐term survivors (<35 months) according to median OS. For feature selection, molecules included in the random forest (RF) analysis were selected based on their significance level in the log‐rank test (P < 0.05). After performing the RF analysis, the importance of each molecule was calculated as the total reduction of the criterion brought by that feature (Gini importance). The top 100 most important metabolites according to feature selection were re‐selected as the final variables included in the RF classifiers. The analysis for each RF classifier was repeated 100 times, and a majority vote determined the final prediction model. The mean accuracy, sensitivity, and specificity were used to evaluate the performance of each classifier and MPR. Accuracy refers to the percentage of positive predictions made by the classifier that were correct. The correct prediction is determined by classifying patients with an OS longer than median OS in the long‐term group and that shorter than median OS in the short‐term group. Accuracy was expressed as the ratio of true positives and true negatives to the total observations. Sensitivity was calculated by dividing true positives by all observations in the actual class, representing the percentage of actual positive predictions that were correctly classified by the classifier. Specificity refers to the ratio of true negatives to total negatives. The RF classifiers were calculated using the R package randomForest. 2.5. Further statistical analyses All analyses were performed using R software (version 4.0.2, [122]https://cran.r‐project.org) with suitable packages. Survival analysis was performed using Kaplan‐Meier analysis and Cox proportional hazards regression, with 95% confidence interval (95% CI) estimates (R 4.0.2, survival). Variables in the multivariate Cox regression were included based on their significance in the log‐rank test (P < 0.05) during univariate Cox regression analyses. Comparisons between tumor and adjacent normal lung tissue were performed using the rank‐based Mann‐Whitney U‐test and Spearman's rank‐order correlation for continuous data (R 4.0.2, corrplot). The log‐rank test was used to assess differences. P values < 0.05 were considered significant. 3. RESULTS 3.1. Patients’ characteristics The cohort of consecutive patients with resected NSCLC following NAC initially included 117 patients who received at least one cycle of platinum‐based chemotherapy before surgery, as previously described [[123]20]. After all inclusion and exclusion criteria were applied, 88 NAC and 85 chemotherapy‐naïve patients were identified with sufficient materials available for metabolic analyses. The median ages were 62 years (interquartile range [IQR], 42‐77 years) and 63 years (IQR, 39‐84 years) for the NAC and chemotherapy‐naïve cohorts, respectively. No significant differences in the distribution of histological subtypes, median age, or sex were observed between the NAC and chemotherapy‐naïve cohorts. Detailed clinicopathological patient characteristics are summarized in Table [124]1. A total of 59 adjacent normal lung tissue samples from patients with NSCLC in the NAC cohort were included for analyses. 3.2. Metabolic classifiers established for stratifying patients into prognostic risk groups Metabolites from tumor and stroma regions were automatically extracted using spatial metabolomics analysis, resulting in 5014 distinct molecular features detected in tissues from all patient samples. In the NAC cohort, metabolic classifiers were trained separately on metabolites identified within the tumor and those identified within stroma tissues. The top 100 metabolites for each classifier, ranked in descending order according to feature importance, are shown for tumors in Figure [125]2A and for the stroma in Figure [126]2B. Postulated endogenous annotations for the top 100 molecules are listed in Supplementary Table [127]S1. FIGURE 2. FIGURE 2 [128]Open in a new tab The metabolic classifiers were established for tumor cells (A) and the stroma (B) for stratifying patients with NSCLC who received NAC followed by resection into prognostic risk groups. The forest plot shows the hazard ratio and 95% confidence intervals achieved for best‐performing metabolites for categorizing patients into prognostic risk groups. Metabolites with P < 0.05 are highlighted and ranked in descending order of importance. The feature importance value from 0 to 6, the higher values indicated more impacted on the prediction model. Abbreviations: NSCLC, non‐small cell lung cancer; NAC, neoadjuvant chemotherapy; PA, phosphatidic acid; LysoPI, lysophosphatidylinositol; CPA, cyclic phosphatidic acid; LysoPA, lysophosphatidic acid; PE, phosphatidylethanolamine; SM, sphingomyelin; PC, phosphatidylcholine; LysoPE, lysophosphatidylethanolamine; LysoPC, lysophosphatidylcholine; PGP, phosphatidylglycerophosphate; TG, triglyceride; PI, phosphoinositol; Cer, ceramide; PE‐Nme, dimethylphosphatidylethanolamine; SAM, S‐adenosylmethioninamine