Targeting the TCA cycle can ameliorate widespread axonal energy deficiency in neuroinflammatory lesions

Inflammation in the central nervous system can impair the function of neuronal mitochondria and contributes to axon degeneration in the common neuroinflammatory disease multiple sclerosis (MS). Here we combine cell-type-specific mitochondrial proteomics with in vivo biosensor imaging to dissect how inflammation alters the molecular composition and functional capacity of neuronal mitochondria. We show that neuroinflammatory lesions in the mouse spinal cord cause widespread and persisting axonal ATP deficiency, which precedes mitochondrial oxidation and calcium overload. This axonal energy deficiency is associated with impaired electron transport chain function, but also an upstream imbalance of tricarboxylic acid (TCA) cycle enzymes, with several, including key rate-limiting, enzymes being depleted in neuronal mitochondria in experimental models and in MS lesions. Notably, viral overexpression of individual TCA enzymes can ameliorate the axonal energy deficits in neuroinflammatory lesions, suggesting that TCA cycle dysfunction in MS may be amendable to therapy.

Inflammation in the central nervous system can impair the function of neuronal mitochondria and contributes to axon degeneration in the c om mon n eu ro in fl am matory disease multiple sclerosis (MS). Here we combine celltypespecific mitochondrial proteomics with in vivo biosensor imaging to dissect how inflammation alters the molecular composition and functional capacity of neuronal mitochondria. We show that n eu ro in fl am matory lesions in the mouse spinal cord cause widespread and persisting axonal ATP deficiency, which precedes mitochondrial oxidation and calcium overload. This axonal energy deficiency is associated with impaired electron transport chain function, but also an upstream imbalance of tricarboxylic acid (TCA) cycle enzymes, with several, including key ratelimiting, enzymes being depleted in neuronal mitochondria in experimental models and in MS lesions. Notably, viral overexpression of individual TCA enzymes can ameliorate the axonal energy deficits in neuroinflammatory lesions, suggesting that TCA cycle dysfunction in MS may be amendable to therapy.
MS is a common neurological disease in which central nervous system (CNS) inflammation results in myelin loss and progressive neurodegen eration. While the importance of such neurodegenerative processes for longterm disability of patients with MS is well established [1][2][3] , the mechanistic links between CNS inflammation and neurodegenera tion remain to be resolved. Emerging evidence from patients with MS and the corresponding animal models indicates that neuronal mitochondria could be critical hubs that render inflammatory sig nals into neurodegenerative consequences. This concept is not only supported by the essential function of mitochondria in neuronal energy homeostasis 4 , but also by previous studies showing (1) that damaged mitochondria accumulate in axons located in experimental and human neuroinflammatory lesions 5-7 ; (2) that neuronal mitochon dria acquire DNA deletions over the course of the disease 8 ; and (3) that such inflammationinduced mitochondrial damage impairs electron transport chain (ETC) function 2,9 . Mitochondrial damage could thus be initiated already in the highly inflamed lesions of early MS and then further amplify over the course of the disease. This process would provide a link not only between inflammation and neurodegeneration, but also between initial relapsing-remitting and later progressive pathology 2,9 . As a result of these insights, mitochondria have emerged as a promising therapeutic Article https://doi.org/10.1038/s42255-023-00838-3 immunizing Thy1PercevalHR mice with myelinoligodendrocyte gly coprotein (MOG). In vivo twophoton imaging then allowed us to record the ATP/ADP ratio in axons that cross through a neuroinflammatory lesion and at the same time to stage axonal morphology, a predictor of axonal fate as we established previously 6 . We found widespread axonal ATP deficits that not only affected axons that were swollen (stage 1 axons) or fragmented (stage 2) and had hence already entered the focal axonal degeneration (FAD) process 6,14 , but also morphologically nor mal axons (stage 0; Fig. 1a-e). Axons were not only showing a reduced ATP/ADP ratio in acute lesions (analyzed 2 or 3 d after EAE symptom onset), but ATP deficits persisted in axons of all FAD stages in chronic lesions (analyzed an additional 20 d later; Fig. 1d,e). To explore, whether neurons in their entirety lacked ATP or whether this was a local axonal deficit, we traced individual sensory axons from the lesion back into the dorsal roots, where neuroinflammation is typically not present in EAE. Comparing the ATP/ADP ratios in the same axons close to and far from the intraspinal lesion area revealed that ATP deficits were most pronounced at the sites of CNS inflammation (Fig. 1f,g). To indepen dently corroborate these results with another biosensor, we intrave nously injected recombinant adenoassociated virus (rAAV.PHP.eB) to neuronally express ATeam, a fluorescence resonance energy transfer (FRET)based ATP level sensor (rAAV.hSyn:ATeam) 15 . In vivo imaging of ATeamexpressing axons confirmed a pervasive reduction in ATP availability in axons within spinal EAE lesions (Extended Data Fig. 2a-d), whereas in converse, a pH biosensor (SypHer3s; rAAV.hSyn:SypHer3s) 16 , failed to identify changes in axoplasmic pH that could interfere with the PercevalHR and ATeam sensors (Extended Data Fig. 2e). Hence, our data indicate that axons in neuroinflammatory lesions exist in a sustained state of reduced ATP availability, a finding that is in line with the localized accumulation and damage of mitochondria that we previously described in spinal cord lesions of the same MS model 6,17 .

