Accumulation of long-chain fatty acids in the tumor microenvironment drives dysfunction in intrapancreatic CD8⁺ T cells

The advent of immunotherapy has revolutionized cancer treatment by inducing, providing, and/or reactivating antitumor T cells. Therapeutic results have been heterogeneous, with better outcomes generally correlating with the ability of tumor-specific T cells to infiltrate the tumor, persist, and retain effector functions. Complete and durable clinical responses have been achieved in patients whose cancers were resistant to available standard treatments (Mellman et al., 2011). Yet, it has met with limited success in most patients with solid tumors (Joyce and Fearon, 2015; Menon et al., 2016), including pancreatic ductal adenocarcinoma (PDA; Guo et al., 2017). Although the overall survival rate in PDA patients appears to correlate with CD8⁺ T cell infiltration (Ene-Obong et al., 2013; Fukunaga et al., 2004; Ino et al., 2013), our knowledge of the mechanisms that regulate the function of these infiltrating T cells in the context of the tumor microenvironment (TME) is still limited.

CD8⁺ T cells are key effectors of antitumor immunity. However, tumor-infiltrating CD8⁺ T cells often acquire an altered state of differentiation referred to as “exhaustion” (Wherry and Kurachi, 2015) and, as a result, fail to control tumor outgrowth. Several studies in both murine models of pancreatic cancer and PDA patients have demonstrated that CD8⁺ T cells are often scarce and, if present, have become dysfunctional (Bailey et al., 2016; Beatty et al., 2015; Clark et al., 2007; Stromnes et al., 2015). These observations are perceived to reflect, at least in part, the unique immunosuppressive TME of PDA.

The metabolic state and nutrient availability in the TME are among the main contributing determinants of the functional fate of CD8⁺ T cells infiltrating solid tumors (Wei et al., 2017). Indeed, recent findings suggested that CD8⁺ T cell activation and differentiation are tightly intertwined with metabolic reprogramming. Resting CD8⁺ T cells rely on oxidative phosphorylation (OXPHOS) to fuel their metabolic needs. Upon activation, CD8⁺ T cells engage glycolysis to fulfill their energetic and biosynthetic demands. In contrast, tumor cells are highly glycolytic and support their overgrowth by consuming large amounts of key nutrients, especially glucose (Vander Heiden et al., 2009), which are also required for optimal T cell activation (Siska and Rathmell, 2015). When deprived of glucose, CD8⁺ T cells are compromised in their ability to produce effector cytokines (Cham et al., 2008; Cham and Gajewski, 2005; Jacobs et al., 2008). However, they can still sustain their proliferation and effector functions via mitochondrial OXPHOS (Chang et al., 2015; Sena et al., 2013), using long-chain fatty acids (LCFAs) to fuel the tricarboxylic acid cycle through the process of fatty acid (FA) β-oxidation (FAO) as an alternative to glucose-derived pyruvate, thereby maximizing production of energy and reducing equivalents (Zhang et al., 2017). Therefore, to mount an effective antitumor response, tumor-infiltrating CD8⁺ T cells must retain the metabolic flexibility to adjust in response to the availability of nutrients. It remains unclear how nutrient compositional changes in the TME correlate with the observed dysfunction of tumor-infiltrating CD8⁺ T cells. In a preclinical mouse model of PDA as well as in human tumor specimens, we show that PDA progression is characterized by enrichment of specific lipid accumulation in the TME. In such a lipid-rich TME, intrapancreatic CD8⁺ T cells become exhausted and reduce their ability to metabolize the available lipid substrates. Moreover, the combination of increased lipid levels and decreased mitochondrial function and FAO is toxic to intrapancreatic CD8⁺ T cells, resulting in increased cell death. Our data suggest that metabolic reprogramming of tumor-specific CD8⁺ T cells might represent a strategy to promote survival in the metabolically hostile TME as part of an approach to enhance the clinical efficacy of immunotherapy.

In PDA, the TME is known to be highly immunosuppressive, limiting T cell infiltration and functions (Liyanage et al., 2002; Mukherjee et al., 2001). Most solid tumors have been reported to be deprived of crucial metabolites needed for T cell function (Ecker and Riley, 2018). To address whether compositional changes in PDA TME might directly contribute to CD8⁺ T cell impairment, we characterized the TME in the KC mouse model, which closely recapitulates human PDA progression (Hingorani et al., 2003). Cohorts of mice were sacrificed at different time points between 9 and 30 wk of age and, based on cytological and architectural abnormalities, divided into early stage of tumor development with pancreatic intraepithelial neoplasia (early PanIN) versus high-grade preneoplastic lesions (late PanIN; Fig. S1 A). Staining of tissue sections revealed that neutral lipids accumulate in the pancreatic tissue during the progression from early to late PanIN, detected with both red oil staining (Fig. S1 B, red arrowheads) and LipidTOX (Fig. S1 C). Therefore, to analyze low molecular weight compounds, we performed a temporally and spatially resolved compositional analysis of the TME by combining confocal imaging on longitudinal samples with a high-resolution, label-free, imaging mass spectrometry (MS) approach (namely high-resolution matrix-assisted laser desorption/ionization Fourier transform ion cyclotron resonance imaging MS, or MALDI FT-ICR IMS, hereafter referred to as IMS). Frozen pancreata were sectioned on a gold-coated stainless-steel target for IMS. IMS was performed in either positive or negative ion mode, generating a mass spectrum at each pixel. Thus, a mean spectrum was generated by segmenting the data through k-means clustering and positive correlation with CD8⁺ T cell distribution on consecutive sections upon coregistering the IMS with the confocal images. The peak list of each segment was then searched against the LIPID MAPS database, and mean intensity values for each accurate mass lipid identification were extracted and filtered to retain only those lipids with sufficient signal-to-noise ratios and that represented large fold changes between time points. Finally, data were visualized in a heatmap format sorted by head group, fatty acyl chain length, and level of fatty acyl chain unsaturation (Fig. S1, D and E). This technology has been used previously to survey pancreatic cancer imaging MS to discriminate breast from pancreatic cancer metastasis in formalin-fixed, paraffin-embedded tissues (Casadonte et al., 2014; Eberlin et al., 2014; Grüner et al., 2012) and to survey the molecular composition of multifunctional tissues such as the pancreas (Prentice et al., 2019), but not to characterize the metabolic state (Caprioli et al., 1997; Cornett et al., 2008). Fig. 1 A shows the IMS profiles (right panels) of one representative lipid species, PE(42:6) (a polyunsaturated phosphatidylethanolamine with a total carbon FA chain length of 42 carbons), which was found to be enriched during progression in the areas (one of which is highlighted in orange) populated by CD8⁺ T cells (Fig. 1 B, right panels, insets showing magnifications of the areas highlighted in magenta). Overall, our study revealed, in the areas populated by infiltrating CD8⁺ T cells, an enrichment of specific glycerophospholipids containing very-long-chain FAs (VLCFAs) in PanINs. Results of the global glycerophospholipids analysis are summarized in bar plots representing the relative fraction of lipids of the indicated total carbon FA chain length detected at indicated time points (Fig. 1, C and D; and Fig. S1, F–H). We detected predominantly an increase of PEs and phosphatidylinositols (PIs; Fig. 1, C and D, respectively, with dotted lines highlighting an ∼10% increase), and enrichment of specific species was already evident in early lesions (Fig. 1, E and F; and Fig. S1 E). Global differences were less apparent for phosphatidylserines (Fig. S1 F) and phosphatidylcholines (PCs; Fig. S1 G). Very-long-chain (VLC) sphingomyelins, by contrast, were relatively decreased at late disease time points compared with control (Fig. S1 H).

Quantification of the extracellular glucose in pancreata using a colorimetric assay followed by normalization against the weight of each pancreas highlighted a decrease in glucose availability in late PanINs (Fig. S1 I), suggesting that infiltrating CD8⁺ T cells are restricted to the available long-chain (LC) and VLC glycerophospholipids for fulfilling their metabolic needs in the TME. This model was further supported by the IMS analysis for the first metabolic intermediate of the glycolytic pathway, hexose-6-phosphate (H6P), which was progressively reduced from Early to late PanINs, indicating a decline of the overall glycolytic activity (Fig. S1 J). Altogether, these data suggest that PDA progression is accompanied by distinctive compositional alterations of the TME that may subvert the metabolic needs of infiltrating CD8⁺ T cells.

