Deficiency of metabolic regulator PKM2 activates the pentose phosphate pathway and generates TCF1+ progenitor CD8+ T cells to improve checkpoint blockade

TCF1high progenitor CD8+ T cells mediate the efficacy of PD-1 blockade, however the mechanisms that govern their generation and maintenance are poorly understood. Here, we show that targeting glycolysis through deletion of pyruvate kinase muscle 2 (PKM2) results in elevated pentose phosphate pathway (PPP) activity, leading to enrichment of a TCF1high central memory-like phenotype and increased responsiveness to PD-1 blockade in vivo. PKM2KO CD8+ T cells showed reduced glycolytic flux, accumulation of glycolytic intermediates and PPP metabolites, and increased PPP cycling as determined by 1,2 13C glucose carbon tracing. Small molecule agonism of the PPP without acute glycolytic impairment skewed CD8+ T cells towards a TCF1high population, generated a unique transcriptional landscape, enhanced tumor control in mice in combination with PD-1 blockade, and promoted tumor killing in patient-derived tumor organoids. Our study demonstrates a new metabolic reprogramming that contributes to a progenitor-like T cell state amenable to checkpoint blockade.

These therapies target inhibitory receptors expressed on the CD8+ T cells including PD-1, which is upregulated upon activation and expressed as cells accumulate dysfunctional characteristics 5,6 . Recent studies have demonstrated that progenitor-exhausted CD8+ T cells expressing transcription factor TCF1 (T cell factor 1, encoded by Tcf7), are key mediators of anti-PD-1 therapy to inhibit tumor progression 7,8 . While TCF1 has been associated with progenitor-exhausted T cell populations, it is also a critical regulator of memory CD8+ T cell development [9][10][11] . TCF1 expressing CD8+ T cells mediate more durable anti-tumor immunity compared to their TCF1effector-like counterparts 7,8,11,12 . Upon antigenic stimulation, TCR ligation drives differentiation of CD8+ T cells associated with marked alterations in the transcriptional and metabolic landscape [13][14][15][16] , with effector cells upregulating aerobic glycolysis to support their proliferative and cytotoxic phenotypes. Recent studies have begun to assess the potential of manipulating metabolic pathways to alter differentiation status of CD8+ T cells [17][18][19] , but it remains unclear how these transcriptional and metabolic changes affect T cell differentiation to mediate anti-tumor e cacy. In this study, we interrogated the metabolic landscape of tumor-in ltrating CD8+ T cells and show that loss of metabolic regulator PKM2 serves as a key hub controlling TCF1+ central memory-like progenitor cells to improve responsiveness to PD-1 checkpoint blockade. Our results demonstrate that while glycolysis is critical for effector cell activity, a temporary dampening may serve to skew T cell differentiation to provide durable anti-tumor immunity and checkpoint blockade synergy.

Results
A screen of glycolytic genes in intratumoral CD8 + T cells identi es Pyruvate Kinase M (PKM) as a potential regulator of cell fate.
To investigate the evolving transcriptome as a function of tumor progression, we used RNA sequencing analysis of tumor-in ltrating CD8 + T cells from the HKP1 orthotopic mouse model of NSCLC (Fig. 1A) 20 . Principal component analyses (PCA) showed clustering of the samples as a function of time and tumor burden with CD8 + T cells from large tumors exhibiting a distinct transcriptional landscape (Fig. 1B). We integrated this treatment-Next, we used the co-culture system to determine the impact of PKM de ciency on T cell phenotype. We infected activated OT-I + T cells with shRNA targeting either PKM or a negative control, CD4 (Extended Data Fig. 3C-3D). Following co-culture with HKP1-ova-GFP cells, at 6 days after initial stimulation, shPKM T cells showed increased expression of SlamF6 and TCF1, whereas GzmB, Tim 3,and CD39 were reduced (Fig. 2D;. Concurrently in these cultures, we observed reduced cytokine-producing populations with PKM knockdown and increased abundance of a TCF1 high population ( Fig. 2E-2G). To speci cally target the PKM2 isoform, we used CRISPR/Cas9 ribonucleoprotein complexes with isoform-speci c guides to target exon 10 in activated OT-I + T cells; this resulted in robust loss of PKM2, and compensatory elevation of PKM1 expression (Extended Data Fig. 3E-3G). PKM2 knockout (PKM2 KO ) T cells following co-culture with HKP1-ova-GFP tumor cells showed higher levels SlamF6 and TCF1 with concomitant decreases in GzmB, IFNγ, and TNFa ( Fig. 2H-K). As before, enhanced TCF1 levels were due to increased abundance of the TCF1 high population, while knockout resulted in reduced cytokine-producing populations ( Fig. 2I-2K). PKM2 loss had a durable effect on altering differentiation, also displaying increased proportions of TCF1 high cells at day 9 post-initial-stimulation, when a more profound exhaustion phenotype 22 had emerged in PKM2 wild-type (PKM2 WT ) T cells (Extended Data Fig. 5). Taken together, these data suggest that loss of PKM2 after T cell activation reduces effector differentiation.
PKM2 loss alters CD8 + T cell differentiation states in NSCLC and melanoma models in vivo.
We next determined the consequence of PKM2 de ciency on CD8 + T cell phenotypes in the tumor microenvironment (TME) and their draining lymph nodes (dLNs) in vivo. To control for possible phenotypic variation due to inter-mouse heterogeneity, we performed adoptive co-transfers of a mix of PKM2 KO  To determine if the effects of PKM2 deletion on T cell phenotype were not limited to a Kras G12D/+ p53 −/− model of NSCLC, we employed the B16F10 melanoma model expressing Ova 257 − 264 and GFP, B16F10-ova-GFP. We showed dominance of PKM2 WT T cells in both the tumor and dLN tissue, with higher cytokine expression in PKM2 WT cells (Extended Data Fig. 6E-6I). As in NSCLC, PKM2 KO T cells exhibited a central memory-like phenotype, with elevated CD44 + CD62L + proportions at both timepoints in both tissues and enhanced TCF1 and Eomes expression at the later timepoint, marginally in the draining lymph node and signi cantly in the tumor (Extended Data Fig. 6J-6O). Together, the data demonstrate that PKM2 is expressed in tumor-speci c T cells upon activation, and its loss results in a central memory-like state across multiple tumor types but with potential cancer-speci c differences in magnitude of effect.
In vivo PKM2 loss synergizes with checkpoint blockade to promote anti-tumor activity.
We observed that PKM2 loss in tumor-speci c CD8 + T cells altered their differentiation state, resulting in a TCF1 high central memory-like phenotype. However, the ensuing CD8 + T cell phenotypes were not su cient to impair tumor growth or improve overall survival (Fig. 4A-4C). Importantly, intratumoral TCF1 + CD8 + T cells have been shown to mediate the e cacy of PD-1 checkpoint blockade 7,8,12 . To determine if the TCF1 high CD8 + T cells resulting from PKM2 de ciency respond to PD-1 inhibition to exert anti-tumor effects, we tested a combination of anti-PD-1 and PKM2 loss in CD8 + T cells in the HKP1-ova-GFP mouse model. PKM2 WT or PKM2 KO OT-I + T cells were adoptively transferred into cohorts of HKP1-ova-GFP mice, and subsequently administered 3 doses of anti-PD-1 antibody as described before 4,27 . PKM2 KO  To further investigate the effects of PD-1 inhibition, we adoptively co-transferred PKM2 WT and PKM2 KO T cells into tumor-bearing mice, followed by administration of anti-PD-1 or IgG, and performed comprehensive ow cytometric phenotyping of donor populations. We again observed rapid tumor control by anti-PD-1 treatment (Extended Data Fig. 7D). Furthermore, T cell phenotypes from anti-PD-1 treated mice largely mirrored their IgG or non-treated counterparts (Fig. 3). In comparison to PKM2 WT T cells, we observed 1) decreased proportions of  Fig. 7M-7N). In both PKM2 WT and PKM2 KO T cells, TCF1 high cells displayed increased Ki67 expression compared with their TCF1 low counterpart, suggesting that TCF1 high cells were more proliferative (Extended Data Fig. 7O-7P). Anti-PD-1 treatment resulted in similar phenotypes in both genotypes of T cells: increased proportions of CD44 + CD62L-cells in tumors at later timepoints, a more rapid progression to the TCF1 low phenotype in the tumor, and a trend for more Ki67 expression in TCF1 high cells in the tumor at the latest timepoint. Importantly, we observed maintained higher proportions of memory-like T cells in PKM2 KO even with anti-PD-1 treatment, a population critical to the checkpoint blockade response 7,8 . These data suggest that the anti-PD-1 response machinery and mechanism is largely retained with PKM2 loss, with a sustained memory skewing in the presence of checkpoint blockade.
