MPM pleural exudates (n = 19) were obtained from Mesobank, a Research Ethics Committee-approved Research Tissue Bank. The MPMs were of epithelioid histology, except for one with biphasic histology. Demographic and clinical data of the patients are reported in Table 1. All patients provided written informed consent, and samples were anonymized. Mesobank is supported by Asthma and Lung UK, The Victor Dahdaleh Foundation, and the June Hancock Mesothelioma Research Fund.

Lipid mediators were quantified in the pleural fluid of patients with or without Pemetrexed chemotherapy. All samples were fast frozen and stored at -80 °C for a maximum of 6 months before lipidomic analysis to prevent alteration of lipid content. Fifty-six oxylipins and metabolites of AA, EPA, and DHA were quantified by liquid chromatography-tandem mass spectrometry (LC-MS/MS) [Supplementary Table 1]. The extraction and analysis were performed as previously described^[21]. Briefly, lipid mediators were extracted from samples by solid-phase extraction (SPE) and eluted with a mixture of methyl formate and methanol (MeOH). After solvent evaporation, samples were dissolved in MeOH:H₂O, 50:50 (v/v), and injected into an ultra-high-performance liquid chromatography (U-HPLC) Exion LCAD system coupled to a 6500+ QTRAP (AB SCIEX GmbH, Zug, Switzerland) mass spectrometer. The lipids were separated on a Kinetex C18 column (100 mm × 2.1 mm, 1.7 µm, Phenomenex) and detected by negative electrospray ionization in multiple reaction monitoring (MRM) mode. When authentic standards were available, the identification of the lipid mediators in the samples was based on their retention times with those of the pure standards injected under the same conditions. Calibration curves were obtained using the same authentic standards, and quantification was carried out based on peak areas from MRM transitions using a linear model weighted by 1/X (X, the nominal concentration). For 21-HDOHE, 4(S),14(S)-DiHDOHE, RvD6, 18(S)-RvE3, 18(R)-RvE3 identification was based on theoretical MRM transitions established on the basis of the molecule’s structure, and quantification was carried out using a pure standard with a similar structure. Peak detection, integration, and quantitative analysis were performed using Sciex OS MQ software. Quantification results were expressed in pg/mL and detailed in Supplementary Table 2.

The Mesothelioma gene expression datasets used in this study were downloaded from The Cancer Genome Atlas (TCGA) (https://www.cancer.gov/tcga), specifically from the GDC TCGA Mesothelioma (MESO) project. Gene expression data were provided as FPKM values. Survival information was defined as the OS of patients, i.e., the time from diagnosis to death, regardless of cause. Deconvolution and Statistical Analysis in TCGA Mesothelioma Dataset. The gene expression data were used to perform bulk RNA-seq deconvolution using CIBERSORTx, a tool for cell-type profiling. CIBERSORTx was employed with the LM22 gene signature, which is a predefined gene signature matrix for 22 immune cell types, to estimate the relative proportions of different immune cell populations in the bulk tumor samples. Following deconvolution, the patients were stratified into “low-expressing” and “high-expressing” groups, based on the composite signature calculated in the previous steps. To compare immune cell-type differences between the two groups, a Mann-Whitney U test was performed using Python’s scipy stats module. Boxplots were drawn for immune cell types whenever the Mann-Whitney U test revealed statistically significant differences (P < 0.05).

To further investigate the transcriptional differences between the highly/lowly Composite Signature-expressing groups, a differential gene expression analysis was performed on the FPKM-UP expression data with the LIMMA package. To identify differentially expressed genes (DEGs), the following hard filters were applied: |log2FC| > 0.5 and adjusted P-value < 0.05. P-values were adjusted using the Benjamini-Hochberg (BH) method. IPA QIAGEN Inc.(https://www.qiagenbioinformatics.com/products/ingenuity-pathway-analysis) and KEGG pathway enrichment analysis were performed to explore the biological implications of the resulting DEGs.

