Integrated stress response couples mitochondrial fitness with lineage reprogramming to drive cancer evolution Authors: Shiqi Diao, Jia Yi Zou, Shuo Wang, Jason E Chan, Roderik M Kortlever, Nicolas Poulain, Nour Ghaddar, Hyungdong Kim, Gerard I Evan, Constantinos Koumenis, Maria Hatzoglou, Peter Walter, Nahum Sonenberg, John Le Quesne, Tuomas Tammela, Antonis E Koromilas Published: 2026 ABSTRACT AbstractTumour progression towards dedifferentiated cell clusters plays a critical role in intratumour heterogeneity and therapy resistance. While tumour microenvironmental stress has been implicated, the underlying mechanisms remain poorly defined. Using mouse models of lung adenocarcinoma, we demonstrate that activation of the integrated stress response (ISR)—marked by phosphorylation of eIF2 (p-eIF2) and ATF4 induction—drives tumour heterogeneity. ISR activation facilitates the emergence of high-plasticity, undifferentiated and pre-epithelial-to-mesenchymal transition clusters characterized by elevated ATF4 and MYC activity. This process is MYC dependent and involves ISR-mediated repression of NKX2-1, a key determinant of alveolar identity, and induction of CHCHD10, a regulator of mitochondrial integrity and metabolic fitness. Disruption of the p-eIF2–ATF4 axis induces mitochondrial dysfunction, limits dedifferentiation and suppresses tumour growth. In human lung adenocarcinoma, ISR-driven dedifferentiation correlates with advanced disease and poor prognosis, identifying the ISR as a central driver of lineage reprogramming and metabolic fitness in tumour progression.Subject terms: Energy metabolism, Stress signalling, Cancer metabolism, Non-small-cell lung cancer Main A defining feature of cancer is the ‘stress phenotype’, arising from increased DNA damage, proteotoxicity and metabolic stress caused by oncogenic mutations or factors in the tumour microenvironment (TME)1. The integrated stress response (ISR) enables adaptation to diverse stressors, including nutrient deprivation, hypoxia and misfolded protein accumulation2. By modulating protein synthesis, the ISR promotes survival and eliminates irreparably damaged cells2. Four kinases (PERK, GCN2, PKR and HRI) activate the ISR by phosphorylating the α subunit of eIF2 at serine 51 (p-eIF2α)3,4. This phosphorylation reduces global protein synthesis while selectively enhancing translation of stress-related mRNAs3,5. The ISR protects normal cells from stress-induced damage, thereby suppressing tumourigenesis2,3. However, cancer cells can co-opt the ISR to survive the hostile TME3. ISR-mediated reprogramming supports tumour cell survival and growth, making it a promising therapeutic target6. Although ISR’s role in stem cell phenotypes is established7, its function in lineage reprogramming, the process by which cells dynamically transition into distinct phenotypic states in a stressful microenvironment in vivo, remains unclear. Using lineage tracing and single-cell transcriptomics, we demonstrate that ISR activation, driven by increased p-eIF2α, promotes lung adenocarcinoma (LUAD) tumour expansion towards high-plasticity, highly undifferentiated and pre-epithelial-to-mesenchymal transition (EMT) phenotypic states. This transition is mediated by suppression of the lineage-specifying transcription factor NK2 homeobox 1 (NKX2-1; thyroid transcription factor-1, TTF-1), accompanied by elevated expression of activating transcription factor 4 (ATF4) and enhanced MYC activity within these cell clusters. We further show that ISR suppression leads to the emergence of a developmentally arrested cluster characterized by impaired mitochondrial function. Together, these findings reveal that the ISR promotes LUAD evolution towards undifferentiated, high-plasticity cell states by preserving mitochondrial fitness. Results ISR inhibition disrupts LUAD tumour evolution in mice We used a mouse LUAD model combining Cre-loxP-mediated oncogenic Kras G12D expression and Tp53 suppression via short hairpin RNA (shRNA), with a conditional homozygous S51A knock-in mutation of Eif2s1 (fTg/0; eIF2αA/A), which prevents eIF2α phosphorylation at serine 51 and thereby blocks ISR activation6. Wild-type Eif2s1 mice (fTg/0; eIF2αS/S) served as controls6. Lung tumours were induced by intratracheal intubation of Cre-expressing lentiviruses driven by the carbonic anhydrase 2 (CA2) promoter, which selectively targets alveolar lung epithelial cells8, with co-expression of Trp53 shRNA to accelerate tumour formation8 (Fig. 1a). Because CA2 expression in the alveolar compartment is restricted to alveolar type (AT)2 cells9, CA2-Cre-driven expression occurs in AT2 cells8. GFP expression enabled tumour visualization and flow cytometric sorting6 (Fig. 1a). Ultrasound imaging revealed that loss of p-eIF2α reduced lung tumour formation and growth6 (Fig. 1b). ISR drives mouse LUAD progression to high-fitness lineages, leading to increased tumour heterogeneity. a, Schematic overview of the experimental design and analysis workflow. t-SNE, t-distributed stochastic neighbor editing. b, Representative ultrasound images of lung tumours detected in eIF2αS/S (n = 6) and eIF2αA/A (n = 7) mice 35 weeks after Cre-lentivirus intubation. Tumours are outlined by dashed yellow lines. Quantification of tumour size based on ultrasound is shown. Mean ± s.d.; dots represent individual mice. Two-sided Mann–Whitney U test; P = 0.0175. c, Uniform manifold approximation and projection (UMAP) of tumour cells coloured by annotated cell types across genotypes, with relative cluster percentages (n = 17,919 cells). d, Heatmap showing scaled average expression of selected marker genes across identified clusters. e, Signature scores of selected gene sets and hallmark pathways across annotated cell types. f, Comparison of ISR scores across annotated cell types from merged eIF2αS/S and eIF2αA/A tumour datasets. Schematic in a created in BioRender; Zou, J. Y. https://biorender.com/3l8gg4t (2026). Source data To investigate the role of the ISR, particularly p-eIF2α, in LUAD evolution, GFP+/CD45− lung tumour cells were isolated by flow cytometry and subjected to single-cell RNA sequencing (scRNA-seq) (Fig. 1a). Tumours were collected at 35 weeks to ensure sufficient representation of p-eIF2α-deficient cells given their growth impairment6 (Fig. 1b). We identified nine transcriptionally distinct clusters, including AT2-like, AT1-like, club-like, hybrid lung/gastric-like, high-plasticity and pre-EMT states, consistent with previously reported LUAD cell populations10–13 (Fig. 1c,d). We also defined a ‘lung transition’ cluster representing a continuum between AT1-like and high-plasticity states (Fig. 1c,d). In addition, scRNA-seq revealed a highly undifferentiated cell cluster that does not correspond to a defined lung epithelial lineage but instead exhibits moderate activation of basal stem cell programmes14 (Fig. 1e). This state is characterized by elevated unfolded protein response (UPR), mTOR signalling and MYC-dependent transcriptional programmes15,16 (Fig. 1e and Extended Data Fig. 1a). Transcriptomically, this cluster closely resembles the pre-EMT state (Extended Data Fig. 1b), although the latter exhibits higher expression of neuronal reprogramming genes, including high-mobility group AT-hook 2 (Hmga2)17 and one cut homeobox 2 (Onecut2)18, as well as EMT markers such as zinc finger E-box-binding homeobox 2 (Zeb2)19 (Fig. 1d and Supplementary Table 1). Characterization of eIF2αS/S and eIF2αA/A LUAD clusters. a, Dot plot analysis of lineage-associated genes across tumor cell clusters. b, Box plots showing Bhattacharyya distances quantifying transcriptomic similarity between the highly undifferentiated cluster and other LUAD cell clusters. Distances were calculated in PCA space (first 50 principal components) using n = 10 independent random subsampling iterations (100 cells per cluster per iteration), with distances averaged across principal components for each iteration. Boxes indicate median and IQR; whiskers extending to 1.5 × IQR. c, Heatmap showing scaled feature scores of the Molecular Signatures Database (MSigDB) canonical gene sets across annotated cell types, ordered by inferred tumor progression. eIF2αA/A tumours contained AT2-like, AT1-like, lung transition, hybrid lung/gastric-like and club-like clusters, but lacked late-stage populations (Fig. 1c). Instead, they harboured a distinct cluster, termed Mito-Dysf, characterized by dysregulation of nuclear-encoded mitochondrial genes, with concurrent upregulation of oxidative phosphorylation programmes (Fig. 1c–e). Compared with highly undifferentiated and pre-EMT cells, Mito-Dysf cells retain higher Sftpc and Nkx2-1 expression, indicating greater lineage fidelity and a more differentiated state20 (Extended Data Fig. 1a,c). Consistently, Mito-Dysf cells show reduced expression of developmental regulators: hepatocyte nuclear factor 4 alpha (Hnf4a)21 and Hmga217 (Extended Data Fig. 1a), and do not align with established tumour evolutionary fates based on fitness module scoring12 (Fig. 1e, second panel from the top), suggesting a failed cellular reprogramming process. ISR signature analysis22 revealed the lowest scores in the Mito-Dysf cluster, and high scores in high-plasticity and pre-EMT populations, with the highest levels observed in the highly undifferentiated cluster (Fig. 1f). Together, these findings indicate that ISR promotes lineage reprogramming towards late-stage clusters, whereas its genetic inactivation restricts this progression, resulting in a more differentiated, mitochondrially dysregulated Mito-Dysf fate. ISR-upregulated genes drive lineage transition in LUAD tumours To model tumour cell lineage transitions, we applied CellRank for single-cell fate mapping23, identifying AT2-like cells as the initial state and four terminal fates: AT1-like, hybrid lung/gastric-like, pre-EMT and Mito-Dysf (Fig. 2a–d). AT2-like cells from both genotypes occupy the same trajectory space, with divergence arising only beyond this shared starting point (Figs. 1c and 2d). The lung transition cluster emerges as a key decision point along this trajectory (Extended Data Fig. 2a,b), present in both genotypes but with distinct transcriptional programmes. eIF2αS/S lung transition cells show increased expression of ISR target genes such as asparagine synthetase (Asns), apolipoprotein B mRNA editing enzyme catalytic subunit 3G (Apobec3)24, and eukaryotic initiation factor 4E-binding protein 1 (Eif4ebp1)25 and are enriched for chromatin remodelling and mRNA translation pathways. By contrast, eIF2αA/A cells upregulate mitochondrial biogenesis genes including RNA binding motif protein 3 (Rbm3)26, ubiquitin A-52 residue ribosomal protein fusion product 1 (Uba52)27 and eukaryotic translation initiation factor 5A (Eif5a)28, with enrichment of mitochondrial function and apoptosis pathways (Extended Data Fig. 2c,d and Supplementary Table 1). ISR acts as the primary determinant of the high-fitness lineage trajectory. a, CellRank-computed macrostates highlight the most confidently assigned cells for each state. Names and colours were assigned based on overlap with clusters from Fig. 1c, with AT2-like cells representing the initial cell state. b, Circular projection of cells based on fate probabilities towards the macrostates shown in a, coloured by cluster annotations from Fig. 1c. Macrostates are positioned at the periphery, and cells are placed according to their probability of transitioning towards each terminal state. c, Pseudotime scores and predicted ordering of the nine clusters. d, Partition-based graph abstraction (PAGA) plot illustrating directional transitions between tumour cell types based on RNA velocity. The velocity field is projected onto the UMAP, with arrows indicating local average velocity. e, Smoothed gene expression trends along pseudotime, with each coloured trend weighted by CellRank fate probabilities for the indicated lineages. f, Genes positively correlated with the pre-EMT fate are negatively correlated with the Mito-Dysf fate, and vice versa. The top driver Eif4ebp1 of pre-EMT fate is highlighted in red. g, Functional annotation of CellRank-predicted driver genes for pre-EMT and Mito-Dysf states. Bar graph shows top non-redundant enriched ontology clusters. Bar length (x axis) represents −log10P from a hypergeometric test; colour indicates statistical significance (raw P values), with darker colours denoting progressively stronger enrichment. h, Evaluation of cellular developmental potential using CytoTRACE2. Higher scores indicate increased undifferentiation. Left: box plots showing CytoTRACE2 scores across tumour cell clusters (median and interquartile range (IQR); whiskers extend to 1.5× IQR; dots represent outliers). Right: violin plots comparing CytoTRACE2 scores between eIF2αS/S and eIF2αA/A tumour cells. Data include n = 11,167 cells from a single pooled sample of 5 eIF2αS/S mice, and n = 6,752 cells from a pooled sample of 5 eIF2αA/A mice. i, Correlation of developmental potential with ISR and AT2 signatures. Scatter plots show CytoTRACE2 scores versus ISR scores (left) or AT2 scores (right). Each dot