Integrated analysis of transcriptional and metabolic responses to mitochondrial stress

Summary

Mitochondrial stress arises from a variety of sources, including mutations to mitochondrial DNA, the generation of reactive oxygen species, and an insufficient supply of oxygen or fuel. Mitochondrial stress induces a range of dedicated responses that repair damage and restore mitochondrial health. However, a systematic characterization of transcriptional and metabolic signatures induced by distinct types of mitochondrial stress is lacking. Here, we defined how primary human fibroblasts respond to a panel of mitochondrial inhibitors to trigger adaptive stress responses. Using metabolomic and transcriptomic analyses, we established integrated signatures of mitochondrial stress. We developed a tool, stress quantification using integrated datasets (SQUID), to deconvolute mitochondrial stress signatures from existing datasets. Using SQUID, we profiled mitochondrial stress in The Cancer Genome Atlas (TCGA) PanCancer Atlas, identifying a signature of pyruvate import deficiency in IDH1-mutant glioma. Thus, this study defines a tool to identify specific mitochondrial stress signatures, which may be applied to a range of systems.

Graphical abstract

Highlights

• •Multi-omics characterization of mitochondrial inhibitors generates stress signatures
• •SQUID allows the deconvolution of mitochondrial stress signatures from existing datasets
• •SQUID is used to define metabolic changes in IDH1-mutant glioma

Motivation

Mitochondrial stress may be caused by impairments in numerous aspects of mitochondrial biology. Yet, there is an incomplete understanding of how inhibition of specific components of mitochondrial biology affects cellular transcription and metabolism. Previous studies have focused on specific transcriptional responses, but the use of integrated multi-omics approaches to study mitochondrial stress has been limited in scope. Finally, no tool for investigating mitochondrial stress signatures across existing datasets is available. Thus, we developed stress quantification using integrated datasets (SQUID) to define distinct mitochondrial stress response signatures that integrate transcriptomic and metabolomic responses to specific types of mitochondrial inhibition, yielding a resource that can be used to deconvolve underlying drivers of mitochondrial dysfunction in complex biological contexts.

The mitochondrial stress response is the mechanism by which cells respond to dysfunctional mitochondria. Kelley et al. describe the characterization of cellular responses to mitochondrial inhibitors and generate a tool, termed SQUID, that can deconvolve signatures of mitochondrial stress from RNA sequencing or metabolic data.

Introduction

Mitochondria play a pivotal role within cells by coordinating essential cellular processes, including fuel metabolism, ATP synthesis, generation of molecules for biomass, and cell signaling.¹ These activities are needed to support cell growth and survival, but the concentrated biochemical activity and unique evolutionary origin of the mitochondria, where the local proteome is derived from both nuclear and mitochondrial sources, render this critical organelle susceptible to stress and dysfunction. Mitochondrial stress can arise from cell-intrinsic changes, such as disruption of mitochondrial protein import or mutations in mitochondrial enzymes, or cell-extrinsic changes, such as hypoxia, which limits oxygen.² Changes to mitochondrial function are associated with aging, neurodegenerative diseases, cancer, and mitochondrial diseases.^3,4,5,6

Since mitochondrial fitness is vital to support cell growth and survival, cells have evolved a network of protective pathways that minimize damage caused by normal mitochondrial processes (such as reactive oxygen species-mediated oxidative damage) and support repair processes. Among the best-studied responses to mitochondrial stress in mammals are the interconnected mitochondrial unfolded protein response (UPR^mt) and the mitochondrial integrated stress response (ISR).^7,8 Unfolded proteins can accumulate in mitochondria because of local damage from reactive molecules as well as from an imbalance between the import of proteins encoded in the nucleus versus mitochondria. The UPR^mt can be induced by mitochondrially localized sensors such as OMA1, which activates the cytosolic arm of the ISR by raising ATF4 expression.⁹ Mitochondrial unfolded protein stress can also transmit a retrograde signal, resulting in nuclear re-localization of the ATF5 transcription and subsequent expression of mitochondrial protein quality control machinery.¹⁰ MTORC1 has also been described as a mitochondrial stress sensor that activates ATF4.^11,12 Once activated, ATF4, ATF5, and C/EBP homologous protein (CHOP) promote translation of genes that reprogram cellular metabolism and induce the expression of mitochondrial chaperones and proteases to mitigate proteotoxic stress in the mitochondria.^10,13,14,15 Most studies on mitochondrial stress response pathways have focused on how mitochondria respond to the disruption of electron transport and/or mitochondrial proteostasis.^9,13,16,17 As a result, it is still unclear whether mitochondria coordinate distinct responses depending on the kind of mitochondrial stress. Defining transcriptional and metabolic signatures induced by different mitochondrial stressors (such as metabolic stress, oxidative stress, proteotoxic stress, and others) will reveal insight into specific adaptations to maintain mitochondrial homeostasis under diverse stress conditions. The resulting signatures may be applied to identify mitochondrial stressors characterizing different disease states, such as cancer.

In this study, we investigated transcriptional and metabolic responses to mitochondrial dysfunction induced through targeted inhibition of core mitochondrial pathways, including the import of metabolites into the mitochondria, metabolic enzymes, mitochondrial translation, proton gradient maintenance, and multiple steps in the electron transport chain. Our aim was to define distinct transcriptional and metabolic fingerprints characterizing cellular responses across mitochondrial stress conditions, ultimately yielding a comprehensive resource that can be used to deconvolve underlying drivers of mitochondrial dysfunction in complex biological contexts. We combined RNA sequencing with steady-state metabolomics to identify cellular responses to mitochondrial stress at the gene and metabolite levels, which we combined to generate multi-condition signatures of genes and metabolites induced by distinct mitochondrial inhibitors. Finally, we generated a tool for calculating scores based on these signatures in RNA sequencing or metabolomics datasets and used this approach to investigate the role of IDH1 mutations on pyruvate metabolism in glioma. This workflow, which we have termed SQUID (stress quantification using integrated datasets) can be used on existing datasets to identify and investigate signs of mitochondrial stress.

Results

Metabolic stress induces large-scale transcriptional and metabolic changes in human fibroblasts

We sought to investigate the cellular response to mitochondrial stress using a targeted multi-omics approach. We performed studies in primary human fibroblasts, which we chose as a non-transformed model system to interrogate mitochondrial stress responses. Fibroblasts were treated with mitochondrial inhibitors before combined transcriptomic and metabolomic analysis, followed by integrated analysis on the three datasets (Figure 1A). Drugs were chosen that inhibit diverse arms of mitochondrial biology, including electron transport chain function (antimycin A, metformin, rotenone, and 2,4-dinitrophenol [DNP]), fuel uptake and utilization within the mitochondria (BPTES, etomoxir, and UK-5099), mitochondrial protein synthesis (chloramphenicol and doxycycline), and NAD⁺ synthesis (FK866) (Figure 1B). Drug dosages were selected to inhibit the targeted pathway without causing significant cell death, with multiple dosages assayed for effects on cell viability, cell growth, and mitochondrial function (Figures S1A-S1F). Thus, our overall goal was to impair distinct arms of mitochondrial biology and define the resulting adaptive and specific transcriptional and metabolic responses.

Figure 1.

Generating a multi-omics atlas of mitochondrial stress in human fibroblasts

(A) Diagram showing the workflow of the multi-omics mitochondrial stress panel, incorporating both transcriptional and metabolic profiling.