ATP deficits precede mitochondrial redox and Ca 2+ changes
Oxidative stress to axons had previously been proposed as a mediator of mitochondrial damage in neuroinflammatory lesions [1][2][3]9,18 . Thus, we set out to determine whether and when an altered redox status of axonal mitochondria would emerge in relation to the onset of neuroinflammationinduced ATP deficit. We used Thy1mitoGrxroGFP mice, which allow recording of the glutathione redox potential in neu ronal mitochondria, as we previously established 19 . After crossing these redox reporter mice to Thy1-OFP mice 20 to allow staging of axonal mor phology, we induced EAE and performed ratiometric in vivo confocal imaging using singlephoton excitation at 405 and 488 nm at the peak of acute disease symptoms. An oxidative shift in mitochondrial redox state was only apparent in fragmented (stage 2) axons, whereas mito chondria in normalappearing (stage 0) and swollen (stage 1) axons were unaltered compared to control axons in a healthy spinal cord ( Fig. 2a-d). This is notable, as the analysis of individual mitochondria confirmed that damagerelated mitochondrial shape changes, such as rounding up (reduced shape factor), are already present at the onset of FAD (Fig. 2e) 6,17 . Decompensated mitochondrial redox stress thus seems to be a late event during inflammatory axon degeneration and affects target to prevent neurodegeneration throughout the disease course of MS. Unfortunately, so far, such strategies have failed to provide robust clinical benefits, at least in large clinical trials 10 . One of the reasons underlying this failure is that our understanding of how the molecular machinery and functional capability of mitochondria is impaired in neuroinflammatory lesions is still incomplete. Moreover, most stud ies so far have focused on ETC impairments 8,11 , which might be hard to target therapeutically given the complex structure of the underlying macromolecular complexes and their genetics.
Here we apply new in vivo imaging strategies combined with selec tive ex vivo proteomic analysis of neuronal mitochondria in a murine MS model (experimental autoimmune encephalomyelitis; EAE) to investigate the molecular underpinnings and functional consequences of inflammationinduced mitochondrial pathology. We reveal that widespread axonal ATP deficiency is initiated in acute neuroinflamma tory lesions and persists in the chronic disease stage. These bioener getic deficits precede mitochondrial redox or calcium dyshomeostasis. Selective MitoTagbased proteomic analysis of neuronal mitochondria isolated from acute and chronic EAE spinal cords instead revealed an imbalance of critical TCA cycle enzymes, with a prominent loss of several enzymes, including isocitrate dehydrogenase 3 (Idh3) and malate dehydrogenase 2 (Mdh2). Notably, the depletion of these key TCA cycle enzymes in neuronal mitochondria is also apparent in human MS lesions. Finally, we show that viral gene therapy, overexpressing the catalytic subunit of Idh3 or Mdh2, can partially reverse axonal ATP deficits in neuroinflammatory lesions. Our study thus provides a refined understanding of how the molecular composition of neuronal mitochondria changes in response to neuroinflammation. We identify the depletion of TCA cycle enzymes as a critical mediator of the axonal energy crisis that occurs in neuroinflammatory lesions, thus defining a potential target for therapeutic intervention.

Pervasive axonal ATP deficits emerge early in EAE lesions
As ATP production is a central function of mitochondria that is key in axons 12 , we established an in vivo imaging approach to record the ATP/ ADP ratio of individual axons in neuroinflammatory lesions. We gener ated Thy1PercevalHR mice, in which neurons express PercevalHR, a genetically encoded excitation ratiometric biosensor that monitors the cytoplasmic ATP/ADP ratio 13 . Histological characterization confirmed the targeting of the sensor to a broad range of neuronal populations in the brain and spinal cord (Extended Data Fig. 1a). To assess whether the sensor was indeed capable of recording ATP deficits in spinal axons, we pharmacologically interfered with ATP production and recorded the ATP/ ADP ratio by measuring fluorescence emission after excitation at 950 and 840 nm using in vivo twophoton microscopy. Application of either the glycolysis inhibitor iodoacetic acid (IAA) (10 mM) or the mitochondrial uncoupling agent carbonyl cyanide mchlorophenyl hydrazone (CCCP) (100 µM) to the exposed dorsal spinal cord resulted in a swift reduction of the ATP/ADP ratio in spinal axons (Extended Data Fig. 1b,c).
To assess the emergence of axonal energy deficits in neuroinflam matory lesions, we next induced EAE, a widely used model of MS, by Article https://doi.org/10.1038/s42255-023-00838-3 only a small proportion of nonfragmented axons in neuroinflamma tory lesions. Oxidative mitochondrial damage per se is hence unlikely to underlie the widespread axonal ATP deficits in EAE.
In addition to oxidative damage, calcium overload was proposed to mediate mitochondrial dysfunction 1,21 . To monitor calcium handling of axonal mitochondria in vivo, we generated a transgenic reporter mouse line that expresses the ratiometric calcium sensor in the spinal cord demonstrated that the sensor can record relevant pathological calcium signals in vivo (Extended Data Fig. 1d-f). We then crossed these reporter mice to Thy1OFP mice to relate axonal morphology to mitochondrial calcium levels and shape in acute EAE lesions. As with our analysis of mitochondrial redox state, we found that mitochondrial calcium levels are primarily increased in frag mented (stage 2) axons, that is, during the end stage of FAD, whereas normalappearing (stage 0) and swollen (stage 1) axons showed no pronounced calcium dyshomeostasis, despite clear morphological signs of organelle damage (Fig. 2f-j). Taken together, these experiments indicate that overt dysregula tion of the mitochondrial redox state and calcium handling are rather late events during inflammatory axon degeneration. Hence, a molecu larly distinct dysregulation of mitochondrial function and molecular composition must explain the prodromal and pervasive ATP deficits in EAE axons.