For the metabolic tumor classifier, sphingomyelin (SM, d18:1/15:0 or d16:1/17:0) was shown as an example of a prognosis marker in patients with NSCLC. A high mass intensity for SM was significantly associated with a good prognosis (Figure [129]3A). Ion distribution maps revealed the specific localization of metabolites in tumor cell regions (Figure [130]3B). Boxplots displayed the variance in mass intensity values measured for SM (Figure [131]3C). For the metabolic stroma classifier, a high mass intensity of m/z 480.3091, which can be postulated annotated as either lysophosphatidylcholine (LysoPC, 15:0/0:0) or lysophosphatidylethanolamine (LysoPE, 18:0/0:0), was significantly associated with long survival (Figure [132]3D) and demonstrated distinct distribution patterns (Figure [133]3E). Boxplots displayed the variance in mass intensity values measured for LysoPC/LysoPE (Figure [134]3F). The other important endogenous metabolites used to distinguish between good and poor prognosis in patients with NSCLC following NAC are presented in Supplementary Figure [135]S2. FIGURE 3. FIGURE 3 [136]Open in a new tab Endogenous metabolites included in the classifiers were used to distinguish between good and poor prognosis groups among NSCLC patients who received neoadjuvant therapy. For tumor classifier, high level of m/z 687.5425 [SM (d18:1/15:0 or d16:1/17:0)] was associated with a good prognosis (A). Ion distribution maps revealed the specific localization of SM in tumor cell regions for high and low mass intensity. The corresponding H&E stains of the same tissue core were shown on the right (B). Boxplot with individual points was shown for variance in different groups for SM (C). For stroma classifier, a high mass intensity of m/z 480.3091 [LysoPC(15:0/0:0)/LysoPE(18:0/0:0)] was associated with a good prognosis (D). The ion map revealed the specific localization of LysoPC/LysoPE in stroma regions (E) and the boxplot was shown for variance in different groups for LysoPC/LysoPE (F). *** P < 0.001. Abbreviations: NSCLC, non‐small cell lung cancer; SM, sphingomyelin; LysoPC, lysophosphatidylcholine; LysoPE, lysophosphatidylethanolamine We have identified all of the odd‐chain fatty acids (OCFAs) included among the 100 most important molecules used for the metabolic classifier (Supplementary Table [137]S1). The following OCFAs were identified: m/z 687.5425 [SM (d18:1/15:0 or d16:1/17:0)], m/z 480.3091 [LysoPC (15:0/0:0)], m/z 508.3405 [LysoPC (17:0/0:0)], and m/z 852.6496 [PC (15:0/24:0)]. In addition, we performed Spearman's rank‐order correlation analysis between these OCFAs. The results indicated that LysoPC (17:0/0:0) was positively correlated with SM (d18:1/15:0 or d16:1/17:0) and LysoPC (15:0/0:0), whereas PC (15:0/24:0) was negatively correlated with the other OCFAs (Supplementary Figure [138]S3). 3.3. Spatial metabolomics improved the evaluation of therapy response compared with major pathological response Comparing prognostic efficacy between the two metabolic classifiers and clinicopathological features revealed significant prognostic power for both the tumor (P < 0.001) and stroma (P < 0.001) metabolic classifiers in the Kaplan‐Meier survival analysis, resulting in better prognostic stratification performance than either MPR or TNM staging alone (Figure [139]4). The tumor metabolic classifier displayed a prediction accuracy of 81.6%, which was similar to that of the stroma metabolic classifier (accuracy = 78.4%) in the NAC cohort (Figure [140]4A and [141]4B). The accuracy, sensitivity, and specificity for MPR were all lower than those for either metabolic classifier (Figure [142]4C). The accuracies achieved when combining the metabolic classifiers with MPR (Supplementary Figure [143]S4) were slightly lower than those achieved by either metabolic classifier alone. In the chemotherapy‐naïve cohort, the accuracies of the tumor (Figure [144]4D) and stroma metabolic classifier (Figure [145]4E) were higher than that of TNM staging (Figure [146]4F). FIGURE 4. FIGURE 4 [147]Open in a new tab The performance of metabolic classifiers and pathological parameters for stratifying patients with NSCLC into prognostic risk groups. For the NAC cohort, both tumor (A) and stromal (B) metabolic classifiers showed better performance for stratifying prognostic risk groups than MPR (C). For the Chemo‐naïve cohort, the performance of the tumor (D) and stroma (E) metabolic classifiers were superior to TNM staging (F). Kaplan‐Meier survival analyses were used to evaluate differences in patient overall survival. Abbreviations: NSCLC, non‐small cell lung cancer; MPR, major pathological response; NAC, neoadjuvant chemotherapy; Chemo, chemotherapy; TNM, tumor‐node‐metastasis