To obtain further insights into how PD-1 inhibition impacts PKM2 KO CD8 + T cells, we performed RNA sequencing (RNA-seq) analyses of sorted tumor-in ltrating T cells from HKP1-ova-GFP tumor-bearing mice adoptively cotransferred with a mix of PKM2 KO and PKM2 WT OT-I + T cells and treated with IgG control or anti-PD-1 (Fig. 4G, Extended Data Fig. 8A-8C). As expected, PKM2 WT cells increased in numbers and PKM2 KO cells lacking PKM2 expression were associated with increased proportions of CD44 + CD62L + cells (Extended Data Fig. 8D-8E). When normalized for inter-mouse heterogeneity, samples clustered based on genotype, treatment status, and time ( Fig. 4H). Differential gene expression and gene set enrichment analysis (GSEA) showed enrichment of a memory signature 28 in PKM2 KO cells and enrichment of an effector signature in PKM2 WT cells, consistent with ow cytometry-based phenotyping ( Fig. 4I-4J, Extended Data Table 3). These differential enrichments were present at both the early and later timepoint, demonstrating a persistent skewing towards a memory-like phenotype with PKM2 loss (Extended Data Fig. 8F). Further interrogation of transcriptome data using hallmark datasets 29 displayed a more proliferative phenotype in PKM2 WT cells, consistent with their elevated abundance compared with PKM2 KO cells, as well as an enhanced complement signature (Fig. 4K). The differential signature enrichment by genotype shifted over time: at early timepoints, PKM2 WT cells displayed enrichment of signatures associated with cell proliferation and glycolysis (Extended Data Fig. 8G, Extended Data Table 4), while at later timepoints these cells had more in ammation-associated signatures (Extended Data Fig. 8H, Extended Data Table 4). Anti-PD-1 treatment meanwhile enriched for a proliferative and metabolically active state compared with IgG (Fig. 4L); these trends also shifted over time, re ecting similar patterns to the differences observed between PKM2 WT and PKM2 KO cells, with early differences in proliferation-associated signatures and later differences in in ammationassociated signatures (Extended Data Fig. 8I-8J, Extended Data Tables 5-6). Finally, to interrogate how PKM2 WT and PKM2 KO cells responded differently to anti-PD-1 treatment, we performed gene set enrichment analysis (GSEA) to identify pathways which were differentially affected based on genotype upon anti-PD-1 treatment (Fig. 4M, Extended Data Table 7). A subset of the response was conserved between genotypes, including activation of the hallmark signatures E2F targets, G2M checkpoint, and glycolysis, among others (Extended Data Table 7). There were also divergently enriched signatures, suggesting a qualitative difference to these cellular populations. In PKM2 WT cells, anti-PD-1 treatment induced enrichment of hallmark pathways associated with proliferation, including apical junction and mitotic spindle, while PKM2 KO cells showed enrichment of oxidative phosphorylation (OXPHOS), suggesting a metabolic shift upon anti-PD-1 therapy (Fig. 4M, Extended Data Table 7). This was further supported by gene ontology: biological process (GO:BP) datasets, where PKM2 KO cells showed enrichment for terms such as ATP synthesis coupled electron transport, mitochondrial respiratory chain complex assembly, OXPHOS, and electron transport chain, along with other similar terms (Extended Data   Table 7). Intriguingly, both OXPHOS and fatty acid oxidation are associated with memory cell fuel utilization, supporting a memory cell phenotype for PKM2 KO cells and suggesting an active metabolic response to anti-PD-1 therapy 14,18,19,30 . Taken together, these data con rm the altered differentiation status induced by PKM2 loss, with robust effects on proliferation and in ammatory processes, and demonstrate the impact of this altered state on response to checkpoint blockade.
PKM2 loss results in reduced glycolytic ux, accumulation of glycolytic intermediates and activation of the pentose phosphate pathway.
PKM2 is a rate-limiting enzyme in glycolysis, catalyzing the conversion of phosphoenolpyruvate (PEP) to pyruvate 26 . In addition to its metabolic role, PKM2 can serve as a transcriptional regulator, modify activity of several signaling cascades such as HIF1α, mTOR, STAT1, STAT3, STAT5, and TGFβ/Smad2/3 [31][32][33][34] . Given these multiple roles, we examined the subcellular localization of PKM2 in activated CD8 + T cells isolated from cocultures and observed robust cytoplasmic expression and little to no nuclear PKM2, suggesting an important metabolic role for PKM2 in CD8 + T cells (Extended Data Fig. 3G). To investigate the effects of PKM2 loss on T cell metabolism, we sorted T cells from co-cultures of HKP1-ova-GFP cells with either PKM2 WT or PKM2 KO OT-I + T cells at 6 days post-initial-stimulation and performed a glycolysis stress test using a Seahorse XF bioanalyzer. PKM2 KO T cells showed reduced glycolysis and glycolytic capacity compared with PKM2 WT counterparts, as expected ( Fig. 5A-5B). This PKM2 de ciency resulting in reduced glycolysis did not impact fatty acid and glutamine oxidation (Extended Data Fig. 9A-9D). Therefore, to further characterize the evolving metabolic alterations resulting from PKM2 de ciency, we conducted steady-state polar metabolite pro ling by liquid chromatography/mass spectrometry (LC/MS) at multiple timepoints following CRISPR/Cas9 RNP-mediated deletion of PKM2 (Extended Data Fig. 9E-9L, Extended Data Table 8). Compared with PKM2 WT , PKM2 KO T cells sorted from co-cultures showed accumulation of various glycolytic intermediates over time (Extended Data Fig. 9E-9H). Both PEP, which has been shown to interfere with other glycolytic reactions 35 , and 3phosphoglycerate, the substrate of phosphoglycerate mutase, an enzyme negatively regulated by PEP, demonstrated robust accumulation at early timepoints in PKM2 KO T cells in co-culture (Extended Data Fig. 9F).
PKM2 KO T cells further displayed coincident enrichment of metabolites of the pentose phosphate pathway (PPP) with this accumulation of glycolytic intermediates (Extended Data Fig. 9I-9L). Glycolysis and the PPP share both glucose as a fuel source as well as several metabolites, including glucose 6-phosphate, fructose-6-phosphate, and glyceraldehyde 3-phosphate, which can be utilized by the PPP to generate NADPH and ribose-5-phosphate in activated T cells 16 . We observed consistent enrichment in sedoheptulose 7-phosphate in co-culture (Extended Data Fig. 9I-9L), with varying accumulation of other metabolites. Notably, fructose 6-phosphate, a product of another reaction negatively regulated by PEP, is also an important product of transketolase and transaldolase activity in the non-oxidative phase of the PPP; it was particularly enriched at later timepoints (Extended Data Fig. 9G-9H). These data demonstrate that PKM2 loss results in reduced glycolysis, and an increase in abundance of PPP metabolites.
To examine the effect of PKM2 loss on glucose processing more directly, we performed metabolic pro ling using 1,2 13 C-labelled glucose 36,37 . Use of glucose labelled at positions 1 and 2 allows for discrimination of metabolites generated by glycolytic or PPP activity, while also allowing for observation of multiple rounds of PPP activity through differential labelling patterns 36,37 . We activated OT-I + T cells as before, then either deleted or retained PKM2 using electroporation of CRISPR/Cas9 RNPs, and subsequently co-cultured these T cells with HKP1-ova-GFP cells. At days 4 and 6 post-initial-stimulation, we sorted T cells from co-cultures, incubated them for 2 hours with 1,2 13 C-labelled glucose, then isolated metabolites and performed polar metabolite pro ling by LC/MS (Extended Data Table 9). Metaboanalyst 38,39 analyses of labelled metabolites demonstrated robust impact of PKM2 loss on the PPP at both timepoints (Fig. 5C, Extended Data Fig. 9M). At day 6 post-initial-stimulation, in PKM2 KO T cells we observed accumulation of labelled 3-phosphoglycerate and phosphoenolpyruvate, two metabolites upstream of PKM2, suggesting a blockade from the knockout (Fig. 5D-5G). We also observed increased proportions of labelled gluconate, ribose 5-phosphate, and sedoheptulose 7-phosphate (Fig. 5D, 5H-5K); gluconate is an entry point into the oxidative PPP, while ribose 5-phosphate and sedoheptulose 7-phosphate are two critical products generated in the non-oxidative PPP. Interestingly, we observed increased proportions of different order isotope charges in these glycolytic and PPP products than would be generated by a single round of glycolysis or the PPP, suggesting multiple PPP cycles in PKM2 knockout T cells ( Fig. 5F-5G, 5I-5K). Day 4 postinitial-simulation data display similarities and some differences, one of which being a more signi cant impact on the TCA Cycle (Extended Data Fig. 9M). PKM2 loss resulted in signi cant differences in labelled glyceraldehyde 3-phosphate, 3-phosphoglycerate, phosphoenolpyruvate, ribose 5-phosphate, sedoheptulose 7-phosphate, and oxoglutarate, with increased normalized isotope counts observed in knockouts in each of these metabolites except for sedoheptulose 7-phosphate, which was decreased (Extended Data Fig. 9N-9V). Once again, labelling patterns show isotope charges indicating multiple rounds of PPP activity (Extended Data Fig. 9Q-9V). Taken together, these data substantiate our steady-state ndings that PKM2 loss results in altered glycolytic ux and an increase in PPP metabolite generation, with multiple rounds of PPP activity.