represents an individual cell. Solid lines indicate linear regression. Two-sided Spearman’s rank correlation test: Spearman correlation coefficients (R) and P values are indicated. Critical turning points in eIF2αS/S and eIF2αA/A LUAD evolution. a, CellRank analysis showing fate entropy across tumor cell clusters. Left: UMAP colored by fate entropy. Right: Violin plot showing distribution of entropy values across annotated cell clusters. b, Total Bhattacharyya distance quantifying transcriptomic similarity between a selected cluster and all other inferred cell lineages. c, UMAP showing the lung transition cluster from eIF2αS/S and eIF2αA/A tumors (n = 3,909 cells; left). Dot plot of top differentially expressed genes between eIF2αS/S and eIF2αA/A tumors (right). d, Pathway enrichment of the top 100 genes in the lung transition cluster for eIF2αS/S (top) or eIF2αA/A (bottom). Bar plot shows top non-redundant enriched ontology clusters. Bar length represents-log10(P value) from a hypergeometric test. e, Smoothed gene expression trends of Tomm20, Eif5a, Hmga2 and Onecut2 along pseudotime, weighted by CellRank fate probabilities. Tomm20, Eif5a are key driver genes for Mito-Dysf cluster, while Onecut2 and Hmga2 are the key driver genes for pre-EMT cluster. f, Scatter plot showing gene correlation with different fate probabilities. Each point represents one gene, colored by average expression. g, Correlation of developmental potential with ISR activity and MYC activity. Scatter plots show MYC scores versus ISR scores (top) or CytoTRACE2 scores (bottom). Each dot represents an individual cell. Solid lines indicate linear regression. Two-sided Spearman’s rank correlation test: Spearman correlation coefficients (R) and P values indicated. Pseudotime analysis revealed that ISR-upregulated genes, such as Atf4 and Eif4ebp1, increase progressively along pre-EMT trajectory, alongside established LUAD evolution markers Hmga2 and Onecut211 (Fig. 2e and Extended Data Fig. 2e). Eif4ebp1 emerged as a top driver of the pre-EMT fate (Fig. 2f and Supplementary Table 2), while homeodomain-only protein homeobox (Hopx) and Hnf4a drove AT1-like cell fate and hybrid lung/gastric-like fates, respectively13,21 (Fig. 2e). Conversely, mitochondrial genes such as Rbm3, ATP synthase membrane subunit K (Atp5k), Eif5a and translocase of outer mitochondrial membrane 20 (Tomm20) increased along the Mito-Dysf trajectory in eIF2αA/A tumours (Fig. 2e and Extended Data Fig. 2e). The pre-EMT and Mito-Dysf states were mutually exclusive, with genes positively correlated with one negatively correlated with the other (Fig. 2f). Pathway enrichment highlighted this divergence; pre-EMT top genes associate with mRNA translation, amino acid biosynthesis and stress response, while Mito-Dysf top genes associate with mitochondrial protein turnover (Fig. 2g). Consistently, AT1-like drivers negatively correlated with pre-EMT drivers (Extended Data Fig. 2f), in line with pre-EMT dedifferentiation features (Extended Data Fig. 1c). CytoTRACE analysis29 confirmed that highly undifferentiated and pre-EMT clusters exhibit the highest undifferentiation scores, whereas the Mito-Dysf cluster—despite occupying a terminal position—retains moderate differentiation (Fig. 2h). Overall, eIF2αS/S tumours show higher undifferentiation than eIF2αA/A tumours (Fig. 2h), and ISR signature scores positively correlate with both undifferentiation and MYC signature scores (Fig. 2i and Extended Data Fig. 2g), implicating MYC in ISR-mediated LUAD evolution. To identify ISR-dependent drivers of the AT2-like to pre-EMT transition, we extracted the top 100 CellRank transition genes (Supplementary Table 2), of which 40 were ISR dependent, including Eif4ebp1 and Psph30 (Extended Data Fig. 3a–c). Immunohistochemistry (IHC) confirmed higher 4E-BP1 and ONECUT2 protein levels in eIF2αS/S relative to eIF2αA/A tumours (Extended Data Fig. 3d), supporting the role of ISR-dependent translation programmes in driving progression towards pre-EMT states. ISR signaling drives the transition of AT2-like cells into pre-EMT states. a, Venn diagram showing overlap between the top 100 pre-EMT fate driver genes identified by CellRank and the ISR gene set. b, Pathway enrichment analysis of the 40 genes shared between the ISR gene set and pre-EMT fate driver genes identified in panel a. Bar length represents -log10(P value) from a hypergeometric test; color indicates statistical significance, with darker colors denoting progressively stronger enrichment. c, Protein-protein interaction (PPI) network of the 40 shared genes. Atf4 and Eif2s1 were manually added to the network for visualization. Node size and color intensity represent interaction degree; nodes without detected interactions were excluded. d, IHC analysis of 4E-BP1 (n = 4 mice per genotype) and ONECUT2 (n = 3 mice per genotype) in eIF2αS/S and eIF2αA/A tumors. Scale bars, 50 μm. Mean ± s.d.; two-sided unpaired Student’s t-test; P values indicated. e, Partition-based graph abstraction (PAGA) velocity graph showing connectivity among major epithelial cell states across integrated eIF2αS/A and eIF2αA/A datasets. Nodes represent annotated clusters; edge thickness indicates connectivity strength. Bottom: tumor cells colored by annotated cell types in each genotype. f, Integrated UMAP of p-eIF2α and ATF4-deficient (Atf4+/−) tumors showing major cell states and progression toward the Mito-Dysf fate. Cells are colored by cell type and genotype (Atf4+/– in green, eIF2αA/A in blue). The inferred ‘bridge’ population along the trajectory is highlighted, and PAGA velocity graph (Leiden-based) overlay indicates the main trajectory toward the Mito-Dysf branch. Source data Together, early lineage transitions occur independently of the ISR, but the lung transition cluster represents a critical decision point at which ISR activity-through effectors including Atf4 and Eif4ebp1-drives divergence towards high-plasticity and pre-EMT fates. Loss of p-eIF2α redirects cells towards the mutually exclusive Mito-Dysf fate, constraining tumour progression. Partial ISR disruption impedes LUAD tumour progression To determine whether partial ISR reduction impairs LUAD evolution, we analysed tumours expressing a heterozygous Eif2s1 S51A knock-in allele (eIF2αS/A). Ultrasound imaging confirmed reduced tumour growth relative to eIF2αS/S controls (Fig. 3a). scRNA-seq revealed that eIF2αS/A tumours retained all clusters present in eIF2αS/S tumours but lacked the Mito-Dysf cluster, with marked reductions in high-plasticity and pre-EMT clusters and expansion of AT2-like and hybrid lung/gastric-like clusters (Fig. 3b). This is accompanied by decreased expression of highly undifferentiated cluster genes in eIF2αS/A compared with eIF2αS/S, including MYC-dependent programmes in the highly undifferentiated cluster (Fig. 3c,d), indicating impaired acquisition of the highly undifferentiated phenotype (Fig. 3b and Supplementary Table 3). p-eIF2 and ATF4 haploinsufficiency restrains LUAD evolution. a, Ultrasound imaging and quantification of tumour burden. Representative ultrasound images of lung tumours at 35 weeks after Cre-lentivirus intubation in eIF2αS/S and eIF2αS/A mice. Tumours are outlined by dashed yellow lines. Right: quantification of tumour size based on the ultrasound imaging. Mean ± s.d.; dots represent individual mice (n = 7 per group). Two-sided Mann–Whitney U test; P = 0.0023. b, UMAP showing tumour cells from eIF2αS/S and eIF2αS/A groups (n = 15,588 cells), coloured by annotated cell types. Density overlay indicates cell distribution. Relative cluster percentages are shown. c,d, Scaled signature scores for the highly undifferentiated cluster (c) and MYC upregulated (UP) signature (d) across cell types in eIF2αS/S, eIF2αA/A and eIF2αS/A models. Values represent the average score of single-cell transcriptomes within each group. Greyed-out indicate conditions comprising <5% of cells within the corresponding dataset. e, Distribution of Atf4 (+) regulon activity scores calculated using SCENIC across annotated cell types. f, Representative histology and quantification of ATF4 nuclear localization. Left: haematoxylin and eosin (H&E) and ATF4 IHC staining of LUAD tumour sections from eIF2αS/S and eIF2αA/A mice. Scale bars, 50 μm. Right: quantification of the log2-transformed nuclear-to-cytoplasmic (N/C) ratio of ATF4 staining intensity. Mean ± s.d.; dots represent individual mice (n = 3 per group). Two-tailed unpaired Student’s t-test; P value indicated. g, Ultrasound imaging (left) and quantification (right) of tumour burden. Representative ultrasound images of lung tumours at 35 weeks after Cre-lentivirus intubation in eIF2αS/S Atf4+/+ and eIF2αS/S Atf4+/− mice. Tumours are outlined by dashed yellow lines. Quantification of tumour size based on ultrasound imaging. Mean ± s.d.; dots represent individual mice (n = 5 per group). Two-sided Mann–Whitney U test; P = 0.0317. h, UMAP showing tumour cells from eIF2αS/S Atf4+/+ and eIF2αS/S Atf4+/− groups (n = 19,943 cells), coloured by annotated cell types (left). The density overlay indicates the cell distribution (middle), and relative cluster percentages (right) are shown. Source data To assess whether the reduced highly undifferentiated population in eIF2αS/A tumours might transition towards the Mito-Dysf fate, we performed integrated trajectory analysis combining eIF2αS/A and eIF2αA/A datasets (Extended Data Fig. 3e). This showed that the Mito-Dysf fate arises from a separate branch rooted in the lung transition cluster and becomes established only upon complete ISR loss, and that the highly undifferentiated cluster does not form a trajectory towards Mito-Dysf. Thus, reduction of highly undifferentiated cells and emergence of the Mito-Dysf cluster represent parallel consequences of ISR impairment rather than a precursor–product relationship, and full p-eIF2α abrogation is required to unmask the latent lung transition to Mito-Dysf trajectory, Taken together, these data demonstrate that intact ISR is required for MYC-driven progression towards highly undifferentiated and pre-EMT states, whereas partial ISR reduction constrains tumours progression before commitment to the Mito-Dysf fate. ATF4 drives the emergence of undifferentiated clusters We performed single-cell regulatory network inference and clustering (SCENIC)31 to identify transcriptional factors governing cluster-specific gene programmes. ATF4 activity was linked to the progression of lung transition cells towards highly undifferentiated and pre-EMT clusters in eIF2αS/S tumours, while it remained low during lung transition to Mito-Dysf conversion in eIF2αA/A tumours, consistent with retention of AT2-like and AT1-like features in the Mito-Dysf cluster (Figs. 2e and 3e). Given that ATF4 expression is reduced in eIF2αA/A tumours (Fig. 3f), we assessed its function using a conditional Atf4f/f allele32. Heterozygous deletion of Atf4+/− in eIF2αS/S tumours decreased tumour growth (Fig. 3g), whereas homozygous deletion abrogated detectable tumour formation by ultrasound imaging. scRNA-seq of Atf4+/− tumours showed loss of the highly undifferentiated, high-plasticity and pre-EMT clusters, with enrichment of AT2-like, lung transition, and hybrid lung/gastric-like clusters (Fig. 3h). Integrated analysis revealed partial overlap between eIF2αA/A and Atf4+/− transcriptional landscapes, with a continuum linking late-stage clusters in Atf4+/− tumours to the Mito-Dysf state in eIF2αA/A tumours (Extended Data Fig. 3f). However, despite this proximity, cells do not fully transition into the Mito-Dysf fate in the presence of residual ATF4 activity, suggesting that ATF4 constrains access to this alternative trajectory. Therefore, ATF4 is a critical effector of ISR-driven LUAD progression, required for transition to aggressive undifferentiated states; its loss redirects tumours towards more differentiated fates and impairs growth. MYC is required to initiate ISR-driven lineage evolution in LUAD The positive correlation between ISR, MYC and undifferentiation scores (Extended Data Fig. 2g) prompted us to directly assess MYC’s role in ISR-driven lineage evolution using a conditional MycON/OFF LUAD model (KM model)33, in which Myc expression in KRAS G12D-driven LUAD is reversibly activated by tamoxifen33,34 (Fig. 4a). MYC drives ISR-linked lineage progression and ATF4-coupled transcription in LUAD. a, Schematic timeline of the KM LUAD experiment. Mice harbouring KrasG12D and MycER alleles were treated with tamoxifen 12 weeks after mutant KRAS expression to induce MYC activity (MycON). After 4 weeks, tamoxifen was withdrawn to switch Myc activity off (MycOFF). Tumours were collected at defined timepoints (T) corresponding to MycON (8 h, 24 h and 7 days; n = 2–4 mice per timepoint) and MycOFF states (8 h, 12 h and 24 h after withdrawal; n = 3–4 mice per timepoint). b, UMAP of integrated single-cell data from eIF2αS/S and MycON/OFF LUAD tumours. Green and blue dots indicate lung transition and high-plasticity clusters from eIF2αS/S tumours, respectively, while red dots indicate the corresponding population in MycON/OFF tumours (left). UMAP of MycON/OFF tumours alone shows major tumour cell populations, with each dot representing one cell coloured by annotated state (right). c, Violin plots showing ISR and MYC module expression across the indicated clusters in pooled MycON/OFF datasets. d, Heatmaps of average MYC, ISR and pre-EMT module scores across indicated clusters under MycON and MycOFF conditions. Module activity is Z-score normalized within each module. Greyed-out cells indicate conditions comprising <5% of cells within the corresponding tumour dataset. e, Box plots showing the Z-score-normalized ratio of lung transition to high-plasticity cells across MycON and MycOFF timepoints. Data represent individual mice that passed single-cell quality control (n = 2–4 mice per group; see the Methods for exclusion criteria). Boxes indicate the median and interquartile range; whiskers extend to the minimum and maximum values. f, IHC analysis for 4E-BP1 in KM tumours. Left: MycON (16 weeks of tumour age after 4 weeks of tamoxifen). Right: MycOFF (17 weeks of tumour age after 1 week of withdrawal). Quantification shown as percentage of field of view (% FoV) occupied by 4E-BP1-positive staining. Mean ± s.d.; n = 3 mice per group. Two-sided unpaired Student’s t-test; P value indicated. Scale bars, 100 μm (full view) and 50 μm (magnified view). Schematic in a created in BioRender; Zou, J. Y. https://biorender.com/irzq51m (2026). Source data scRNA-seq of Myc-activated KM tumours revealed AT2-like, AT1-like and lung transition/high-plasticity populations, identified by integration with eIF2αS/S tumours in a shared UMAP space (Fig. 4b). Terminally dedifferentiated states were absent, probably due to intact Tp53 in this model. Gene set scoring confirmed elevated ISR and MYC activity in the lung transition/high-plasticity cluster (Fig. 4c). Across multiple MycON timepoints, ISR signature scores and pre-EMT gene expression increased following Myc activation and remained elevated through 7 days (Fig. 4d). Upon Myc shutdown both ISR and MYC signatures, together with pre-EMT markers, declined within 8 h after tamoxifen withdrawal (Fig. 4d). Consistent with these transcriptional dynamics, lung transition/high-plasticity cluster abundance increased with MycON duration and decreased upon Myc inactivation (Fig. 4e). IHC confirmed 4E-BP1 expression in MycON tumours, and reduced expression levels in MycOFF tumours (Fig. 4f). These findings show that MYC–ISR cooperation drives LUAD lineage evolution towards high-plasticity states, and that MYC inactivation collapses ISR signalling and reverts tumours towards a differentiated AT2-like identity. The lung transition cluster is the fate-decision checkpoint of ISR-driven tumour evolution To test whether the lung transition cluster is a true developmental checkpoint, we reconstructed lineage relationships among tumour epithelial populations. Although CellRank identifies the AT1-like population as a potential differentiation endpoint (Fig. 2a,b), RNA velocity vectors reveal a dominant flux channelling cells through this state into the lung transition cluster (Fig. 2d), indicating that the lung transition cluster functions as an active evolutionary gateway, rather than a parallel differentiation branch. Mapping tumour cell states onto normal adult mouse lung reference atlases35 showed that AT2-like tumour cells align with normal AT2 cells, while the AT1-like cluster aligns with the normal ‘AT2/AT1’ intermediate, involved in alveolar repair22 (Extended Data Fig. 4a). The lung transition cluster lies immediately downstream of this intermediate, linking the regenerative state to the high-plasticity cluster (Extended Data Fig. 4a), indicating that LUAD progression recapitulates regeneration only up to this point before diverging towards plastic states. The lung transition cluster is a key node in ISR-driven AT2-like cell dedifferentiation. a, UMAP showing integration of tumor samples with normal lung atlas datasets (n = 65,357 cells). The bridge connecting AT2/AT2-like and lung transition clusters is circled in red. b, Pseudotime analysis of single cells colored by cell state. Solid lines indicate smoothed mean ISR scores for eIF2αS/S (red) and eIF2αA/A (blue) tumors. Shaded regions represent the 5th–95th percentile intervals of cells within each state. c, Pathway enrichment analysis of genes upregulated in the lung transition cluster relative to the AT1-like state in eIF2αS/S tumor cells. Bar plot shows top non-redundant enriched ontology clusters. Bar length represents -log10(P value) from a hypergeometric test. d, UMAP of integrated datasets from eIF2αS/S, eIF2αA/A, eIF2αS/A and eIF2αS/S Atf4+/− tumors (left), with cells colored by KAC marker expression (right). e, Pathway enrichment analysis comparing human KRT8+ alveolar intermediate cells and alveolar intermediate cells from the GSE222901 dataset. Bar plot shows top non-redundant enriched ontology clusters. Bar length represents -log10(P value) from a hypergeometric test. f, Heatmap plot shows Nkx2-1 expression across early and late-stage cell clusters in ISR tumor datasets. Greyed-out cells indicate conditions comprising <5% of cells within the corresponding dataset. g, IHC staining of NKX2-1 in tumors from eIF2αS/S and eIF2αA/A mice (n = 3 mice per genotype). Scale bars, 50 μm. Mean ± s.d.; two-sided unpaired Student’s t-test; P values indicated. Source data ISR activity along this trajectory was low in AT1-like cells but high in the lung transition cluster (Fig. 1f and Extended Data Fig. 4b), accompanied by enrichment of gastrulation and organ development pathways (Extended Data Fig. 4c). This defines a low ISR differentiation-like state (AT1-like) from a high ISR transition state (lung transition), conceptually aligning with prior frameworks distinguishing regenerative alveolar intermediate cells from stress-adaptive, tumour-initiating KRT8+ alveolar intermediate cells (KACs)10,36. Consistently, projection of the human and mouse KAC transcriptional signature10 onto our dataset revealed strong overlap with the lung transition cluster in mouse LUAD tumours (Extended Data Fig. 4d), and pathway analysis confirmed that human KACs are distinguished from alveolar intermediate cells by hyperactivation of ISR-dependent programmes (Extended Data Fig. 4e), identifying the lung transition cluster as the murine counterpart of human KACs. Downstream of this checkpoint, eIF2αS/S lung transition cells diverge along two routes: transdifferentiation towards a hybrid lung/gastric-like lineage or dedifferentiation towards high-plasticity states (Fig. 2d). Both trajectories were accompanied by progressive repression of the lung lineage determinant Nkx2-1 (Extended Data Fig. 1a), suggesting that loss of stable lung identity is a defining feature of ISR-dependent progression. In eIF2αS/S tumours, Nkx2-1 expression declined progressively with advancing tumour stage, whereas it remained persistently elevated in eIF2αA/A tumours—consistent with diversion into the developmentally arrested Mito-Dysf state (Extended Data Fig. 4f). Partial ISR attenuation in eIF2αS/A and Atf4+/− tumours produced only modest early reductions in Nkx2-1 (Extended Data Fig. 4f), indicating incomplete engagement of the dedifferentiation programme. IHC confirmed higher NKX2-1 protein levels in eIF2αA/A relative to eIF2αS/S tumours (Extended Data Fig. 4g). Together, these analyses define the lung transition cluster as a conserved, stress-adaptive checkpoint in LUAD evolution. ISR activation at this node licenses Nkx2-1 repression and progression towards dedifferentiated, plastic states; conversely, ISR disruption preserves lung lineage identity and redirects tumour cells towards the non-progressive Mito-Dysf fate. ITGA2 is enriched in late-stage tumour clusters and defines late-stage identity ex vivo Integrin α2 (Itga2 or Cd49b), a component of the integrin α2β1 collagen receptor marks high-plasticity cells in human and mouse LUAD11. In eIF2αS/S tumours, Itga2 expression was enriched in the late-stage clusters, whereas in eIF2αA/A tumours, it was restricted to the Mito-Dysf cluster (Extended Data Fig. 5a). Collectively, Itga2high cells accounted for ~30% of tumour cells in eIF2αS/S tumours (Extended Data Fig. 5b). Ex vivo cultured eIF2αS/S and eIF2αA/A LUAD cells retain the transcriptional profiles of their primary tumors. a, UMAP of the merged dataset (n = 44,780 cells), colored by cell type (left), genotype (middle), or Itga2 expression (right). b, Bar plot showing cell type percentages by genotype, colored as in panel a. Boxes indicate clusters with high Itga2 expression. c, Flow cytometry analysis of ITGA2 staining. Left: Histograms showing ITGA2 distribution across tumor types, with threshold for high expression defined as the top 30% of cells in the eIF2αS/S group. Right: Quantification of ITGA2high cells by genotype. Mean ± s.d.; eIF2αS/S (n = 7), eIF2αS/A (n = 7), eIF2αA/A (n = 3), eIF2αS/S Atf4+/− (n = 4). Two-sided unpaired Student’s t-test; P values indicated. Results not reaching significance are labeled as ‘ns’ (not significant). d, Flow cytometry of ITGA2 staining in cultured eIF2αS/S and eIF2αA/A tumor cells. Fluorescence-minus-one (FMO) control was used as negative control. e, Heatmap showing top differentially expressed genes between eIF2αS/S and eIF2αA/A cultured cells. Arrows indicate alveolar marker genes. f, GSEA of cellular state signatures comparing eIF2αS/S and eIF2αA/A cultured LUAD cells. Normalized Enrichment Scores (NES) and P values are indicated within the individual panels; negative NES indicates enrichment in eIF2αS/S cells, and positive NES indicates enrichment in eIF2αA/A cells. Statistical significance was determined using a permutation-based test (1,000 permutations). g, Western blot analysis of NKX2-1 expression in eIF2αS/S and eIF2αA/A cultured LUAD cells. TUBULIN serves as a loading control. Representative of n = 3 independent experiments. Source data Flow cytometry of freshly isolated LUAD tumours confirmed this pattern: ITGA2high cells were most abundant in eIF2αS/S tumours, reduced in eIF2αA/A tumours and lowest in eIF2αS/A and Atf4+/− tumours (Extended Data Fig. 5c). This gradient is consistent with impaired dedifferentiation in eIF2αS/A and Atf4+/− tumours, which fail to progress to the late-stage clusters consisting of high-plasticity, highly undifferentiated, pre-EMT or Mito-Dysf seen in eIF2αS/S and eIF2αA/A tumours, respectively (Extended Data Fig. 5c). Thus, ITGA2 levels mark tumours capable of advancing through developmental reprogramming versus those that cannot. Flow cytometry of cultured LUAD cells confirmed ITGA2high expression in both eIF2αS/S and eIF2αA/A cultured cells (Extended Data Fig. 5d), indicating retention of late-stage characteristics ex vivo. RNA sequencing (RNA-seq) further showed upregulation of differentiated alveolar markers, including Nkx2-1, Hopx, Sftpc, advanced glycosylation end-product specific receptor (Ager) and lysozyme 2 (Lyz2), in eIF2αA/A cells compared with eIF2αS/S cells (fold-change >2, adjusted P < 0.05), consistent with an impaired evolutionary trajectory in p-eIF2α-deficient cells (Extended Data Fig. 5e). Gene set enrichment analysis (GSEA) using CellRank fate-driving gene sets indicated that eIF2αS/S cells occupy a terminally dedifferentiated state with highly undifferentiated and pre-EMT characteristics, along with enrichment of neuroendocrine markers. By contrast, eIF2αA/A cells were enriched for AT2 lineage markers (Extended Data Fig. 5f and Supplementary Table 2). Consistent with scRNA-seq (Extended Data Fig. 4f), NKX2-1 expression was lower in eIF2αS/S than in eIF2αA/A cultured cells (Extended Data Fig. 5g). These findings show that cultured LUAD cells retain in vivo tumour states, with ITGA2 marking late-stage populations and ISR activity governing tumour cell identity, dedifferentiation and lineage plasticity. Loss of p-eIF2α or ATF4 impairs mitochondrial function in LUAD cells To assess ISR’s role in mitochondrial function, we analysed cultured eIF2αS/S and eIF2αA/A LUAD cells. Transmission electron microscopy (TEM) showed increased mitochondrial numbers in eIF2αA/A cells (Fig. 5a), consistent with upregulated mitochondrial protein-encoding genes in the Mito-Dysf cluster (Fig. 2g). Despite this, eIF2αA/A cells displayed reduced oxygen consumption rate (OCR), including lower basal, ATP-linked, maximal respiration and spare respiratory capacity, alongside reduced extracellular acidification rate (ECAR), indicating metabolic dysfunction (Fig. 5b). ISR-deficient LUAD cells show mitochondrial dysfunction with reduced CHCHD10. a, TEM images of mitochondria in eIF2αS/S and eIF2αA/A LUAD cells ex vivo, with mitochondrial density quantification (number per μm2). Yellow arrows indicate mitochondria. Scale bars, 2 µm (full cell) and 500 nm (enlarged views). Mean ± s.d.; n = 5 cells per genotype. Two-sided unpaired Student’s t-test; P values indicated. b, Seahorse assay of mitochondrial function in eIF2αS/S and eIF2αA/A LUAD cells. OCR and ECAR were assessed, including basal respiration, ATP-linked respiration, maximal respiration and spare respiratory capacity. Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. c,d, Seahorse assays of mitochondrial function: eIF2αS/S tumour cells transduced with shCtrl or Atf4-targeting shRNA (shATF4_1) (c); eIF2αS/S and eIF2αA/A tumour cells treated with shCtrl or Chchd10-targeting siRNAs (d). Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. e, Clonogenic survival of CHCHD10-deficient LUAD cells. Top: clonogenic assays in eIF2αS/S tumour cells transduced with shCtrl or Chchd10-targeting shRNAs (shC10_1 and shC10_2). Bottom left: colony number as percentage of control. Mean ± s.d.; n = 3 biological replicates. Two-tailed unpaired Student’s t-test; P values indicated. Bottom right: immunoblotting for CHCHD10 and ACTIN (representative of n = 3 biological replicates). f, Ultrastructural analysis of mitochondria. Left: TEM images of eIF2αS/S tumour cells transduced with shCtrl or Chchd10-targeting shRNAs (shC10_1 and shC10_2). Yellow arrows indicate mitochondria. Insets: normal and abnormal morphology. Right: quantification of mitochondrial density (number per μm2) and percentage of abnormal mitochondria. Mean ± s.d.; n = 5 cells per condition. Two-sided unpaired Student’s t-test; P values indicated. Scale bars, 500 nm. g,h, Pathway alterations in Chchd10-deficient LUAD cells: transcriptomic analysis of mouse LUAD cells treated with control or Chchd10 siRNAs, as well as GSEA of differentially expressed genes (‘High in C10 KD’ indicates upregulated pathways in CHCHD10 KD, whereas ‘High in Ctrl’ indicates downregulated pathways) (g); proteomic analysis of eIF2αS/S tumour cells treated with control or Chchd10-targeting shRNAs identifying top altered pathways (h). Data derived from n = 3 biological replicates per condition. FCCP, carbonyl cyanide-p-trifluoromethoxyphenylhydrazone; AA, antimycin A; Rot, rotenone. Source data Immunoblot analysis showed decreased expression of Rieske iron–sulfur protein (RISP), a catalytic subunit of respiratory complex III, and cytochrome c in eIF2αA/A cells (Extended Data Fig. 6a), further supporting mitochondrial impairment. Accordingly, eIF2αA/A cells exhibited reduced mitochondrial membrane potential (Extended Data Fig. 6b) and decreased active mitochondrial mass, assessed by MitoTracker Red staining (Extended Data Fig. 6c,d). The ISR sustains mitochondrial integrity and function in LUAD tumor cells. a, Immunoblot of mitochondrial proteins RISP and cytochrome c in eIF2αS/S and eIF2αA/A LUAD cells, with ACTIN as a loading control. b, TMRM fluorescence in eIF2αS/S and eIF2αA/A cells. Left: Flow cytometry plots. Right: Mean fluorescence intensity relative to control. Mean ± s.d.; n = 4 biological replicates. Two-sided unpaired Student’s t-test; P < 0.0001. c, MitoTracker fluorescence in eIF2αS/S and eIF2αA/A cells. Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P = 0.0080. d, Confocal microscopy of eIF2αS/S and eIF2αA/A cells stained with MitoTracker Red and DAPI. Mitochondrial activity ratio: active mitochondria per nucleus. Scale bars, 10 μm. Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P = 0.0031. e, Clonogenic assays in eIF2αS/S cells transduced with shCtrl or Atf4-targeting shRNAs (shATF4_1/shATF4_2). Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. Bottom left: ATF4 immunoblotting after thapsigargin (1 μM, 2 h). f, TMRM fluorescence in eIF2αS/S cells transduced with shCtrl or Atf4-targeting shRNAs (shATF4_1 and shATF4_2). Left: Flow cytometry histograms. Right: Mean fluorescence intensity relative to control. Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. g, Confocal microscopy of control and Atf4-deficient eIF2αS/S cells stained with MitoTracker Red and DAPI. Scale bars, 20 μm. Mean ± s.d.; n = 4 biological replicates. P values indicated. h, TEM images of mitochondria in eIF2αS/S (n = 68 cells) and eIF2αA/A (n = 83 cells) tumor cells. Right: Mitochondrial cross-sectional area (μm2). Scale bars, 2 μm (full view) and 100 nm (magnified view). Mean ± s.d.; two-sided unpaired Student’s t-test; P < 0.0001. i, Immunoblotting of indicated proteins in eIF2αS/S and eIF2αA/A cultured cells from independent cell preparations. j, Left: Confocal microscopy showing DRP1 (green) and mitochondrial staining (purple) in eIF2αS/S (n = 51 cells) and eIF2αA/A (n = 47 cells) tumor cells. Right: Co-localization (Manders’ coefficient) per cell. Scale bars, 20 μm. Mean ± s.d.; two-sided unpaired Student’s t-test; P < 0.0001. Source data Given the central role of ATF4 in ISR signalling, we examined its contribution to mitochondrial homeostasis. Atf4 knockdown (KD) in p-eIF2α-proficient cells recapitulated the effects of p-eIF2α loss, impairing clonogenic growth (Extended Data Fig. 6e) and reducing OCR, ECAR (Fig. 5c), mitochondrial membrane potential (Extended Data Fig. 6f) and active mitochondrial mass (Extended Data Fig. 6g). These data show that the ISR, through p-eIF2α and ATF4, is required to maintain mitochondrial integrity and metabolic fitness in LUAD cells. ISR maintains mitochondrial homeostasis in LUAD cells To determine how ISR impairment alters mitochondrial homeostasis, we examined mitochondrial morphology, mitophagy markers and dynamics regulators in eIF2αS/S and eIF2αA/A LUAD cells. TEM showed that, although eIF2αA/A cells contain more mitochondria (Fig. 5a), their size was reduced, indicating increased fragmentation (Extended Data Fig. 6h). Levels of the mitophagy markers, including phosphatase and tensin homologue-induced kinase (PINK1) and E3 ubiquitin protein ligase PARKIN were unchanged between genotypes (Extended Data Fig. 6i), arguing against increased mitochondrial turnover. Instead, phosphorylation of dynamin-related protein 1 (DRP1) at S616, a modification that promotes mitochondrial recruitment and fission37, was elevated in eIF2αA/A cells, alongside increased mitochondrial DRP1 immunofluorescence intensity (Extended Data Fig. 6i,j). Together, these data indicate that ISR impairment drives pathological mitochondrial fragmentation through excessive DRP1-mediated fission. To identify ISR effectors involved in mitochondrial regulation, we intersected bulk RNA-seq data from Atf4-KD LUAD cells with the MitoCarta mitochondrial gene database38 and a published ATF4 chromatin immunoprecipitation sequencing (ChIP-seq) dataset39. The analysis identified three mitochondrial-associated genes, coiled-coil-helix-coiled-coil-helix domain containing 10 (Chchd10), aldehyde dehydrogenase 1 family member L2 (Aldh1l2) and glutamate-pyruvate transaminase 2 (Gpt2), emerged as potential ATF4 targets (Extended Data Fig. 7a). Among these, only elevated Chchd10 expression was associated with poor prognosis in KRAS-mutant human LUAD (Extended Data Fig. 7b). CHCHD10 enhances mitochondrial integrity and predicts worse LUAD outcomes. a, Venn diagram showing overlap between ATF4 Chip-seq peaks39 and ATF4-dependent target genes in eIF2αS/S cultured tumor cells, with comparison to the Mito Carta 3.0 mitochondrial gene database38. b, Kaplan–Meier analysis of overall survival in the TCGA KRAS-mutant LUAD cohort (n = 137), stratified into CHCHD10-high (n = 73) and CHCHD10-low (n = 64) groups based on median expression. Two-sided log-rank test; P = 0.048. c, Heatmap showing Chchd10 expression across eIF2αS/S and eIF2αA/A tumor cell clusters. d, IHC staining of CHCHD10 in eIF2αS/S and eIF2αA/A LUAD tumors. Scale bars, 50 μm. Mean ± s.d.; n = 3 mice per group. Two-sided unpaired Student’s t-test; P < 0.0001. e, Immunoblot analysis of CHCHD10 in eIF2αS/S, eIF2αA/A and eIF2αS/S cells transcuded with control (shCtrl) or Atf4-targeting shRNAs (shATF4_1, shATF4_2). f, Confocal microscopy of eIF2αS/S cells transduced with shCtrl or Chchd10-targeting shRNAs (shC10_1 and shC10_2), stained with MitoTracker Red and DAPI. Quantification of mitochondrial activity (MitoTracker intensity normalized to DAPI) is shown. Scale bars, 20 μm. Mean ± s.d.; n = 4 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. g, Flow cytometry analysis of mitochondrial membrane potential using TMRM fluorescence in cells described in (f). Quantification of mean fluorescence intensity relative to control is shown. Mean ± s.d.; n = 3 biological replicates. Two-sided unpaired Student’s t-test; P values indicated. Source data Chchd10 encodes a protein required for inner mitochondrial membrane integrity and respiration40 and was highly expressed in the highly undifferentiated cluster of eIF2αS/S tumours relative to the Mito-Dysf cluster of eIF2αA/A tumours (Extended Data Fig. 7c). CHCHD10 levels were reduced in both p-eIF2α-deficient tumours and cultured eIF2αA/A cells, as well as in ATF4-deficient cells, by IHC and immunoblotting (Extended Data Fig. 7d,e). Chchd10 KD in eIF2αS/S cells reduced OCR and ECAR (Fig. 5d), decreased cell survival (Fig. 5e), diminished active mitochondrial mass (Extended Data Fig. 7f) and membrane potential (Extended Data Fig. 7g), and caused pronounced ultrastructural defects, including cristae fragmentation by TEM (Fig. 5f). Despite increased mitochondrial density, the proportion of morphologically abnormal mitochondria was substantially elevated (Fig. 5f). Transcriptomic profiling showed that CHCHD10 promotes TNF signalling via NFκB, hypoxia response and EMT programmes, while repressing oxidative phosphorylation, peroxisome signalling and differentiation markers (Fig. 5g). Complementary proteomic analysis further implicated CHCHD10 in steroid metabolism, IL-6 pro-inflammatory signalling, mRNA metabolism and mitochondrial oxidative phosphorylation (Fig. 5h). Collectively, these findings identify CHCHD10 as a central mediator of the ISR-dependent mitochondrial homeostasis in LUAD, supporting survival and oncogenic signalling, with elevated expression associated with poor prognosis in KRAS-mutant LUAD. Pharmacological ISR inhibition halts LUAD evolution and reduces tumour growth We assessed the effects of the ISR inhibitor ISRIB, a small molecule that counteracts p-eIF2α-mediated translational repression2, on LUAD progression in mice. After 35 weeks of tumour development, eIF2αS/S mice were treated with ISRIB or vehicle for 4 weeks. ISRIB significantly reduced tumour growth as measured by ultrasound imaging (Fig. 6a). scRNA-seq analysis showed that ISRIB suppressed expansion of lung transition, high-plasticity and pre-EMT clusters (Fig. 6b), while increasing the AT2-like population in residual tumours (Fig. 6b). Flow cytometry confirmed a shift from ITGA2high towards ITGA2low cells following treatment (Extended Data Fig. 8a,b). ISRIB blocks LUAD evolution and primes mitochondrial defects. a, Representative ultrasound images (left) and quantification of tumour size (right) after 30 days of vehicle (control) or ISRIB treatment, initiated 35 weeks post-Cre-lentivirus intubation. Tumours are outlined by dashed yellow lines. Mean ± s.d.; dots represent individual mice (n = 5 per group). Two-sided Mann–Whitney U test; P = 0.0079. b, UMAP showing tumour cells from vehicle- and ISRIB-treated mice (n = 8,580 cells), coloured by annotated cell types. The density overlay indicates the cell distribution, and relative cluster percentages are shown. c, Prioritization of ISRIB-responsive genes using Augur. Dot plot showing top genes driving the response to ISRIB in eIF2αS/S tumours. The x axis represents the mean effect size, with negative and positive values indicating inhibition and promotion by ISRIB, respectively. The dot size indicates the standard error, and colour encodes the permutation P value. Red arrows highlight key genes of interest (Lcn2 and Onecut2). P values from the permutation test (100 label-shuffling iterations) are shown. d, Pathway enrichment analysis of ISRIB-responsive genes. Functional enrichment of genes upregulated or downregulated by ISRIB in high-plasticity and pre-EMT clusters. The bar plot shows the top enriched ontology clusters. The bar length represents −log10P from a hypergeometric test. e, ISRIB inhibits growth of LUAD PDX tumours in vivo. Final tumour volume in immunodeficient NOG mice following subcutaneous transplantation of LUAD PDX tumours treated with vehicle (n = 17) or ISRIB (10 mg kg−1; n = 18) for 8 weeks. Mean ± s.d.; dots represent individual mice. Two-sided unpaired Student’s t-test; P value indicated. Two data points in the vehicle group exceeding the plotted range are not shown. f,g, IHC analysis of ISR pathway markers in PDX tumours following ISRIB treatment. Nuclear ATF4 (f) and cytoplasmic 4E-BP1 (g) in vehicle- and ISRIB-treated tumours after 4 weeks of treatment. Scale bars, 50 μm. Quantification was performed using tumours from n = 4 mice per group. Mean ± s.d.; two-sided unpaired Student’s t-test; P values indicated. Source data ISRIB increases mitochondrial and AT2 markers while reducing dedifferentiation markers. a, ISRIB treatment reduces ITGA2 expression in eIF2αS/S tumors. Left: Representative density plots. Right: Quantification of ITGA2-positive cells in vehicle (control) and ISRIB-treated tumors. The threshold for positivity is indicated by the dashed line. Mean ± s.d.; dots represent individual mice (n = 3 per group). Two-sided unpaired Student’s t-test; P = 