Next, we sought to understand whether the compositional changes leading to a lipid-rich but otherwise nutrient-poor TME had an impact on CD8⁺ T function. First, we conducted a longitudinal analysis of the T cell infiltration both in situ, using immunohistochemistry (Fig. 2, A and B), and ex vivo, by flow cytometry of isolated T cells (Fig. 2, C and D). We found that both CD4⁺ (Fig. 2 A, upper panels) and CD8⁺ (Fig. 2 A, lower panels) T cells penetrated into the pancreas at early stages of disease (Fig. 2 A, left panels) and also persisted until late stages (Fig. 2 A, right panels). Flow cytometric analysis of the immune cells infiltrating the pancreatic tissue showed that the percentage of total CD3⁺ T cells peaked at early stages of inflammation (early PanINs) and remained at a similar frequency in high-grade lesions (Fig. 2 D). In line with previous studies (De Monte et al., 2011; McAllister et al., 2014; Stromnes et al., 2015; Zhang et al., 2014), among CD3⁺ T cells, the absolute number of Foxp3⁺ regulatory T cells (T regs; Fig. 2 B) and the percentage of total CD4⁺ T cells (Fig. 2 C) were increased. On the contrary, both absolute number (Fig. 2 B) and frequencies (Fig. 2 C) of CD8⁺ T cells significantly declined over time while they progressively acquired characteristics of functional exhaustion, as indicated by up-regulation of exhaustion markers, including PD1 and TIM3 (Fig. 2, E and F) and reduced ability to produce effector cytokines and molecules such as IFNγ and Granzyme β (Fig. 2, G and H). Thus, PDA is able to recruit an initial CD8⁺ T cell immune response, but the CD8⁺ T cells become functionally impaired during tumor progression.

We have described that the glycerophospholipid composition of the TME changes and, concurrently, CD8⁺ T cells infiltrating PDA become progressively dysfunctional. Because T cells can use LCFA to fuel FAO (O’Sullivan et al., 2014; van der Windt et al., 2012), which may be especially useful in a low-glucose environment such as PDA, we queried whether intrapancreatic CD8⁺ T cells effectively acquire free FAs from the external microenvironment. Having found an enrichment in neutral lipids and glycerophospholipids containing VLCFA in the TME, we injected mice i.p. with a fluorescently labeled LCFA (palmitate, Bodipy-C16) and analyzed intrapancreatic CD8⁺ T cells 1 h later by flow cytometry.

Early and late PanINs infiltrating CD8⁺ T cells (in bold gray and black lines, respectively, Fig. 3, A and B) acquired significantly higher amounts of Bodipy-C16 compared with control splenic CD8⁺ T cells (in light gray, Fig. 3, A and B), while the uptake of a fluorescent glucose analogue, 2-[N-(7-nitrobenz-2-oxa-1,3-diazol-4-yl)amino]-2-deoxy-d-glucose (2-NBDG), was comparable between CD8⁺ T cell subpopulations (Fig. S2 A). Because FAs can passively cross the cell membrane in a limited extent, we next looked at surface expression of CD36, a receptor that binds LCFAs and facilitates their transport into the cells. We found that a significantly larger population of CD8⁺ T cells sorted from early and late PanINs expressed CD36 on their membrane compared with CD8⁺ T cells sorted from spleen (Fig. S2 B), consistent with an increased uptake of LC lipids by infiltrating CD8⁺ T cells. Of note, we observed higher Bodipy-C16 uptake also in CD4⁺ T cells and total CD45⁺CD3⁻ myeloid cells compared with their splenic counterparts (Fig. S2 C), suggesting a general mechanism by which immune cells infiltrating the pancreatic TME adapt to its peculiar nutrient composition by attempting to switch from a glucose- to a lipid-based metabolism.

To evaluate if and to what extent the nature of the lipid species acquired from the TME could impact the global lipid compositions of T cells, intrapancreatic CD8⁺ and CD4⁺ T cells were flow-sorted from early and late PanINs, and total lipids were analyzed by liquid chromatography–MS. As indicated from principal component analysis (PCA; Fig. S2 D), using abundance levels for all metabolite species across each sample to determine the principal axes of abundance variation, splenic naive T cells and early and late PanINs infiltrating CD8⁺ T cells have different metabolic profiles. Unsupervised clustering of the lipid expression profiles of CD8⁺ T cells revealed that the lipid composition of infiltrating CD8⁺ T cells was already significantly different from naive cells at early time points (Fig. 3 C). However, compared with early stages, CD8⁺ T cells infiltrating late PanINs exhibited even more pronounced compositional changes, with increase (399 and 165 ions detected in positive and negative mode, respectively) or decrease (14 and 106 ions detected in positive and negative mode, respectively) of a large number of lipid species (Fig. S2 E). In particular, annotation of statistically different lipid metabolic compounds over the time course of disease progression confirmed an overall increase of lipids in infiltrating CD8⁺ T cells, which was more marked in late PanINs in the superclass level of the FAs (Fig. S2 G, red column). These compositional changes were peculiar for CD8⁺ T cells infiltrating late PanINs. Indeed, infiltrating CD4⁺ T cells sharing the same TME showed minimal differences in early versus late PanINs (increase of 24 and 25 ions and decrease of 12 and 44 ions in positive and negative mode, respectively; Fig. S2 F), suggesting that, unlike CD8⁺ T cells, CD4⁺ T cells did not experience major lipid accumulation during tumor progression. Accordingly, when comparing the lipid compositional profiles of CD8⁺ and CD4⁺ T cells infiltrating late PanINs, we observed an overall increment of lipids at all superclass levels except glycerophospholipids (Fig. 3 D), which are the main structural components of cellular membranes. The greatest increase involved the FA superclass (Fig. 3 D, FA), and analysis of fold changes of average peak areas of putatively identified FA metabolites in either CD8⁺ or CD4⁺ T cells revealed an enrichment of different FAs at late time points (Fig. 3 E). All these FAs are LC lipids, with chain lengths ranging from C12 to C16. In summary, our lipidomic analysis confirmed that tumor-infiltrating CD8⁺ T cells encounter major changes in their lipid composition and accumulate mostly LCFAs at late stages.

We have shown that the pancreatic TME is enriched for LC and VLC glycerophospholipids and that infiltrating CD8⁺ T cells increase their uptake and accumulate LCFAs in high-grade lesions. Therefore, we asked whether an excess of specific LCFAs might contribute to the induction of T cell dysfunction. Having found palmitamide in the list of the FAs that accumulate in CD8⁺ T cells isolated from late PanINs (Fig. 3 E), we cultured CD8⁺ T cells in the presence of BSA-palmitic acid (Palm; palmitamide is a primary FA amide coming from palmitic acid). We found that Palm decreased proliferation (Fig. 3, F and G) and ability to secrete effector molecules such as IFNγ, Granzyme β, and TNFα (Fig. 3 H), mimicking the dysfunctional state of CD8⁺ T cells seen in vivo (Fig. 2, G and H). Of note, neither a medium-chain FA such as hexanoic acid (Fig. S2 H) nor an LCFA such as BSA-linoleic acid, which we did not find to accumulate (Lino; Fig. S2, I and J), had any relevant impact on CD8⁺ T cells in the same experimental setup, supporting the idea that CD8⁺ T dysfunction may be associated with the accumulation of specific lipid metabolites.

An excess of intracellular free FAs is known to damage cellular membranes (Aon et al., 2014), and to limit such cytotoxicity, they are stored in lipid droplets (LDs), which function as both an energetic reservoir and protective mechanism at once (Nguyen and Olzmann, 2017; Thiam and Beller, 2017; Welte, 2015). Although we did see formation of LDs in CD8⁺ T cells treated with Lino, we observed a significant defect in LD accumulation after Palm treatment, as assessed by both Red Oil and LipidTOX staining (Fig. S2, K and L, respectively), suggesting that this protective mechanism depends on the quality rather than quantity of LCFAs taken up and, in agreement with previous observations (Listenberger et al., 2003; Plötz et al., 2016), does not occur in response to excess palmitate. Based on these observations, we reasoned that effective storage of LCFAs in LDs could also be relevant for limiting lipotoxicity in CD8⁺ T cells infiltrating late PanINs. Therefore, we examined the ability of CD8⁺ T cells to form LDs in vivo during tumor progression. Even though >20% of the CD8⁺ T cells infiltrating early lesions contain LDs, this percentage was significantly reduced at late time points (Fig. 3, I and J; and Fig. S2 M). Altogether, these data suggest that CD8⁺ T cells infiltrating late PanINs have limited capacity to store specific LCFAs in LDs, and this may induce higher lipotoxicity and contribute to their exhaustion.

Elevated levels of free FAs are associated with apoptosis in many cell types (Sharma and Alonso, 2014; Ulloth et al., 2003). Moreover, palmitic acid has been reported to directly induce mitochondrial dysfunction (de Pablo et al., 1999), potentially by disrupting the mitochondria-associated ER membrane (Shinjo et al., 2017). We confirmed these findings in our in vitro optimally activated CD8⁺ T cells by looking at the effect of palmitic acid on their mitochondrial function. As expected, CD8⁺ T cells treated with palmitic acid showed decreased mitochondrial functionality (detected as a drop in the mitochondrial potential with MitoTracker Orange CMTMRos staining; Fig. 4 A). Conversely, treatment with linoleic acid did not induce any mitochondrial dysfunction (Fig. 4 B).