Pentose phosphate pathway agonism results in a TCF1 + state independent of glycolytic blockade and divergent from the phenotype induced by hexokinase inhibition.
Our data demonstrated that PKM2 loss results in accumulation of PPP activity. We therefore asked if induction of this altered metabolic state was su cient to phenocopy the altered differentiation observed upon PKM2 loss. Recent studies have interrogated effects of manipulation of the oxidative phase of the PPP on CD8 + T cell effector function, with con icting outcomes [40][41][42][43] . To test the impact of elevated PPP activity upon T cell differentiation and function, we treated T cells with AG1, a small molecule agonist of glucose-6-phosphate dehydrogenase (G6PD) 44 , which catalyzes the rst and committed step in the oxidative phase of the PPP 35  suggesting a skewing towards the TCF1 high progenitor population. We subsequently asked if PPP agonism induced this phenotype through a loss in glycolysis, which is known to result in inhibited effector differentiation 19 . To test this, we activated OT-I + Thy1.1 + T cells and subsequently co-cultured them with HKP1ova-GFP cells until 6 days post-initial-stimulation. We then sorted the T cells from co-culture and treated them with either vehicle DMSO, AG1, or the hexokinase inhibitor 2-DG for 2 hours, then performed a glycolysis stress test. While acute treatment with 2-DG resulted in almost complete glycolytic shutdown, as expected, AG1 treatment had little effect on glycolysis, suggesting that PPP agonism may induce an altered differentiation state independent of glycolysis loss (Fig. 6E).
To evaluate the differential effects of these two different metabolic manipulations on CD8 + T cell differentiation, we bred a TCF1 eGFP reporter (Tcf7 GFP , Reference 45 ) mouse strain to the OT-I + strain, allowing us to track TCF1 expression in live cells. We activated the cells for 1 day, then initiated treatment with either DMSO, AG1, or 2-DG.
We expanded the cells for 1 more day, then co-cultured them with HKP1-ova-GFP cells with continuing drug or vehicle treatment, passing cells every two days. At day 6 post-initial-stimulation, we sorted eGFP + and eGFP-cells from each condition and performed RNA sequencing. Both AG1 and 2-DG treatment resulted in increased proportions of eGFP + cells over the proportion emerging in the DMSO treated samples, indicating their abilities to impact differentiation ( Fig. 6F-6G). 2-DG treatment resulted in a near complete loss of eGFP-cells, not permitting analysis of eGFP-cells under that treatment condition (Fig. 6F). Analyzing principal components and the top 1000 variable genes of eGFP + and eGFP-cells from the different treatments indicated large gene expression differences, both between samples based on TCF1 status and between eGFP + or eGFP-cells based on treatment, with 2-DG-eGFP + cells being the most transcriptionally distinct group (Fig. 6H-6I). Comparing eGFP + and eGFPcells revealed expected gene expression differences, including increased Sell, Slamf6, Id3, Tcf7, and Nsg2 in eGFP + cells, and increased Gzma, Gzmb, Id2, Cd244a, and Havcr2 in eGFP-cells, validating the system and identifying key machinery conserved in TCF1 + cells across multiple metabolic modalities (Fig. 6J, Extended Data Table 10).

Page 10/53
We subsequently asked how gene expression varied between eGFP + cells under different metabolic stimuli: baseline (via DMSO), PPP agonism (via AG1), or glucose blockade (via 2-DG). When compared to DMSO, both AG1 and 2-DG induced robust alterations to the transcriptional landscape, with 2-DG inducing more widespread differential gene expression (Fig. 6K, Extended Data Table 11). There was some overlap between treatments in genes over-or under-expressed compared with DMSO-treated eGFP + cells: 53.5% of up-regulated genes upon AG1 treatment were also up-regulated in 2-DG-treated cells, while 73.1% of down-regulated genes upon AG1 treatment were also down-regulated in 2-DG-treated cells (Fig. 6K). Given the larger disparity in gene expression between 2-DG treated samples and DMSO-eGFP + cells compared with the differential expression between AG1and DMSO-treated eGFP + cells, this overlap in genes altered by treatment corresponded to only 9.1% of upregulated genes and 25.5% of down-regulated genes in 2-DG treated samples (Fig. 6K). Further analyses comparing AG1-eGFP + and 2-DG-eGFP + samples supported a transcriptional landscape difference in TCF1 + cells generated by either PPP agonism or glucose blockade (Fig. 6L). Hallmark analysis showed an enrichment of proliferation (E2F targets, mitotic spindle) and in ammatory pathways (allograft rejection, complement, interferon gamma response) in AG1-eGFP + samples compared with enrichment of the unfolded protein response and hypoxia related pathways (mTORC1, hypoxia) in 2-DG-eGFP + samples (Fig. 6M). Taken together, these data indicate that PPP agonism and glucose blockade both induce TCF1 + populations, but with substantial differences in the transcriptional landscape.
Finally, we explored upstream regulators of gene expression, searching for conserved factors driving gene expression differences between PKM2 KO and PKM2 WT in our adoptive co-transfer experiment (Extended Data Table 3) and between AG1 and DMSO treatment in this in vitro co-culture experiment (Extended Data Table 12). Analysis showed a limited number of upstream regulators overlapping between these disparate experimental conditions, among them being Bach2 and Foxo1 (Fig. 6N, Extended Data Table 13). Both Bach2 and Foxo1 have been reported to be upstream inducers of TCF1 expression and regulators of memory formation and maintenance 46-50 , and had increased expression in both RNA sequencing datasets. We therefore examined their expression in our co-culture system by ow cytometry and found increased expression of both Foxo1 and Bach2 by both PKM2 KO and AG1 treatment at 4 days post-initial-stimulation compared to their respective controls, validating the sequencing results ( Fig. 6O). Taken together, these data suggest shared expression of transcription factor machinery between PKM2 KO and PPP agonism previously reported to induce TCF1 expression.
Pentose phosphate pathway agonism results in enhanced tumor control in combination with PD-1 checkpoint blockade.
As PPP agonism resulted in a TCF1 high population in vitro, similar to the effects of PKM2 knockout, we further tested if anti-tumor e cacy in vivo would be similarly enhanced in combination with anti-PD-1 therapy. Activated OT-I + T cells were pre-treated with either DMSO or AG1 for 3 days, adoptively transferred into cohorts of HKP1ova-GFP mice, and subsequently administered 3 doses of anti-PD-1 antibody as described previously (Fig. 7A).