MitoTag analysis reveals depletion of ETC and TCA cycle components
To obtain an unbiased characterization of the neuroinflammatory changes to the molecular makeup of neuronal mitochondria, we adapted a celltypespecific proteomics approach that we recently established 23 . For this, we performed intrathecal injections of a rAAV. hSyn:Cre virus into the ventricles of MitoTag neonates, which resulted in the tagging of the outer membrane of neuronal mitochondria with green fluorescent protein (GFP). This enabled selective isolation and subsequent mass spectrometry analysis of these mitochondria from the spinal cords of healthy mice, as well as from different stages of EAE ( Fig. 3a Table 1), which revealed pronounced changes to the mitochondrial proteome of neurons. These changes could not be predicted by protein lifetime, location inside the mitochondria or whether a protein was encoded in the nucleus versus the mitochondrial DNA (Extended Data Fig. 3d-f). Gene set analysis of the major dysregulated pathways converged on the ETC and TCA cycle, which are abundant in healthy neuronal mitochon dria, but depleted in acute neuroinflammatory lesions (Fig. 3c,d and Extended Data Fig. 3g,h). The ETC complexes were uniformly affected, typically more strongly in acute than in chronic EAE lesions ( Fig. 3d and Extended Data Figs. 3g,h and 4a). This corresponded with reduced axon complex IV (COX) activity as measured in situ by a histochemical assay 24 (Extended Data Fig. 4b,c) and corroborated the previously described neuronal ETC dysfunction in neuroinflammatory lesions 5 . Notably, however, TCA cycle enzymes also showed pronounced dysregulation (Fig. 3d,e). To corroborate these proteomic findings in EAE, expand them to MS and probe their bioenergetic significance, we focused our further analysis on Idh3, which mediates the irreversible oxida tive decarboxylation of isocitrate and Mdh2, which oxidizes malate to oxaloacetate. Both enzymes generate NADH, which is subsequently used for ATP generation by the ETC, with Idh3 being ratelimiting for the TCA cycle, while Mdh2 catalyzes a key reaction that links the TCA cycle to anaplerotic reactions 25,26 . At the same time, Idh2, which accelerates the reversible NADPdependent conversion of isocitrate to αketoglutarate, seems to be differentially affected from Idh3 and Mdh2, as its abundance is unchanged in proteomes from acute and increased in chronic neuroinflammatory lesions (Fig. 3d,e). Expres sion changes of key TCAcyclerelated enzymes, including Idh3, Idh2 and Mdh2, were also present at the transcriptional level as revealed by analysis of published RiboTag translatomes of spinal motoneurons in EAE 27 (Fig. 3f). Hence, neuroinflammation caused a marked alteration in the expression of TCA cycle enzymes, including the ratelimiting enzyme Idh3, which likely hinders neuronal mitochondria from gen erating sufficient amounts of ATP.

Axonal Idh3 and Mdh2 are depleted in EAE and MS lesions
We next aimed to confirm the proteomics results by immunofluo rescence analysis of neuronal mitochondria in EAE, as well as in MS lesions. First, we quantified the expression levels of a subset of TCA cycle enzymes, for which the neuronal MitoTag analysis predicted dysregulation in EAE. We performed immunostainings on spinal cord sections from Thy1mitoRFP mice 19 , where analysis can be restricted to red fluorescent protein (RFP)tagged neuronal somata mitochondria. Indeed, the levels of Idh3a (Fig. 4b,d), as well as Mdh2 (Fig. 4e,g) were decreased by more than 50% in EAE lesionadjacent areas compared to controls. As predicted by our proteomics analysis, the expression of Idh2 was slightly increased (Fig. 4h,j). We further confirmed the deple tion of Idh3a and Mdh2 in axons per se, the neuronal compartment where we detected ATP depletion (Fig. 4c,d,f,g). Second, to explore whether similar expression changes of TCA cycle enzymes were also present in MS lesions, we performed immunofluorescence analysis of axonal Idh3a, Idh2 and Mdh2 in brain biopsy and autopsy sections from seven patients with MS (Extended Data Table 1). As fixation protocols are typically variable in patientderived material, we internally normal ized the expression of a given TCA cycle enzyme between the lesion area and the adjacent normalappearing white matter (NAWM) on the same section (Fig. 5a). This analysis revealed a reduction of Idh3a and Mdh2 in MS lesions ( Fig. 5b-e), which was also apparent in chronic active MS lesions, indicating that depletion of TCA cycle enzymes persists long term. Notably, Idh2 expression was not markedly affected, as predicted from the animal model ( Fig. 5f,g). Taken together, our findings suggest that TCA cycle disruption is a persistent neuronal change during MS lesion formation and progression.

Idh3a or Mdh2 overexpression ameliorate axonal ATP deficits
Considering the pronounced depletion of Idh3a in EAE and MS, as well as its critical role in the TCA cycle, we explored whether overexpres sion of individual depleted TCA cycle components, such as Idh3a,  the enzyme's catalytic subunit 28 , could restore axonal ATP levels in neuroinflammatory lesions. We initially focused on Idh3 because of its role as a pacemaker enzyme of the TCA cycle 29 and the substan tial translational efforts that are underway to target Idh enzymes as these are mutated in several cancers [30][31][32][33][34][35] . For overexpressing Idh3a together with a fluorescent marker (tdTomato), we used the systemic injection of a rAAV.PHP.eB virus, which drives transgene expression via a panneuronal promoter (human synapsin) 36 . Confocal analysis confirmed that a substantial fraction of CNS neurons was transduced with the rAAV.hSyn:Idh3atdTomato virus as indicated by tdTomato expression. Moreover, tdTomatolabeled neurons in EAE spinal cords showed a marked increase in mitochondrial Idh3a expression (Extended Data Fig. 5). We then directly assessed the effects of Idh3a restoration on axonal energy deficits in neuroinflammatory lesions by 49    inducing EAE in Thy1PercevalHR mice injected with either the rAAV. hSyn:Idh3atdTomato virus or a control rAAV.hSyn:CretdTomato virus, carrying a similarly sized, but in our setting inert transgene. In vivo imaging of the surgically exposed spinal cord in EAEinduced animals again confirmed strong tdTomato expression in a fraction of axons, while other PercevalHRexpressing axons were tdTomatonegative and provided an internal control (Fig. 6a,b). Compared to this control axon population (which rules out general effects on lesion activity by the virus), transduction with a rAAV.Syn:Idh3atdTomato virus sig nificantly increased the ATP/ADP ratio and thus partially restored the axonal energy deficiency in neuroinflammatory lesions (Fig. 6c). No such rescue was observed in axons transduced with the control rAAV. hSyn:CretdTomato (Fig. 6d). Notably, ATP/ADP ratios rose in axons of all damage stages, suggesting that correcting Idh3a levels could also counteract ATP deficiencies in more advanced stages of FAD. By extending the observation period from the acute phase after the EAE peak by 3 weeks, we confirmed that both the reduction in ATP/ADP ratio and a reversing effect of Idh3a overexpression persisted, albeit the latter's effects seemed to abate (Extended Data Fig. 6a-c). Next, we asked whether the effect of overexpressing Idh3a would be unique to this enzyme by targeting Mdh2 in a similar way. We observed that also viral gene transfer of Mdh2 partially restored axonal ATP/ADP levels in neuroinflammatory lesions, at least in advanced stages of axon damage (Extended Data Fig. 7).
Overall, these results suggest that a combinatorial targeting of the TCA cycle, or rather a metabolic approach to boosting TCA cycle func tion, will be needed to more optimally remedy axonal bioenergetics in neuroinflammation; indeed, when we explored the effects of Idh3a overexpression on axonal morphologies in tdTomatopositive versus tdTomatonegative axons, we could not detect robust differences in axonal swelling or fragmentation at 3 weeks after disease onset (Extended Data Fig. 6d).