Because mitochondrial function strictly depends on their structural integrity, we employed transmission EM (TEM) to assess the state of mitochondria based on morphology in intrapancreatic CD8⁺ T cells sorted from early and late PanINs. CD4⁺ T cells from the same animals were used as a control population experiencing the same hostile TME. Morphometric analysis revealed that intrapancreatic CD8⁺ T cells experienced severe mitochondrial defects. These abnormalities were more pronounced than in CD4⁺ T cells and were already evident in cells infiltrating early PanINs but worsened during disease progression (Figs. 4, C and D). Indeed, CD8⁺ T cells infiltrating late PanINs displayed higher frequencies of mitochondria with irregular cristae and dilated intercrista spaces (swollen), as well as hypercondensed mitochondria with a distinct electron dense matrix (Fig. 4, C and D), which are usually found in preapoptotic cells (Green and Reed, 1998). Together with decreased mitochondrial mass (as assessed by MitoTracker Green FM; Fig. 4 E) and functionality (Fig. 4 F), these results highlight a severe mitochondrial impairment in CD8⁺ T cells infiltrating late PanINs and support the hypothesis that accumulated lipids may have a detrimental impact on mitochondrial integrity in these cells.

To address causality between uptake of specific LC lipids and acquired mitochondrial dysfunction, we activated CD8⁺ T cells in vitro in the presence of Palm or Lino and assessed mitochondrial integrity (MitoTracker⁺) in the cells accumulating LC lipids (Bodipy FL C₁₆⁺). As predicted, in the fraction of CD8⁺ T cells accumulating Bodipy FL C₁₆⁺, mitochondrial staining significantly decreased only upon Palm treatment (Fig. 4 G). Because CD8⁺ T cells accumulating neutral lipids (LipidTOX⁺) did not show similar mitochondrial defects even upon Palm treatment (Fig. 4 H), we concluded that mitochondrial defects tightly correlated with the type of lipid taken up and the ability to form LDs.

This link was further validated in vivo in tumor-infiltrating CD8⁺ T cells. Indeed, we injected Bodipy FL C₁₆, isolated CD8⁺ T cells from the tumor, and stained with MitoTracker Orange CMTMRos. This allowed us to track in vivo T cell lipid uptake and concurrently assess mitochondrial function. FACS analysis revealed a significantly higher frequency of Bodipy FL C₁₆–positive and MitoTracker-negative cells in the CD8⁺ compartment (in red) compared with the CD4⁺ compartment (in blue; Figs. 4, I and J). These results suggested that, although most CD4⁺ T cells could tolerate the uptake of LC lipids, a large fraction of the CD8⁺ T cells taking up LC lipids bore damaged mitochondria. Similar results were obtained when comparing CD8⁺ T cells with the myeloid cell compartment (CD45⁺ CD3⁻; Fig. 4 J), suggesting that lipid accumulation and mitochondrial dysfunction could be intertwined specifically within infiltrating CD8⁺ T cells.

Because T cells have to rely on mitochondrial OXPHOS to use lipids as an energy source, we predicted that the severe mitochondrial dysfunction experienced by CD8⁺ T cells infiltrating late PanINs could have an impact on their bioenergetic profile. Thus, we analyzed the metabolic flux of infiltrating CD8⁺ and CD4⁺ T cells sorted from early and late PanINs in real time using an extracellular flux analyzer (Seahorse XF Analyzers), which monitors the O₂ consumption rate (OCR) as an indicator of OXPHOS and the extracellular acidification rate (ECAR) as a readout for glycolysis. As expected, T cells (both CD4⁺ and CD8⁺) after activation with CD3/CD28 ligation dramatically increased OCR compared with naive T cells (Fig. 4, K and L), but preferentially used aerobic glycolysis (lower OCR/ECAR ratio; Fig. 4 M). By contrast, OXPHOS was vital for early infiltrating CD8⁺ and CD4⁺ T cells, as indicated by their increased OCR (Fig. 4, K and L) and OCR/ECAR ratio (Fig. 4 M) compared with resting naive T cells, likely as a means to meet the high metabolic demand necessary to exert their function within the tumor. However, CD8⁺ T cells progressively lost their ability to engage OXPHOS, as demonstrated by the significantly lower OCR (Fig. 4, K and L, red column) and OCR/ECAR ratio (Fig. 4 M, red column) at late time points compared with cells in early PanINs. In contrast, infiltrating CD4⁺ T cells increased OCR rate during tumor progression (Figs. 4, K and L, blue column), and the frequency of CD4⁺ T cells increased over time (Fig. 2 C). These data suggested that, unlike CD4⁺ cells, CD8⁺ T cells cannot adapt to the lipid-rich TME by switching their metabolism from glycolysis to OXPHOS. Together, our data show that the lipid-rich TME is populated with metabolically defective CD8⁺ T cells, which fail to effectively engage the OXPHOS pathway in response to limited glycolysis, an obligate step to sustain viability and exert effector functions.

To probe the mechanistic details of the metabolic dysfunction in intrapancreatic CD8⁺ T cells infiltrating late PanINs, we profiled the transcriptome of flow-sorted intrapancreatic CD8⁺ T cells. RNA sequencing (RNA-seq) highlighted distinct signatures associated with early and late PanINs. In particular, PCA (Fig. S3 A) and unsupervised clustering (Fig. 5 A) of expression profiles showed that infiltrating CD8⁺ T cells isolated from late PanINs were transcriptionally distinct from those isolated from the early PanINs. Compared with CD8⁺ T cells isolated from early PanIns, cells isolated from late lesions were characterized by down-regulation of key genes involved in cell-mediated cytotoxicity (orange dots, Gzma and Gzmb) and mitochondrial function (blue dots, Ndufa3, Ndufa9, Ndufa11, and Vdac3; Fig. 5 B), consistent with what we observed studying phenotypes. Moreover, there was up-regulation of genes involved in peroxisome biogenesis (Pex11g and Pex13) and ER stress (Atf6, Txndc11, and Txndc12; magenta dots, Fig. 5 B). Furthermore, gene set enrichment analysis (GSEA) of significantly differentially regulated genes demonstrated altered metabolic signatures in the different T cell populations, including gene subsets associated with FA homeostasis (Figs. S3 B).

Among these genes, the transcriptional down-regulation of the VLC acyl-CoA dehydrogenase (ACADVL) caught our attention, because this enzyme catalyzes the initial step of mitochondrial β-oxidation and targets esters of LCFAs and VLCFAs (with a chain length of 12–20 carbons; Tucci et al., 2012), which we had initially found to be accumulated in the areas infiltrated by CD8⁺ T cells during tumor progression.

Thus, we employed quantitative confocal microscopy to validate ACADVL down-regulation at protein levels. Enzyme levels were measured as mean fluorescence intensity (MFI) inside each T cell, and specific antibodies targeting CD8⁺ (Fig. 5 C) or CD4⁺ (Fig. 5 D) cells were used to distinguish between the two subpopulations at early and late time points (Fig. 5, C and D, left and right panels, respectively). Our analysis confirmed that CD8⁺ T cells had lower levels of VLC acyl-CoA dehydrogenase (VLCAD) when infiltrating late PanINs compared with early (as highlighted by differences in the number of yellow merged cells in Fig. 5 C compared with Fig. 5 D, and in the MFI in Fig. 5 E). By contrast, in the same lipid-rich TME, CD4⁺ T cells did up-regulate the level of VLCAD (Fig. 5, D and E), consistent with their ability to engage OXPHOS and adapt to the changing TME.

To investigate whether the accumulation of specific LCFAs in the TME was sufficient to induce the observed transcriptional changes, we assayed the expression of a subset of genes (n = 84) involved in the regulation and enzymatic pathways of FA metabolism in CD8⁺ T cells activated in vitro in the presence of palmitic or linoleic acid. In agreement with previous findings, we found that activated CD8⁺ T cells differentially regulate a large fraction (n = 58) of the tested genes compared with resting naive T cells (Fig. S3, C and D, left panel). However, the activation in the presence of Palm substantially changed their transcriptional profile (n = 54; Fig. S3, C and D, central panel), in particular of genes involved in FA metabolism. This recapitulates the changes observed in the RNA-seq analysis of flow-sorted infiltrating CD8⁺ T cells (Fig. 5, A and B). Interestingly, CD8⁺ T cells responded similarly to treatment with either Lino or Palm (Fig. S3, C and D, right panel), revealing a general transcriptional response to the excess of LCFAs. Of note, ACADVL was significantly down-regulated in LCFA-treated cells compared with control (Fig. S3, C and E), confirming that its expression is tightly dependent on LCFA accumulation.

Altogether, our data indicate that CD8⁺ T cells undergo major transcriptional changes after entering the TME, which involve down-regulation of mitochondrial function and lipid homeostasis in response to lipid accumulation. In the peculiar PDA TME, which is enriched in lipids containing LCFAs, the down-regulation of ACADVL is particularly relevant, because it impairs the ability of CD8⁺ T cells to metabolize the excess of specific LCFA lipids, amplifying their cytotoxic effect.