Similar to results when co-cultured with tumor cells, pre-treatment with AG1 resulted in increased proportions of CD44 + CD62L + cells and TCF1 + cells prior to adoptive transfer ( Pentose phosphate pathway agonism promotes anti-tumor immunity in ex vivo immunocompetent human patient-derived tumor organoids. Our studies in mouse T cells showed a potent ability of PPP agonism to induce a TCF1 + progenitor state and combine with checkpoint blockade to yield a signi cant tumor control and survival bene t. We therefore tested the effects of AG1 in a human immunocompetent patient-derived tumor organoid system 51,52 (Fig. 7G). Following standard methodologies 51,52 , we generated patient-derived tumor organoids (PDTOs) from NSCLC patient specimens. We rapidly expanded 53 T cells from autologous peripheral blood mononuclear cells (PBMCs) donated by the patients for two weeks, either in the presence or absence of AG1 (Fig. 7G), then assessed TCF1 expression. Similar to our ndings in mice, AG1 treatment signi cantly increased TCF1 expression in samples from 4 out of 4 patients (Fig. 7H). Expanded T cells were co-cultured with autologous PDTOs for two weeks, then restimulated for analysis for tumor reactivity by IFNγ production and tumor killing by cleaved caspase-3 production in tumor organoids. While IFNγ production was variable, AG1-treated T cells produced higher levels of IFNγ than their DMSO-treated counterparts in samples from 3 out of 4 patients (Fig. 7I). Importantly, AG1-treated T cells consistently induced signi cantly higher proportions of apoptotic organoids than their DMSO-treated counterparts, with a consistent phenotype between all 4 patient samples (Fig. 7J). Taken together, these data indicate that agonism of the pentose phosphate pathway may augment the anti-tumor capabilities of T cells in humans.

Discussion
Our global gene expression pro ling of treatment-naïve and PD-1 blockade-treated HKP1 tumors identi ed metabolic regulatory pathways in intratumoral CD8 + T cells, consistent with the critical roles of metabolism in regulating T cell fate and function [13][14][15][16] . Of the metabolic alterations identi ed, glycolysis was enriched in T cells of progressing tumors. Glycolytic function is critical for optimal effector activity of T cells 13,14,[54][55][56][57][58] . Our shRNA screen targeting a majority of differentially expressed glycolytic enzymes in intratumoral T cells displayed a spectrum of phenotypes, including loss of effector cytokine production or proliferative status, consistent with recent reports 17,19,[54][55][56][57][58][59][60][61][62] , whereas knockdown of other glycolytic genes resulted in upregulation of checkpoint proteins, implicating a T cell dysfunction phenotype. Intriguingly, the screen showed that loss of metabolic regulator PKM2 induced a progenitor-exhaustion like state, associated with increased expression of the transcription factor TCF1. This altered differentiation of activated T cells with PKM2 loss generated a central memory-like phenotype in vivo, which combined with anti-PD-1 to yield signi cant tumor control and improved overall survival. Further mechanistic explorations detailed a decreased glycolytic metabolism and an increased utilization of the pentose phosphate pathway (PPP). Subsequent exploration of the rami cations of increased PPP activity revealed its capacity to induce a TCF1 high transcriptional state independent of and distinct from glycolytic blockade, and therapeutic utility controlling tumors both in an in vivo mouse lung cancer model in combination with anti-PD-1, and in an ex vivo immunocompetent human patient-derived tumor organoid system. These data describe a new metabolic reprogramming that contributes to a progenitor-like T cell state amenable to therapeutic checkpoint blockade.
PKM2 is expressed in various cell types, and both normal and disease states. Several reports have investigated PKM2 in CD4 + T cells in different non-tumor in ammatory contexts, indicating a role for PKM2 in glycolysis and a resultant activation phenotype [31][32][33][34]63 . One study reported metabolic data demonstrating a dampening of glycolysis and pentose phosphate pathway activity upon PKM2 loss 63 . Other studies have focused on PKM2's transcriptional regulation during Th differentiation 31,34 , and effects of pharmacological inhibition or activation on Th polarization [32][33][34] . Our results in CD8 + T cells critically diverge from these ndings. First, in contrast to effects of its loss in CD4 + T cells, we observe enhanced PPP activity upon PKM2 knockout in CD8 + T cells, suggesting a different metabolic compensation. Second, PKM2 plays an important role in the nucleus of CD4 + T cells upon their activation and during differentiation; in contrast, we observe little to no accumulation in the nucleus, suggesting most if not all PKM2 is cytoplasmic. Furthermore, knockout of PKM2 in naïve CD8 + T cells results in upregulation of PKM1 upon activation and normal differentiation to effector CD8 + T cells, suggesting that metabolic pyruvate kinase activity is su cient for this process (data not shown). While we hesitate to completely exclude the other regulatory effects of PKM2, its subcellular localization, the relative abundance of cytoplasmic and nuclear PKM2, and the ability of PKM1 to substitute and allow normal effector differentiation suggest a markedly different functionality for PKM2 in CD8 + T cells compared with that delineated in CD4 + T cells.
Our steady-state metabolomics and 1,2 13 C glucose carbon tracing data demonstrate that PKM2 loss in CD8 + T cells resulted in accumulation of PPP metabolites. The presence of increased amounts of alternatively labelled isotopes of metabolites in PKM2 KO T cells indicated more cycles of PPP activity 36,37 compared with that observed in PKM2 WT cells. Small molecule agonism of glucose 6-phosphate dehydrogenase, the rst enzyme in the oxidative PPP, to induce the altered metabolic state observed upon PKM2 loss resulted in a similar altered differentiation towards a TCF1 + progenitor-like state. These data suggest that PKM2 loss may lead to this TCF1 + progenitor-like phenotype by amplifying PPP activity. A reasonable experiment to test this hypothesis would be to abrogate PPP activity and examine the effects of PKM2 manipulation on differentiation. However, this experiment is complicated by the reliance of CD8 + T cell differentiation on the PPP, which has been previously reported 41,43 . Based on those data, we tested the effects of inhibition of the PPP using small molecule inhibitors targeting different enzymes in the pathway: G6PDi-1, inhibiting glucose 6-phosphate dehydrogenase; 6aminonicotinamide (6-AN), inhibiting 6-phosphogluconate dehydrogenase; and oxythiamine, inhibiting transketolase. We evaluated a variety of doses and treatment windows for these inhibitors, and consistently found with both short and long-term treatments that both G6PDi-1 and 6-AN, which impact the oxidative PPP, resulted in signi cantly reduced effector T cell populations in both PKM2 WT and PKM2 KO T cells, while blocking the non-oxidative phase of the PPP with oxythiamine had a negligible effect (Extended Data Fig. 10).
Furthermore, long-term blockade of the oxidative PPP resulted in precipitous cell loss, suggesting a failure to either differentiate or thrive. In our hands, treatment of equivalent initial numbers of cells resulted in a cell yield in the G6PDi-1 group of only 20.7% of the DMSO-treated cell number at 1 day after treatment initiation. This loss continued, with G6PDi-1-treated cell numbers falling to 7.9% and 2.9% of timepoint-matched DMSO-treated control cell numbers after 3 days and 5 days respectively of treatment. These data are consistent with the previous studies 41,43 showing negative effects of PPP inhibition on CD8 + T cell differentiation, and along with our ndings using the PPP agonist AG1 and the hexokinase inhibitor 2-DG, suggest that balanced glycolysis and pentose phosphate pathway activity is necessary for robust effector differentiation.
Loss of PKM2 in CD8 + T cells resulted in a TCF1 + progenitor-like state. Similar phenotypes have also been observed following inhibition of other glycolytic enzymes including hexokinase (HK) 19 and lactate dehydrogenase (LDH) 17 . However, this progenitor-or memory-like state induced by disruption of PKM2, HK, or LDH activity is mediated by disparate molecular mechanisms: PKM2 loss results in enhanced PPP activity, while pharmacological HK inhibition results in ampli cation of fatty acid oxidation (FAO) 19 and transient LDH inhibition results in increased tricarboxylic acid cycle (TCA) activity 17 . This differential metabolic rerouting suggests multiple redundant mechanisms converge on a TCF1 + differentiation state similar in some crucial respects, including TCF1 and cytokine expression patterns and cell proliferation, and display overlap with progenitorexhausted populations described in the literature 5-8,64−66 . However, the utilization of divergent metabolic pathways in these cells suggested there may be corresponding transcriptional heterogeneity. Our sequencing experiments using a TCF1 reporter to sort viable TCF1 + and TCF1-cells differentiated under PPP agonism via AG1 or glucose blockade via 2-DG demonstrated signi cant differences in the transcriptional landscape of TCF1 + cells. AG1-treated cells were more proliferative, more transcriptionally active, and with more in ammatory signatures, while 2-DG-treated cells had signi cantly higher unfolded protein response and hypoxia signatures.
These differences require further exploration, but may in turn be leveraged therapeutically, allowing selection of different metabolic modalities and cognate transcriptional pro les for use in different tissue, disease, or diseasestage contexts.