Discussion
That axons in neuroinflammatory lesions exist in a state of 'energy fail ure' linked to mitochondrial damage has long been inferred based on (1) abnormalities in mitochondrial density, dynamics and morphology; (2) mitochondrial dysfunction as revealed by in situ assays of respiratory complex function or mitochondrial potential; and (3) the accumulation of mitochondrial DNA mutations in inflamed CNS tissue 1,2,18 . Further more, there is molecular evidence that axons in MS lesions might exist in a hypoxic state, which would further restrict mitochondrial respira tion 37 . These data have directed most attention to the ETC, which covers most of the uniquely high energy demands of neurons and determines the mitochondrial potential and other aspects of respiration that can be assayed in situ, for example, using enzymatic assays. Thus, the focus on the ETC is both well justified by its importance, but also biased by experimental accessibility. At the same time, the ETC is not necessarily a good therapeutic target, given its complex molecular makeup at the protein, as well as at the genome level, the extremely low turnover of some of its components and its Janusfaced nature as the major source not only of ATP, but also of oxygen radicals. Finally, the ETC needs to be fed with a steady and sufficient flow of redox equivalents and other substrates, most of which derive from the TCA cycle. Hence the TCA cycle, as the 'upstream' bioenergetic hub of metabolism, also deserves attention in the context of the presumed 'energy failure' of axons in neuroinflammation 38 . Especially so, as the TCA cycle is increasingly emerging as a metabolic hub relevant to neurodegeneration 25 and a 'druggable' target, for example in cancer 31,39 .
Against this backdrop, the present study makes the following con tributions: (1) We demonstrate directly by in vivo biosensor measure ments (both using PercevalHR to measure the ATP/ADP ratio, as well as ATeam, which assays ATP levels) that most axons in neuroinflammatory lesion exist in a state of depressed cytoplasmic ATP availability, already during early stages of the axonal degeneration pathway. Moreover, this change precedes overt signs of mitochondrial redox or calcium dyshomeostasis. (2) We show by molecular profiling of neuronal mito chondria not only a disruption of the ETC, but a similarly profound alteration of the 'upstream' TCA cycle, where a number of enzymes, including Idh3 and Mdh2, are depleted in neuronal mitochondria in EAE and MS lesions. (3) Finally, we demonstrate that overexpressing Idh3a (the catalytic subunit of the enzyme) or Mdh2 suffices to increase axonal ATP provision, albeit to a limited extent. This suggests, on the one hand that the axonal energy crisis caused by neuroinflammation is due to a dual deficiency, both of redox substrate provision from the TCA cycle to the ETC, as well of oxidative phosphorylation itself. On the other hand, a full rectification of these neuroenergetic deficits will likely require a manipulation of either a master regulator or multiple enzyme targets in parallel.
Together, the results we present here expand our understanding of the molecular pathogenesis of immunemediated mitochondrial damage and propose new avenues for therapeutic intervention. They also raise the question of where the dysregulation of the TCA cycle originates. Overall, the pattern of mitochondrial proteome changes does not relate to any simple pattern that we could discern. While the overall mitochondrial mass in neurons cannot be estimated eas ily by this method, the data suggest that within the mitochondrial proteome of neurons, specific dysregulation happens, as for example some biosynthetic pathways (such as mitochondrial translation) are upregulated, while most TCA cycle and ETC components are sup pressed. Indeed, the parallel and progressive upregulation of Idh2, while Idh3 is downregulated, suggests a specific metabolic 'rewiring' response 40 , rather than just a global loss of mitochondrial biogen esis or increase in mitophagy. Most likely, this response involves biosynthetic, as well as proteostatic mechanisms, as the observed dysregulation of mitochondrial proteins does neither correspond to whether a protein is encoded in the mitochondrial DNA (where mutations are known to accumulate in MS) 2,8 versus in the nucleus, nor with the estimated lifetime 41 . Indeed, the correlation of published neuronal translatome data from a similar model 27 with our neuronal proteomes provides direct support for an element of 'anterograde regulation' from the nucleus, while the long lifetimes of some down regulated enzymes (including subunits of Idh3 (ref. 42)) argue for a parallel degradation, given the swift development of the proteomic phenotype in our acute EAE model (which is present in acute lesions 2 d after disease onset).  Irrespective of the origin of the reduced ATP availability that our biosensor measurements reveal, the further question arises of what this implies for axonal survival and hence, given the central role of axon degeneration in MS disease course 43 , for possible lasting functional consequences. While limited ATP availability is often cited as a suf ficient argument to assume axonal demise and profound and lasting ATP deficiency can clearly cause axon degeneration 44  collapse of axons in neuroinflammation. Indeed, we find reduced ATP/ ADP ratios, ATP levels and TCA cycle enzyme expression in the major ity of axons, even those with a normal morphology. From previous work, we know that only a subset of these axons is at immediate risk of degeneration and some even recover morphological and mitochondrial integrity 6 . Notably, a similarly pervasive pattern of dysfunction has been apparent in our previous studies of axonal transport 17 . As axonal transport is highly ATP dependent and in converse, mitochondria are a major axonal transport cargo 4 , a vicious cycle could result in the longterm breakdown of the ATP supply chain, which might not cause immediate axon degeneration, but rather subacute dysfunction. Fur thermore, previous reports of in situ ATP measurements in mouse white matter tracts have shown that neuronal activity puts an immediate strain on energy supply 45 , which could sensitize specific activityrelated    patterns of degeneration. To explore this notion, but also to better understand the relative importance and metabolic consequences of the two aspects of the axonal energy deficiency emerging from the dual dysregulation of the ETC and the TCA cycle, we also performed metabolic modeling 46 (Methods provides details) based on our analysis of mitochondrial proteomes from control and EAE spinal cords. The modeling predicted a diminished absolute availability of ATP, as well as a reduced ATP/ADP ratio, and also suggested that indeed, with growing 'metabolic load' (an increased ATP consumption rate, which in neurons could for example, relate to increased rates of action potential firing and neurotransmission), these deficits should progress (Extended Data Fig. 8a). Furthermore, in silico rectification of groups of proteins to their control levels revealed that while 'normalizing' the ETC had some effect on ATP levels, the TCA cycle enzymes are predicted to be the more efficacious target (Extended Data Fig. 8b). So how to counterbalance the axonal energy deficiency? Indeed, 'brain energy rescue' has been hailed as a possible therapeutic strategy across neurological disorders 47 . The most direct approach would be the provision of energy substrates or key cofactors for the function of bioenergetic enzymes. Indeed, first trials along these lines, for example providing biotin, have been conducted in MS, but failed 48 . Notably, biotin is required for the activity of enzymes that feed into the TCA cycle upstream of Idh3 and Mdh2. Our data on TCA cycle disruptions, com bined with the previous insights on ETC dysfunction, would argue that simply feeding a substrate or cofactor into a broken supply chain will indeed not be efficient. Therefore, key bottlenecks in the ATPproviding metabolic pathways need to be unblocked in addition. Our data show, in accordance with our modeling predictions, that while addressing single bottlenecks, such as Idh3 (a key 'pacemaker' enzyme of the TCA cycle 29 ) might have some benefits, such a focused intervention would only result in limited changes to the neuronal energy state. While the fact that Idh3 activity can be directly allosterically stimulated by a low ATP/ADP ratio 49 and activated by calcium 50 , two changes that are induced in axons in response to an inflammatory challenge, argue that Idh3a overexpression might have a comparably strong unblocking effect on the TCA cycle, our results with Mdh2 overexpression show that multiple target points exist within this pathway upstream of the ETC. For actual benefits on axonal survival and neurological deficits, we assume that a combinatorial intervention will be needed; in accordance with our observation that even if we extended Idh3 overexpression for several weeks after EAE onset, no clear reduction in axon pathology was apparent. Moreover, we expect the effects of neuroenergetic dystrophy to really only manifest in the slowevolving late states of chronic neuro inflammation that likely drive progression in MS. Unfortunately, animal models of this key phase of MS largely remain elusive. Thus, while the data presented here provide proof of the principle for TCA cycle target ing, they probably do not yet foreshadow a true translational strategy.
Still, the fact that the TCA cycle enzymes are the focus of sub stantial biomedical interest 51,52 , given the key role of some of them (such as Idh1 and Idh2, but also Mdh2) in cancer biology, suggests that multipronged targeting of TCAcyclerelated pathways in neuroinflam mation might be a viable translational development. Notably, Idh3 seems to be dysregulated in some neoplasms, including glioblasto mas 53 . This has resulted in the development of Idh targeting drugs, for example for certain leukemia and brain tumor subtypes 52,53 . In cancer, the detrimental consequence of Idh mutations seems to be metabolic rerouting that enhances tumor cell proliferation and immune cell dysregulation 34 . Indeed, also in neuroinflammation some aspects of axonal dysfunction resulting from TCA cycle disruptions could be due to metabolite imbalances rather than mere lack of efficient fueling of the ETC and hence ATP deficiency. In this context, the concomitant upregulation of Idh2 in chronic stages of our model is notable and raises important questions for future exploration, for example by metabo lomic analysis of celltypespecific mitochondria. Such analysis could be combined with pharmacological intervention, to reroute substrates into the diminished bioenergetic pathways of axons in an inflammatory milieu. For instance, could the available Idh blockers, most of which are deliberately not targeting Idh3, still be useful in forcing metabolic flow back into the TCA cycle? Similarly, could approaches of increasing mitochondrial mass (such as overexpression of PGC1α, which is benefi cial in inflammatory demyelination and regulates Idh3a expression as part of its broad transcriptional effects 11,52 ) be synergistic with targeting the TCA cycle to unblock energy provision in axons and resolve their dystrophy in conditions like MS? The identification of the TCA cycle and some of its key enzymes as a disease contributor and a potential therapy target indeed opens the possibility for such a comprehensive evaluation of strategies to mitigate the neuronal energy crisis in the inflamed CNS.