From the GDC-TCGA mesothelioma dataset described above, a panel of genes associated with lipid metabolism and oxidative stress pathways was selected for analysis: PTGR1, ALOX12, LTA4H, EPHX2, EPHX1, CYP3A4, ALOX15, GPX4, GSTM4, and CYP4F11. Categorical variables, such as “treatment status” and “diagnosis”, were converted into binary variables through one-hot encoding. A binary variable was created to define the survival outcome based on the median OS time, classifying patients with survival times above or below the median. The dataset was stratified into two groups based on treatment status: patients receiving and not receiving therapy. Lasso regression with 10-fold cross-validation was applied separately to the treated and untreated groups to identify genes significantly associated with the binary survival outcome. Lasso regression was implemented using the sklearn Python library v1.2.2. Genes with non-zero Lasso coefficients were considered predictive of survival. The intersection of genes selected from the treated and untreated groups was computed to identify shared prognostic features. A bootstrap analysis with 100 resampling iterations was performed to evaluate the selected genes’ robustness. For each iteration, the dataset was resampled, and the Lasso model was refitted. The coefficients of the genes identified in both groups were recorded across iterations. The mean of the coefficients was calculated to assess their stability and significance. A composite prognostic score (CPS) was derived as a linear combination of the expression values of the overlapping genes, weighted by their mean Lasso coefficients obtained from the bootstrap analysis. The CPS was used to stratify patients into high- and low- risk groups based on the median value of the signature.

KM-plotter (http://kmplot.com/analysis/) was employed to assess the impact of mRNA levels of representative SPM-producing enzymes in MPM patients from the GDC-TGCA mesothelioma database. Additionally, Kaplan-Meier survival analysis was performed to compare survival outcomes between high-risk and low-risk groups. The Log-Rank test was used to evaluate the statistical significance of differences in survival curves between the two groups. The Kaplan-Meier survival curves were visualized to illustrate the separation between risk groups. To further confirm the prognostic significance of the composite signature, a Cox proportional hazards model was fitted with survival time, event status, and the composite signature as predictors. Kaplan-Meier curves, Log-Rank tests, and Cox regression were conducted using the lifelines Python library v0.27.8. To evaluate the predictive performance of the CPS, we performed time-dependent receiver operating characteristic (ROC) analysis at 12, 24, and 36 months of overall survival. Risk scores were derived from the multivariate Cox model, with higher partial hazard scores indicating increased risk. The 24-month timepoint was chosen as the primary threshold for further diagnostic assessment, as it ensured sufficient observed deaths (n = 50) and patients still at risk (n = 31). By comparison, the 12-month analysis had fewer events, and at 36 months, only 14 patients, limiting robustness for calibration and sensitivity/specificity estimation. To further characterize the model’s performance at 24 months, sensitivity and specificity were computed using Youden’s index to determine the optimal threshold, and a calibration plot was generated.

Results are expressed as means ± SEM and compared using a nonparametric Mann-Whitney two-sided test in GraphPad Prism 9 (GraphPad Software, San Diego, CA). Given the exploratory nature of the lipidomic analysis, which involved 56 lipid species and multiple pairwise comparisons among the three experimental groups (Naive, Chemo 2 cycles, and Chemo > 4 cycles), P values obtained from Mann-Whitney U tests were corrected for multiple testing using the Benjamini-Hochberg false discovery rate (FDR) method. The adjustment was performed in R software (version 4.6.0). Lipids with an adjusted P-value (FDR) < 0.05 were considered statistically significant. Principal component analysis (PCA) was created using the freely available R software. For multiple correlation analysis, the free StatCalculator was used (https://www.statscalculators.com/), with control for multicollinearity. For the OS analysis, Kaplan-Meier survival curves were generated and compared between untreated and chemotherapy-treated patients using the log-rank test, implemented in the lifelines Python library (v0.27.8). For multiple correlation analysis, the free StatCalculator was used (https://www.statscalculators.com/), with control for multicollinearity.