0.0079. b, Flow cytometry analysis of ITGA2 expression in primary tumor cells of the indicated genotypes. c, Split violin plots showing expression of Atp5k, Lcn2, Hmga2, and Hnf4a in eIF2αS/S control and ISRIB-treated tumor cells (total n = 8,580 cells). Data are derived from pooled samples of n = 3 mice per group. Inner box plots indicate median and interquartile range; whiskers extend to 1.5 × IQR. d, Heatmap showing mean expression of Onecut2 (top) and Eif4ebp1 (bottom) across clusters in vehicle- and ISRIB-treated tumors. e, Cluster proportions in LUAD tumors of the indicated genotypes following integration of ISR sequencing datasets. Source data At the gene expression level, the mitochondrial marker Atp5k and AT2 marker Lcn2 were upregulated across clusters, while Hmga2 was reduced in pre-EMT and high-plasticity clusters and Hnf4a decreased in hybrid lung/gastric clusters (Extended Data Fig. 8c). Global analysis identified the AT2 marker Lcn2 as one of the most upregulated genes and pre-EMT marker Onecut2 as the most downregulated gene post-treatment (Fig. 6c). Gene Ontology analysis showed enrichment of mitochondrial component genes and apoptosis-related pathways in late-stage clusters and downregulation of EMT, as well as MAPK signalling pathways (Fig. 6d and Supplementary Table 4). The Mito-Dysf cluster was absent from ISRIB-treated samples, probably reflecting incomplete ISR inhibition. Consistently, Onecut2, Hnf4a and Eif4ebp1 expression decreased following ISRIB treatment, indicating effective targeting of late-stage tumour populations (Fig. 6c; Extended Data Fig. 8c,d and Supplementary Table 4). Overall, ISRIB induced a transcriptional shift towards AT2 and mitochondrial programmes and away from pre-EMT and hybrid lung/gastric states stronger than in eIF2αS/A tumours but weaker than in Atf4+/− tumours (Extended Data Fig. 8e). We next evaluated ISRIB in a human LUAD patient-derived xenograft (PDX) harbouring a KRAS G12C mutation, grown in immunodeficient NOD/Shi-scid IL2rγ (null) (NOG) mice (Fig. 6e). Daily oral ISRIB administration for 8 weeks reduced tumour growth by ~30% relative to vehicle controls (Fig. 6e). IHC analysis showed reduced nuclear ATF4 and 4E-BP1 levels (Fig. 6f,g), consistent with suppression of late-stage dedifferentiated states. Overall, the data demonstrate that ISRIB suppresses LUAD progression by limiting ISR-driven lineage plasticity and reducing dedifferentiated tumour populations, while promoting a shift towards a more differentiated, Mito-Dysf-like states in both mouse and human Kras-mutant LUAD models. The ISR marks tumour evolution and represents a therapeutic target in human LUAD Analysis of 539 LUAD samples from The Cancer Genome Atlas (TCGA) and 706 non-mucinous LUAD cases from surgically treated patients showed that ISR-dependent gene expression, together with late-stage programmes, was associated with poorer survival (Fig. 7a,b, Extended Data Fig. 9 and Supplementary Table 5). Conversely, high expressions of AT2-like, AT1-like and lung transition signatures correlated with longer overall survival (Fig. 7a,b, Extended Data Fig. 9 and Supplementary Table 5). This pattern was also observed in KRAS-mutant tumours (Fig. 7c and Supplementary Table 5). ISR-high, dedifferentiated LUAD cells associate with worse patient outcomes. a–c, Prognostic significance of cluster markers in human LUAD cohorts. Hazard ratios (HRs) were calculated using Cox proportional hazards models, treating cluster marker activity scores as continuous variables. Forest plots show HRs with 95% confidence intervals (error bars). TCGA LUAD cohort (n = 539 patients) (a); independent TempO-seq LUAD cohort (n = 706 patients) (b); KRAS-mutant LUAD subcohort (n = 162 patients) (c). Two-sided Wald test; P values indicated. NS, not significant. d, Association of cluster marker activity with histological tumour grades. Box plots show activity scores across clinical grades G1 (n = 41), G2 (n = 388) and G3 (n = 277) in the TempO-seq cohort (n = 706). Boxes indicate the median and IQR; whiskers extend to 1.5× IQR; dots represent outliers. Two-tailed unpaired Student’s t-test; P values indicated (*P < 0.05, **P < 0.01, ***P < 0.001). e,f, UMAP of human LUAD tumour cells: cells coloured by annotated cell types (e); cells coloured by ISR signature intensity (f). The heatmap shows scaled ISR expression across Union for International Cancer Control (UICC) stages (I–IV). The analysis includes 83,439 cells. MSLN, mesothelin; NE, neuroendocrine. g, Model of ISR function in LUAD tumour evolution. LUAD tumours arise from AT2-like cells with low ISR activity. Progressive activation of the p-eIF2α/ATF4 axis drives dedifferentiation towards high-plasticity, highly undifferentiated and pre-EMT states. ATF4 haploinsufficiency (Atf4+/−) restricts cells to AT2-like, AT1-like, club-like and lung transition states. Attenuated p-eIF2α signalling (eIF2αS/A) or ISRIB treatment limits dedifferentiation to a hybrid lung/gastric-like cluster with mitochondrial dysregulation, whereas complete loss (eIF2αA/A) produces a mitochondrially dysfunctional (Mito-Dysf) cluster. ATF4, in cooperation with MYC, promotes late-stage high-plasticity and highly undifferentiated states, with ATF4-dependent CHCHD10 supporting mitochondrial fitness. High ITGA2 marks late-stage LUAD clusters. LUSC, lung squamous cell carcinoma; NSCLC, non-small cell lung cancer. Schematic in f created in BioRender; Zou, J. Y. https://biorender.com/sz2vkvo (2026). Source data Marker genes of ISR-driven dedifferentiated clusters are identified in LUAD patient samples and associated with poor survival outcomes. (a-e) Kaplan-Meier plots of overall survival impact in a cohort of 706 non-mucinous LUAD patients, stratified by expression of ISR (a), high-plasticity (b), highly undifferentiated (c), pre-EMT (d), and Mito-Dysf (e) signatures. Two-sided log-rank test; P values indicated. Dashed lines represent confidence intervals. Across LUAD patient samples, favourable prognosis signatures (AT1-like, AT2-like, club-like and lung transition) declined with increasing tumour grade, while poor prognosis signatures (highly undifferentiated and elevated ISR) increased progressively (Fig. 7d). Analysis of a large-scale lung cancer scRNA-seq atlas comprising 232 patients with non-small cell lung cancer and 86 non-cancer controls41 confirmed elevated ISR activity in stage IV tumours (Fig. 7e,f). These findings demonstrate that elevated ISR activity is associated with poor prognosis in LUAD, particularly in advanced-stage disease, reflecting its link to aggressive, dedifferentiated states. The inverse association between differentiated cell signatures and ISR activity supports the ISR as a key driver of tumour progression and a candidate therapeutic target. Discussion The ISR drives LUAD evolution while suppressing dysfunctional mitochondrial fate Longitudinal single-cell analyses demonstrate that the ISR drives a conserved lineage trajectory from AT2-like cells towards high-plasticity, pre-EMT and highly undifferentiated states (Figs. 2d,h and 7g), consistent with intratumoural heterogeneity and malignant progression. The pre-EMT population displays neuroendocrine-like transcriptional traits (Fig. 1d and Extended Data Fig. 1c), in line with prior reports linking such programmes to histological transformation in lung cancer42. ISR disruption reveals a Mito-Dysf fate marked by impaired metabolic fitness and reduced proliferative potential—a maladaptive state suppressed by ISR-dependent maintenance of mitochondrial integrity through CHCHD10 (Fig. 5d–h and Extended Data Fig. 7). The ISR promotes plasticity while preventing collapse into a dysfunctional state, providing a framework for therapeutic strategies that reroute tumour evolution (Fig. 7g). MYC-ISR cooperation drives LUAD dedifferentiation and late-lineage evolution MYC and ISR signalling are interdependent: MYC suppression abrogates ISR activation and arrests tumours in a low-ISR, early AT2-like state (Fig. 4), while ISR disruption impairs MYC-dependent gene expression (Figs. 1e and 3d). MYC upregulation in the highly undifferentiated population may further amplify ISR activity through the eIF2α kinases GCN2 and PERK43, reinforcing this feedback loop. Poorly differentiated clusters exhibit distinct MYC-dependent programmes. The highly undifferentiated cluster shows the highest MYC activity with upregulated mTOR signalling, enhancing ATF4 expression and MYC-driven translation. By contrast, high-plasticity and pre-EMT clusters rely more on p-eIF2α-dependent translational control under lower MYC and mTOR activity16, consistent with their selective depletion upon p-eIF2α haploinsufficiency (Fig. 3b). ATF4 is essential within this axis: Atf4+/− loss arrests cells in the lung transition cluster blocking progression to undifferentiated, pre-EMT or high-plasticity states (Fig. 3h). The ATF4–MYC axis drives this progression partly through 4E-BP1 upregulation, which suppresses cap-dependent translation to limit proteotoxicity under metabolic stress43; high 4E-BP1 expression associates with poor prognosis in LUAD and across TCGA cancer types44. Integration of single-cell data from EGFR-driven42 and KRAS-mutant LUAD models reveals that small-cell lung cancer precursor cells correspond to the highly undifferentiated cluster and carry elevated ISR expression (Extended Data Fig. 10a,b), consistent with MYC’s requirement for histological transformation in EGFR-driven models42. This suggests that the ISR plays a conserved role in managing MYC-induced stress and supporting dedifferentiation across LUAD subtypes, with potential relevance to small-cell lung cancer transformation. The MYC–ISR interdependence thus represents a therapeutically exploitable vulnerability whereby ISR targeting may attenuate MYC-driven plasticity. ISR signature defines histological plasticity and therapy resistance across distinct LUAD models. a, UMAP integrating the mouse ISR-LUAD model from this study with an EGFR-driven LUAD histological transformation model42 (total n = 25,828 cells). Cells are colored by annotated cell types from this study (left) and the EGFR-driven model (right). b, UMAP of integrated ISR-LUAD and EGFR-driven LUAD datasets42, colored by ISR signature expression. c, Heatmap showing column-scaled ISR signature expression across epithelial cell states from EGFR-mutant lung tumors during EGFR withdrawal. Rows correspond to AT2 cells, LUAD, residual LUAD, and SCLC-like states. Color intensity reflects relative ISR signature expression within each state. d, ISR signature across annotated epithelial tumor cell states from time-resolved single-cell profiling of KRAS G12C-mutant lung tumors treated with multiple RAS inhibitor (MultiRASi). Datasets include tumors treated with the pan-RAS(ON) inhibitor RMC-7977 in combination with the RAS(ON) G12C-selective inhibitor RMC-4998, as well as tumors treated with RMC-7977, RMC-4998, and the SHP2 inhibitor RMC-455048. Dot color represents scaled ISR-dependent gene expression, and dot size indicates the fraction of cells expressing ISR signature. Mitochondrial dysfunction as a barrier to LUAD tumour evolution In eIF2αA/A tumours, cells accumulate in the Mito-Dysf state with elevated Nkx2-1 expression resembling the lung transition cluster (Fig. 1d), indicating failure to complete the dedifferentiation required for high-plasticity state11. The Mito-Dysf cluster is characterized by upregulation of mitochondrial ribosomal biogenesis and oxidative phosphorylation genes (Figs. 1e and 2g), reflecting a compensatory yet maladaptive metabolic response. Although how the ISR suppresses Nkx2-1 remains unclear, HIPPO signalling regulated by the ISR during endoplasmic reticulum stress45 may modulate NKX2-1 activity through YAP and TAZ46, representing a potential mechanistic link. Loss of p-eIF2α increases mitochondrial number yet paradoxically impairs function, disrupting membrane potential and respiratory activity (Fig. 5a). CHCHD10 mediates ISR-driven mitochondrial fitness in this context40 (Fig. 5d–h and Extended Data Fig. 7), supporting oncogenic TNF–NFκB, EMT and IL-6 signalling in LUAD cells (Fig. 5g,h). Increased mitochondrial abundance in p-eIF2α-deficient cells reflects enhanced fission (Fig. 5a and Extended Data Fig. 6h–j), consistent with PERK-mediated mitochondrial hyperfusion and ISR-dependent mitochondrial homeostasis during ER stress47. Collectively, ISR loss in LUAD triggers excessive fission and respiratory failure, trapping tumours in a non-evolving Mito-Dysf fate. The dual role of the ISR, driving plasticity through MYC-ATF4 while safeguarding mitochondrial integrity through CHCHD10, suggests that its inhibition may both limit tumour evolution and expose a mitochondrial vulnerability. Targeting ISR as a strategy to impair LUAD progression Cell plasticity is a major driver of therapy resistance in LUAD11, and ISR-driven dedifferentiation promotes resistant subpopulations in both mouse models and human tumours. Analysis of independent single-cell datasets under targeted therapy pressure revealed enrichment of ISR activity in persisting