Animals
All experiments were performed on either postnatal day 3 pups or adult mice according to the protocols on a C57BL/6 (strain designa tion C57BL/6J, Jackson Laboratories) background at age 2-6 months. All animals were bred and housed under standard conditions with a 12h light-dark cycle at a temperature of 22 ± 2 °C and 55 ± 10% relative humidity. Food and water for mice were provided ad libitum. Female and male were equally allocated into control and experimental groups if not explicitly mentioned otherwise. The experiments were not powered for independent analysis of male and female mice, but separate analysis of female and male mice did not reveal any major sexspecific effects in our analyses of ATP/ADP levels. All experimental procedures were conducted in accordance with regulations of the relevant animal wel fare acts and protocols approved by the responsible regulatory office.
To measure neuronal mitochondrial redox changes in spinal axons, we used the Thy1mitoGrxroGFP mouse line, which selectively expresses the redox sensor GrxroGFP2 in neuronal mitochondria 19 . To visualize axonal morphology, these mice were crossed to Thy1OFP mice 20 , in which neurons are labeled by cytoplasmic expression of an OFP. To investigate the proteomic profile of neuronal mitochon dria in the context of neuroinflammation, we used the Gt(ROSA)26Sor MitoTag knockin mouse line generated by recombinasemediated cas sette (loxPflanked STOP) exchange into the Rosa26 locus that allows a Credependent expression of GFP targeted to the outer mitochondrial membrane. The MitoTag mouse line is available from The Jackson Labo ratory as JAX#032675 (Rosa26CAGLSLGFPOMM) 23 . Thy1mitoRFP mice, in which tagRFP is localized to the matrix of neuronal mitochon dria 19 were used to ascertain the localization of immunofluorescence signals to neuronal mitochondria.