(B) Pathway diagram showing major mitochondrial pathways targeted in the panel, including fuel import and energy generation via electron transport, alongside mitochondrial protein synthesis.

(C) Total number of differentially expressed genes (DEGs) identified in each condition when compared to the DMSO control. Each bar shows the fraction of DEGs significant in a given number of conditions.

(D and E) Total number of metabolites significantly changed vs. the DMSO control at t = 1 h time point (D) or t = 6 h time point (E) in each condition. Again, each bar shows the fraction of metabolites significant in a given number of conditions.

See also Figure S1 and Tables S1 and S2.

We first investigated the scale of perturbation for each drug in the panel. We calculated the total numbers of significantly altered transcripts and metabolites induced by each condition. In the transcriptional data, this ranged from 42 differentially expressed genes (DEGs) upon chloramphenicol treatment to 772 DEGs after etomoxir treatment (Figures 1C; Table S1). We also compared DEGs across conditions to identify shared and unique responses. A core set of 371 genes changed across ≥4 conditions (Figure S1G). Notably, rotenone and doxycycline primarily induced shared transcriptional changes, while the inhibitors UK-5099 and DNP elicited largely unique transcriptional responses (Figures S1H and S1I). Thus, a core transcriptional response was present across many mitochondrial stressors, though certain treatments triggered unique responses.

Next, we investigated how mitochondrial inhibitor treatment rewired the cellular metabolome, incorporating two time points to delineate acute (1 h) and adaptive (6 h) metabolic changes. After 1 h of treatment, metabolite levels were largely unchanged across most conditions. However, antimycin A treatment significantly altered 55 metabolites (Figure 1D; Table S2). By the 6 h time point, all conditions exhibited changes in metabolite levels. Metformin induced the largest metabolic perturbation, where 93 metabolites changed in abundance (Figure 1E; Table S2). The number of significant metabolites was greater at 6 h than at 1 h for most conditions across the panel of mitochondrial drugs (Figure S1J). Therefore, both the effect size and the time of onset of metabolic changes varied depending on treatment, highlighting that mitochondrial inhibitors trigger unique metabolic changes and do so on different timescales.

Transcriptional analysis highlights a shared metabolic response not mediated by the UPR^mt

To identify common transcriptional responses to mitochondrial inhibitors, we clustered conditions using the correlation between each condition’s log₂ fold changes versus DMSO. We identified a tight grouping of drugs targeting the electron transport chain (ETC), including rotenone, antimycin A, and metformin. FK866 and doxycycline also clustered with these ETC-targeting drugs, highlighting the dependency of ETC function on NAD and mitochondrial protein synthesis. Meanwhile, UK-5099 clustered distinctly apart from the other conditions, consistent with its high fraction of unique DEGs (Figure 2A). Thus, correlation analysis suggests that inhibiting targets within a shared pathway induced similar transcriptional responses.

Figure 2.

Transcriptional profiling of mitochondrial stress identifies shared metabolic gene responses across mitochondrial stress conditions

(A) Clustered heatmap of conditions using Spearman correlation of log₂ fold change vs. DMSO of all filtered genes.

(B) DEG counts for the condition combinations with the most overlapping genes. Conditions that contribute to the overlap are denoted with black circles.

(C) Heatmap depicting the number of genes with the specified number of significant upregulated (x axis) and downregulated (y axis) conditions.

(D and E) Log₂ fold change vs. DMSO for each condition for genes in the Hallmark glycolysis (D) or oxidative phosphorylation (E) gene sets. DEGs are highlighted with the number of total conditions in which they are differentially expressed.

(F) Clustering heatmap of log₂ fold change vs. DMSO for genes in the Hallmark oxidative phosphorylation and glycolysis gene sets.

(G) Transcription factors (TFs) in the Molecular Signatures Database (MSigDb) regulatory gene set that are significantly enriched in the most conditions.

(H) Dot plot of enrichment results for the top TFs from (G). Dot size denotes −log₁₀ adjusted p value, while dot color denotes net enrichment score (NES).

(I and J) Log₂ fold change vs. DMSO for each condition for genes in the target gene sets for ATF5 (I) and ATF4 (J) from MSigDb.

(K and L) Log₂ fold change vs. DMSO for each condition for ATF4 (K) and its downstream target DDIT3 (L).

(M) Western blot of ISR and UPR^mt pathway members upon 1, 6, and 24 h treatment with antimycin A, doxycycline, DNP, and tunicamycin. Amino acid starvation serves as a positive control for ISR activation.

For (K) and (L), data are presented as log₂ fold change vs. DMSO ± standard error of log₂ fold change, both calculated by DESeq2. ∗Adjusted p value (padj) < 0.05, ∗∗∗∗padj < 10⁻⁴.

See also Figure S2 and Table S1.

Given similarities between induced transcriptional changes across conditions, we investigated in which conditions DEGs overlapped the most frequently. Consistent with the correlation analysis, large-scale overlap was evident among DEGs induced by treatment with etomoxir, antimycin A, metformin, or FK866 (Figure 2B). A total of 271 genes were shared across DEGS induced by each of those four inhibitors (Figure S2D). We next investigated the patterns of overlapping DEGs in the dataset. The majority of DEGs arising in more than one condition displayed consistent directional changes in gene expression (Figure 2C). To characterize pathways enriched in these core genes, we performed Hallmark gene set enrichment among genes significant in ≥1 condition (the union of all DEGs) through ≥9 conditions (the highest condition overlap observed). Several pathways, such as Wnt signaling, were significant across the union of all DEGs but were not further enriched among DEGs shared across many conditions. In contrast, metabolic pathways such as glycolysis and oxidative phosphorylation became increasingly enriched as the gene list narrowed to include only DEGs shared across multiple conditions (Figure S2E). Indeed, a subset of glycolysis and oxidative phosphorylation genes was consistently upregulated across multiple conditions (Figures 2D and 2E). While a similar pattern emerged among apoptosis genes (Figure S2F), G2M checkpoint genes showed very few overlapping DEGs, consistent with reducing enrichment as the number of significant conditions increased (Figure S2G). In addition, analysis of glycolysis and oxidative phosphorylation genes highlighted a discrete cluster of genes sharply upregulated across all conditions except UK-5099 (Figure 2F). As a result, these data suggest that a shared metabolic gene response is present across multiple types of mitochondrial stress.

We also investigated the activities of ATF4 and ATF5, transcription factors known to mobilize cellular responses to mitochondrial stress.^{7,10,13,14,18,19} In transcription factor target enrichment, ATF5 activity was significantly altered in the most conditions. Strikingly, however, this was due to decreased activity in five conditions (Figure 2G). This enrichment was present among conditions that targeted electron transport, matching the shared transcriptional responses in those conditions (Figure 2H). Consistent with enrichment results, ATF5 targets were largely slightly downregulated across conditions (Figure 2I). Despite lower ATF5 activity in five conditions, its overall expression was unchanged (Figure S2H). Furthermore, targets of ATF4 and ATF3 (ATF5 family members) were not significantly altered in any condition (Figures 2J and S2H). Notably, however, expression of ATF4 and its downstream gene DDIT3 was markedly increased in the DNP condition, showing increased ATF4 expression and activity upon DNP treatment (Figures 2K and 2L). In contrast, ATF5 showed no change in expression (Figure S2I). Therefore, only the DNP condition showed signs of activation of either the ISR or UPR^mt pathways. The lack of ATF4 and ATF5 induction across most panel conditions raises two possible explanations: either the mitochondrial stressors tested do not activate these transcription factors under these conditions, distinct from mitochondrial proteotoxic stress, or the early time points examined (1 and 6 h) precede the induction of these stress response pathways.