Our data demonstrate that PKM2 loss resulted in accumulation of glycolytic intermediates and increased PPP activity. Agonism of G6PD activity phenocopied PKM2 loss. This suggested a conserved mechanism could be used in both contexts to result in generation of the TCF1 + progenitor populations. Upstream regulator analysis comparing RNA sequencing data from PKM2 KO and PKM2 WT T cells isolated back from mouse tumors after adoptive co-transfers and from AG1 and DMSO treated T cells isolated after in vitro co-culture with tumor cells demonstrated conserved Foxo1 and Bach2 signatures. Importantly, expression of both Foxo1 and Bach2 was increased by PKM2 loss or AG1 treatment at 4 days post-initial-stimulation, suggesting that these factors may play a mechanistic role downstream of PPP activity inducing TCF1. Interestingly, while a Bach2 upstream regulator signature was observed in TCF1 + cells induced by 2-DG treatment compared with TCF1 + cells in the DMSO control, a Foxo1 signature was not observed (Extended Data Table 13), suggesting differential transcription factor usage for generation and maintenance of this distinct TCF1 + state. Foxo1 and Bach2 are well-known regulators of stem and memory characteristics and regulate TCF1 expression 46-50 . Ongoing experiments are exploring how these two proteins and other core memory/progenitor-like machinery are directly regulated by PPP activity. One intriguing possibility is through epigenetic control. NADPH, an important bioactive reducing agent produced in the oxidative PPP 35 , has been recently shown to bind to HDAC3, resulting in an inability to interact with its co-activators NCOR1 and NCOR2 67 . These data suggest that PPP activity may impact epigenetic control of T cell differentiation. Indeed, recent work has shown that HDAC3 controls CD8 + T cell cytotoxicity programs 68 , while other studies have shown important roles for EZH2 69 and SUV39H1 70 in controlling T cell differentiation and effector function. Future experiments will explore the rami cations of ampli ed PPP activity on transcriptional and epigenetic control of T cell fate.
Finally, our work raises interesting possibilities for the treatment of human disease. PKM2 displayed similar expression patterns, and its knockout had similar phenotypic results in tumor-speci c T cells in both a non-small cell lung cancer model and a melanoma model, suggesting a conserved mechanistic response regardless of tumor type, although with potential differences in kinetics and magnitude. Using both ow cytometry and RNA sequencing, we observed that anti-PD-1 treatment resulted in induction of similar phenotypes in PKM2 WT and PKM2 KO T cells, including greater proliferation and a transition to a more effector-like phenotype, suggesting that the machinery for response to the therapy was largely intact. Loss of PKM2 did however result in a sustained higher proportion of a memory-like cellular compartment, with a limited terminal effector response, which has been shown to be important for superior long-term protection 71 . Treatment with the G6PD agonist AG1 demonstrated that PPP ampli cation resulted in a similar phenotypic skewing to a TCF1 + progenitor-like state in co-culture with tumor cells, and in vitro pre-treatment of tumor-speci c T cells with AG1 resulted in enhanced tumor control upon adoptive transfer in combination with PD-1 blockade. Using a human immunocompetent patient-derived tumor organoid (PDTO) platform, we demonstrated that treatment with AG1 during T cell rapid expansion from autologous PBMCs resulted in elevated TCF1 expression, mimicking our mouse data. After culture of these expanded T cells with autologous PDTOs to provide tumor-speci c T cell stimulation, assays to assess T cell reactivity and tumor killing potency displayed elevated T cell effector function in AG1-treated cells.
Together, these suggest that PPP ampli cation may be a useful therapeutic modality in adoptive cell transfer immunotherapy approaches, and may synergize with PD-1 blockade to provide durable clinical bene ts.
150,000 HKP1 cells or 250,000 HKP1-ova-GFP cells suspended in sterile PBS were administered via the tail vein into syngeneic female 8-week-old C57BL/6J mice. Tumor growth in vivo was evaluated twice weekly via bioluminescence imaging (BLI) using a Xenogen IVIS system. To divide mice into cohorts for different treatments, BLI data were collected 1 day prior to treatment and mice were grouped such that similar mean tumor burdens were present in each treatment cohort. For treatment with anti-PD-1 or IgG2a, mice were injected intraperitoneally with 250µg of anti-PD-1 or IgG2a control on days 10, 13, and 17 post-tumor implantation. For all survival studies, mouse tumor burden was evaluated bi-weekly, and mice were euthanized when tumor burden reached humane endpoints.
Prior to the administration of cells, mice were anesthetized with iso urane and fur removed on the dorsal right ank. Depilatory cream for fur removal was left on the ank for 30-60 seconds and promptly wiped off using warm water and gauze. 500,000 B16F10-ova-GFP cells suspended in 100 µL sterile PBS were administered via subcutaneous injection into syngeneic female 8-week-old C57BL/6J mice. Tumor growth in vivo was evaluated twice weekly via caliper measurements (Fisherbrand Traceable Digital Caliper) and tumor volumes was calculated with the formula V = L(W 2 )/2, where L and W are tumor length and width (mm), respectively.

Retrovirus production.
Retroviruses were packaged by calcium-phosphate transfection of packaging cells similar to previously published methodologies 72 . For retrovirus production with the pMFG-Ova-N4-EGFP construct, HEK293T cells were plated, treated with chloroquine, then transfected with pMFG-Ova-N4-EGFP and the packaging plasmid Phoenix. Cells were incubated for 10-14 hours at 37℃, 5% CO 2 , and then media exchanged for fresh antibiotic-free media with sodium butyrate and cells moved to a tissue culture incubator at 32℃, 5% CO 2 . Virus was collected at 48 hours and 72 hours post-transfection.
Cells were incubated for 10-14 hours at 37℃, 5% CO 2 , and then media exchanged for fresh antibiotic-free media with sodium butyrate and cells moved to a tissue culture incubator at 32℃, 5% CO 2 . Virus was collected at 48 hours and 72 hours post-transfection.
T cell isolation and stimulation.
One day prior to T cell isolation, plates were coated sequentially with biotinylated poly-L-lysine (Sigma, Thermo Scienti c), Streptavidin-Plus (Prozyme), and nally with biotin anti-CD3ε and biotin anti-CD28. Coated plates were wrapped in para lm to prevent evaporation and incubated overnight at 4℃.
On the day of T cell isolation, mice of speci ed genotypes were euthanized, and spleens dissected and ground Antigen-speci c splenic CD8 + T cells were isolated and stimulated with plate-bound anti-CD3ε and anti-CD28 (2µg/mL each) as described above. 18 hours after plating, media was replaced with retrovirus-containing media supplemented with 4µg/mL polybrene (hexadimethrine bromide, Sigma-Aldrich) and 50U/mL IL-2. T cells were spin-transduced at 1000 x g for 90 minutes at 30℃, and subsequently moved to a 32℃ incubator for 4 hours.
Retrovirus-containing media was replaced with fresh Complete T Cell Media supplemented with 50U/mL IL-2.
Cells were returned to the 32℃ incubator for another 7.5 hours, then transferred to a 37℃ incubator. T cells were sorted approximately 48-50 hours after initial plating on a Becton-Dickinson FACS Aria II sorter for successfully transfected viable T cells (DAPI − ZsGreen + ).
Antigen-speci c splenic CD8 + T cells were isolated and stimulated as above. 24 hours after plating, T cells were transfected with CRISPR/Cas9 ribonucleoprotein via electroporation utilizing the Neon Transfection System (Thermo Fisher Scienti c) and ATTO-conjugated tracrRNA, crRNAs, and Cas9 protein purchased from IDT similar to previously published methodologies 73 . RNPs were formed by mixing Cas9 and duplexed ATTO-tracrRNA:crRNA at a 1:1.2 molar ratio (1:1 volume ratio), and incubated at room temperature for 20 minutes. Activated viable T cells were counted and resuspended in Neon Buffer T, then mixed with RNP and electroporation enhancer. The mixture was pipetted using the Neon pipette and Neon tips and the pipette plugged into position in the Neon transfection device. Electroporation was performed at 3 pulses, 10ms pulse width, 1600V. T cells were subsequently rested in Complete T Cell Media with 50U/mL IL-2 for at least 2 hours, then successfully transfected viable T cells (DAPI − ATTO550 + ) were sorted on a Becton-Dickinson FACS Aria II, and subsequently expanded on anti-CD3ε and anti-CD28 (2µg/mL each) coated plates in Complete T Cell Media with 50U/mL IL-2 for 24 hours.
Knockout e ciency was validated 72 hours later by ow cytometry.
T cell/tumor cell co-culture and staining.