Generation of reporter mouse lines
To measure the neuronal ATP/ADP ratio, a Thy1PercevalHR mouse lines were generated using a bluntend cloning strategy with Perceval GW1HR plasmid (Addgene, #49082) purchased from Addgene 13 . In brief, the GW1PercevalHR plasmid was digested with Xba1 and EcoRI enzymes, blunted and inserted into the Thy1vector cut beforehand with XhoI, blunted and dephosphorylated. After ligation and electropo ration, minipreps of ampicillinselected Escherichia coli cultures were performed, followed by verifying the correct plasmid orientation and enzyme digestions. The construct was transfected in human embryonic kidney 293T (HEK) cells and cytoplasmic expression was confirmed by confocal microscopy. Maxipreps of the adequate cloning candidates were carried out using QIAGEN kits. The Thy1PercevalHR plasmid was further linearized using PvuI and EcoRI restrictions enzymes.
To measure the calcium level in neuronal mitochondria, Thy1mitoTwitch2b mouse lines were generated. The Twitch2bpcDNA3 plasmid 22 was kindly provided by O. Griesbeck (Max Planck Institute of Neurobiology). The Twitch2b sensor was extracted using NotI restriction enzyme and inserted into the pCMV:mycmito plasmid (Addgene, #71542) cut beforehand with NotI restriction enzyme and Article https://doi.org/10.1038/s42255-023-00838-3 dephosphorylated. After ligation and electroporation, minipreps of ampicillinselected E. coli cultures were performed, followed by veri fying the correct plasmid orientation and correct enzyme digestions. The construct was transfected in HEK cells and verified by live imaging. After retesting the construct and its functional responsiveness in HEK cell culture, maxipreps of the adequate cloning candidates were carried out using QIAGEN kits. The pCMVmycmitoTwitch2b sequence was further extracted using PmI and XbaI, blunted and inserted into the Thy1vector cut beforehand with XhoI, blunted and dephosphoryl ated. The Thy1mitoTwitch2b plasmid was linearized using ZraI and AflIII restrictions enzymes. Sequencing of Thy1mitoTwitch2b and Thy1PercevalHR constructs was performed by Eurofins. The sample purity of the linearized DNA was determined using the absorbance ratio at 260 nm versus 280 nm. The concentration of the linearized DNA used for pronuclear injections was 45.2 ng µl −1 in 60 µl for Thy1PercevalHR (with a ratio A260:280 of 1.83) and 36.5 ng µl −1 in 120 µl (with a ratio A260:280 of 1.83) for Thy1mitoTwitch2b. The gen eration of the Thy1PercevalHR and Thy1mitoTwitch2b was conducted at the Transgenic Core Facility of the Max Planck Institute for Molecular Cell Biology and Genetics in Dresden by standard pronuclear injec tions into pseudopregnant host mice. For analysis of mitochondrial calcium level in spinal axons Thy1mitoTwitch2b mice were crossed to Thy1OFP mice.

AAV vector generation and virus production
The AAV.hSyn:Cre plasmid, which encodes for Cre recombinase expression specifically in neuronal populations under the control of the human synapsin promoter, was generated by replacing the DIOGFP sequence (Addgene, #50457) to create the AAV.hSyn back bone plasmid. In brief, the AAV.hSyn:DIOGFP plasmid was initially digested with SalI/XhoI, followed by incubating with BglII to remove DIO.GFP and ligated with GFP.ires.Cre, which had been excised from AAV.CMV:GFP.ires.Cre with BglII, using Quick Ligase according to the manufacturer's instruction. Further from AAV.hSyn:GFP.ires.Cre, GFP. ires was excised with BmgBI and the backbone was ligated to acquire the AAV.hSyn:Cre plasmid.
The AAV.hSyn:Cre.P2A.tdTomato plasmid used to express Cre and RFP dTomato specifically in neurons was purchased from Addgene (Addgene, #107738). The construct with the correctly oriented insert was introduced to Stellar Competent cells (Clontech, 636763) and the plasmid purification was performed with a QIAGEN Plasmid Maxi kit according to the manufacturer's protocol.
AAV vector packaging was performed using HEK 293T (ATCC, crl 3216) and produced as described. HEK 293T cells were transfected with pADhelper, AAVcapsid 9 (Addgene, #112865) or PHP.eB 36 (Addgene, #103005) and the AAVconstruct (molar ratio 1:1:1) using an RPMI: PEI incubation protocol. AAV vectors were collected from the supernatant with polyethylene glycol (SigmaAldrich, no. 25322683) solution and the cell pellets. Freeze-thaw cycles were performed to lyse the cells and the residual DNA from the packaging was further degraded with benzonase (SigmaAldrich, E1014). Following the purification procedure using iodixanol gradient ultracentrifugation, the virus was concentrated by subsequent centrifugation and incubation with formulation buffer (PluronicF68 0.001% in saline PBS). The product was then further collected (with a genomic titer of ~10 12 to 10 14 ) and stored in small aliquots at −80 °C.