To investigate whether a later time point was required for ISR activation, we examined later time points via western blotting. Following antimycin A, doxycycline, DNP, and tunicamycin treatment, we evaluated ATF4 and CHOP expression alongside eIF2α phosphorylation. Antimycin A did not induce ATF4 activation even after 24 h, while doxycycline, DNP, and tunicamycin did after 6 and 24 h. In addition, the positive control of amino acid starvation induced ATF4 expression after 6 h. Notably, however, eIF2α phosphorylation was largely consistent across conditions (Figure 2M). Thus, it appears that ATF4 expression is not universally induced in these specific conditions at the 6 h time point used for RNA sequencing.

Taken together, the transcriptional data highlight shared pathways induced upon mitochondrial stress across conditions, particularly subsets of glycolytic and oxidative phosphorylation genes. In addition, we find that antimycin A, etomoxir, FK866, and metformin treatment conditions show strong overlap in induced genes, suggesting a shared transcriptional response upon ETC disruption. Together with the finding that UPR^mt is not induced at the 6 h time point used for RNA sequencing, these data suggest that the identified transcriptional response represents a general metabolic rather than transcription factor-mediated response to mitochondrial stress.

Metabolomics analysis identifies time-dependent differences in metabolic adaptation upon mitochondrial stress

We next investigated how cellular metabolites changed in response to our panel of mitochondrial inhibitors. We evaluated two time points post treatment, capturing acute (1 h) and adaptive (6 h) metabolic changes. These time points were cleanly separated by principal-component analysis (PCA), while time point-specific PCA analysis showed differential clustering dependent on the time point (Figures S3A–S3C). We began by clustering the conditions at each time point. Clustering differed by time point, as evidenced by rotenone clustering with antimycin A after 1 h but with metformin and UK-5099 after 6 h (Figures 3A and 3B). Notably, condition clustering was distinct in the transcriptional data or either metabolic time point, suggesting an uncoupling between metabolic and transcriptional responses to mitochondrial stress.

Figure 3.

Metabolite profiling of mitochondrial stress identifies shared pathway responses upon mitochondrial pathway inhibition

(A and B) Clustered heatmap of conditions using Spearman correlation of log₂ fold change vs. DMSO of all metabolites at t = 1 h (A) and t = 6 h (B).

(C and D) Metabolite dot plots for the top significant metabolites at t = 1 h (C) and t = 6 h (D), sorted by number of significant conditions. Dot size denotes −log₁₀ adjusted p value, while dot color denotes log₂ fold change vs. DMSO.

(E) Log₂ fold change plots for lactate, glycerol 3-phosphate, sedoheptulose 7-phosphate, isocitrate, and hypoxanthine across conditions. Time point is denoted by bar opacity, while significance is denoted within the bar.

(F) MEAN enrichment results for the t = 1 h and t = 6 h datasets. Dot size denotes −log₁₀ adjusted p value, while dot color denotes NES. Pathways are clustered across both time points.

(G) Enrichment random walk plots from MEAN analysis for glycolysis, tricarboxylic acid (TCA) cycle, and glutathione metabolism at the 1 h and 6 h time points.

For (E), data are presented as log₂ fold change vs. DMSO ± standard error of log₂ fold change. ∗padj < 0.05, ∗∗padj < 0.01, ∗∗∗padj < 0.001, ∗∗∗∗padj < 10⁻⁴.

See also Figure S3 and Table S2.

We next examined metabolites changed across multiple conditions at each time point. At the 1 h time point, isocitrate was altered in the most conditions, with many other top metabolites coming from the tricarboxylic acid (TCA) cycle, glycolysis, or the pentose phosphate pathway, suggesting that central carbon metabolism is altered upon acute mitochondrial stress (Figure 3C). At 6 h, these central carbon metabolites were still altered; however, metabolites in other pathways, including purine and pyrimidine metabolism, were also frequently altered (Figure 3D). Thus, changes to central carbon metabolism may precede changes in other pathways.

We investigated metabolites that could potentially serve as common markers of mitochondrial stress response. Two glycolytic intermediates, lactate and glycerol 3-phosphate, were increased across multiple types of mitochondrial inhibition, suggesting that they may represent alternative endpoints of glycolysis upon mitochondrial dysfunction. Both sedoheptulose 7-phosphate and isocitrate changed across both time points, though, notably, isocitrate increased in abundance after 1 h before a decrease after 6 h. Finally, the purine metabolite hypoxanthine was notably increased in all 10 conditions at 6 h but unchanged at 1 h in all conditions (Figure 3E). As a result, several metabolites are consistently altered across multiple conditions in a time-dependent manner, suggesting that changes occur in central carbon metabolism endpoints upon mitochondrial stress.

To identify patterns of pathway alteration across the conditions and time points, we adopted an approach inspired by gene set enrichment analysis (GSEA). The approach, which we term metabolite enrichment of annotated networks (MEAN), combines the GSEA algorithm with manual annotation of quantified dataset metabolites (Figure S3D). MEAN identified shared pathway perturbations across conditions at both time points. A striking pattern emerged with amino acid metabolism, where most pathways within amino acid metabolism were strongly upregulated after 6 h (Figures 3F, S3E, and S3F). We then investigated major pathways for time point-dependent changes in enrichment. Glycolysis and the TCA cycle were sharply and consistently downregulated at 1 and 6 h, respectively, with varied changes at the other time points. In contrast, glutathione metabolism was universally upregulated at both time points, highlighting the central role of redox metabolism in the stress response (Figure 3G). Taken together, MEAN identifies metabolic pathway perturbations shared across mitochondrial stressors tested and provides insight into time point-specific changes to major pathways.

Similar to the transcriptional analysis, these metabolomic data demonstrate that diverse mitochondrial stressors lead to a core set of consistent metabolic changes. We represented metabolic changes observed after 6 h of drug treatment on a metabolic map that summarizes and graphically depicts metabolic adaptations to mitochondrial stress (Figure 4). This map highlights consistent patterns that emerge across metabolic pathways in response to mitochondrial stress, including a depletion of hexose phosphates and early TCA cycle intermediates by UK-5099 treatment as well as conditions targeting electron transport, an increase in amino acids and purine monophosphates across multiple conditions, and a consistent reduction in most pyrimidine species upon BPTES or rotenone treatment. In addition, this map highlights expected signs of metabolic stress, including increases in lactate and oxidized glutathione across multiple conditions. Taken together, this pathway map provides deep insight into metabolic alterations induced upon distinct arms of mitochondrial stress.

Figure 4.

An integrated map of metabolic pathways shows key markers of metabolic response to mitochondrial stress

Metabolic pathway map generated using the t = 6 h time point for each condition. Metabolites in black were quantified, while those in gray were not detected or quantified. Major pathways and metabolite categories are denoted in colored zones. Dashed arrows denote multiple reactions. Reactions utilizing NAD⁺/NADH are marked with orange circles, while reactions utilizing FAD/FADH₂ are marked with teal circles. For each quantified metabolite, if significantly altered in at least one condition, significance for all 10 conditions is denoted using a pattern of 10 circles next to the metabolite, with red denoting upregulation and blue denoting downregulation. Metabolites displayed in black with no significance circles had no significant changes.