For baseline co-culture dynamics and phenotypes, antigen-speci c CD8 + T cells were activated either via antigen presentation from splenic antigen-presenting cells (APCs) or via stimulation with plate-bound anti-CD3ε and anti-CD28 as above. Both methodologies yielded similar results. For antigen presentation from APCs, red blood cell lysed single cell suspensions were made from spleens from donor mice, cultured with 1µg/mL Ova 257 − 264 peptide (InvivoGen), and plated at 2mL per well in 12 well tissue culture plates. Two days later, activated antigenspeci c CD8 + T cells were isolated from the culture using the Militenyi MACS system for untouched CD8 + T cell isolation. The alternative activation method utilizing plate-bound anti-CD3ε and anti-CD28 was performed as described above: untouched CD8 + T cell isolation via Militenyi MACS sorting, and culturing on plate-bound anti-CD3ε and anti-CD28 at 2µg/mL each in Complete T Cell Media with 50U/mL IL-2. These cells were also cultured for two days. In both experiments, T cells were subsequently either co-cultured with tumor cells or passed to plates coated with anti-CD3ε and anti-CD28 at 2µg/mL each. In experiments with genetic manipulations, T cells were modi ed as described above, and co-cultures or sequential plate-bound stimulations initiated 48 hours after initial T cell activation. In experiments with treatment with the G6PD agonist AG1, T cells were isolated and activated on anti-CD3ε and anti-CD28 coated plates (2µg/mL each) as described above, cultured in Complete T Cell Media with 50U/mL IL-2 for 24 hours, then treated with either 3µM AG1 or an equivalent volume of DMSO and cultured for an additional 24 hours. Co-cultures with tumor cells were then initiated as above and treatment continued with 3µM AG1 or an equivalent volume of DMSO in Complete T Cell Media with 50U/mL IL-2 for the duration of the experiment.
In experiments with the TCF1 eGFP reporter (Tcf7 GFP ) mice, T cells were isolated from Tcf7 GFP+ OT-I + mice and activated on anti-CD3ε and anti-CD28 coated plates (2µg/mL each) as described above, cultured in Complete T Cell Media with 50U/mL IL-2 for 24 hours, then treated with 3µM AG1, 2mM 2-deoxyglucose (2-DG), or an equivalent volume of DMSO and cultured for an additional 24 hours. Co-cultures with tumor cells were then initiated as above and treatment continued with 3µM AG1, 2mM 2-DG, or an equivalent volume of DMSO in Complete T Cell Media with 50U/mL IL-2 for the duration of the experiment, being passed every two days and replenished with drug or DMSO as before.
In experiments with treatment with the Pentose Phosphate Pathway inhibitors G6PDi-1, 6-aminonicotinamide, and oxythiamine, T cells were isolated and activated on anti-CD3ε and anti-CD28 coated plates (2µg/mL each) as described above, cultured in Complete T Cell Media with 50U/mL IL-2 for 2 days, then co-cultured with tumor cells as above for 2 days, then co-cultures passed and drugs or an equivalent volume of DMSO added at indicated doses in Complete T Cell Media with 50U/mL IL-2 for the last 2 days.
T cell subcellular fractionation and Western blotting.
To examine subcellular localization of PKM isoforms, T cells were harvested, activated, genetically modi ed, and co-cultured with tumor cells as described above. At day 4 post-initial-stimulation, CD8 + Thy1.1 + T cells were isolated from co-culture using the EasySep Mouse CD90.1 Positive Selection Kit from STEMCELL Technologies according to the manufacturer's instructions. Brie y, co-cultures were aspirated, ltered through 40 micron lters, resuspended at 1x10 8 cells/mL, transferred into 5 mL polystyrene round-bottom tubes, and then selection cocktail was added to sample at 50µL/mL sample, and mixed and incubated for 3 minutes at room temperature.
RapidSpheres were then added to the sample at 40µL/mL sample, and mixed and incubated for 3 minutes at room temperature. Samples were topped up to 2.5mL in FACS Buffer, gently pipetted, and placed on an EasyEights separation magnet for 10 minutes at room temperature. Supernatant was subsequently pipetted off, and the washing with FACS Buffer and incubation on the magnet performed twice more for a total of 3 x 10 minute separations. Cells were then aliquoted and protein subsequently extracted.
Cytoplasmic, nuclear, and whole cell extracts of the isolated T cells were collected using the PARIS (Protein and RNA Isolation System) Kit from Thermo Fisher according to the manufacturer's protocol. Brie y, cells were isolated from co-culture, pelleted, washed once with PBS, and re-pelleted. Aliquots for whole cell extracts were resuspended in Cell Disruption Buffer supplemented with protease (cOmplete protease inhibitor cocktail tablets, Roche) and phosphatase (PhosSTOP phosphatase inhibitor cocktail tablets, Roche) inhibitors, vortexed for 1 minute, and incubated on ice for 10 minutes before being centrifuged at max speed for 2 minutes at 4℃, and supernatant aspirated and stored at -80℃ until use. Cell aliquots for cytoplasmic and nuclear extract generation were resuspended by gently pipetting 5 times in Cell Fractionation Buffer supplemented with protease and phosphatase inhibitors, and incubated on ice for 10 minutes. Cells were then centrifuged at 500xg for 5 minutes at 4℃, and the supernatant (the cytoplasmic extract) was aspirated and stored at -80℃ until use. The remaining pellet was washed again with the Cell Fractionation Buffer supplemented with protease and phosphatase inhibitors, centrifuged at 500xg for 2 minutes at 4℃, supernatant discarded, and pellet resuspended in Cell Disruption Buffer supplemented with protease and phosphatase inhibitors. The resuspended pellet was vortexed for 1 minute and pipetted for 2 minutes, then incubated on ice for 10 minutes before being centrifuged at max speed for 2 minutes at 4℃, and supernatant (the nuclear extract) aspirated and stored at -80℃ until use.
Western blotting was performed using standard methodologies. Brie y, protein content from extracts was determined by a Pierce BCA Protein Assay (Thermo Scienti c), and equivalent amounts of protein were mixed with loading dye (NuPAGE LDS Sample Buffer), boiled at 95℃ for 5 minutes, then loaded on 4-15% precast gels (Mini-PROTEAN TGX Precast Protein Gels, Bio-Rad), stacked at 80V for 20-30 minutes, then run for 120V for another 20-30 minutes. Protein was transferred wet onto PVDF membranes at 250mA for 1.5 hours, then membranes were washed brie y in TBS-T, blocked with 5% milk in TBS-T at room temperature for 1 hour under gentle agitation, brie y washed in 5% BSA in TBS-T, then incubated with primary antibodies at indicated concentrations in indicated diluents overnight at 4℃ under gentle agitation. After at least 16 hours, membranes were washed 3 times for 5 minutes with TBS-T under gentle agitation, then incubated with secondary antibody diluted in 5% milk in TBS-T at room temperature for 1 hour under gentle agitation. Membranes were then washed 3 times for 5 minutes with TBS-T at slightly faster agitation, then membranes were developed using Amersham ECL Prime Western Blotting Detection Reagent (GE Healthcare) and imaged using a Bio-Rad ChemiDoc XRS + with Image Lab Software. Where necessary, blots were then washed once for 5 minutes with TBS-T under gentle agitation, stripped for 1 hour at room temperature under gentle agitation using Restore Western Blot Stripping Buffer (Thermo Scienti c), then blocked, incubated with primary antibody, washed, incubated with secondary antibody, washed, and developed and imaged as before.
Flow cytometry staining. Glycolytic stress tests were performed on sorted T cells from co-cultures using the Agilent Seahorse XF Glycolysis Stress Test Kit (Agilent, 103020-100). Brie y, antigen-speci c T cells were isolated, activated, genetically modi ed by CRISPR/Cas9, and co-cultured, then resuspended, stained for DAPI, CD8β, and Thy1.1, and viable cells sorted into Complete T cell Media as described above at 6 days post-initial-stimulation. T cells were washed with prewarmed assay media (Seahorse XF DMEM Medium, pH 7.4, 2mM glutamine), and resuspended at 1.25-1.5x10 5 cells per 50µL assay media. Resuspended T cells were plated on Seahorse XF96 Cell Culture Microplates coated with 22.4 µg/mL Cell-Tak (Corning) by centrifugation with no brake, put into a 37℃ non-CO 2 incubator for 30 minutes, supplemented with an additional 130µL of assay media, and returned to the incubator. Sensor cartridges in utility plates were pre-hydrated with Seahorse XF Calibrant in a 37℃ non-CO 2 incubator, and appropriate ports loaded: Port A-Glucose (10X port concentration: 100mM), Port B-Oligomycin (10X port concentration: 10µM, Port C-2-DG (10X port concentration: 500mM). Assays were initiated in the XF software, sensor cartridges loaded and calibrated, then cell culture microplates loaded and assays run on an XFe96 Extracellular Flux Analyzer (Agilent).