Spinal cord laminectomy
Mice were imaged at the acute stage of EAE, 2 d after onset of disease (only animals with an EAE score ≥2.5 were included) and at onset plus 22 d for the chronic stage. The mice were anesthetized by i.p. injection of medetomidine (0.5 mg kg −1 ), midazolam (5.0 mg kg −1 ) and fentanyl (0.05 mg kg −1 ) and tracheotomized and intubated to minimize breath ing if needed. The dorsal spinal cord was surgically exposed as previ ously described 6 . In short, the skin was disinfected and incised along the spinal column. The lumbardorsal spinal cord was surgically exposed at the position of vertebrae L3 and L4 by performing a laminectomy and the surgical field was constantly superfused with prewarmed artificial CSF (148.2 NaCl, 3.0 KCl, 1.4 CaCl 2 , 0.8 MgCl 2 , 0.8 Na 2 HPO 4 and 0.2 NaH 2 PO 4 in mM) to stay moisturized. The vertebral column was then positionfixed on a spinal clamping device (Narishige STSa), allowing controlled movement in the x, y and z directions during the imaging session. A 3.5% agarose well was built up around the spinal opening and filled with artificial CSF. A standardized imaging protocol was used and at least three images were acquired in volume stacks penetrating up to 100 µm into the tissue from the spinal cord surface (meningeal level). Acute EAE was studied in animals that showed a clinical score ≥2.5 with confluent lesions present and image regions were defined based on the accumulation of infiltrated immune cells. During the imaging ses sions, mice were kept under constant anesthesia and their breathing and reflexes were controlled every 30 min. Animals showing signs of traumatic damage after laminectomy were excluded from the analysis.

In vivo multi-photon and confocal imaging
Imaging of Thy1-PercevalHR mice. To measure the axonal ATP/ADP ratio, the genetically encoded ATP/ADP indicator PercevalHR was imaged using sequential twophoton excitation at 950 nm and 840 nm. Emission was collected simultaneously in the cyan and yellow channels using emission barrier filter pairs (bandwidth of 455-490 nm and 526-557 nm) on the Olympus MPEresonant scanner system. Iodoacetic acid (IAA) (10 mM) or carbonyl cyanide mchlorophenyl hydrazone (CCCP) (100 µM) at a final concentration was added to the imaging solution to obtain the baseline signal of the PercevalHR sensor. For imaging of the effects of Idh3atdTomato and Mdh2tdTomato overexpression on the ATP/ADP ratio, we used the same settings as described above. In addition, tdTomato signals were recorded using a wavelength of 1,040 and collected in the red channel using a barrier filter of band width 655-725 nm.
Imaging of AAV.PHPeB-hSyn:ATeam injected mice. To determine the axonal ATP level, the genetically encoded ATP sensor, ATeam, was imaged using a wavelength of 840 nm to excite mseCFP and cp173mVenus fluorescent proteins. The emission signals were col lected simultaneously in cyan and yellow channels using emission barrier filter pairs (bandwidth of 455-490 nm and 526-557 nm) on the Olympus MPEresonant scanner system.
Imaging of Thy1-mitoTwitch2b × Thy1-OFP mice. To measure mito chondrial calcium levels in axons, the genetically encoded calcium indicator Twitch2b was excited using a wavelength of 840 nm to excite mCerulean3 and cpVenusCD fluorescent proteins simultaneously 22 . The signals were collected in the cyan and yellow channels. To reveal axonal morphology, the orange fluorescent signal was excited with a wavelength of 750 nm. The customized emission barrier filter pairs with bandwidths of 457487 nm, 500-540 nm (GaAsP detectors) and 560-600 nm (RXD2 detector) were used on the Olympus MPEFV1200 system. Images (12 bit) were acquired with a ×25/1.05 dipping cone waterimmersion objective, a pixel size of 0.26 mm per pixel or smaller, a dwell time of 2.0 µs per pixel and a laser power of 30 mW measured in the back focal plane. Volume stacks penetrating ~50 µm into the dorsal spinal cord from the surface were acquired with a Z spacing of 1 µm.

Imaging of Thy1-mitoGrx-roGFP × Thy1-OFP mice.
To determine the mitochondrial redox state in spinal axons, the genetically encoded redox indicator GrxroGFP was sequentially excited using onephoton excitation at 405 and 488 nm wavelengths as previously described 19 . The signals were collected with emission barrier filter pairs (bandwidth 492-592 nm) in separate channels using a 50/50 beam splitter on an Olympus FV1000 confocal system. Smaller stacks were acquired with a Z spacing of 1 µm at a frame rate of 0.1-0.2 Hz.