In univariate analysis, the tumor and stroma metabolic classifiers, MPR, and TNM staging all demonstrated significant predictive efficacy in the NAC cohort (Table [148]2). However, in the multivariate regression analysis, which included the metabolite‐based classifiers and clinicopathological parameters, the metabolic classifiers were the only independent prognostic factors (tumor metabolic classifier: P = 0.001; stroma metabolic classifier: P = 0.049), whereas MPR and TNM staging were no longer significantly associated with prognosis. In the chemotherapy‐naïve cohort, multivariate analysis confirmed the stroma metabolic classifier as the only independent prognostic factor (P < 0.001), whereas the tumor metabolic classifier and TNM staging were no longer identified as significant factors (Table [149]2). TABLE 2. Univariate and multivariate Cox proportional analyses to identify the OS predictors for patients with NSCLC Terms Univariate Cox analysis Multivariate Cox analysis HR 95% CI P value HR 95% CI P value NAC cohort Tumor metabolic classifier 5.591 2.678‐11.670 <0.001 3.823 1.716‐8.514 0.001 Stroma metabolic classifier 4.626 2.446‐8.749 <0.001 2.180 1.004‐4.737 0.049 MPR 0.465 0.260‐0.830 0.010 0.913 0.445‐1.874 0.804 TNM staging 1.304 1.082‐1.571 0.005 1.223 0.977‐1.550 0.078 Chemo‐naïve cohort Tumor metabolic classifier 2.836 1.603‐5.016 <0.001 1.767 0.971‐3.217 0.063 Stroma metabolic classifier 6.118 3.219‐11.630 <0.001 4.953 2.517‐9.747 <0.001 TNM staging 2.570 1.200‐5.502 0.015 1.417 0.653‐3.075 0.379 [150]Open in a new tab Abbreviations: OS, overall survival; NSCLC, non‐small cell lung cancer; HR, hazards ratio; CI, confidence interval; NAC, neoadjuvant chemotherapy; MPR, major pathologic response; TNM, tumor‐node‐metastasis; Chemo, chemotherapy In addition, we evaluated the abilities of the metabolic classifiers to predict patient survival in relation to the pathological response of MPR. The Kaplan‐Meier survival analyses indicated that the metabolic classifiers were able to further stratify patients according to MPR outcomes. Both tumor and stroma metabolic classifiers predicted significant differences in OS between NAC responders (MPR present; P < 0.001) and non‐responders (MPR absent; P < 0.001) (Figure [151]5). FIGURE 5. FIGURE 5 [152]Open in a new tab Metabolic classifiers sub‐stratified NSCLC patients with different pathological responses into prognostic risk groups. The tumor (A) and stroma metabolic classifiers (B) could further stratify responder (MPR present) and non‐responder patients (MPR absent) into short‐ and long‐term survivors using the Kaplan‐Meier analysis. Abbreviations: NSCLC, non‐small cell lung cancer; MPR, major pathological response 3.4. Metabolic classifiers were specific to neoadjuvant therapy response To validate whether the classifiers are specific to NAC response, the metabolite levels measured in the NAC cohort were compared with those measured in the chemotherapy‐naïve cohort (Figure [153]6). The calculated hazard ratios of the top 100 identified metabolites defined by the NAC classifier were shown for the chemotherapy‐naïve cohort. Significant differences in metabolite levels were identified between the chemotherapy‐naïve and NAC cohorts. Of the 100 included metabolites in the tumor classifier, 93 were significantly associated with OS in the NAC cohort, whereas only 35 were associated with OS in the chemotherapy‐naïve cohort (Figure [154]6). Of the 100 metabolites included in the stroma classifier, 98 were significantly associated with OS in the NAC cohort, whereas only 32 were significantly associated with OS in the chemotherapy‐naïve cohort (Figure [155]6). We also compared the 100 most important metabolites between tumor and normal tissues, which showed that 19 metabolites were significantly upregulated and 4 were significantly downregulated in tumor tissues compared with normal tissues. Most of the upregulated metabolites were identified as lipids, including phosphatidylethanolamine (PE), phosphatidic acid (PA), phosphatidylglycerophosphate (PGP), lysophosphatidic acid (LysoPA), lysophosphatidylinositol (LysoPI), phosphoinositol (PI), LysoPC, LysoPE, and SM (Supplementary Figure [156]S5). FIGURE 6. FIGURE 6 [157]Open in a new tab Prognostic significance of the metabolites included in the RF classifiers informed by the NAC cohort, compared with the chemotherapy‐naïve cohort. The specificities of the metabolic classifiers were assessed for predicting survival in the NAC and chemotherapy‐naïve cohorts based on the metabolites detected in tumor (A) and stroma tissues (B).The univariate hazard ratios from Cox regression models were calculated for the top 100 metabolites in patients with chemotherapy‐naïve NSCLC, using metabolites in patients treated with NAC as references. Each line in the plot represents the