DNA content was measured for Seahorse data normalization post-run by Hoechst 33342 staining and uorescence read at 361/486nm. A standard curve using reference DNA stained simultaneously allowed for calculation of DNA content in ng and was used to normalize Seahorse data.
For experiments testing the acute effects of drugs on glycolysis, antigen-speci c T cells were isolated, activated, and co-cultured with HKP1-ova-GFP as described previously, then resuspended, stained for DAPI, CD8β, and Thy1.1, and viable cells sorted into Complete T cell Media as described above at 6 days post-initial-stimulation.
Sorted cells were then washed with pre-warmed assay media (Seahorse XF DMEM Medium, pH 7.4, 2mM glutamine), and resuspended at 1.5x10 5 cells per 50µL assay media supplemented with indicated doses of drugs.
Resuspended T cells were plated on Seahorse XF96 Cell Culture Microplates coated with 22.4 µg/mL Cell-Tak (Corning) by centrifugation with no brake, put into a 37℃ non-CO 2 incubator for 30 minutes, supplemented with an additional 130µL of assay media supplemented with drugs, and returned to the incubator for an additional 1.5 hours. Glycolytic stress tests were subsequently performed as described above.
T cell adoptive transfer.
For adoptive transfer of naïve antigen-speci c CD8 + T cells, untouched splenic CD8 + T cells from donor mice were isolated as above using the Militenyi MACS system. Cells were then washed twice with PBS and resuspended at 1x10 6 cells/mL in sterile PBS. Syngeneic female 8-week-old C57BL/6J recipient mice were anesthetized with iso urane, and 5x10 5 cells in 50µL PBS were transferred retro-orbitally.
For adoptive transfer of in vitro activated or genetically modi ed or drug treated antigen-speci c CD8 + T cells, syngeneic female 8-week-old C57BL/6J received tumor implantation as described above 7 days prior to adoptive transfer. Tumor-bearing mice received 5 Gy X-ray irradiation (Rad Source Technologies RS 2000 Biological Research X-ray Irradiator) on the same day as adoptive transfer of the T cells. Both in vitro activated and/or genetically modi ed antigen-speci c CD8 + T cells were adoptively transferred 48 hours after initial activation on anti-CD3ε and anti-CD28 (2µg/mL each) coated plates; drug treated T cells were transferred 96 hours after initial activation, with drug or DMSO treatment for the last 72 hours prior to transfer. Drug treated cells were replated at 1x10 6 cells/mL on fresh anti-CD3ε and anti-CD28 (2µg/mL each) coated plates 48 hours prior to transfer. All transferred cells were washed twice with PBS and resuspended at 20x10 6 cells/mL in sterile PBS. Irradiated tumor-bearing recipient mice were anesthetized with iso urane, and 1x10 6 cells in 50µL PBS were transferred retro-orbitally.
For genetically modi ed antigen-speci c CD8 + T cells, initial ow cytometric phenotyping and ratio and knockout con rmation were performed post-transfer. An aliquot of input cells were stained for CD8α, Thy1.1, Thy1.2, CD44, and CD62L to con rm ratios and phenotype transferred cells. Knockout was con rmed after another 48 hours on anti-CD3ε and anti-CD28 (2µg/mL each) coated plates in Complete T Cell Media with 50U/mL IL-2, then staining for the same surface markers as well as PKM1 and PKM2. Flow cytometry data were acquired on a Becton-Dickinson LSRFortessa and analyzed with FlowJo 10 (FlowJo, LLC).

Mouse tissue harvest.
At speci ed time points, mice were euthanized and perfused with PBS. Lungs were then dissected, diced, ground through a 140µm wire mesh (Cell Screen/100mesh, Bellco Glass, Inc.) then ltered through 70µm lters. Lymph nodes were dissected and ltered through 70µm lters. Resultant single-cell suspensions had red blood cells lysed with ACK Lysing Buffer, and cell pellets were washed and resuspended in FACS Buffer. Surface stains, nuclear stains, and stains requiring methanol permeabilization were performed as described above. Intracellular stains for cytokines and effector proteins required stimulation with 1µg/mL Ova 257 − 264 peptide (InvivoGen) and 2 million freshly isolated C57Bl/6J splenocytes for 5 hours in Complete T Cell Media at 37°C in a humidi ed incubator, with simultaneous Golgi blocking with Brefeldin A and Monensin. Subsequent staining and acquisition was performed as described above.
Sorting of mouse lung-in ltrating T cells and TCF1 eGFP reporter-expressing T cells from in vitro co-culture, RNA extraction, and RNA Sequencing analysis.
At speci ed time points, mice were euthanized and tissue harvested and red blood cells lysed as above. Samples For RNA Sequencing of cells from the OT-I + Tcf7 GFP+ strain expressing different levels of a TCF1 eGFP reporter after in vitro co-culture with HKP1-ova-GFP tumor cells and treatment with either DMSO, AG1, or 2-DG, cells were harvested, ltered, and washed, and stained as described above, and DAPI − CD8β + Thy1.2 + eGFP + or DAPI − CD8β + Thy1.2 + eGFP-cells were sorted into Complete T cell Media, washed, resuspended into RLT Plus lysis buffer supplemented with b-mercaptoethanol, and frozen at -80℃ until RNA extraction using RNeasy Plus Micro Kits with gDNA elimination via gDNA Eliminator spin column as per the manufacturer's protocols.
For the treatment-naïve dataset, cDNA libraries were generated using the Illumina TruSeq RNA Sample Preparation kit and sequenced single-end 50 bps on the HiSeq2500 sequencer. Tophat2 75 was used to align raw sequencing reads to the mm9 mouse reference genome. Cu inks 76,77 was used to measure transcript abundances in Fragments Per Kilobase of exon model per Million mapped reads (FPKM) with upper-quartile normalization and sequence-speci c bias correction. Inter-sample relationships within the CD8 + T cell dataset were evaluated by principal component analysis in R 78 and visualized using ggplot 79 . Differential gene expression was assessed by utilizing limma 80 , with pairwise comparisons of isolated CD8 + T cells in groups identi ed by differential tumor burden (Naïve, Day 7,High,and Low). Signi cance cutoff values were set at absolute log 2 fold-change ≥ 1, p-value < 0.05, and adjusted p < 0.2. Heatmaps were made using pheatmap and RColorBrewer after normalizing log 2 -transformed FPKM values by the maximum FPKM value for each transcript.
Heatmaps in Fig. 1E and Extended Data Fig. 1B  For analysis of the previously published anti-PD-1 treatment dataset 4 (GSE114300), differential gene expression was performed as previously published, and pairwise comparisons performed between BLI groups 4 and 6 for the Venn diagram found in Fig. 1C using the same signi cance cutoffs. For GSEA, comparisons were performed between BLI groups 2 and 3, 4 and 6, and 5 and 6 using fgsea 84 .
For the adoptive co-transfer dataset, cDNA libraries were generated using the SMART-Seq v4 Ultra Low Input RNA plus Nextera XT DNA Sample preparation kit and sequenced with paired-end 2x100 cycles on the NovaSeq 6000 sequencer. Raw sequencing reads in BCL format were processed through bcl2fastq 2.19 (Illumina) for FASTQ conversion and demultiplexing. Adaptors were trimmed with cutadapt (version1.18), RNA reads were aligned and mapped to the GRCm38 mouse reference genome by STAR 85 (version 2.5.2), and transcriptome reconstruction was performed by Cu inks (version 2.1.1). Raw read counts per gene were extracted using HTSeq-count v0.11.2 85 . Differential gene expression was determined using DESeq2 86 . Principal component analyses were performed using DESeq2. The mouse-speci c effect was regressed out using the removeBatchEffect from limma, and PCA was performed on the corrected values. GSEA was performed as above.
For the Tcf7 GFP reporter dataset, cDNA libraries were generated using the Illumina Stranded mRNA Prep kit and sequenced with paired-end 2x100 cycles on the NovaSeq 6000 sequencer. Raw sequencing reads in BCL format were processed through bcl2fastq 2.20 (Illumina) for FASTQ conversion and demultiplexing. RNA reads were aligned and mapped to the GRCm39 mouse reference genome by STAR 85 (version 2.7.10b). The sequencing and mapping quality was evaluated using FastQC (v0.12.1) 87 . Raw read counts per gene were extracted using HTSeqcount (version 2.0.1) 85 . Differential gene expression was determined using a Wald test from DESeq2 86 . Genes with adjusted p-value ≤ 0.05 and fold change ≥ 1.5 were considered signi cantly differentially expressed.