See also Figure S4.

In addition, we generated pathway-specific maps to compare metabolic changes at the 1 and 6 h time points. In glycolysis, 6-carbon intermediates show depletion at both time points, while after 6 h, upregulation of 3-carbon intermediates is evident across multiple conditions. This buildup of 3-carbon intermediates could be indicative of decreased flux into acetyl-CoA over time (Figure S4A). Notably, the pattern of changes in the TCA cycle looks largely similar despite differences in enrichment as determined by MEAN (Figure S4B). Finally, pyrimidine metabolism shows stark differences, where most metabolites are unchanged or increased at 1 h but consistently depleted at 6 h (Figure S4C). Together, these maps further highlight time point-dependent changes in major metabolic pathways upon mitochondrial stress.

Multi-omics integration generates signatures of mitochondrial stress

To define signatures associated with specific types of mitochondrial inhibition, we leveraged the multi-omics analysis. We performed multi-omics sparse partial least-squares differential analysis (sPLS-DA) using mixOmics DIABLO, which allows for the generation of mRNA and metabolite weights through the analysis of variance across conditions rather than individual samples (Figure 5A).²⁰ As DIABLO requires paired samples across data types, we used random pseudo-pairing between mRNA and metabolite samples of the same condition. To ensure that the output was not affected by the pseudo-pairing used, we compared the first 10 sets of random pairs and found that consistent signatures were generated regardless of the exact pairing used (Figure S5A).

Figure 5.

Integrated analysis of transcriptomics and metabolomics allows for the generation of multi-omics signatures of mitochondrial stress

(A) Schematic workflow for data integration between transcriptomics and metabolomics datasets using the mixOmics package and pseudo-pairing of samples.

(B) DIABLO output for the first four variates of sPLS analysis, termed as stress signatures 1–4. Each signature is plotted as mean ± standard error per condition.

(C) Top gene and metabolite loadings by absolute value for each signature, as determined by DIABLO. Color denotes positive or negative loading weight.

(D) Hallmark overrepresentation enrichment analysis of metabolite loadings for signatures 1–4 from DIABLO analysis. Dot size denotes −log₁₀ adjusted p value, and dot color denotes gene enrichment ratio.

(E) MEAN enrichment of 1-h and 6-h metabolite loadings from DIABLO analysis. Dot size denotes −log₁₀ adjusted p value, while dot color denotes NES.

See also Figure S5.

DIABLO analysis provided insight into the major multi-condition patterns that are present in the three datasets. Using DIABLO, we defined four major sets of genes and metabolites that form patterns of conditions in the data, which we termed signatures of the mitochondrial stress response. These signatures serve as discrete patterns of a cellular metabolic stress response and are consistent across the three datasets (Figures S5B-S5C). Signature 1 is associated with a high score in UK-5099 and low scores in the drugs antimycin A, BPTES, etomoxir, FK866, and metformin, highlighting the opposing cellular responses induced by these inhibitors. Meanwhile, signature 2 is dominated by the mitochondrial uncoupler DNP. Finally, both signatures 3 and 4 bifurcate the conditions into two discrete groups (Figure 5B). Given the disparate targets of the conditions separated by signatures 3 and 4, these results suggest that the signatures could be driven by shared molecular responses rather than shared molecular targets. The signatures explained higher variance in the metabolite datasets compared to the mRNA dataset (Figure S5D). As a result, integrated analysis allows for the identification of multi-component stress signatures that leverage both transcriptional and metabolic data.

Each of these signatures is defined as a set of weighted genes and metabolites, which are used to calculate each sample’s score. We therefore investigated the top genes and metabolites in each signature to understand the molecular changes captured by the signature. Signature 1 highlights genes downregulated upon UK-5099 treatment (Figure S5E), alongside multiple nucleotides, particularly at the 6 h time point. Signature 2 includes multiple canonical ATF4 targets, including GOT1, ASNS, and GARS. Notably, signature 3 genes include TXNIP, a well-studied marker of oxidative and ETC dysfunction, alongside its arrestin family member ARRDC4. Furthermore, glycolysis intermediates are present in the top loadings at both time points. Finally, signature 4 shows multiple TCA cycle and nucleotide intermediates at the 1 h time point (Figure 5C). Together, the loadings suggest potential driving mechanisms behind the patterns observed in the four signatures generated by DIABLO.

To further characterize the stress signatures generated by DIABLO analysis, we performed gene enrichment and MEAN analysis on the top loadings for each signature. Hallmark overrepresentation analysis identified several enriched pathways among signature 2 loadings, including mTORC1 signaling and the UPR, consistent with the ATF4 targets in the top loadings. In addition, hypoxia and apoptosis genes were enriched in signature 1 and 3 loadings, respectively (Figure 5D). MEAN analysis highlighted enrichment of metabolic pathways in signatures 1, 3, and 4, including a strong downregulation of central carbon pathway intermediates for signature 3 at both 1 and 6 h time points (Figure 5E). Therefore, the signatures generated by DIABLO are associated with specific transcriptional and metabolic pathway changes that differ between conditions.

SQUID allows for the identification of altered mitochondrial stress in IDH1-mutant glioma

Ultimately, we aimed to utilize the generated gene signatures to probe mitochondrial stress in other biological contexts, such as cancer or diabetes. Numerous studies associate various diseases with mitochondrial perturbations, yet a method does not exist to infer upstream molecular processes that drive mitochondrial stress states in cancer datasets. To bridge this gap, we generated a method, termed SQUID, to apply our identified mitochondrial stress signatures to existing datasets. SQUID performs DIABLO integrated analysis on the overlap of genes and/or metabolites between the dataset of interest and our mitochondrial stress dataset. This generates stress signatures, alongside condition-specific signatures, tailored to the dataset. Stress scores are then calculated for the target dataset by summing up the product of gene (or metabolite) abundances and weights calculated from DIABLO, akin to how sPLS-DA scores are calculated. As a result, a high score in a given signature indicates gene expression or metabolite abundance resembling that of the condition in the original panel of drug perturbations (Figure 6A). Using SQUID, we can score external datasets using our mitochondrial stress signatures derived from precise drug perturbations. This enables us to quantify the pattern of mitochondrial dysfunction present in unknown samples and predict molecular drivers of mitochondrial dysfunction.

Figure 6.

Application of mitochondrial stress signatures to the TCGA PanCancer Atlas identifies a pyruvate metabolism phenotype in IDH1 R132H-mutant glioma

(A) Schematic for the application of integrated mitochondrial stress dataset to outside datasets using sPLS-DA, using the overlap of genes and/or metabolites between the two datasets.

(B) Median mitochondrial stress score for each signature and individual condition score across cancer types in the TCGA PanCancer Atlas. Glioma samples are highlighted in maroon.

(C–E) Stress scores for glioma samples classified by IDH1 status for signature 1 (C), metformin (D), and UK-5099 (E).

(F) Kaplan-Meier survival curve for IDH1 WT samples separated into high- and low-score status for UK-5099, where high represents the top quartile of scores, and low represents the bottom quartile. Data are presented with 95% confidence intervals. The Cox proportional hazards model was used to determine HR and significance between groups.

(G) Kaplan-Meier survival curve for IDH1 R132H samples separated into high- and low-score status for UK-5099.

(H and I) Cell growth curve for HOG IDH1 EV (H) and HOG IDH1 R132H (I) cells. Significance is compared to 0 μM control.