Single-cell RNA sequencing (scRNA-seq) data processing and in silico integration into Inflammation-resolution networks. Single-cell RNA-seq profiles (GSE243446 and GSE190597) of human pleura with preprocessed and library-size-normalized read counts (q) were downloaded from published studies^[22]. Based on canonical markers and signature genes cross-referenced to known markers, the dataset includes many different mesothelial, immune, and stromal cells. To ensure the analysis aligns with the data generated in the present study, patients 1 and 3 from the GSE243446 dataset were excluded due to comorbidities. The dataset GSE190597 includes 13 samples from three histological types of tumors: sarcomatoid, biphasic, and epithelioid, and encompasses 15 distinct cell types, including both malignant and mesothelial cells. Log1p normalization, total-count normalization, and scaling of the data were performed using the scanpy-python package. We applied a Gaussian-Mixed-Model (GMM) before scaling to divide the samples into expressing and unexpressed cells for each gene, identifying a suitable read-count threshold that defines a gene as expressed or not. We downloaded manually curated networks from the “Biosynthesis of PIM and SPM from AA”, “Biosynthesis of PIM and SPM from DHA”, and “Biosynthesis of PIM and SPM from EPA” pathways of our previously published “Atlas of Inflammation Resolution” (AIR) platform^[23]. We manually mapped the lipids analyzed in the present study to molecules in the network to avoid naming-convention mismatches. For each of the lipids and, additionally, for each lipid class that is represented as phenotype elements in the network, we identified all sets of genes corresponding to enzymes that could catalyze their synthesis from either one of the substrates AA, DHA, or EPA as described in previous works^[24]. Finally, we considered a lipid mediator to be synthesizable in a given cell if the cell expresses all genes in at least one of the gene sets. To identify cell clusters of similar lipid mediator (LM) synthesis, we first reduced the data to only genes of enzymes occurring in the lipid mediator network and the samples to only cells where at least one lipid mediator was identified as synthesizable. We then applied Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction from the scanpy Python package on the reduced dataset. Clusters in the UMAP embedding were identified using Leiden graph-clustering with a manually selected resolution parameter of 0.275. A dot plot was generated using the DotPlot function from the Seurat package to visualize the expression of selected genes across the mesothelial and malignant cells.

We performed targeted metabololipidomics by liquid chromatography-tandem mass spectrometry (LC-MS/MS) on pleural fluid samples collected from 19 patients with MPM, including untreated individuals (n = 8) and patients who had received two to six cycles of chemotherapy (n = 11; cisplatin + pemetrexed, except for one patient treated with carboplatin) [Table 1 and Supplementary Figure 1]. All but one patient exhibited epithelioid histology [Table 1]. All patients in the naïve group had anamnestically reported asbestos exposure, while three of eleven patients in the treated group were not recorded as having asbestos exposure. In total, we quantified 56 oxylipins, comprising 22 PIMs derived from AA and 34 SPMs derived from AA, EPA, or DHA [Supplementary Table 1, Figure 1 and Supplementary Figure 2 for examples of MS spectra].

Principal component analysis (PCA) revealed clustering of chemotherapy-treated patients by lipid class and precursor fatty acids (AA, EPA, DHA) [Figure 1A-C]. We next compared oxylipin levels across three patient groups (naïve, 2 cycles, ≥ 4 cycles) to identify chemotherapy cycle-dependent differences [Figure 1 and Supplementary Table 1]. Among significantly modulated PIMs (P < 0.05), prostaglandins PGD₂, PGE₂, and PGF₂α were markedly increased in patients who had received ≥ 4 cycles [Figure 1D]. Conversely, patients treated with two cycles exhibited significantly lower levels of these prostaglandins compared with naïve individuals. Leukotrienes (LXB₄, LTC₄) and hydroxy-/epoxyeicosanoids (5-HETE, 5,6-EET) decreased in both chemotherapy-treated groups. Overall, chemotherapy induced a net increase in PIMs, with prostaglandins showing a clear cycle-dependent pattern. Longer treatment duration was associated with increased prostaglandins and decreased leukotrienes [Figure 1D].

We next examined EPA- and DHA-derived SPMs. Among the resolvins, RvD1 increased after two cycles but decreased in the ≥ 4-cycle group [Figure 1E]. The E-resolvin RvE1, the protectins (PDx, AT-PD1), and the AA-derived 5,6-EET decreased in both treated groups [Figure 1E]. Other SPMs, such as LXB4 and RvE3, exhibited a similar pattern without reaching significance [Supplementary Table 1]. Maresin2 (MaR2) was significantly reduced only in the high-cycle group [Figure 1E]. The SPM precursors showed no statistically significant changes, although EPA-derived 18-HEPE trended downward in both treated groups, and DHA-derived 17-HDOHE and 14-HDOHE trended upward in the high-cycle group [Supplementary Table 1]. Collectively, longer-term chemotherapy reduced RvD1, RvE1, Mar2, and the protectins PDx and AT-PD1, without significantly altering precursor levels.