Principal component analyses were performed using DESeq2. Volcano plots were prodused using the R package EnhancedVolcano 88 . GSEA was performed as above.
For identi cation of upstream regulators, Qiagen Ingenuity Pathway Analysis (IPA, Qiagen, Inc.) software was used. Relevant differential gene expression analyses generated by DESeq2 were uploaded to IPA, then core analysis run using lters of: Species-Mouse; Tissues & Cell Lines-T Lymphocytes; baseMean Cutoff-2000; and pvalue Cutoff-0.1. Upstream Regulators were identi ed and ltered by absolute Activation z-score > 2.
Steady-state polar metabolite pro ling was performed according to a method described in a previous publication 89 . T cells were sorted from co-culture or off CD3ε and anti-CD28 plates as described above, washed twice with ice-cold PBS, then supernatants aspirated and cell pellets frozen at -80℃ until analysis. All metabolomics samples were analyzed at the same time. Metabolites were extracted using pre-chilled 80% methanol (-80°C). The extract was dried completely with a Speedvac. The dried sample was redissolved in HPLC grade water before it was applied to the hydrophilic interaction chromatography LC-MS. Metabolites were measured on a Q Exactive Orbitrap mass spectrometer (Thermo Scienti c), which was coupled to a Vanquish For glucose isotope labelling experiments, T cells were activated, genetically modi ed, and co-cultured as described above. At 4 days and 6 days post-initial stimulation, viable CD8 + T cells were sorted from co-culture by FACS as described above into glucose-free RPMI 1640 Medium with L-Glutamine and without Glucose (Gibco) supplemented with 10% heat-inactivated dialyzed (12-14kD) FBS (Atlanta Biologicals), insulin-transferrinselenium-ethanolamine (ITS-X, Gibco), 1mM sodium pyruvate (Gibco), 100 U/mL penicillin with 100µg/mL streptomycin (Corning), 50µM β-mercaptoethanol (Sigma), and 50U/mL IL-2 (designated Tracer Sort Media). Immunocompetent patient-derived tumor organoid (PDTO) analysis.
PDTOs were developed as previously with minor modi cations 90  Heregulin Beta-1 (Peprotech), and 500nM A-83-01 (Tocris)). Up to ten 100µL drops of Matrigel/cell suspension were distributed into a 6 well cell suspension culture plate (Gibco). The drops were allowed to polymerize for 30 minutes inside the incubator at 37°C and 5% CO 2 and afterwards, 3mL tumor type-speci c primary culture media were added per well. Fresh culture media was replaced every 3 to 4 days. PDTOs at approximately 300 to 500µm were passaged using TrypLE Express (Gibco) for 10-12 minutes in the water bath at 37°C. Single cells and small cell clusters were replated according to the procedure described above. Monthly mycoplasma screening was performed using the PCR Mycoplasma Detection Kit (Applied Biological Materials). PDTOs were cryopreserved in Recovery Cell Culture Freezing Medium (Gibco) in liquid nitrogen.
Human T cell rapid expansion protocol (REP).
REP was performed following methods published by Jin and Rosenberg 53 . Brie y, patient peripheral blood mononuclear cell (PBMC) T cells were cultured with allogeneic PBMCs irradiated at 40 or 50Gy at a ratio of 1:200 in a large volume of media (T-75 ask, multiple wells in 6 well plates), supplemented with 3000U/mL recombinant human IL-2 and 1µg/mL anti-CD3. 3µM AG1 or DMSO vehicle control were also included in the media. Half of the media was refreshed every 3 days. After 14 days of culture, cells were harvested, assayed, and cryopreserved as above. An aliquot of these cells was collected and intracellular staining for TCF1 was performed at 48 hours after a dose of drug.
Human T cell and NSCLC PDTO co-culture.
As previously described in Djskstra et al 52  Human PDTO killing assay.
T cells recovered after 2 weeks of co-culture with human NSCLC PDTOs were rechallenged with PDTOs to assess their cytotoxic potential. PDTOs were stained with Cell Trace Far Red (Thermo Fisher Scienti c) and seeded in 96 well plates at 5x10 5 cells/mL in 150µL of medium containing 5µM NucView488 caspase-3 substrate (Biotium). T cells were added at a 5:1 effector:target ratio in 50µL of medium. Plates were imaged using Incucyte (Sartorius) every hour, capturing 4 images per well, over 12 hours. Apoptotic PDTOs were identi ed as double positive cells for Cell Trace Far Red and NucView488. Baseline spontaneous apoptotic events were subtracted for each experimental condition.
Statistical analysis.
Results are expressed as mean ± SD. RNA Sequencing gene expression data were analyzed using the limma or DESeq2 packages in R. Gene Set Enrichment Analyses were performed using the fgsea package in R. All statistical tests used are described in the corresponding gure legend. T cells in vitro were analyzed by multiple unpaired two-tailed t-tests with Holm-Šídák multiple comparisons corrections or two-way ANOVAs with Dunnett's multiple comparisons correction. Ex vivo analyses of adoptively co-transferred T cells were performed using multiple paired two-tailed t-tests with Holm-Šídák multiple comparisons corrections or two-way ANOVAs with Tukey's multiple comparisons correction. Analyses of tumor burden between groups by bioluminescence imaging used two-way ANOVA with Šídák multiple comparisons corrections and Grubbs' outlier test with alpha = 0.0001.
Mouse overall survival was evaluated by Log-rank (Mantel-Cox) test. Metabolomics data were analyzed using multiple unpaired two-tailed t-tests. All statistical analyses not performed in R were done using the GraphPad Prism 9 statistical program (GraphPad Software, LLC   f, Representative contour plots for Tox and TCF1 staining in T cells infected with two hairpins targeting CD4 (red, CD4-4E and CD4-5C) and two hairpins targeting PKM (blue, PKM-2B and PKM-2D) at 6 days post-initialstimulation.
g, Quanti cation of populations of T cells infected with shCD4 (red) and shPKM (blue) at 6 days post-initialstimulation.
Metabolite abundance was normalized by total isotope counts, with resultant data calculated as fold of average NTC-2 abundance for the given batch.
c, Metaboanalyst identi cation of the pentose phosphate pathway as the most impacted pathway.
d, Labeling pattern and quanti cation for glycolysis and pentose phosphate pathway metabolites after one round, with statistically signi cantly enriched (p<0.05) labelled metabolites shown in dark blue and labelled metabolites with p<0.1 in purple in (d). Multiple rounds of the pentose phosphate pathway can occur, with F6P isomerizing with G6P or integration of glycolysis-derived G3P, leading to different labeling patterns and isotopes.
Numbers: (a-b) n=4 biological replicates per guide, experiment repeated three times; (c-k) n=9-10 biological replicates per group, aggregate of three experiments.

Figure 6
Pentose phosphate pathway agonism in T cells generates a memory-like phenotype distinct from that induced by blockade of glucose utilization.
a-d, Flow cytometry analysis of activated OT-I+ Thy1.1+ T cells treated with either DMSO control (red) or 3µM glucose-6-phosphate dehydrogenase agonist AG1 (blue) and co-cultured with HKP1-ova-GFP tumor cells. T cells were activated for 24 hours, treated with DMSO or AG1 for another 24 hours, then co-cultured with HKP1-ova-GFP tumor cells at a 5:1 effector:target ratio until 6 days post-initial-stimulation with continuing DMSO or AG1 treatment.

Figure 7
Pentose phosphate pathway agonism results in tumor control in murine and human model systems.
a-f, OT-I+ Thy1.1+ CD8+ T cells were activated for 1 day, treated with either DMSO or 3 µM AG1 for 3 days, then adoptively transferred into lymphodepleted C57Bl/6 mice 7 days after orthotopic implantation of HKP1-ova-GFP tumors. 3 doses of anti-PD-1 were administered on days 10, 14, and 17 after tumor implantation. Phenotype of treated adoptively transferred T cells was evaluated, and tumor burden and overall survival monitored.
a, Experimental schematic.
b-c, Flow cytometric analyses of transferred T cells for CD44 and CD62L expression (b) and Tox and TCF1 expression (c).