(J) Endpoint values for growth curves presented in (L) and (M), normalized to the 0 μM control.

(K and L) Basal respiration oxygen consumption (K) and ATP-dependent oxygen consumption (L) for HOG IDH1 EV and R132H as determined by Seahorse mitochondrial stress kit.

∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. Significance was determined via two-tailed t test for (C)–(E) and (H)–(L) and Cox proportional hazards model for (F) and (G). Data in (H)–(L) are presented as means ± standard error.

See also Figure S6.

We used SQUID to investigate mitochondrial stress signatures in The Cancer Genome Atlas (TCGA) PanCancer Atlas dataset due to its combination of patient data coupled with sample assays (including RNA sequencing [RNA-seq] and mutational analysis) across many cancer types.²¹ Stress signatures generated using PanCancer Atlas genes closely resembled the original stress signatures (Figure S6A). Given that the TCGA PanCancer Atlas contains patient mutation data, we investigated the role of specific mutations on mitochondrial stress scores. We decided to focus on IDH1 in glioma, as it has well-studied changes to cellular metabolism and epigenetic regulation through synthesis of the oncometabolite D-2HG.^22,23,24,25 IDH1 mutations were strongly enriched in glioma samples, with nearly 80% of glioma samples having mutations in IDH1. Consistent with past studies, these IDH1 mutations largely occurred at R132 (Figure S6B).^24,25,26,27

We hypothesized that the high incidence of the IDH1 R132H mutation would affect mitochondrial stress scores in glioma. Compared to other cancer types, glioma had the second-highest score for signature 1 and UK-5099 and the second-lowest score for metformin (Figure 6B). Signature 1 is driven by UK-5099, which targets the mitochondrial pyruvate carrier, while metformin has numerous putative targets in cells, including complex I. As a result, these stress signatures suggest IDH1-dependent changes to pyruvate import and complex I activity. We investigated whether these stress scores were dependent on IDH1 mutation status. Signature 1 and UK-5099 scores were significantly higher in IDH1 R132-mutant samples than in IDH1 WT samples, while metformin scores were significantly lower (Figures 6C–6E). This is consistent with reports on IDH1 altering pyruvate metabolism and ETC function.^23,26,27 Therefore, IDH1 mutation status is associated with altered stress signatures in glioma samples, suggesting that these mutations potentially drive changes in stress.

We next investigated whether altered stress signature scores were associated with patient survival. Given that IDH1-mutant gliomas have better survival outcomes than IDH1 wild-type (WT) gliomas in the TCGA data, consistent with previous studies (Figure S6C), we split our analysis by IDH1 genotype. In IDH1 WT samples, patients in the top quartile of UK-5099 scores had increased survival compared to patients in the bottom quartile (hazard ratio [HR] = 3.8, p = 0.002) (Figure 6F). In contrast, in IDH1 R132H samples, there was no difference in survival outcomes (Figure 6G). Similar patterns were observed in signature 3 and metformin, with significant survival differences only among IDH1 WT samples (Figures S6D-S6G). In addition, in each case, the samples with stress scores resembling those of the R132H mutants showed improved survival, suggesting a link between stress and survival even in WT patient samples. Therefore, the IDH1 R132H mutation alters mitochondrial stress in glioma, and mitochondrial stress score is correlated to survival outcome in glioma samples across IDH1 genotypes.

Following our analysis of UK-5099 signature scores observed in glioma samples, we then directly investigated the UK-5099 response in glioma cell lines. The UK-5099 scores were higher in IDH1 R132H patient samples compared to patients with WT IDH1, suggestive of deficient pyruvate import. We therefore hypothesized that IDH1 R132H cells would be more susceptible to UK-5099 treatment. We utilized empty vector (EV) and IDH1 R132H-expressing human oligodendroglioma (HOG) cells, a frequently used model for studying mutant IDH1 biology.²⁸ UK-5099 reduced growth in HOG cells expressing either IDH1 R132H or EV (Figures 6H and 6I). In particular, the R132H-expressing cells were more sensitive to UK-5099 at both 50 and 100 μM than the EV-expressing controls (Figure 6J). We hypothesized that this reduction in cell growth was due to reduced respiratory capacity in the cells. Analysis of mitochondrial oxygen consumption using a Seahorse assay demonstrated much higher basal respiration and ATP production rates in the HOG EV cells compared to HOG R132H cells, with significant decreases in ATP production upon UK-5099 treatment in both cell lines (Figures S6H and S6I). Furthermore, R132H cells exhibited larger relative decreases in both basal respiration and ATP production at 20 μM UK-5099 compared to the EV cells, highlighting an increased susceptibility in the mutants (Figures 6K and 6L). Therefore, the IDH1 R132H mutation in glioma cells shows an increased susceptibility to UK-5099 than IDH1 WT cells, highlighting the ability of SQUID to identify potential metabolic vulnerabilities in cancer types.

Together, these results highlight the power of SQUID to elucidate and quantify mitochondrial stress signatures in outside datasets. In this application, SQUID identified a mutation-specific form of mitochondrial stress in glioma that predicted a difference in drug susceptibility. These results demonstrate the potential applications of SQUID to study mitochondrial stress in many contexts.

Discussion

The response to mitochondrial stress in mammalian cells has been a burgeoning field of research for the past decade. These studies have illuminated transcription factors and signaling pathways that restore mitochondrial homeostasis during proteotoxic stress, where unfolded proteins accumulate in the mitochondria. However, it is not known whether there are distinct signatures associated with the inhibition of specific arms of mitochondrial function. To this end, we utilized transcriptional and metabolic analyses to characterize the effects of a panel of inhibitors of mitochondrial major pathways on primary human fibroblasts. We identified shared transcriptional and metabolic changes that highlight the central role of rewired metabolism in the response to mitochondrial stress and generated integrated signatures of mitochondrial stress that couple both genes and metabolites altered upon mitochondrial stress. We then developed SQUID, a workflow for applying these mitochondrial stress signatures to outside datasets and used SQUID to identify a signature of dysregulated pyruvate metabolism in IDH1-mutant glioma.

The results of this study were enabled by the integration of transcriptional and metabolomics data. While previous studies on mitochondrial stress have largely focused on gene expression changes, our study included metabolomics performed at two time points to characterize another factor of the mitochondrial stress response. We performed integrated analysis on the three datasets to identify stress signatures defined by gene and metabolite changes. The signatures generated by this approach differed from those generated from transcriptional or metabolics data alone (data available upon request), highlighting that the signatures are informed by all datasets. For instance, the loadings for signature 3 are predominantly composed of glycolytic and TCA cycle intermediates, making the signature a readout of central carbon metabolism (Figure 5B). As a result, integrated analysis highlights the interrelation between gene and metabolite levels upon mitochondrial stress.