subpopulations under EGFR inhibition42 (Extended Data Fig. 10c) and mutant KRAS-targeted treatment48 (Extended Data Fig. 10d). These ISR-high cells retain adenocarcinoma identity while exhibiting heightened plasticity and pre-EMT features, with full lineage conversion occurring only after prolonged selective pressure. We propose a two-stage model in which ISR-driven plasticity first generates dedifferentiated, lineage-primed intermediates, represented by the Nkx2-1-low/ISR-high state, followed by secondary events enabling full phenotypic conversion. This model predicts that ISR inhibition preferentially disrupts these drug-tolerant intermediates before lineage conversion. Consistent with this, ISRIB reduces LUAD tumour growth6, limits resistant subpopulations and shifts cells towards less tumorigenic states in both mouse and PDX models with enhanced efficacy in immune-competent settings, reflecting ISR-immune regulatory links49 (Fig. 6). ISR inhibition also suppresses Onecut2 and Hnf4a (Fig. 6c and Extended Data Fig. 8c,d), key pre-EMT markers that persist under mutant KRAS inhibitor treatment48. Given the dependence of KRAS-driven LUAD on mitochondrial metabolism50, disruption of mitochondrial fitness provides an additional mechanism to limit intratumoural heterogeneity. Future directions merit further investigation. First, although ISR–ATF4 signalling is required in KRAS-driven LUAD, the temporal window during which it influences early tumourigenesis remains unclear; perturbation at pre-neoplastic stages will help define its role in initiation and lineage commitment. Second, our conclusions are based on loss-of-function approaches; complementary gain-of-function strategies, such as phosphomimetic eIF2α variants or alternative pharmacological ISR activation (for example, eIF2 phosphatase inhibitors), will refine the conditions driving these states. Third, while our analyses focus on LUAD, the broader relevance of ISR–ATF4 signalling across cancers remains to be established; extending these studies to diverse cohorts and multimodal clinical datasets will clarify its role across distinct oncogenic contexts. Methods Ethics statement This study was conducted in compliance with all relevant ethical regulations for animal research. Mouse studies performed at the Lady Davis Institute were approved by the Animal Care Committee of the Lady Davis Institute (McGill University) (protocol #JGH-5754) and conducted in accordance with institutional guidelines. Experiments involving KM mice at The Francis Crick Institute were performed in accordance with UK Home Office regulations under project licence PP9124501 issued to G.I.E. PDX models were obtained from The Jackson Laboratory (JAX) PDX repository. All patient tumour samples were collected by contributing institutions under institutional review board-approved protocols with informed consent from participants. Samples were de-identified before receipt by JAX, and all data used in this study are fully anonymized. Access to detailed patient-level metadata is restricted to protect participant confidentiality and is managed by the originating institutions. The Leicester Archival Thoracic Tumour Investigatory Cohort—Adenocarcinoma cohort was originally constructed under REC 14/EM/1159 (East Midlands Research Ethics Committee), and the ongoing management of the resource by the National Health Service Greater Glasgow and Clyde Biorepository was approved under an amendment granted by the Leicester South REC. Ongoing use of the collection is now managed under REC 16/WS/0207. Animal studies All animal experiments were performed in mice. Unless otherwise indicated, experiments used male and female mice on a C57BL/6 background. Autochthonous lung tumour studies were performed in age- and sex-matched littermates, and mice were 8–12 weeks old at the time of tumour induction. Mice were randomly assigned to experimental groups. Exact sample sizes for each experiment are provided in the corresponding figure legends. Mouse strains, alleles, genetic background and source information are described below. Transgenic mouse models and treatments Mice on a C57BL/6 background were used in all experiments. For lung tumour induction, 2-month-old male and female mice were randomly assigned to experimental groups. Mice were maintained in ventilated cages under controlled conditions (30% humidity, 12-h light–dark cycle, lights on from 06:00 to 18:00, 23 °C). Lung tumourigenesis was induced in KRAS+/LSL-G12D;fTg/0;eIF2αS/S and KRAS+/LSL-G12D;fTg/0;eIF2αA/A mice on C57BL/6 background6 via intratracheal intubation with lentiviruses expressing Cre recombinase and TP53 shRNA under the control of U6/H1 promoter6. Lung tumour development was monitored using ultrasound imaging on a VisualSonics VEVO 3100 high-frequency ultrasound system51. ISRIB treatment was administered daily by oral gavage at a dose of 10 mg kg−1 (ref. 6) using a solution containing 0.5% hydroxypropylmethylcellulose and 0.1% Tween 80 (pH 4.0) for the indicated duration. To impair ATF4 in mouse LUAD tumours, KRAS+/LSL-G12D;fTg/0;eIF2αS/S mice were crossed with ATF4f/f mice on a C57BL/6 background43. KM mice were maintained on regular diet in a pathogen-free facility under a 12-h light/dark cycle (lights on from 07:00 to 19:00), in individually ventilated cages at 23 °C and 61% humidity, with continuous access to food, water and cage enrichment. Krastm4Tyj (LSL-KrasG12D)52 and Gt(ROSA)26Sortm1(MYC/ERT2)Gev (LSL-Rosa26MIE/MIE)33 mice have been described previously. For all experiments, the mice were kept with a heterozygous Lsl-KrasG12D allele and homozygous LSL-Rosa26MIE/MIE alleles. Age and sex-matched littermate mice were divided equally over the treatment groups. Mouse weight, health and body condition were monitored daily during tamoxifen treatment. For activation of MycERT2, tamoxifen (Sigma; TS648) dissolved in peanut oil (Sigma; P2144) was administered daily by intraperitoneal injection for a maximum of 2 weeks at a dose of 1 mg per 25 g body mass, or by tamoxifen diet (Harlan Laboratories UK, TAM400 diet) for a maximum of 4 weeks as previously described53. Deactivation of MycERT2 was achieved by switching mice from a tamoxifen diet back to a regular diet. To deliver adenovirus-Cre recombinase (AdV-Cre), 8–12-week-old mice were anaesthetized with isoflurane (Covetrus, Isofane 250 ml UK, 2800025; 100% w/w inhalation vapour with 0.2% oxygen), and 5 × 107 plaque-forming units of AdenoCre (Ad5CMVCre, Viral Vector Core, University of Iowa) were administered as described previously54. PDX assay The PDX tumour model of LUAD carrying KRAS G12C was obtained from JAX (tumour model: TM00186 LG0418F). The PDX was amplified by subcutaneous transplantations in NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice according to the provider’s specifications. Tumour growth in mice was measured with digital calipers two times per week, and the volume was calculated by the following formula: tumour volume [mm3] = [(length [mm]) × (width [mm])2]/2. ISRIB treatment was administered daily by oral gavage, following the same dosing regimen as in the autochthonous model, starting 5 weeks after tumour implantation. Tumour monitoring was terminated when tumours reached a maximum size of approximately 2.0 cm (20 mm) in any dimension. Tumour measurements remained within the limits approved by the Animal Welfare Committee of McGill University for all animal experiments. Immunoblotting Cells were washed twice with ice-cold phosphate-buffered saline (PBS), and proteins were extracted in ice-cold radioimmunoprecipitation assay buffer containing 10 mM Tris–HCl, pH 7.5, 50 mM KCl, 2 mM MgCl2, 1% Triton X-100, 3 μg ml−1 aprotinin, 1 μg ml−1 pepstatin, 1 μg ml−1 leupeptin, 1 mM dithiothreitol, 0.1 mM Na3VO4 and 1 mM phenylmethylsulfonyl fluoride. Cell lysates were kept on ice for 15 min, then centrifuged at 13,200g for 15 min at 4 °C, and supernatants were stored at −80 °C. Protein concentrations were measured using the Bradford assay (Bio-Rad). Protein expression was analysed by immunoblotting after loading 50 µg of the protein extracts from the same set of samples on 10% sodium dodecyl sulfate–polyacrylamide gels. Following electrophoresis and transfer to Immobilon-P membrane (Millipore), membranes were blocked for 1 h at room temperature in 5% skimmed milk before incubation with antibodies against phospho-eIF2α (Abcam, clone E90, cat. no. Ab32157, 1:1,000), eIF2α (Cell Signaling Technology, clone L57A5, cat. no. 2103, 1:1,000), ATF4 (Cell Signaling Technologies, cat. no. 11815S, 1:1,000), CHCHD10 (Proteintech, cat. no. 25671-1-AP, 1:1,000), NKX2-1 (Abcam, clone EP1584Y, cat. no. ab76013, 1:2,000), Cytochrome c (Cell Signaling Technologies, clone D18C7, cat. no. 11940, 1:1,000), UQCRFS1/RISP (Cell Signaling Technologies, cat. no. 95231, 1:1,000), TUBULIN (Sigma-Aldrich, cat. no. T5168, 1:4,000), ACTIN (MP Biomedicals, cat. no. ICN691001, 1:4,000), DRP1 (Cell Signaling Technologies, cat. no. 8570, 1:1,000), phospho-DRP1 (Ser616) (Cell Signaling Technologies, clone D9A1, cat. no. 4494, 1:1,000), PARK2/Parkin (Proteintech, cat. no. 14060-1-AP, 1:1,000) and PINK1 (Proteintech, cat. no. 23274-1-AP, 1:1,000). Membranes were then incubated with the corresponding secondary antibody: mouse IgG-horseradish peroxidase (HRP)-conjugated secondary antibody (KPL, cat. no. 474-1806, 1:2,000) for TUBULIN and ACTIN and rabbit IgG-HRP-conjugated secondary antibody (Cell Signaling Technologies, cat. no. 7074, 1:1,000) for all other primary antibodies. Protein signals were visualized by enhanced chemiluminescence (Thermo Fisher Scientific, cat. no. 32106) according to the manufacturer’s instructions. TEM Cells were plated 48 h before collection to achieve 80–90% confluency on the day of collection. Cells were detached using PBS–EDTA and centrifuged at 300g for 10 min. After removal of the supernatant, electron microscopy samples were fixed overnight at 4 °C in 2.5% glutaraldehyde in 0.2 M cacodylate buffer (pH 7.2). Following fixation, samples were washed three times in cacodylate buffer supplemented with 3% sucrose. Samples were then post-fixed for 2 h at 21 °C in 1.33% osmium tetroxide prepared in s-collidine buffer supplemented with sucrose. After post-fixation, samples were washed in distilled water and counterstained with 1% uranyl acetate for 30 min protected from light. Dehydration was performed through a graded acetone series (25%, 50%, 75%, 95%, 100% and 100%) for 15–30 min per step at room temperature. Samples were then infiltrated overnight in a 1:1 mixture of acetone and SPURR resin, followed by incubation in pure SPURR resin for 3 h and then for an additional 2 h at room temperature. Samples were transferred to BEEM capsules containing fresh SPURR resin and polymerized at 60 °C for 24–48 h. After cooling, blocks were demoulded, and ultrathin sections were prepared using an ultramicrotome. Sections were collected on Formvar-carbon 200 mesh copper grids (Sigma-Aldrich), stained with uranyl acetate and lead citrate, and examined using a Hitachi H-7100 transmission electron microscope equipped with AMT Image Capture Engine software (version 600.147) at the INRS-CAFSB platform. Mitochondrial function assays Mitochondrial membrane potential was assessed using tetramethylrhodamine methyl ester (TMRM; Thermo Fisher Scientific, cat. no. M20036) under non-quenching conditions by flow cytometry. Cells were incubated with TMRM at a final concentration of 20 nM for 30 min, washed with PBS and analysed on a Sony Spectral Cell Analyzer ID7000. 