To evaluate whether enforcing overexpression of ACADVL has the potential to enhance T cell metabolic fitness in a clinically relevant model, CD8⁺ T cells, transduced to express a high-affinity TCR (designated TCR₁₀₄₅) specific for mesothelin (Msln), a protein tumor target overexpressed by human and murine PDA cells, were additionally engineered to also overexpress ACADVL. We transduced CD8⁺ T cells with a construct containing either TCR₁₀₄₅ alone or TCR₁₀₄₅ and ACADVL linked by a P2A element to ensure equimolar expression (Fig. 6 A). Transduced T cells expressed the engineered TCR (Vβ9⁺) and bound H2-D^b tetramers, the presenting MHC I allele, containing Msln_406–414 (Fig. S4 A). ACADVL overexpression in Msln-specific T cells resulted in elevated enzymatic activity in cells transduced with ACADVL-TCR₁₀₄₅ compared with TCR₁₀₄₅ alone (Fig. 6 B). We also observed higher mitochondrial function in the T cells overexpressing ACADVL (Fig. 6, C and D). Of note, 7 d after in vitro stimulation and transduction, both cell types similarly expressed activation markers CD69 and 4-1BB (Fig. S4 B) and produced similar amounts of the antitumor cytokines IFNγ and TNFα in response to stimulation with the cognate Msln_406–414 peptide (Fig. S4 C).

To assess the effect of ACADVL overexpression on the energetic profile of tumor-specific CD8⁺ T cells, Msln-specific T cells were evaluated for changes in OCR and ECAR 7 d after transduction and in vitro stimulation using the Seahorse Mito Stress Test. ACADVL-TCR₁₀₄₅ cells exhibited elevated basal OCR and spare respiratory capacity (Fig. 6, E and F) relative to TCR₁₀₄₅ T cells. These data suggest that enhanced ACADVL enzymatic activity in CD8⁺ T cells confers enhanced mitochondrial metabolic fitness and ability to respond to mitochondrial stress, which has the potential to improve T cell survival in a metabolically harsh TME, particularly one with excess VLCFAs.

To test this hypothesis, we evaluated the in vivo activity of these metabolically reprogrammed T cells in a model of adoptive T cell therapy (ATC) for genetically engineered PDA in KPC mice. In these mice, infused Msln-specific CD8⁺ T cells have been shown to infiltrate PDA and kill pancreatic tumor cells but to undergo severe contraction in cell number within 3 wk of infusion, and the cells that do persist become progressively dysfunctional, thereby failing to sustain long-term antitumor activity (Stromnes et al., 2015). Therefore, we monitored KPC mice until development of a defined pancreatic tumor mass (usually ∼10–12 wk, assessed by monitoring with serial high-resolution ultrasound imaging) and then adoptively cotransferred equal numbers of TCR₁₀₄₅ and ACADVL-TCR₁₀₄₅ T cells expressing different congenic markers to facilitate tracking each population of the engineered T cells over time in the same treated host within the same TME (Fig. 6 G). We found that, 10 d after infusion, ACADVL-TCR₁₀₄₅ T cells were present in the tumor at significantly higher numbers than TCR₁₀₄₅ T cells (Fig. 6 H), showing a better T cell metabolic fitness and survival advantage. This preferential survival of ACADVL-TCR₁₀₄₅ T cells was even more pronounced at 21 d after infusion, suggesting that T cells expressing ACADVL have improved survival and persistence in a metabolically hostile TME. ACADVL-TCR₁₀₄₅ T cells were also found in greater numbers in the spleen and blood relative to TCR₁₀₄₅ T cells (Fig. S4 D), suggesting that expression of ACADVL may confer an overall survival advantage to CD8⁺ T cells.

TCR₁₀₄₅ T cells infiltrating the TME have been shown to already exhibit reduced ability to produce antitumor cytokines by day 8 and to then become progressively dysfunctional and express high levels of inhibitory receptors, consistent with an exhausted phenotype by day 28 after entering KPC tumors (Nieman et al., 2011; Stromnes et al., 2015). Therefore, we investigated whether ACADVL-TCR₁₀₄₅ T cells preferentially retained effector function in the TME. By 10 d after injection, ACADVL-TCR₁₀₄₅ T cells expressed only low levels of PD-1, Tim-3, Lag-3, and TIGIT (T cell immunoreceptor with Ig and ITIM domains; Fig. S4 E), similar to what was reported for TCR_1,045 T cells. Further, both T cell populations isolated from the tumor 10 d after transfer already produced reduced amounts of IFNγ and TNFα compared with their splenic counterparts, but similar amounts of cytokine relative to each other (Fig. S4 F).

Together, these data suggest that metabolic reprogramming of engineered TCR₁₀₄₅ T cells by increasing ACADVL expression can improve the persistence of effector CD8⁺ T cells in the hostile TME, a substantive obstacle for CD8⁺ T cells targeting solid tumors. However, ACADVL expression in CD8⁺ T cells does not directly address the suppressive features of the PDA TME, and achieving sustained therapeutic benefit will likely require strategies not only to enhance metabolic fitness and survival, such as by increasing ACADVL expression, but also modulation of inhibitory pathways, such as by checkpoint blockade.

To assess the translational potential of our findings, we initially evaluated seven matched pairs of human PDA and adjacent normal tissues. Red Oil staining confirmed the tendency of human PDA tumors to accumulate neutral lipids (Fig. S5 A), and we then performed IMS to discern the compositional features of the human TME, as we had done above in the mouse studies. Similar to the murine model, we found that LC glycerophospholipids accumulated in five of the seven human PDA samples (Fig. 7 A), and that this accumulation was most apparent for PEs and PIs (∼10% increase, as highlighted by dotted lines in Fig. 7 B, and less for other classes; Fig. S5 B). As in murine samples, VLC sphingomyelins were by contrast relatively decreased (Fig. S5 B).

Motivated by these similarities with our murine model, we examined whether human CD8⁺ T cells also undergo transcriptional reprogramming and metabolic dysfunction when infiltrating the lipid-rich TME. Single-cell RNA-seq on 12 human PDA tumors and 9 matched adjacent normal pancreatic tissues confirmed that CD8⁺ T cells infiltrating these two different pancreatic microenvironments were transcriptionally distinct (Fig. 7 C and Fig. S5, C and D). In particular, human PDA-infiltrating CD8⁺ T cells had reduced expression of genes associated with a cytotoxic signature (Fig. 7 D) and transcriptionally resembled previously described exhausted T cells (Wherry et al., 2007; Fig. S5 F), as observed with the CD8⁺ T cells infiltrating the murine late PanINs. Of note, CD8⁺ T cells infiltrating PDA showed diminished expression of ACADVL (Fig. S5 E).

Multispectral image analysis of 22 samples of chronic pancreatitis and 96 PDA tumors probed with specific fluorescence-conjugated antibodies (Fig. 7 E) affirmed at the protein level the specific down-regulation of ACADVL in tumor-infiltrating CD8⁺ T cells (Fig. 7 F), demonstrating this to be a metabolic deficit that cytotoxic T cells acquire when infiltrating human PDA tumors. Altogether, these data reinforce a new model of tumor-specific T cells becoming metabolically dysfunctional in the TME of PDA and support future clinical studies of therapeutic modulation of ACADVL expression as a tool to increase metabolic fitness of tumor-specific T cells, and thus ultimately the efficacy of ACT.

The initial hypothesis that the density of the desmoplastic pancreatic TME prevented the influx of T lymphocytes, has been weakened by the fact that CD4⁺ T cells are abundant within the tumor (De Monte et al., 2011; McAllister et al., 2014; Zhang et al., 2014). In agreement, we show here, in a genetically engineered tumor model, that an immune response is recruited to the PDA TME early during progression. However, such a response is significantly contracted in numbers and functionality at late time points, in part as a consequence of being subjected to a combination of immunosuppressive mechanisms. Among such mechanisms, blockade of the immune checkpoint (ICB) has unprecedented therapeutic impact on a subset of tumors, motivating increased efforts to target the interactions between immune cells and the TME. Despite data from us and others (Diana et al., 2016; Shen et al., 2017; Winograd et al., 2015) showing that inhibitory checkpoint molecules are up-regulated in CD8⁺ T cells infiltrating PDA, ICB has lacked clinical efficacy in pancreatic cancer (Brahmer et al., 2012), suggesting that additional mechanisms account for the lack of response. Because the mediators of immune surveillance are metabolically interconnected in the TME with tumor cells, including competing for the same nutrients (Kishton et al., 2017), a deeper understanding of the metabolic state of the PDA TME may offer essential insights into the obstacles to PDA immunity and suggest complementary approaches to improve the efficacy of immunotherapy of PDA, including ICB.

Along with a limited availability of extracellular glucose and the decrease of glycolytic hexose intermediates, our time- and space-resolved compositional analysis of the PDA TME suggested that CD8⁺ impairment could be further rooted in metabolic dysfunction (Chang et al., 2015; Ho et al., 2015; Sukumar et al., 2015). Indeed, when glycolysis is limited, CD8⁺ T cells should have the metabolic flexibility to switch to mitochondrial OXPHOS (Chang et al., 2015; Sena et al., 2013), as a means to maximize the production of energy and reducing equivalents necessary to exert their effector functions (Zhang et al., 2017). This might be particularly useful for CD8⁺ T cells infiltrating the PDA TME, in which the low availability of glucose is paralleled by accumulation of VLCFAs. However, whereas tumor cells evolved to survive by feeding their bioenergetic needs with FAΟ (Vander Heiden et al., 2009), our study reveals that the transcriptional reprogramming experienced by intrapancreatic CD8⁺ T cells interferes with their ability to metabolize LCFAs, further contributing to their exhaustion. In this context, our data pinpoint ACADVL down-regulation as a crucial event. Indeed, impaired VLCAD enzymatic activity elicited accumulation of specific LC lipids, pronounced mitochondrial insufficiency, and failure of the effector functions. Remarkably, this closely recapitulates symptoms and biochemical findings in patients who inherited VLCAD deficiency. Indeed, situations of increased energy demand may induce lethargy, hypoketotic hypoglycemia, coma, and death, especially during fasting, viral illnesses, and prolonged physical exercise, because of a severe energy deficiency and the accumulations of toxic acylcarnitines (Wanders et al., 1999; Spiekerkoetter et al., 2003).