We identified shared mRNA and metabolomic responses across multiple drug treatments, highlighting a core set of changes that are common across mitochondrial stressors. For example, shared transcriptional responses were identified between several drugs that targeted the ETC. Likewise, genes involved in glycolysis and oxidative phosphorylation were generally over-represented, highlighting a shared transcriptional response to mitochondrial stress. Notably, the known mitochondrial stress marker TXNIP was downregulated in seven conditions, consistent with previous studies.^29,30 Strikingly, multiple complex I subunits were upregulated across most conditions, including NDUFS7, altered in nine conditions, and NDUFA11 and NDUFB7, both altered in seven conditions. These results suggest that restoration of complex I activity may be vital for the response to mitochondrial stress. In addition, the lipid catabolism gene PNPLA2 was upregulated in eight conditions, suggesting that fatty acids could serve as an alternative fuel source during mitochondrial stress. We also identified common metabolomic changes across multiple mitochondrial drug treatments, highlighting conserved metabolic adaptations to diverse mitochondrial stressors. Central carbon metabolites were dysregulated across several conditions, with strong changes observed in glyceraldehyde 3-phosphate and succinate (Figure 3G). In addition, numerous amino acids increased in abundance upon treatment with BPTES, chloramphenicol, metformin, and UK-5099. These metabolic changes may serve as common regulatory nodes during mitochondrial repair processes, and future work may explore whether they can serve as biomarkers of mitochondrial dysfunction.

We leveraged the integrated nature of the study to generate a workflow, termed SQUID, for quantifying mitochondrial stress signatures in datasets containing mRNA and/or metabolite data. There is a relative dearth of metabolite profiling of human patient samples, while RNA-seq has become quite abundant. By leveraging multi-omics stress signatures, SQUID provides metabolic insight into datasets lacking metabolic data. This was demonstrated using the TCGA PanCancer Atlas, which includes RNA-seq information but no metabolomics data. We identified an alteration in pyruvate metabolism using SQUID that was then validated in vitro. We believe that SQUID can serve similar informative roles in the study of other human diseases and conditions, such as aging, neurodegeneration, and diabetes, which are all closely linked to changes in cellular metabolism.^{31,32,33,34,35} This approach could be bolstered by the inclusion of additional data types, such as proteomics, which would further increase the utility of SQUID.

In summary, our work provides a detailed characterization of cellular responses to mitochondrial stress using both transcriptional and metabolic analyses. We leveraged these findings to produce SQUID, a tool that quantifies mitochondrial stress signatures in transcriptional or metabolic datasets. Using SQUID, mitochondrial stress can be investigated in a wide array of contexts, including existing patient samples for many diseases.

Limitations of the study

Here, we developed SQUID as a set of mitochondrial stress response signatures that leverage transcriptional and metabolic changes upon a panel of mitochondrial inhibitors. We used primary human fibroblasts as a non-cancerous cell model, as they have non-damaged mitochondria. However, cell type-specific dependencies on the mitochondrial stress response are known, making it impossible for any single-cell model to capture all aspects of the mitochondrial stress response. That choice, and the early time points utilized in the study, are likely causes of the limited UPR^mt signature in the dataset. Validation of the dataset was performed using the TCGA PanCancer Atlas using transcriptional data, which independently identified a pyruvate import deficiency in IDH1-mutant glioma, consistent with the literature. However, the use of SQUID on metabolic datasets has so far been limited to internal studies, and a large-scale comparison of mitochondrial stresses using metabolic data has not been performed. In addition, our study is limited to human genes, though SQUID does support homolog conversion between mouse and human.

Resource availability

Lead contact

Requests for further information and resources should be directed to the lead contact, Marcia Haigis (marcia_haigis@hms.harvard.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The RNA-seq dataset generated for this study is available at GEO under series GSE241261. All other datasets used in this study are publicly available and are listed in the key resources table.
• •The scripts produced for this study are available at Mendeley Data (DOI: https://www.doi.org/10.17632/4769fjr4yd.2), as listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

L.P.K. was supported by the Joslin Diabetes Center T32 Training Program in Diabetes and Metabolism (NIH grant 2T32DK007260). S.-H.H. was supported by the American Cancer Society Postdoctoral Fellowship (PF-24-1249214-01-TBE). A.E.R was supported by NIH K22 CA266150, as well as postdoctoral fellowships from the American Cancer Society (130373-PF-17-132-01-CCG) and the Cell Biology Education and Fellowship Fund at Harvard Medical School. M.C.H. was supported by funding from NIH R01CA273461, the Ludwig Center at Harvard Medical School, the Glenn Foundation for Medical Research, the Massachusetts Life Sciences Center, and Agilent Technologies. The authors are grateful for generous support from the David Liposarcoma Research Initiative at the Dana-Farber Cancer Institute, supported by KBF Canada via the Rossy Foundation Fund. This project was also supported by Research Computing (Harvard Medical School) and the Harvard Chan Bioinformatics Core (Harvard T.H. Chan School of Public Health). The authors also thank Kiran Kurmi, Shakchhi Joshi, Joseph Crowley, and Conghui Yao for their assistance and guidance with metabolomics data acquisition and analysis. The figures include artwork from BioRender.com.

Author contributions

Conceptualization, M.C.H. and A.E.R.; methodology, A.E.R. and L.P.K.; software, L.P.K.; formal analysis, L.P.K.; investigation, M.C.H., A.E.R., S.A.B., L.P.K., and S.-H.H.; writing – original draft, L.P.K.; writing – review & editing, L.P.K., A.E.R., and M.C.H.; resources, A.E.R. and M.C.H.; funding acquisition, L.P.K., A.E.R., and M.C.H.; supervision, A.E.R., M.C.H., and P.K.S.

Declaration of interests

M.C.H. received funding from Agilent Technologies. M.C.H. is on advisory boards for Alixia Therapeutics, MitoQ, and Minovia. M.C.H. is a co-founder of and advisor for and received funding from ReFuel Bio on unrelated projects. M.C.H. is on the advisory board of Cell Metabolism and Molecular Cell.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-ATF4	ProteinTech	10835-1-AP
Anti-CHOP	Cell Signaling	Cat# 2895; RRID: AB_2089254
Anti-phospho-eIF2α	Cell Signaling	Cat# 3398; RRID: AB_2096481
Anti-eIF2α	Cell Signaling	Cat# 5324; RRID: AB_10692650
Anti-Tubulin	Cell Signaling	Cat# 2125; RRID: AB_2619646
Chemicals, peptides, and recombinant proteins
Fibroblast Basal Medium	ATCC	PCS-201-030
Fibroblast Growth Kit-Low serum	ATCC	PCS-201-041
Trypsin-EDTA for Primary Cells	ATCC	PCS-999-003
Trypsin Neutralizing Solution	ATCC	PCS-999-004
Penicillin-Streptomycin (10,000 U/mL)	Gibco	15140122
RPMI 1640 Medium w/o Amino Acids, Sodium Phosphate	US Biological	R8999-04A-25L
Antimycin A	Sigma-Aldrich	A8674-50MG
BPTES	Sigma-Aldrich	SML0601-5MG
Chloramphenicol	Sigma-Aldrich	C0378-25G
2,4-Dinitrophenol	Sigma-Aldrich	D198501-100G
Doxycycline hyclate	Sigma-Aldrich	D9891-5G
(+)-Etomoxir (sodium salt)	Cayman Chemical	11969
Daporinad (FK866)	Selleck Chemicals	S2799
Metformin hydrochloride	Sigma-Aldrich	PHR1084-500MG
Rotenone	Sigma-Aldrich	R8875-1G
Tunicamycin	Sigma-Aldrich	T7765-5MG
UK-5099	Sigma-Aldrich	PZ0160-5MG
Seahorse XF RPMI medium	Agilent	103576–100
XF Mito Stress Test	Agilent	103015–100
Critical commercial assays
Direct-zol RNA Miniprep	Zymo Research	R2050
TruSeq Stranded mRNA Library Prep	Illumina	20020595
Deposited data
Gene expression from fibroblasts treated with mitochondrial inhibitors	This paper	GEO: GSE241261
TCGA PanCancer Atlas	Liu et al.21	Accessed via http://www.cbioportal.org
Experimental models: Cell lines
Primary Dermal Fibroblast; Normal, Human, Adult (HDFa)	ATCC	PCS-201-012
Software and algorithms
R (4.3.1)	The R Project for Statistical Computing	https://www.r-project.org
Bioconductor (3.17)	Bioconductor project	https://www.bioconductor.org
Salmon (1.4.0)	Patro et al.36	https://github.com/COMBINE-lab/salmon
Tximeta (1.24.0)	Love et al.37	https://bioconductor.org/packages/release/bioc/html/tximeta.html
DESeq2 (1.40.2)	Love et al.38	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
pheatmap (1.0.12)	Kolde et al.39	https://github.com/raivokolde/pheatmap
clusterProfiler (4.9.2)	Yu et al.40	http://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
mixOmics (6.24.0)	Rohart et al.20	https://www.bioconductor.org/packages/release/bioc/html/mixOmics.html
TraceFinder	ThermoFisher	https://www.thermofisher.com/us/en/home/industrial/mass-spectrometry/liquid-chromatography-mass-spectrometry-lc-ms/lc-ms-software/lc-ms-data-acquisition-software/tracefinder-software.html
Original code	This paper	Mendeley Data: https://doi.org/10.17632/4769fjr4yd.2