4′,6-Diamidino-2-phenylindole (DAPI; Roche, cat. no. 10236276001) was added immediately before acquisition to exclude dead cells. The mean fluorescence intensity was quantified using FlowJo software (v10.10.0) and normalized to vehicle control. Mitochondrial content was assessed using MitoTracker Red CMXRos (Thermo Fisher Scientific, cat. no. M7512) for both flow cytometry and confocal microscopy. Cells were incubated with MitoTracker Red at a final concentration of 10 nM for 45 min. For flow cytometry, stained cells were analysed immediately following the addition of DAPI (Roche, cat. no. 10236276001) to exclude dead cells, and mean fluorescence intensity was used as a measure of mitochondrial content. For confocal microscopy, cells were fixed in acetone for 5 min at room temperature following staining, counterstained with DAPI (Roche, cat. no. 10236276001) to visualize nuclei, and imaged using a Zeiss LSM 800 confocal microscope to assess mitochondrial distribution and morphology. Gating strategies for mitochondrial membrane potential and mitochondrial content analyses are shown in Supplementary Figs. 3–6. Mitochondrial respiration and glycolytic activity were assessed using a Seahorse XFe96 Analyzer (Agilent Technologies). Tumour cells were seeded in Seahorse XFe96 cell culture plates at a density of 4,000 cells per well 48 h before the assay in Seahorse XF assay medium supplemented with 2 mM L-glutamine, 25 mM glucose and 1 mM sodium pyruvate. Baseline OCR and ECAR were measured three times before sequential injection of mitochondrial inhibitors: oligomycin (1 µM, Sigma-Aldrich, cat. no. O4876), carbonyl cyanide-p-trifluoromethoxyphenylhydrazone (FCCP, 2 µM, Sigma-Aldrich, cat. no. C2920) and a combination of antimycin A (0.5 µM, Sigma-Aldrich, cat. no. A8674) and rotenone (0.5 µM, Sigma-Aldrich, cat. no. R8875). Three measurements were recorded following each injection. Basal respiration was calculated as the final basal OCR measurement minus non-mitochondrial respiration following antimycin A/rotenone treatment. ATP-linked respiration was determined by the decrease in OCR following oligomycin injection, and spare respiratory capacity was defined as the increase in OCR after FCCP injection relative to basal respiration. All measurements were normalized to protein content using a sulforhodamine B (SRB; Sigma-Aldrich, cat. no. S1402) assay. OCR and ECAR data were analysed using Agilent Seahorse Analytics software and GraphPad Prism (v10.3.1). Immunofluorescence and confocal microscopy For assessment of DRP1 colocalization with mitochondria, cells were stained with MitoTracker Deep Red (Thermo Fisher Scientific, cat. no. M22426) according to the manufacturer’s instructions. Cells were then fixed in acetone for 5 min at room temperature, permeabilized with 0.2% Triton X-100 and blocked in 1% bovine serum albumin for 30 min before antibody staining. Samples were incubated with anti-DRP1 primary antibody (Cell Signaling Technology, cat. no. 8570; 1:50), followed by a donkey anti-rabbit IgG (H + L), highly cross-adsorbed Alexa Fluor 568–conjugated secondary antibody (Invitrogen, cat. no. A10042; 1:1,000). Nuclei were counterstained with DAPI (Roche, cat. no. 10236276001). Imaging was performed on a Nikon AX/R confocal microscope equipped with the NSPARC image scanning microscopy module. Images were acquired using a 60× oil immersion objective (numerical aperture 1.42) under Nyquist sampling conditions, and z-stacks were collected for each field of view. Raw images were preprocessed using background subtraction and deconvolution in NIS-Elements before quantitative analysis. Colocalization was quantified in NIS-Elements using Manders’ coefficients (M1 and M2). M1, representing the fraction of total DRP1 signal localized to mitochondria, was used for analysis as a measure of DRP1 recruitment. Cells were excluded from analysis if they exhibited signs of cell death or were not fully contained within the field of view. Immunohistochemistry Mouse lung tissues were fixed in 10% neutral-buffered formalin, paraffin-embedded and sectioned at 4 µm. Sections were mounted on TOMO slides (VWR), dried overnight at 37 °C and processed for immunohistochemistry using a Discovery XT Autostainer (Ventana Medical Systems). Unless otherwise specified, all reagents were from Ventana Medical Systems (Roche Tissue Diagnostics). Slides underwent automated deparaffinization and heat-induced epitope retrieval using CC1 prediluted solution (standard retrieval; Roche Tissue Diagnostics, cat. no. 06414575001). Immunostaining for ONECUT2, EIF4EBP1, CHCHD10, NKX2-1 and ATF4 was performed using rabbit primary antibodies: anti-ONECUT2 (Proteintech, cat. no. 21916-1-AP; 1:100), anti-4EBP1 (Cell Signaling Technology, cat. no. 9644T; 1:8000), anti-CHCHD10 (Proteintech, cat. no. 25671-1-AP; 1:200), anti-NKX2-1 (Abcam, clone EP1584Y, cat. no. ab76013; 1:50) and anti-ATF4 (Cell Signaling Technology, cat. no. 11815S; 1:75), diluted in antibody diluent (Roche Tissue Diagnostics, cat. no. 06440002001) and applied for 32 min at 37 °C. Detection was performed using OmniMap anti-rabbit HRP (Roche Tissue Diagnostics, cat. no. 05269679001) for 8 min, followed by ChromoMap DAB (Roche Tissue Diagnostics, cat. no. 05266645001). Omission of the primary antibody served as a negative control. Slides were counterstained with hematoxylin (Roche Tissue Diagnostics, cat. no. 05266726001) for 12 min, blued with bluing reagent (Roche Tissue Diagnostics, cat. no. 05266769001) for 4 min, removed from the autostainer, washed in tap water, dehydrated through graded alcohols, cleared in xylene and mounted with Eukitt mounting medium (Electron Microscopy Sciences, cat. no. 15320). Sections were analysed by brightfield microscopy or scanned using an Aperio AT Turbo scanner (Leica Biosystems). Quantification was performed using QuPath (v0.5.0). For MRK mice, animals were euthanized and perfused transcardially with PBS. Lungs were collected, fixed overnight in 10% neutral-buffered formalin (Sigma-Aldrich, cat. no. 501320) and paraffin-embedded. Sections (4 µm) were deparaffinized, rehydrated and subjected to antigen retrieval by boiling in 10 mM citrate buffer (pH 6.0) for 10 min. Sections were incubated overnight at 4 °C with anti-4EBP1 antibody (Cell Signaling Technology, cat. no. 9644T; 1:5,000). Images were acquired using a Zeiss Axio Imager.M2 microscope with Zen 3.2 (blue edition) software. Mass spectrometry-based proteomics Frozen cell samples were transferred to 2-ml Eppendorf tubes and lysed in extraction buffer (1% n-dodecyl-β-D-maltoside in 8 M urea) supplemented with protease inhibitor cocktail. A single 7-mm stainless-steel bead was added to each tube, and samples were homogenized for 3 min at 50 oscillations s−1 using a TissueLyser LT (Qiagen). Lysates were centrifuged, and supernatants were transferred to fresh tubes. Proteins were reduced by adding 10 µl of 0.5 M tris(2-carboxyethyl)phosphine and incubating for 1 h at 37 °C with agitation. Alkylation was performed by adding chloroacetamide to a final concentration of 55 mM, followed by an additional 1-h incubation at 37 °C. Proteins were digested with 1 µg of trypsin for 8 h at 37 °C. Peptides were dried and resuspended in 4% formic acid. Peptides were loaded onto a PepMap Neo C18 precolumn (Thermo Fisher Scientific) and separated on a 25-cm Aurora C18 column (IonOpticks; 75 µm × 250 mm) using a 56-min linear gradient from 10% to 30% acetonitrile in 0.2% formic acid at a flow rate of 300 nl min−1 on a Vanquish Neo UHPLC system coupled to an Orbitrap Ascend Tribrid mass spectrometer (Thermo Fisher Scientific). Full MS spectra were acquired at a resolution of 120,000 with a maximum injection time of 50 ms. Data-independent acquisition was performed using 34 windows of 7 m/z across a mass range of 400–800 m/z. Fragmentation was carried out using higher-energy collisional dissociation with a normalized collision energy of 25%. Tandem mass spectrometry spectra were acquired at a resolution of 45,000 over an m/z range of 200–1,200, with an AGC target of 1,000% and a maximum injection time of 59 ms. Raw data were processed using PEAKS Studio v12 (Bioinformatics Solutions) and searched against a concatenated UniProt mouse protein database. Precursor and fragment mass tolerances were set to 10 ppm and 0.01 Da, respectively. Carbamidomethylation of cysteine residues was set as a fixed modification, while oxidation of methionine and deamidation of asparagine and glutamine were specified as variable modifications. Quantitative proteomics data were analysed using the DEP2 package following a standard workflow55. Isolation of mouse LUAD cells Mice bearing LUAD tumours were euthanized 35 weeks after tumour induction. Lungs were collected and dissociated in RPMI 1640 (Wisent, cat. no. 350700CL) containing Dispase II (Gibco, cat. no. 17105041; 0.6 U ml−1), Collagenase Type IV (Gibco, cat. no. 17104019; 0.083 U ml−1) and DNase I (Thermo Fisher Scientific, cat. no. EN0521; final concentration 10 U ml−1) for 60 min at 37 °C. Dissociated cells were filtered through a 100-µm cell strainer and centrifuged at 300g for 5 min at 4 °C. Red blood cell lysis was performed using ACK Lysing Buffer (Gibco, cat. no. A1049201). Cells were washed, pelleted at 300g for 5 min at 4 °C and resuspended in fluorescence-activated cell sorting (FACS) buffer (PBS supplemented with 2% fetal bovine serum and 2 mM EDTA). Cells were incubated for 20 min at 4 °C with the following antibodies: BUV395 rat anti-mouse CD45 (BD Biosciences, cat. no. 564279, 1:200), PE anti-mouse lineage cocktail (BioLegend, cat. no. 133303, 1:200), APC anti-mouse CD326 (EpCAM) (BioLegend, cat. no. 118214, 1:200) and APC/Cy7 anti-mouse CD49b (BioLegend, cat. no. 108920, 1:200). Cells were washed twice with ice-cold PBS, resuspended in FACS buffer and filtered through a 70-µm strainer. DAPI (Roche, cat. no. 10236276001) was added immediately before acquisition to exclude dead cells. Samples were analysed on a BD FACSAria Fusion flow cytometer. Data were acquired using FACSDiva software and analysed with FlowJo (v10.10.0). Detailed gating strategies for cell isolation are provided in Supplementary Figs. 1, 2, 7 and 8. For KM mice, lungs were collected and processed into single-cell suspensions using a mouse tumour dissociation kit (Miltenyi Biotec) and a gentleMACS Octo Dissociator with Heaters, according to the manufacturer’s instructions. In brief, lung tissues were minced into ~1-mm2 fragments and enzymatically dissociated. Red blood cells were lysed using Red Blood Cell Lysis Solution (Miltenyi Biotec), and cell suspensions were filtered through 70-µm strainers before counting. Cell lines Primary KRAS G12D eIF2αS/S and eIF2αA/A lung tumour cells6 were maintained in RPMI 1640 supplemented with 10% fetal bovine serum (Gibco, cat. no. 12483020), 100 U ml−1 penicillin/streptomycin (Wisent, cat. no. 420-201-EL), 0.075% sodium bicarbonate (NaHCO3; Gibco, cat. no. 25080-094), 1× essential amino acids (Gibco, cat. no. 11130-51) and 1× non-essential amino acids (Gibco, cat. no. 11140-050). shRNA/siRNA-mediated silencing Stable cell pools expressing shRNAs targeting Atf4 and Chchd10 were generated by pLKO.1 lentiviral infection followed by selection with 2.5 µg ml−1 puromycin56. The shRNA sequences used were as follows: shATF4 #1, 5′-CCGGCCAGAGCATTCCTTTAGTTTACTCGAGTAAACTAAA GGAATGCTCTGGTTTTTG-3′; shATF4 #2, 5′-CCGGCGGACAAAGATACCTTCGAGTCTCGA GACTCGAAGGTATCTTTGTCCGTTTTTG-3′; shChchd10 #1, 5′-CCGGCTCAAACAGTGCAA ATACAATCTCGAGATTGTATTTGCACTGTTTGAGTTTTTTG-3′; and shChchd10 #2, 5′-CCG GCAGTGCAAATACAATCACGGTCTCGAGACCGTGATTGTATTTGCACTGTTTTTTG-3′. Small interfering RNA (siRNA)-mediated KD of Chdchd10 was performed using four siRNAs (Horizon Discovery) with the following sequences: 5′-CCUAUGAGAUCAAGCAGUU-3′; 5′-GCAAAUACA AUCACGGUCU-3′; 5′-GCUCAGCUGUAGGGCAUGU-3′; and 5′-GCGACCUAACCCUGUGUG A-3′. A total of 2.5 × 105 cells were seeded in 60-mm tissue culture plates and transfected with 200 pmol siRNA targeting Chchd10 or scrambled control siRNA (Horizon Discovery) using Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instructions. Medium was replaced after 4–6 h, and cells were incubated at 37 °C for 24 h before downstream analyses. For colony formation assays, 2 × 103 cells were seeded and cultured for 7 days. Cells were fixed with 4% (v/v) paraformaldehyde and stained with 0.2% (w/v) crystal violet (Millipore Sigma). Colonies were quantified using an automated colony counter (GelCount; Oxford Optronix). Parallel samples were collected for validation of Chchd10 KD by immunoblotting. Bulk RNA-seq analysis Total RNA of KRAS G12D eIF2αS/S and eIF2αA/A cells (n = 4 per group) was extracted using the RNeasy kit (Qiagen) according to the manufacturer’s instructions. RNA-seq libraries were prepared using the TruSeq Stranded Total RNA Library Prep Kit (Illumina) and sequenced to generate 50-bp single-end reads on a HiSeq 2500 system (Illumina) in rapid mode. Raw FASTQ files were aligned to the mouse reference genome (GRCm38/mm10) using HISAT2. Gene-level count matrices were generated using featureCounts. Differential gene expression analysis was performed using DESeq2 (v1.24.0). Genes with fewer than ten counts in at least two samples were excluded from analysis. Differentially expressed genes were defined as those with an absolute log2 fold-change > 1 and a false discovery rate < 0.05. GSEA (v4.0.3, Broad Institute) was performed on all genes ranked by fold change (FC), using Gene Ontology gene sets v5.2 (MSigDB)57–59. A total of 1,000 permutations were used, and gene sets containing between 15 and 500 genes were considered. cDNA Library Preparation Single-cell libraries were prepared using the Chromium Next GEM Single Cell 3′ Kit v3.1 (10x Genomics) according to the manufacturer’s instructions. In brief, sorted cells were manually counted and loaded onto the Chromium Controller (10x Genomics) with a target recovery of 8,000 cells per sample. Reverse transcription, cDNA amplification, and library construction were performed following the manufacturer’s protocol. cDNA and library quality were assessed using a 2100 Bioanalyzer (Agilent Technologies). Final libraries were sequenced as paired-end 100-bp reads on a NovaSeq platform (Illumina). scRNA-seq FASTQ files from scRNA-seq experiments generated using the 10x Genomics Chromium platform were processed using Cell Ranger (v5.0.0). Reads were aligned to the mouse reference genome (GRCm38/mm10), and gene-cell count matrices were generated. Downstream analysis of tumour datasets was performed using the Scanpy (v1.11.0)/AnnData framework60 with custom scripts. Cells with fewer than 500 UMIs, greater than 20% mitochondrial reads, or low gene