To determine whether altered SPM profiles reflect disease aggressiveness, we evaluated OS in the same patient cohort. Chemotherapy did not improve OS, indicating that treatment-associated SPM reductions did not translate into survival differences (Kaplan-Meier analysis, log-rank test: P = 0.284; Figure 2A). This remained true when low- and high-cycles groups were analyzed separately [Supplementary Figure 3]. To deepen these observations, we performed multiple linear correlation analyses to evaluate whether any of the identified metabolites, including those trending (0.1 < P < 0.05), correlated with OS across 19 patients (both naïve and treated). We found that none of the PIMs considered exhibited a statistically significant correlation with patients’ OS. By contrast, we found that the levels of four SPM precursors (15-HEPE, 18-HEPE, 20-HDOHE, 7-HDOHE) and RvE2 and Protectin DX levels were significantly correlated with time to death [Figure 2B]. These findings suggest that impaired SPM biosynthesis may characterize MPM patients with more aggressive disease, independent of treatment.

We analyzed mRNA expression of key SPM-biosynthetic enzymes using a scRNA-seq dataset containing intratumoral mesothelial diploid cells and mesothelioma cells from epithelioid MPM patients^[25]. Mesothelioma cells exhibited reduced expression-both in magnitude and proportion of expressing cells-of several SPM-producing enzymes compared with mesothelial cells [Figure 3A]. These findings were corroborated using the AIR platform, which is based on scRNA-seq datasets from pleural tissues of non-MPM subjects and MPM patients, and revealed disrupted cellular organization in MPM tissues [Supplementary Figures 4 and 5].

We developed a five-gene signature comprising CYP3A4, ALOX15, EPHX1, GSTM4, and GPX4, which stratified TCGA MPM patients into two groups. A CPS was derived as a linear combination of the expression values of the overlapping genes, weighted by their mean Lasso coefficients obtained from the bootstrap analysis. The CPS was used to stratify patients into high-risk and low- risk groups based on the median value of the signature. Those with “high” CPS exhibited better survival rates than those with “low” CPS (P = 0.0002) [Figure 3B]. To detail the prognostic utility of CPS, we fitted a Cox regression model, scoring treatment status, histological subtype, and pathological stage as covariates. Patients with higher CPS values had a 41% reduced risk of mortality compared with those with lower CPS values (univariate HR = 0.59; P < 0.005; 95%CI: 0.43-0.81). This association remained significant in the multivariate Cox model, supporting CPS as an independent predictor of OS (adjusted HR = 0.61; P = 0.002; 95%CI: 0.45-0.83). The time-dependent predictive performance of the CPS, as determined by receiver operating characteristic (ROC) analysis at 12, 24, and 36 months, yielded area under the curve (AUC) values of 0.66, 0.72, and 0.78, respectively, with corresponding 95% confidence intervals of 0.53-0.79, 0.61-0.83, and 0.64-0.90 [Figure 3C]. At 24 months, sensitivity was 0.588 and specificity was 0.833 [Figure 3D]. These results indicate that reduced SPM production or diminished expression of SPM-biosynthetic enzymes identifies prognostically distinct MPM patient subsets [Figure 2A].

Differential expression analysis identified 73 downregulated and 29 upregulated genes between CPS-defined groups [Table 2]. Pathway enrichment analysis [Supplementary Tables 3 and 4] revealed significant enrichment of the retinol metabolism pathway and a trend toward enrichment of the Wnt signaling pathway (P = 0.07) in low-CPS “high risk” MPM patients [Figure 3E].

Immune deconvolution showed a higher proportion of M1 macrophages in the low-CPS group (P = 0.023) and a higher representation of plasma cells in the high-CPS group (P = 0.026) [Figure 3F; Supplementary Tables 5 and 6]. These findings further support the distinct biological identity of the two CPS-defined patient groups and highlight the pro-inflammatory, dysregulated microenvironment associated with low CPS and shorter OS.