Experimental model and study participant details

Cell lines

Primary human fibroblasts were cultured in fibroblast basal media (ATCC) supplemented with fibroblast growth kit low serum (ATCC), which includes FBS, ascorbic acid, recombinant human Insulin, recombinant human FGF-b, hydrocortisone, and L-glutamine. When culturing for metabolomics, fibroblast media also contained 0.1% penicillin/streptomycin. Fibroblasts were passaged using a combination of trypsin for primary cells (ATCC) and trypsin neutralizing solution (ATCC). For drug treatment, media was supplemented with 2 mM additional glutamine and a total of 0.5% DMSO from either vehicle control or drug treatment.

Normal human astrocytes (NHA cells) were cultured in DMEM (Gibco) containing 4.5 g/L glucose, L-glutamine, and 110 mg/L sodium pyruvate, supplemented with 10% FBS (Corning) and 1% penicillin/streptomycin (Gibco). Human oligodendroglioma (HOG) cells were cultured in IMDM (Gibco) containing L-glutamine and 25 mM HEPES, supplemented with 10% FBS and 1% penicillin/streptomycin. Cells were passaged using 0.05% trypsin-EDTA (Gibco). Passage number was monitored for the HOG cells but was unknown for the NHA cells.

Method details

RNA isolation and library generation

RNA was extracted from treated fibroblasts using TRIzol reagent (ThermoFisher) and isolated using an RNA miniprep kit (Zymo Research). Libraries for RNA sequencing were generated using the High Throughput TruSeq Stranded mRNA Library Prep Kit (Illumina) following the manufacturer’s protocol at half reaction volume. Input for each sample consisted of 750ng of RNA and 5ul of 1:500 diluted ERCC spike-in mix 2 (Ambion). Libraries were amplified for 11 cycles during the final amplification step. Library indexing utilized TruSeq Unique Dual Indexes. Sample spot checks were performed on a subset of the samples using a 2100 Bioanalyzer instrument (Agilent). Libraries were then pooled at equimolar concentrations and submitted for quantification and sequencing at the Biopolymers Facility, Harvard Medical School. Sequencing was performed using a NextSeq 500 sequencer (Illumina).

Western blotting

Human primary fibroblasts were treated for 1, 6, or 24 h in the respective conditions in 6 well format. For amino acid starvation treatment, cells were treated using RPMI lacking all 20 proteinogenicc amino acids. Complete RPMI was used as a control for the amino acid starvation treatment to ensure fibroblasts were not adversely affected by culture in RPMI. Cellular protein was isolated using RIPA buffer (150 mM NaCl, 1% NP-40, 0.5% sodium deoxycholate, 0.1% sodium dodecyl sulfate, 50 mM Tris-HCl pH 7.4) supplemented with protease inhibitor (Sigma-Aldrich) and phosphatase inhibitor cocktail (Sigma-Aldrich). The supernatants were collected by centrifugation at > 16,000 g in a microcentrifuge at 4°C for 15 min followed by the BCA assay to quantify protein concentrations (ThermoFisher Scientific). Equal amounts of protein were diluted into 1x SDS-PAGE loading buffer, boiled at 95°C for 10 min, and then loaded onto Criterion TGX 4–20% gels (Bio-Rad) for immunoblot analysis. Proteins were transferred onto nitrocellulose membranes using a Trans-Blot Turbo transfer system (Bio-Rad). Membranes were blocked for 1 h at room temperature in 3% BSA and then incubated overnight at 4°C with the indicated antibodies diluted in 3% BSA: anti-ATF4 (Proteintech, Cat# 10835-1-AP), anti-CHOP (Cell Signaling, Cat# 2895, anti-p-eIF2α (Cell Signaling, Cat# 3398), anti-eIF2α (Cell Signaling, Cat# 5324), anti-Tubulin (Cell Signaling, Cat# 2125). The membrane was washed three times for 15 min with 1x TBST and then incubated for 1 h at room temperature with Rabbit IgG HRP-linked whole antibody (ThermoFisher Scientific, Cat# 45-000-682, 1:5000) diluted in 3% BSA. The membrane was washed three times with 1x TBST and the signal was detected using ECL solution (PerkinElmer).

Metabolite sample preparation

Cell treatment and metabolite isolation were performed in staggered batches, where the DMSO, UK-5099, doxycycline, chloramphenicol, DNP, and rotenone samples were treated at the same time, with the other conditions treated 30 min later. Cell treatment and extraction was also split over two days, with day 1 consisting of treatment and isolation for the 1 h timepoint samples and half of the 6 h timepoint samples, with day 2 consisting of the other half of the 6 h timepoint samples.

Cell dishes were washed successively in PBS and then 150 mM NaCl to remove residual media. Dishes were then quenched in liquid nitrogen. Cells and metabolites were then extracted using 100% methanol followed by scraping into tubes. Cells were lysed using repeated snap freeze and sonication cycles before cell supernatants were collected and dried using a speed-vac. Metabolite pellets for each sample were resuspended in 50% acetonitrile/50% water for mass spectrometry runs.

Resuspended metabolite samples were run on a Vanquish U-HPLC liquid chromatography system (Thermo) before ionization using a heated electrospray ionization (HESI) probe. Ionized metabolites were analyzed on a QExactive HF-X mass spectrometer (Thermo) in negative ion mode. Analytes were separated on an iHILIC column (5 μM, 150 × 2.1 mm internal diameter, Hilicon). Buffers used in liquid chromatography were as follows: buffer A: 20 mM ammonium carbonate and 0.1% ammonium hydroxide in water; buffer B: 100% acetonitrile. HPLC was performed at a flow rate of 0.150 mL/min unless otherwise stated. The chromatographic gradient was as follows: 0–23 min linear gradient from 95% B to 5% B; 23–25 min hold at 5% B, divert to waste from 25–25.5 min, gradient to 95% B at 0.20 mL/min, 25.5–32.5 min hold at 95% B at 0.20 mL/min, and finally 32.5–33 min 95% B at 0.15 mL/min.