complexity were excluded. Doublets were identified and removed using Scrublet61. Highly variable genes were selected using a variance-stabilizing approach. Data were normalized and log-transformed, followed by principal component analysis (PCA). Batch correction was performed using bbknn62. Dimensionality reduction and clustering were performed using UMAP and the Leiden algorithm. To avoid treatment-driven bias in cell state assignment, unsupervised clustering was performed using control cells only (eIF2αS/S or vehicle-treated). A logistic regression classifier was trained on these clusters and applied to assign all cells to defined cell states. This approach ensured consistency with previously defined states in mouse KP lung tumours11. Normalized expression values were MaxAbs-scaled before gene signature scoring using the Scanpy score genes function. Gene signatures were derived from MSigDB57,58,59, and previously defined KP tumour signatures11. Bhattacharyya distances were computed using the distances Python package (v1.5.6). Functional enrichment analysis was performed using Metascape63, incorporating Gene Ontology, Kyoto Encyclopedia of Genes and Genomes (KEGG), and Reactome pathways (P < 0.01). Detailed results are provided in Supplementary Table 1. For KM LUAD tumours, single-cell suspensions were prepared as described above. Approximately 6,000 cells per sample were processed using the 10x Genomics Chromium Single Cell 3′ v3 kit and sequenced on a NovaSeq 6000 platform. Data were processed using Cell Ranger (v6.1.1). Downstream analysis was performed using Seurat (v5.0.0) in R (v4.3.1)64. Cells with fewer than 200 detected genes or greater than 25% mitochondrial reads were excluded. Data integration was performed using Harmony (v1.2.4)65. Dimensionality reduction was conducted using PCA (top 50 components), followed by UMAP visualization. Clustering was performed using the FindClusters function (resolution = 0.5). Cell type annotation was performed using the decoupler Python package66 with curated marker gene sets. Differential gene expression analysis was conducted using the glmGamPoi package (v1.12.2)67. Samples with fewer than 100 tumour cells after quality control were excluded from downstream analyses. Analysis of gene expressions along the LUAD evolution trajectory Cell state transitions were inferred using CellRank23. Gene expression dynamics along trajectories from AT2-like to pre-EMT or Mito-Dysf states were modeled using a generalized additive model (GAM) implemented in CellRank. Imputed gene expression values were used for visualization and normalized to a range of [0,1]. Genes were ordered based on the position of their peak expression along the terminal state probability continuum. Gene expression trends were visualized using built-in CellRank plotting functions. A list of genes associated with trajectory dynamics is provided in Supplementary Table 2. Transcription factor regulatory module (regulon) analysis Regulatory network analysis was performed using pySCENIC workflow31 to identify transcription factor regulons with default parameters and reference databases (mm10 transcription factor targets and 500 bp upstream to 100 bp downstream gene–motif rankings). In addition to identifying condition-specific regulons, cluster-associated regulons were determined based on Pearson correlation between regulon activity scores and cluster membership. Analysis of public datasets Processed scRNA-seq datasets from human LUAD tumours were obtained from the CZI Cellxgene portal41. Analysis was restricted to cells annotated as malignant by the original study authors. For murine datasets, processed scRNA-seq profiles were obtained from GSE12233268, and analysis was limited to alveolar type II (AT2) and alveolar type I (AT1) cells as annotated in the original study. Human LUAD survival analysis TempO-seq targeted RNA-seq was performed on 706 non-mucinous LUAD cores derived from 23 FFPE tissue microarray (TMA) slides. Samples were scraped and lysed using TempO-seq 2× lysis buffer. Libraries were generated using the Human Whole Transcriptome v2.1 panel with standard attenuators and sequenced according to the manufacturer’s instructions. Raw sequencing data were processed using the TempO-seq RTM software package. Reads passing quality control were aligned using the STAR algorithm (v2.7.2a)69 to a pseudo-transcriptome corresponding to the targeted gene panel. Gene isoforms were collapsed at the gene level, and genes detected in fewer than one-third of samples were excluded. Expression data were normalized using the upper-quartile method followed by log2 transformation. Gene signature scores were computed using single-sample GSEA (ssGSEA) implemented in the GSVA R package (v1.51-17)70, based on genes meeting differential expression criteria (false discovery rate <0.05 and logFC >1.5). Cox proportional hazards regression was performed using the survival R package (v3.6-4) to assess associations with patient outcomes, and HRs with 95% confidence intervals were reported. Driver mutation status was available for 469 patients, including 180 with KRAS mutations. Samples were stratified into high and low p-eIF2α groups using the median H-score as a cut-off. Statistics and reproducibility Statistical analyses were performed using R (v4.3.0), QuPath (v0.6.0) and GraphPad Prism (v10), as specified in the relevant Methods sections. All P values were two-sided. For comparisons between two independent groups, statistical significance was assessed using either a two-tailed unpaired Student’s t-test or a two-sided Mann–Whitney U test, as indicated in the corresponding figure legends. Correlations between continuous variables were evaluated using two-sided Spearman’s rank correlation tests. Survival analyses were performed using Cox proportional hazards regression models with two-sided Wald tests, or Kaplan–Meier analysis with two-sided log-rank tests, as indicated in the corresponding figures. No statistical methods were used to predetermine sample sizes; however, sample sizes were similar to those reported in previous studies6. Data distribution was assumed to be normal but was not formally tested. Only exclusions explicitly described in the relevant Methods sections (for example, quality-control filtering for scRNA-seq datasets) were applied; otherwise, no data were excluded. For mouse experiments, animals were randomly assigned to treatment groups using simple randomization. Unless otherwise stated, all findings were validated in at least three independent biological replicates with similar results. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41556-026-01991-z. Supplementary information Supplementary Figs. 1–8. Reporting Summary Peer Review File Differential gene expression and gene set enrichments of eIF2αS/S and eIF2αA/A tumour. Fate probability gene set enrichments. Differential gene expression and gene set enrichments of eIF2αS/A and eIF2αS/S Atf4+/− tumour. Differential gene expression and gene set enrichments of eIF2αS/S tumour with ISRIB or vehicle treatment. Enrichment of the cluster gene signature in human data. Source data Statistical source data. Statistical source data. Statistical source data. Statistical source data. Unprocessed western blots. Statistical source data. Statistical source data. Statistical source data. Statistical source data. Statistical source data. Unprocessed western blots. Statistical source data. Unprocessed western blots. Statistical source data. Unprocessed western blots. Statistical source data. Acknowledgements We thank C. Young, M.-L. Fillion and N. Ben-Limame at the core facility of the Lady Davis Institute and Segal Cancer Centre, Jewish General Hospital, for their support with flow cytometry, confocal microscopy and immunohistochemistry. We thank R. Lambert at the Genomics Platform at the Institut de recherche en immunologie et en cancérologie (IRIC) for assistance with scRNA-seq library construction, and A. Nakamura (AFSB-INRS Electron Microscopy Service) for support with electron microscopy. Proteomics analyses were performed by the Center for Advanced Proteomics Analyses, a node of the Canadian Genomic Innovation Network supported by the Government of Canada through Genome Canada. Extended data Author contributions Conceptualization: A.E.K., G.I.E. and T.T. Methodology: S.D., J.Y.Z., S.W., J.E.C., R.M.K., N.P., N.G. and H.K. Software: S.D., J.E.C., R.M.K. and N.P. Validation: C.K., M.H., G.I.E., P.W., N.S., R.M.K. and J.L.Q. Formal analysis: S.D., J.Y.Z., J.E.C., R.M.K. and N.P. Investigation: S.D., J.Y.Z., S.W., R.M.K., N.G., H.K., J.E.C. and N.P. Data curation: S.D., J.E.C., R.M.K. and N.P. Writing: A.E.K., S.D., J.Y.Z., M.H., P.W., N.S., R.M.K., J.L.Q. and T.T. Visualization: S.D., J.Y.Z., N.G., R.M.K. and N.P. Supervision: A.E.K., T.T., N.S., G.I.E., C.K. and P.W. Project administration: A.E.K. Peer review Peer review information Nature Cell Biology thanks Hongbin Ji and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Funding A.E.K. discloses support for the research of this work from Canadian Institutes of Health Research (CIHR) (PJT-178173, PJT-168864 and PJT-205973). M.H. discloses support for the research of this work from National Institutes of Health (NIH) (DK060569). T.T. discloses support for the research of this work from NIH (R01-CA270116 and from the Josie Robertson and Rita Allen Scholarships. C.K. discloses support for the research of this work from NIH (P01-CA165997; R01-CA268597). G.I.E. discloses support for the research of this work from Cancer Research UK programme grants (C4750/A12077). J.L.Q. discloses support for the research of this work from the Mazumdar-Shaw Chair endowment (University of Glasgow). J.Y.Z. is supported by a studentship award from Le Fonds de recherche du Québec (FRQS) and CIHR. S.D. and H.K. are recipients of studentship awards from the Faculty of Medicine, McGill University. The other authors declare no relevant funding. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript. Data availability RNA-seq data supporting the findings of this study have been deposited in the Gene Expression Omnibus (GEO) under accession codes GSE287513, GSE312048 and GSE312064. Mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE71 partner repository with the dataset identifier PXD072054. Raw data related to the Leicester Archival Thoracic Tumour Investigatory Cohort—Adenocarcinoma Tissue Microarray cohort used in this study are subject to NHSGGC Biorepository ethical approval. To access this data, a biorepository application and material transfer agreement are required. Please contact the Lead Researcher or the NHSGGC Biorepository manager (clare.orange@nhs.scot) for advice on the application process. Source data are provided with this paper. Gating strategies for flow cytometry analyses are provided in Supplementary Figs. 1–8. All other data supporting the findings of this study are available from the corresponding author upon reasonable request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Shiqi Diao, Jia Yi Zou. Extended data is available for this paper at 10.1038/s41556-026-01991-z. Supplementary information The online version contains supplementary material available at 10.1038/s41556-026-01991-z. References Associated Data Supplementary Materials Supplementary Figs. 1–8. Reporting Summary Peer Review File Differential gene expression and gene set enrichments of eIF2αS/S and eIF2αA/A tumour. Fate probability gene set enrichments. Differential gene expression and gene set enrichments of eIF2αS/A and eIF2αS/S Atf4+/− tumour. Differential gene expression and gene set enrichments of eIF2αS/S tumour with ISRIB or vehicle treatment. Enrichment of the cluster gene signature in human data. Statistical source data. Statistical source data. Statistical source data. Statistical source data. Unprocessed western blots. Statistical source data. Statistical source data. Statistical source data. Statistical source data. Statistical source data. Unprocessed western blots. Statistical source data. Unprocessed western blots. Statistical source data. Unprocessed western blots. Statistical source data. Data Availability Statement RNA-seq data supporting the findings of this study have been deposited in the Gene Expression Omnibus (GEO) under accession codes GSE287513, GSE312048 and GSE312064. Mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE71 partner repository with the dataset identifier PXD072054. Raw data related to the Leicester Archival Thoracic Tumour Investigatory Cohort—Adenocarcinoma Tissue Microarray cohort used in this study are subject to NHSGGC Biorepository ethical approval. To access this data, a biorepository application and material transfer agreement are required. Please contact the Lead Researcher or the NHSGGC Biorepository manager (clare.orange@nhs.scot) for advice on the application process. Source data are provided with this paper. Gating strategies for flow cytometry analyses are provided in Supplementary Figs. 1–8. All other data supporting the findings of this study are available from the corresponding author upon reasonable request.