In addition to undermining their energy supply, this metabolic defect may subvert some of the biosynthetic pathways that govern CD8⁺ T cell effector functions, at both structural and signaling levels. For example, cholesterol and glycosphingolipids form signaling hubs in the plasma membranes of T cells known as lipid rafts that are essential for immune cell activation and function (Jury et al., 2007; Lingwood and Simons, 2010; Yang et al., 2016). Changes in the lipid composition affect the biophysical properties of lipid rafts (Janes et al., 2000) and, consequently, influence TCR signaling and contribute to defining functional T cell phenotypes (Köberlin et al., 2015). Moreover, the metabolic reprogramming associated with immune cell activation relies on the mevalonate pathway that generates cholesterol, the ubiquinone necessary for the mitochondrial electron transport chain, and a series of other bioactive intermediates (Gruenbacher and Thurnher, 2017). Interestingly, the classes of lipids that we have found to be enriched in the TME have been suggested to play roles in PDA in cell signaling disruption (e.g., PIs; Martin, 2003) and lipid raft formation (e.g., sphingolipids; Murai, 2012). When combined with the lipidomic analysis, our longitudinal IMS profiling suggests that the enrichment of some glycerophospholipids containing LC and VLC in the TME (starting in early PanINs in a context poor in glucose) temporally coincides with LC lipid uptake while preceding the accumulation of FAs (preferentially LC) and prenol lipids (a product of mevalonate pathway) and the appearance of the metabolically exhausted phenotype in intratumor CD8⁺ T cells. A possible explanation could impute the transcriptional reprogramming to the progressive intracellular accumulation of lipids as an attempt to reestablish lipid homeostasis and limit lipotoxicity. Our data support this model, for example by showing reduced expression in CD8⁺ cell from late PanINs of sterol regulatory element binding transcription factor 1 (SREBF1), which governs the expression of enzymes belonging to the mevalonate pathway, namely MVD and IDI1 genes, and FAO, in particular ACADVL. Mechanistically, the transcriptional reprogramming, which includes ACADVL down-regulation, could result from the prolonged cellular stress induced by lipid accumulation (Fu et al., 2011; Shao and Espenshade, 2012), as indicated by up-regulation of key genes involved in peroxisome biogenesis (Lodhi and Semenkovich, 2014) and ER stress, overall down-regulation of the unfolded protein response (UPR) pathway (Fig. S3 B), and dysfunctional mitochondria (Aon and Camara, 2015). Thus, low VLCAD enzymatic activity could amplify specifically the accumulation of LCFAs, worsening their associated toxicity. Indeed, the initial increase in intracellular lipids might be tolerated by increasing the conversion of diacylglycerol into triacylglycerol species, with their subsequent deposition between the ER membranes and the generation of LDs (Hapala et al., 2011). However, chronic exposure to LCFAs such as palmitate precludes LD formation, increases ER stress, dampens the UPR mechanism, and impacts the activity of mitochondria, in particular at the level of the mitochondria-associated ER membrane (Shinjo et al., 2017), where proteins like VDAC3 (Fig. 5 B) seem to play an important role (Messina et al., 2014). The early appearance of LDs in intrapancreatic CD8⁺ T cells and their decrease coinciding with the dysfunctional state of CD8⁺ T cells further supports this model (Gao and Goodman, 2015; Kim et al., 2016; Vevea et al., 2015), suggesting that changes in the TME may drive the transcriptional responses in infiltrating CD8⁺ T cells and, ultimately, lead to their metabolic exhaustion.

A question that warrants further study is why these metabolic defects are peculiar to CD8⁺ T cells. In our experiments, we found that CD4⁺ T cells and myeloid CD45⁺ CD3⁻ cells adapt to the lipid-rich PDA TME, becoming able to tolerate the increased uptake of LC lipid uptake and still preserving mitochondrial function. In particular, the up-regulation of VLCAD might confer a metabolic advantage to CD4⁺ T cells, the majority of which are likely T reg cells (Carstens et al., 2017; Hwang et al., 2016; Jang et al., 2017; Protti and De Monte, 2013). Indeed, it is known that the TME favors their survival while limiting CD8⁺ T cell activation, and this difference may reflect the ability of T reg cells to engage glycolysis, FA synthesis, and oxidation (Berod et al., 2014; Pacella et al., 2018; Procaccini et al., 2016). In agreement, data from us and others show that T reg cells increase in the pancreas during PDA progression (Fig. 2 B). Previous studies have demonstrated that CD4⁺ T reg cells are reprogrammed by FoxP3 to function in low-glucose environments (Angelin et al., 2017; Howie et al., 2017). Thus, CD8⁺ T cells, compared with other immune infiltrating cells, do not appear to have the necessary metabolic flexibility to survive in the hostile TME.

Our study paves the way for improving the efficacy of immunotherapy based on insights from the metabolic characterization of the TME. For instance, both CD4⁺ T cells (Fig. 5 E) and PDA tumor cells (Bailey et al., 2016) adapt to the lipid-rich TME by up-regulating VLCAD. Based on our data, systemic interventions aiming to inhibit FAO in tumors (Molina et al., 2018) or to activate it in immune cells (Scharping et al., 2016; Zhang et al., 2017) should be carefully weighed, because of the detrimental impact they could have on the CD8⁺ T cell survival and persistence. ATCs with engineered T cells is being evaluated for the treatment of PDA, with pioneering works in mice (Stromnes et al., 2015) and humans (Beatty and Moon, 2014) demonstrating feasibility, safety, and ability to mount therapeutic responses against pancreatic tumors. However, these T cells undergo a progressive loss of function and persist poorly, ultimately failing to maintain long-term activity and tumor eradication. Limited infiltration of infused tumor-reactive T cells, loss of function in cells that do infiltrate, and the inability to survive in the TME are substantive obstacles for ATC against solid tumors. Combining metabolic reprogramming to confer better metabolic fitness, engineered tumor-specific T cells may address some of these issues. Our data demonstrate that enhancing ACADVL expression improves the ability to respond to metabolic stress and, consequently, improves the ability of the T cells to persist and survive even in a metabolically unfavorable environment. Nevertheless, metabolic reprogramming has only a marginal effect on T cell function or the expression of inhibitory receptors. Thus, successful ATC will likely benefit from not only metabolically reprogramming T cells but also directly enhancing function, such as by combining with checkpoint blockade.

Although the first descriptions of fatty infiltration in the pancreas or fatty replacement in pancreas cells dates back nearly a century (Ogilvie, 1933; Schaefer, 1926), the origin of the fatty deposits and prognostic impact of such findings on the development and progression of pancreatic tumors has not been determined. Risk factors for pancreatic cancer such as obesity and metabolic syndrome are associated with increased circulating free FAs, which could infiltrate and accumulate in the pancreas, particularly after the death or damage of normal acinar pancreatic tissue (Smits and van Geenen, 2011). Pancreatic steatosis closely recapitulates the course of nonalcoholic fatty liver disease, with accumulated fat generating high levels of oxidative stress and consequent release of proinflammatory cytokines (Lesmana et al., 2015). Despite being often perceived as a benign condition, fat infiltrates are associated with tumor progression in PDA (Hori et al., 2014; Takahashi et al., 2018; Tomita et al., 2014), as well as other solid tumors such as prostate cancer (Chen et al., 2018; Nieman et al., 2011). Such findings further highlight the importance of characterizing, using the approaches described here, the nature of lipid deposits identified not only in PDA but in other cancer types.

All animal studies and procedures were approved by the Institutional Animal Care and Use Committees in compliance with the Guide for the Care and Use of Laboratory Animals. Animals were housed in a pathogen-free facility on a 12-h light/dark cycle and received food (standard rodent chow) and water ad libitum. All manipulations were performed under Institutional Animal Care and Use Committee–approved protocols. The KC (p48Cre;LSL-Kras^G12D) and KPC (p48Cre; Tp53^LSL-R172H;LSL-Kras^G12D) mouse models have been described (Hingorani et al., 2003, 2005). The mice were housed at MD Anderson Cancer Center (Houston, TX) and Fred Hutchinson Cancer Research Center (Seattle, WA).

Tissues were fixed in 10% formalin and embedded in paraffin. Sections (4 µm) were stained with H&E. Slides were scanned, and images were captured with a Nikon DS-Fi1 digital camera using a wide-field Nikon EclipseCi microscope and analyzed in blind by our pathologist, who scored the pancreatic sections as early and late PanIN lesions.