Cellular growth curve assay

HOG and NHA cells were plated on day 0 in the relevant UK-5099 doses in 12-well plates. Each day, one 12-well plate consisting of 3 replicates each of the 4 doses was trypsinized and counted for each cell line. After trypsinization, cells were counted using a Beckman z series cell counter. Significance was calculated using a t test comparing each timepoint back to the relevant DMSO control.

Seahorse mitochondrial stress assay

The oxygen consumption rate (OCR) of cells was determined using the XFe96 Seahorse extracellular flux analyzer (Seahorse Bioscience) and the Seahorse XF Mito Stress Test kit.

Fibroblasts were plated 24 h before Seahorse analysis using equal cell densities per well. Cells were plated with the relevant drug dosage for treatment overnight. Seahorse runs were performed in Seahorse XF RPMI media (Agilent) supplemented with 10 mM glucose, 2 mM glutamine, and 1 mM sodium pyruvate supplemented with the relevant drug dosage. OCR was monitored during injections of oligomycin (1 μM), FCCP (2 μM), and rotenone/antimycin A (0.5 μM).

HOG and NHA cells were plated 24 h before Seahorse analysis at specific cell densities calculated for 80% confluency at the time of the run. Cells were plated with the relevant UK-5099 dose used during the run, which had previously been shown to not affect confluency after 24 h. Seahorse runs were performed in Seahorse XF DMEM media (Agilent) supplemented with 20 mM glucose, 4 mM glutamine, and 1 mM sodium pyruvate. The media was supplemented with the relevant UK-5099 dose after media change. OCR was monitored during injections of oligomycin (1 μM), FCCP (1.5 μM), and rotenone/antimycin A (0.5 μM). Basal respiration, maximal respiration, and ATP-dependent respiration were calculated using the first timepoints of each relevant stage of the stress test. Significance was calculated using a two-tailed t test.

Quantification and statistical analysis

RNA sequencing analysis

Sequencing reads were pseudo-aligned to an index generated using the Ensembl release 103 human transcriptome using the alignment software Salmon.³⁶ Estimated counts were then imported into R using the tximeta package.³⁷ Raw counts were processed using the DESeq2 package.³⁸ Processing including normalization for sequencing depth, estimation of dispersion, and calculation of fold-change metrics for each condition.

Sample outliers were identified using a combination of principal component analysis and correlation heatmap analysis on each separate condition. Samples that clustered apart from other replicates of the same condition were excluded from further analysis. One replicate was removed from each of the BPTES, rotenone, and tunicamycin conditions due to poor clustering with other samples of that condition.

Gene filtering was performed in two stages: a basic initial filtering step was implemented which removed genes with <10 total reads across all samples. A second filtering step was performed on a per-condition basis when differential expression analysis was performed. For each pairwise comparison (DMSO vs. condition of interest), a gene was kept if there were ≥3 samples in either condition with ≥25 counts for a given gene.

Differentially-expressed genes (DEGs) were identified for a given condition using Wald’s test, as implemented within the DESeq2 package. Each condition was compared pairwise to the DMSO control. Log₂ fold change (LFC) values for gene were shrunk using the adaptive shrinkage estimator (ashr package).⁴¹ p values were adjusted for multiple hypothesis testing using the false discovery rate (FDR) method of Benjamini-Hochberg, as implemented in DESeq2. DEGs were identified using cutoffs for fold change and adjusted p value, with cutoffs of 1.25-fold up or down and an adjusted p value cutoff of 0.05.

Gene set enrichment analysis (GSEA) was performed using the clusterProfiler package.^40,42 Hallmark and transcription factor gene sets from the Molecular Signatures Database (MSigDb) were incorporated into GSEA analysis using the msigdbr package. Transcription factor enrichment was performed using the TFT:GTRD and TFT:TFT_Legacy gene sets.

Metabolite analysis

Targeted metabolite feature extraction and area integration was performed using TraceFinder software (ThermoFisher) using a library of 299 compounds that were pruned to include metabolites quantified in at least 3 samples to remove outlier metabolite identifications. Metabolite analysis was performed using R. Metabolite areas were imported into R and pre-processed before analysis. Peak areas of 0 were imputed using 1/5 of the value of the lowest non-zero area for that metabolite. Peak areas were then normalized using median normalization, where areas for each sample were divided by the median peak area for that sample. Each timepoint and metabolite were analyzed using linear models. For the 1 h timepoint, the formula used was

log₂(area) ∼ condition

while the formula used at the 6 h timepoint was

log₂(area) ∼ condition + batch

to control for batch effects on a per-metabolite basis. This method was used to directly generate log₂ fold change and p values for each metabolite in each condition vs. DMSO. p values were adjusted using FDR correction for each set of condition results separately. Significant metabolites were defined as metabolites with a fold change of 1.25 or higher (or 0.8 or lower) with an adjusted p value cutoff of 0.10.

Metabolite Enrichment of Annotated Networks

MEAN uses a combination of a custom-defined metabolite sets and a generic GSEA function provided by the clusterProfiler package.^40,42 These custom-defined metabolite sets were assembled using a combination of manual annotation of metabolic pathways and pre-defined KEGG pathways. The metabolite sets can be manually adjusted for a researcher’s pathways of interest and for the metabolites that can be quantified using their metabolomics pipeline.

Multi-omics integration using mixOmics

Integration of the transcriptomics and metabolomics datasets was performed using the mixOmics package. Normalized counts from DESeq2 and normalized non-imputed peak areas from metabolomics analysis were transformed using a log₂(x + 1) transformation for use in mixOmics. Samples were pseudo-paired using random pairing for each condition, such that the pairing for each condition was done independently of other conditions. Datasets were integrated using sparse partial least squares differential analysis (sPLS-DA), using the block.splsda function within mixOmics. Sample pseudo-pairing was analyzed using component 1 and 2 plots to identify whether the overall shape of the data was consistent across different seed values for pairing. PLS component scores and loadings were visualized using ggplot.

TCGA PanCancer Atlas analysis

Data from the TCGA PanCancer atlas was downloaded from cBioPortal (cbioportal.org)⁴³ separated by cancer type. Individual cancer type mRNA data was merged by gene name across cancer types. Genes were matched between the mitochondrial stress dataset and the TCGA mRNA data using the clusterProfiler package to convert between Ensembl gene ID and gene symbols. Stress scores were calculated on log₂(x + 1) transformed mRNA counts. For mutational analysis, mutations were considered protein-altering if they were classified as frameshift insertions or deletions, in-frame insertions or deletions, missense mutations, nonsense mutations, nonstop mutations, mutations in splice regions or splice sites, or mutations at translation start sites. Survival analysis was performed using the Cox proportional hazards model and the Kaplan-Meier estimator, as implemented in the survival, condsurv, and ggsurvfit packages in R.

Published: April 14, 2025

Supplemental information

Table S1. Supplemental information for RNA-seq, related to Figures 1 and 2

Sample information for RNA sequencing, normalized counts generated by DESeq2 across all samples, and log2 fold change and significance calculations for each gene and condition compared to DMSO.

Table S2. Supplemental information for metabolic analysis, related to Figures 1, 3, and 4

Sample information for metabolic analysis, raw metabolite areas, normalized and imputed metabolite areas, and log2 fold change and significance calculations for each metabolite, timepoint, and condition compared to DMSO.