Tissues were fixed in 4% formaldehyde for 24 h at room temperature, moved into 70% ethanol for 12 h, and then embedded in paraffin (Leica, ASP300S). After cutting (Leica, RM2235) and baking at 60°C for 20 min for deparaffinization, slides were treated with citric acid–based antigen unmasking solution (Dako) for antigen retrieval according to the manufacturer’s instructions. Nonspecific signals were blocked using 5% goat serum for 1 h, and samples were stained overnight with primary antibodies at 4°C. Slides were incubated with species-specific secondary Alexa-conjugated antibodies (Invitrogen, 1:1,000) and mounted with Prolong Gold anti-fade plus DAPI (Life Technologies). Images were captured with a Nikon high-speed multiphoton confocal microscope (A1 R MP). Quantification was performed blinded by counting labeled cells in three to five high-power fields (40×) from three to eight mice per group. Primary antibodies used included CD3, CD4, CD8a, aSMA, e-cadherin, and ACADVL.

For quantification of neutral lipid by confocal microscopy, pancreatic tissue sections labeled with DAPI, Alexa Fluor 647–conjugated anti-CK19 antibody (Abcam), anti-CD8a primary antibody, Alexa Fluor 488–conjugated anti-rat secondary antibody, and LipidTOX Red (Thermo Fisher Scientific), were imaged on a Leica SP8 (Leica Microsystems) with a 25×/0.95 water-immersion objective lens. Z-stacks of the whole sections were acquired with Leica Navigator software, voxel size 0.3 × 0.3 × 1.5 µm in x, y, and z. The quantification of T cells in tissue and LDs in tissue and tumor areas was done on the maximum projections with a homemade ImageJ (National Institutes of Health) macro. Briefly, the tissue area was segmented according to the tissue autofluorescence in the green channel. The tumor areas were identified and segmented according to the e-cadherin staining, and the LDs were segmented according to the specific fluorescent red signal. The number of LDs and the LD area were calculated in the whole tissue and the tumor, and the T cells were identified and counted according to the specific Alexa Fluor 488 signal by the ImageJ macro. The automatically identified T cells were visually checked in the images, and the total T cell number was eventually corrected. In each T cell, the presence of LDs was evaluated by analyzing the original Z-stack to confirm the occurrence of LDs inside the cell cytoplasm. For each sample, a tissue area of ≥2 mm² was analyzed.

T cells were harvested from spleens and pancreatic samples. Briefly, single-cell suspensions were generated by mincing spleens with mesh screens in TCM medium followed by passing cells through a 40-µm filter (Thermo Fisher Scientific). Red blood cells were lysed with 1× RBC lysis buffer (eBioscience). To obtain tumor-infiltrating lymphocytes, tumors were harvested, cut into small fragments, and treated with 0.8 mg/ml collagenase IV and trypsin inhibitor (Sigma-Alrich) in HBSS 1× (Thermo Fisher Scientific) for 20 min at 37°C. Tumor fragments were homogenized, washed, and strained through a 70-µm filter.

CD8⁺ T cells freshly isolated from spleens of C57BL6 mice by flow cytometry cell sorting were incubated in complete medium supplemented with Palm (Agilent), Lino (Sigma-Aldrich), or a mixture of Palm and BSA-Oleate (Sigma-Aldrich) used at 100-µM concentration.

TEM was performed at the University of Texas MD Anderson Cancer Center High-Resolution Electron Microscopy Facility. CD8⁺ and CD4⁺ T cells were sorted from spleens and pancreata of mice bearing early and late PanIN lesions. Samples were fixed with a solution containing 3% glutaraldehyde plus 2% paraformaldehyde in 0.1 M cacodylate buffer, pH 7.3, for 1 h. After fixation, the samples were washed and treated with 0.1% Millipore-filtered cacodylate-buffered tannic acid, postfixed with 1% buffered osmium tetroxide for 30 min, and stained en bloc with 1% Millipore-filtered uranyl acetate. The samples were dehydrated in increasing concentrations of ethanol, infiltrated, and embedded in LX-112 medium. The samples were polymerized at 60°C for 2 d. Ultrathin sections were cut using a Leica EM FC7 ultramicrotome. After contrasting with uranyl acetate and lead citrate, sections were analyzed with a Leo 912AB transmission electron microscope (Carl Zeiss). Images were acquired with a 2K bottom-mounted slow-scan Proscan camera controlled by EsivisionPro 3.2 software. For morphometrical analysis, for each experimental condition, 25 random whole cellular profiles were acquired by stitching together several images acquired at a nominal magnification of 6,300× (pixel size, 1.77 nm). Using ImageJ, mitochondria present in each cellular profile were classified as normal (narrow and long cristae with uniform and moderate electron-dense matrix), swollen (with increased size and patchy electrolucent matrix), or hypercondensed (very electron-dense matrix with dilated cristae). The few mitochondria for which the assignment to one category was not possible (less than optimal ultrastructural preservation or an unfavorable plane of section) were not considered in the analysis.

Histologic visualization of fat cells and neutral fat was performed using the Oil Red O Kit from Abcam. Briefly, frozen mouse or human pancreatic sections were placed in propylene glycol for 2 min and incubated with Oil Red O Solution. Sections were differentiated in a mixture of 85% propylene glycol in distilled water, stained with hematoxylin, and mounted with aqueous mounting medium.

Glucose concentrations in pancreas supernatant from early and late PanINs were measured using the Glucose Assay Kit (Eton Bioscience). For ex vivo glucose levels, harvested samples were weighed and minced in fixed amounts of PBS. Glucose concentration was quantified relative to the weight of tumors and the volume of collected supernatant, and normalized with glucose concentrations in WT tumors.

Seahorse experiments were performed on sorted intrapancreatic CD8⁺ and CD4⁺ cells using XF Cell Mito Stress kit (Seahorse Bioscience). OCR and ECAR were measured with XF96 Extracellular Flux Analyzers (Seahorse Bioscience) as previously described (van der Windt et al., 2016). Briefly, cells were plated on poly-d-lysine–coated 96-well polystyrene Seahorse plates (100,000 T cells/well), equilibrated for 1 h at 37°C, and assayed for OCR (pmol/min) and ECAR (mpH/min) in basal conditions and after addition of oligomycin (1 µM), carbonyl cyanide-4-phenylhydrazone (1.5 µM), and antimycin A/rotenone (1 µM/0.1 µM).

Fluorophore-conjugated antibodies were purchased from BD Biosciences: CD45 (30-F11), CD3 (145-2C11), CD8α (53-6.7), CD4 (GK1.5), Thy1.1 (OX-7), Thy1.2 (53–2.1), CD69 (H1.2F3), CD44 (IM7), PD1 (29F.1A12), Lag3 (C9B7W), Tim3 (RMT3-23), TIGIT (1G9), 4-1BB (17B5), and Vb9 (MR10-2). Reagents were obtained from BD Bioscience, eBioscience, and BioLegend. The MSLN_406–414/H-2Db tetramer was conjugated to APC by the Fred Hutchinson Immune Monitoring Core. Msln peptides were purchased from Elim Peptide. For intracellular cytokine production, 10⁶ cells isolated from spleen or primary tumors were incubated for 5 h in Golgiplug (BD Biosciences) with or without MSLN_406–414 peptide, stained for surface markers, fixed with BD Biosciences Fix/Perm, permeabilized using BD Biosciences Perm/Wash, and incubated with antibodies to IFNγ (XMG1.2), TNFα (MP6-XT22), IL-2 (JES6-5H4), and Granzyme B (GB11). For transcription factor staining, Tbet (4B10) and Eomes (Ef450) were purchased from BD, and cells were stained for surface markers, fixed with the eBioscience Fix/Perm kit, permeabilized using eBioscience Perm/Wash Buffer, and stained intracellularly. Cell sorting was performed using a FACSAria II (BD) with >95% purity based on the exclusion of dead cells and the expression of CD45, CD3, and CD4/CD8. For mitochondrial studies, cells were incubated with 50 nM MitoTracker Green and/or 25 nM MitoTracker Orange for 30 min at 37°C before staining. Reagents were obtained from Life Technologies. All panels were stained with LIVE/DEAD fixable Aqua stain or Ghost Dyes from TONBO before surface or intracellular staining. UltraComp eBeads were used for all compensation. All FACS data were acquired using a BD-LRSFortessa X-20 and analyzed with FlowJo software.

Mice were injected i.p. with 50 µg Bodipy FL C16 (Thermo Fisher Scientific) in 50 µl DMSO or with 100 µg 2-NBDG (Cayman Chemicals) in PBS, 60 or 20 min, respectively, before sacrifice and cell isolation.

Liquid chromatography–MS lipidomic analyses were performed on sorted intrapancreatic T cells using reverse-phase liquid chromatography (RPLC) coupled with positive- and negative-mode electrospray-ionization high-resolution MS. Lipids from the samples were extracted using the Folch extraction method. Briefly, cell pellets were resuspended in 100 µl ammonium formate buffer, 50 mM, pH 6.8, followed by the addition of 400 µl ice-cold methanol. Each sample was vortexed and sonicated for 10 pulses at 30% power using a sonication probe while cooling on ice. 800 µl chloroform and 100 µl water were added to each sample to a final ratio of 4:2:1 chloroform:methanol:water. Samples were vortexed, incubated on ice for 10 min, and spun down at 1,000 rpm for 10 min. The bottom layer (organic layer containing the lipids) was transferred to a clean Eppendorf tube, and the extraction with chloroform of the aqueous layer was repeated once. The organic layers (lipid layers) were coupled and dried down under vacuum. Before MS analysis, dried lipids were reconstituted in chloroform:methanol (50:50, vol/vol) containing an internal heavy labeled lipid standard (Splash Lipidomix Mass Spec Standard; Avanti Polar Lipids) and centrifuged for 5 min at 15,000 rpm to remove insoluble material. Samples were volume adjusted to ensure that cell count per volume (µl) was normalized. Global untargeted lipidomic analyses were performed using full MS and data-dependent acquisition analyses on a Q-Exactive HF hybrid quadrupole-Orbitrap mass spectrometer (Thermo Fisher Scientific) equipped with a Vanquish UHPLC binary system and auto-sampler (Thermo Fisher Scientific). Extracts (5-µl injection volume for RPLC-POS and 6 µl for RPLC-NEG) were separated by RPLC on a Hypersil Gold 1.9-µm, 2.1 × 100-mm column (Thermo Fisher Scientific) held at 40°C. Reverse-phase separation was performed at 250 µl/min using solvent A (0.1% formic acid in water) and solvent B (0.1% formic acid in isopropanol:acetonitrile:water [60:36:4]) with the following gradient: 30% B for 0.5 min, 30–70% B over 1.5 min, 70–99% B over 13 min, 99–10% B over 1 min, 10% B held 4 min, 10–30% B over 1 min, and 30% B held over 5 min (gradient length 30 min).

Ultra-performance liquid chromatography/tandem MS raw data were imported, processed, normalized, and reviewed using Progenesis QI v.2.1 (Non-linear Dynamics). All sample runs were aligned against a quality control pool reference run, and peak picking was performed on individual aligned runs to create an aggregate dataset. Features (retention time and m/z pairs) were combined using both adduct and isotope deconvolution. Data were normalized to the spike-in lipid standards as an abundance ratio. Statistically significant changes were identified using multivariate statistical analysis, including PCA, and P values were generated using ANOVA or pairwise comparison. Tentative annotations (Schrimpe-Rutledge et al., 2016) were made within Progenesis using accurate mass measurements (<5 ppm error), isotope distribution similarity, and assessment of fragmentation spectrum matching (when applicable) from the Lipid Maps database (Fahy et al., 2007). Four biological replicates and two technical replicates from each sample type were used to calculate the fold change and P value, and features were considered differentially expressed only if they met both criteria of fold change ≥|2| and significance (P ≤ 0.05). Heatmap visualization of individual lipid features was performed using Metaboanalyst 4.0 software (Chong et al., 2018) using Pearson average clustering algorithm. To assess instrument performance and reproducibility throughout our experimental run sequence, the retention times and peak areas of SPLASH Lipidomix standards were monitored.

All studies using human specimens were approved by the University of Texas MD Anderson Cancer Review Board. Tumor tissues were obtained from patients who provided written informed consent for collection and analysis of tumor samples under Institutional Review Board–approved protocols. Informed consent was obtained from each patient before tumor tissue collection. To ensure patient confidentiality, no information on age, sex, or ethnicity was provided to the investigators of this study.

CD8⁺ T cells were sorted with >95% purity using FACSAria from BD directly into RLT lysis buffer (Qiagen). RNA was isolated from purified cells using RNeasy Mini kits (Qiagen), and RNA concentrations were determined using Nanodrop. Fastqc software was used on the fastq files to generate the quality control report, and Tophat software was used to align the sequences to the mouse genome mm10. HTSeq was used to count the read number of genes from each sample. Genes with read counts <10 in all the samples were considered to be noise and therefore removed. The obtained read counts are analogous to the expression level of each gene across all the samples. Differential expression analysis were performed using DESeq2 (Love et al., 2014). Genes with raw mean reads >10 (i.e., ∼9,225 genes) were used for normalization and differential gene expression analysis using DESeq2 package in R. The Wald test defined in the DESeq function of the package was used for differential expression and further downstream analysis. PCA, hierarchical clustering plots (hclust), and XY plots between replicates of the same genotypes were used to examine for nominal amounts of nontechnical variation and other latent factors. PCA and volcano plots were generated using ggplot2 package, and all the heatmaps were generated using pheatmap or heatmap2 package in R environment. Gene overlap analysis was performed against the Molecular Signatures Database (Liberzon et al., 2015; Subramanian et al., 2005) Investigate Gene Sets web program under GSEA.

Gene expression analysis of 84 key genes involved in FA metabolism was assessed on naive or in vitro–activated CD8⁺ T cells using a Mouse Fatty Acid Metabolism RT² Profiler PCR Array (Qiagen; PAMM-007Z) according to manufacturer instructions. Ct values were exported to an Excel file for data analysis using the provided web tools (http://www.qiagen.com/geneglobe). Ct values were normalized based on a panel of reference genes. The data analysis web portal calculates fold changes using using the ΔΔCt method followed by ΔΔCt calculation (ΔCt_{test group} − ΔCt_{control group}). Fold change is then calculated using 2^(−ΔΔCt) formula. Data were plotted using scatter plot and heatmap.

Frozen pancreata were sectioned at 10 µm on a CM3050S Cryostat (Leica Biosystems; chuck temperature of −20°C, chamber temperature of −24°C). Serial sections were alternately thaw mounted onto microscope slides (Thermo Fisher Scientific) for histology and onto a gold-coated stainless-steel target for IMS and stored at −80°C until use.

Gold-coated targets with tissue sections were removed from the freezer and allowed to warm to room temperature in a vacuum desiccator to prevent water condensation on the slide. 2,5-dihydroxyacetophenone (Sigma-Aldrich) was used as a MALDI matrix in both positive and negative ion modes and was applied to the slide using a custom-built sublimation apparatus (110°C, <70 mtorr, 6.5 min), as previously described (Hankin et al., 2007). Slides for histologic analysis were stained with H&E and analyzed using a SCN400 Brightfield Slide Scanner (Leica Microsystems) or analyzed by immunofluorescence.

MALDI IMS was performed on either a 15T or 9.4T SolariX FT-ICR mass spectrometer (Bruker Daltonics) equipped with an Apollo dual MALDI/ESI source. The MALDI source uses a Smartbeam II 2 kHz Nd:YAG (355-nm) laser that has been modified to provide a Gaussian beam laser profile. Positive-ion-mode mouse pancreas lipid images were acquired from m/z 500 to 2,000 on the 15T FT-ICR MS using a time-domain file size of 512K, resulting in an ICR transient length of 0.5767 and a resolving power of ∼100,000 at m/z 796.525. An internal lock mass calibration (m/z 734.5694, [PC(32:0)+H]⁺) was used during image acquisition, and a multipoint external mass calibration was performed after data acquisition using common PC ions generated from the tissue samples. Images were acquired using a 75-µm raster step and 400 shots per pixel. Negative-ion-mode mouse pancreas lipid images were also acquired from m/z 500 to 2,000 on the 15T FT-ICR MS using a time-domain file size of 512K. An internal lock mass calibration (m/z 885.5494, [PI(38:4)-H]⁻) was used during image acquisition, and a multipoint external mass calibration was performed after data acquisition using common glycerophospholipid ions generated from the tissue samples. Images were acquired using a 75-µm raster step and 400 shots per pixel. Negative-ion-mode mouse pancreas metabolite images were acquired from m/z 100 to 1,400 on the 9.4T FT-ICR MS using a time-domain file size of 2M, resulting in an ICR transient length of 0.7340 and a resolving power of ∼215,000 at m/z 259.023. Images were acquired using a 100-µm raster step and 500 shots per pixel. Positive-ion-mode human pancreas lipid images were acquired from m/z 500 to 3,000 on the 9.4T FT-ICR MS using a time-domain file size of 256K, resulting in an ICR transient length of 0.4544 and a resolving power of 40,000 at m/z 796.525 An internal lock mass calibration (m/z 734.5694, [PC(32:0)+H]⁺) was used during image acquisition, and a multipoint external mass calibration was performed after data acquisition using common PC ions generated from the tissue samples. Images were acquired using a 50-µm raster step and 500 shots per pixel. Negative-ion-mode human pancreas lipid images were acquired from m/z 500 to 2,000 on the 9.4T FT-ICR MS using a time-domain file size of 256K. An internal lock mass calibration (m/z 885.5494, [PI(38:4)-H]) was used during image acquisition, and a multipoint external mass calibration was performed after data acquisition using common glycerophospholipid ions generated from the tissue samples. Images were acquired using a 50-µm raster step and 500 shots per pixel. All ion images were generated using FlexImaging 5.0 (Bruker Daltonics) and plotted as ±0.005 daltons using root mean square normalization unless otherwise noted.

Lipids were identified using accurate mass measurements (<5 ppm) by searching against the LIPID MAPS (Lipidomics Gateway, http://www.lipidmaps.org) and METLIN (The Scripps Research Institute, https://metlin.scripps.edu) online databases and are reported using total carbon: double bond nomenclature. Lipid identification data analysis was performed using DataAnalysis and Compass IsotopePattern (Bruker Daltonics) as well as the mMass software package (Strohalm et al., 2008).

To visualize the global lipidomic changes in IMS data across multiple time points, a novel data analysis strategy was developed using both SCiLS (Bruker Daltonics) and R. SCiLS bisecting k-means clustering was used to segment regions of interest in the root mean square–normalized imaging data that positively correlated with CD8 immunostaining performed on serial tissue sections. Early-stage (9 and 13 wk) and late-stage (20 and 30 wk) IMS data were pooled and compared with a control (4 wk). An in-house-written R program that uses the LIPID MAPS database was used to extract the mean intensity values for each accurate mass lipid identification ±0.0009%. The resulting list was exported from SCiLS as a .csv file and imported into R. A second in-house-written R script was used to filter this list and generate a heatmap of log2 fold change intensity values. The filters in the program were used to reduce false-positive results as well as remove lipids with zero intensity in the control group, log2 values between 1 and −1, and intensity values <0.5 of the quantile in the control group. The lipids in the heatmap were sorted by head group, fatty acyl chain length, and level of fatty acyl chain unsaturation. These exported data were also used to make the horizontal bar plots. Briefly, the m/z peak list was searched against the LIPID MAPS database. Isotopic distributions of identified lipids were confirmed manually before importing a list of the lipid identifications and mean intensity values into Excel. Intensities of lipids with different levels of fatty acyl chain unsaturation but the same total fatty acyl chain length were summed within each separate lipid class (e.g., the intensities of PC(34:2) and PC(34:1) were summed). Within each lipid class, the intensity of each total fatty acyl chain length group was normalized to the total intensity of all ions within the lipid class so as to readily compare changes in the relative lipid compositions at multiple time points. For these bar plots, PI(38:4) was removed from the analyses for both sets of samples because it is extremely abundant and distorted the bar plots. The PCs and sphingomyelins were detected in positive ion mode, and the relative intensities reflect the summed distributions of the [M+H]⁺, [M+Na]⁺, and [M+K]⁺ ion types. Ether and vinyl ether (plasmalogen) lipids as well as modified (e.g., by lipid peroxidation) lipids were excluded from the bar plot analyses. IMS quantification was performed by exporting the pixel intensities of the pooled data segments and removing zero-intensity values (attributed to peak picking inaccuracy). These data were used to calculate the average and SEM for this replicate, reflecting the heterogeneity of the disease at each stage.

The ecotropic retroviral packaging cell line, Platinum (Plat)-E cells (Cell Biolabs), were plated (2.2 × 10⁶ cells) in PLAT-E medium (DMEM, 10% FBS, 10 µg/ml blasticidin, 1 µg/ml puromycin, and 100 U/ml penicillin/streptomycin) on 10-mm culture plates. After 24 h, Plat-E cells were transfected with DNA encoding the desired TCR and/or construct using the Effectene transfection kit (Qiagen). At 48 h, medium was replaced with T cell medium (IMDM, 10% FBS, 2 µM l-glutamine, 100 U/ml penicillin/streptomycin, and 25 µM 2-β-mercaptoethanol). On days 2 and 3 after transfection (at 72 and 96 h), virus-containing supernatant was collected. Splenic T cells isolated from P14 Thy1.1⁺ mice were stimulated with α-CD3e (clone 145-2C11) and α-CD28 (clone 37.51; both from BD PharMingen) and IL-2 (50 IU/ml aldesleukin; University of Washington Pharmacy) and transduced with retroviral supernatant by spinfection in polybrene (5 µg/ml, 90 min at 1000 g) 24 and 48 h after T cell activation.

For T cell enrichment experiments, live CD8⁺ Vβ9⁺ T cells were sorted, mixed at a 1:1 ratio, and transferred into KPC tumor-bearing mice. Recipient mice received ≤10⁷ total engineered T cells (stimulated once), cotransferred in equal numbers. For T cell phenotyping and function experiments, cells were sorted as above and cotransferred with ≤5 × 10⁷ antigen-pulsed irradiated splenocytes followed by IL-2 (2 × 10⁴ IU, s.c.) daily for 10 d to promote T cell expansion and survival. In all experiments, mice were treated with one dose of cyclophosphamide (180 mg/kg) i.p. ∼8 h before T cell transfer.

KPC mice underwent serial high-resolution ultrasound imaging (Vevo 2100) at 8 wk of age to monitor autochthonous tumor development. Mice were enrolled based on defined pancreatic mass with ≥3-mm diameter.

Single-cell transcriptome sequencing was performed on human PDA samples on the 10X Genomics Chromium platform using the Single Cell 3′ v2 chemistry per the recommended protocol. Data were processed using Seurat v2.2 in accordance with the recommended analysis for multisample data. All cells were clustered using sharing nearest neighbor, and the clusters most highly expressing CD3 genes were considered T cells and extracted for further analysis. T cells were then classified as CD4 or CD8 based on imputed expression of CD4 and CD8a/CD8b, respectively. Differential gene expression was performed using a Wilcoxon test with Bonferroni P value correction for multiple comparisons.

Data are deposited in the Gene Expression Omnibus database under accession no. GSE147465.

All statistical analyses were conducted using GraphPad Prism 6. Differences between two groups were calculated by unpaired, two-tailed Student’s t test. Multiple comparisons were performed by one-way ANOVA followed by Bonferroni’s multiple comparison tests. Significance was set at P values ≤ 0.05. For all figures: *, P ≤ 0.05–0.01; **, P ≤ 0.01–0.001; ***, P ≤ 0.001–0.0001; ****, P ≤ 0.0001. Unless noted in the figure legend, all data are shown as mean ± SEM. The n numbers for each experiment and the numbers of experiments are noted in the figure legends.

Fig. S1 shows lipid accumulation and glucose and H6P decrease in TME during PDA progression. Fig. S2 shows CD8⁺ T cells accumulating specific LC lipids and becoming dysfunctional during PDA progression. Fig. S3 shows transcriptomic analysis of CD8⁺ T cells in vivo and in vitro.Fig. S4 shows how ACADVL overexpression confers enhanced metabolic fitness. Fig. S5 shows lipid accumulation and ACADVL deregulation in human CD8⁺ T cells infiltrating PDA.

L. Nezi and T. Manzo dedicate this work to Leonardo and Lara Nezi. We thank V. Gopalakrishnan and A. Reuben for reagents; Liam McDonnel for thoughtful discussion; J. Wang and W. Ma for support with RNA-seq experiments; Kenneth Dunner Jr. for helping with electron microscopy experiments; and Caterina Bartolacci for preliminary lipidomic analysis. The lipidomic analyses were provided by the Vanderbilt University Center for Innovative Technology, Vanderbilt Institute of Chemical Biology; the Vanderbilt Institute for Integrative Biosystems Research and Education; and Vanderbilt University (S.D. Sherrod, G.S. Codreanu, and J.A. McLean).

This work was supported by the University of Texas MD Anderson Cancer Center Pancreatic Moon Shot to G. Draetta, Cancer Prevention and Research Institute of Texas grant RP160471 to G. Draetta, US Department of Defense grant W81XWH-11-1-0418 to G. Draetta, Associazione Italiana per la Ricerca sul Cancro StartUp grant to T. Manzo and core funding from IRCCS European Institute of Oncology to L. Nezi. IMS research was performed in the laboratory of R.M. Caprioli and supported in part by National Institutes of Health/National Institute of General Medical Sciences P41 GM103391-08 and National Institutes of Health/National Institute of Diabetes and Digestive and Kidney Diseases F32 FDK105841A (B.M. Prentic). Some of this research was performed at the MD Anderson Cancer Center Flow Cytometry and Cellular Imaging Facility, which is supported by National Cancer Institute grant P30CA016672. The work was also supported by National Cancer Institute grant CA33084 to P.D. Greenberg. This work was partially supported by the Ministero della Salute with Ricerca Corrente and 5x1000 funds.

Author contributions: T. Manzo and L. Nezi designed and coordinated the whole study. T. Manzo, K.G. Anderson, C.B. Nava Lauson, S. Tiberti, B.M. Bates, S. Tucci, and L. Nezi designed, executed, and analyzed experiments; B.M. Prentice, M.A. Jones, M. Reyzer, J.M. Spraggins, N.H. Patterson, and R.M. Caprioli designed and executed the IMS analysis; G.S. Codreanu, J.A. McLean, and S.D. Sherrod designed and executed the lipidomic analysis; A. Raman, A. Schalck, K. Rai, and N.E. Navin performed computational analysis; A. Raimondi and C. Tacchetti performed and analyzed TEM; J.A. Wargo and M. Kim obtained tissues from patients; S. Rodighiero and K. Clise-Dwyer analyzed experiments and provided technical assistance; T. Manzo, P.D. Greenberg, and L. Nezi analyzed the data, wrote, and edited the manuscript. Funding acquisition: T. Manzo, G. Draetta, and L. Nezi.