Image processing and analysis
Images were analyzed with the opensource image analysis software ImageJ (Fiji, http://fiji.sc/Fiji) and Photoshop (Adobe). EAE lesions were identified according to the accumulation of infiltrated immune cells. The morphology of individual axons was traced and assessed through multistacks. To determine PercevalHR, ATeam or SypHer3s signals in axons, background intensities (nonaxonal areas) of both channels were measured and subtracted for every axon in every experiment to cre ate a background mask and pixelbypixel ratios were calculated from the mean over three regions of the same axon and further normalized to the mean value of the control mice to eliminate batch differences. For representative images that do not illustrate intensity variations, maximum intensity projections of image stacks were γadjusted to enhance the visibility of intermediate gray values and processed with a 'Despeckle' filter to lower the detector noise using Photoshop soft ware (Adobe). To measure the redox state of a single mitochondrion or clusters of mitochondria signals in Thy1mitoGrxroGFP × Thy1OFP, the mean intensity values were measured in the 405 and 488 nm chan nels as previously described 19 . The values for the two channels were divided (405/488 nm) to obtain a ratio related to the redox state of the sensor. This ratio was normalized to the mean value of the control mice. Mitochondrial morphology was quantified as the mitochon drial shape factor calculated by dividing the length by the width of a single mitochondrion as previously described 6 . To measure the cal cium levels of a single mitochondrion or clusters of mitochondria in Thy1mitoTwitch2b × Thy1OFP, the mean intensity values were meas ured in the cyan and yellow channels. The FRET signal (YFP channel) was corrected by subtracting the measured crosstalk fraction of the CFP signal and is indicated as cFRET. The mitochondrial calcium levels were then expressed as the backgroundcorrected cFRET:CFP ratio normalized to the mean value of healthy mice as previously established for FRETbased calcium sensors 14 .

Mitochondrial isolation
Mitochondrial samples for mass spectrometry were prepared from adult male mice described in Fecher at al 23 . In brief, AAV. hSyn:Creinjected control and EAE MitoTag mice that thus express the GFPlabeled mitochondrial outer membrane in neurons were anesthe tized with isoflurane and transcardially perfused with 1× PBS/heparin. The lumbar spinal cord was dissected, weighed and homogenized with a Dounce glass homogenizer using three complete up and down cycles with an Atype pestle in isolation buffer on ice. The sample was then transferred to a celldisruption vessel and processed with nitrogen cavitation at 800 p.s.i. and under stirring at 60 r.p.m. for 10 min. After pressure release, a protease inhibitor (cOmplete, EDTAfree Protease Inhibitor Cocktail, SigmaAldrich, 5056489001) was added to the resulting tissue fraction, and subcellular sediments were removed through two times centrifugation at 600g for 10 min. The resulting postnuclear tissue fraction was filtered through a 30µm cell strainer. For immunopurification, the postnuclear tissue fraction was diluted to a maximal concentration of 2 mg tissue per ml in immunopurification Article https://doi.org/10.1038/s42255-023-00838-3 buffer (IPB) and 50 µl microbeads coated with mouse IgG 1 subtype antibody (Miltenyi Biotec, 130091125) against GFP was added to the sample and incubated for 1 h at 4 °C on the shaker. Magneticactivated cell sorting was applied to separate the microbeadcoated mitochon dria. The LS columns (Miltenyi Biotec, 130042401) were placed in a magnetic QuadroMACS separator (Miltenyi Biotec, 130090976) followed by the equilibration step with 3 ml IPB. The samples were applied to the column in repeating 3ml steps and washed three times with IPB. The columns were removed from the magnetic separator and the microbeadcoated mitochondria were gently flushed out with the plunger. Mitochondria were pelleted by centrifugation at 12,000g for 3 min at 4 °C and washed twice with isolation buffer (without BSA and EDTA) and the pellets were immediately stored at −20 °C for further experiments. Subsequently, the protein amount was determined using a BCA assay (Pierce BCA Protein Assay kit; Thermo Fisher Scientific, 23227) according to the manufacturer's instructions. BSA was used as the standard and the sample buffer was used to correct the measure ment alterations caused by detergent or BSA.

Sample preparation for mass spectrometry
Mitochondria were immunocaptured from the spinal cord according to the abovedescribed protocol. Samples were lysed in a modified RIPA buffer (1% Triton X100, 0.5% sodium deoxycholate, 0.1% SDS, 150 mM NaCl, 5 mM EDTA and 50 mM TrisHCl, pH 8). Cell debris and undissolved material were removed by centrifugation at 16,000g for 10 min at 4 °C. Protein concentrations were assessed using a BCA assay. A protein amount of 20 µg was further diluted with distilled water at a 1:2 ratio and 50 mM MgCl 2 was added to a final concentration of 10 mM and DNA/RNA was digested using 12.5 U benzonase. Disulfide bonds were reduced by adding dithiothreitol to a final concentration of 10 mM and incubation for 30 min at 37 °C. Cysteine alkylation was performed by adding iodoacetamide to a final concentration of 40 mM and incubation for 30 min at 24 °C in the dark. The alkylation step was quenched by adding 4 µl 200 mM dithiothreitol.
Protein digestion was performed using the singlepot, solidphase, sample preparation (SP3) protocol 55 . Briefly, 10 µl of a 4 µg µl −1 bead slurry of SeraMag SpeedBeads A and B (GE Healthcare) were added to 20 µg of alkylated protein lysate. Protein binding to the magnetic beads was achieved by adding acetonitrile to a final volume of 70% (v/v) and mixing at 1,200 r.p.m. at 24 °C for 30 min in a Thermomixer (Eppendorf). Magnetic beads were retained in a DynaMag2 magnetic rack (Thermo Fisher Scientific) and the supernatant was discarded. Detergents were removed using four washing steps with 200 µl 80% (v/v) ethanol. Proteins were digested with 0.25 µg LysC (Promega, V1671) at 37 °C for 3 h followed by a second digestion step with 0.25 µg trypsin (Promega, V5111) for 16 h at room temperature. Tubes were placed in a magnetic rack and peptides were transferred to 0.22µm Costar SpinX filter tubes (Corning) to remove the remaining magnetic beads. Samples were dried by vacuum centrifugation. Finally, peptides were dissolved in 20 µl 0.1% (v/v) formic acid. Furthermore, the peptide concentration was estimated using the Qubit protein assay (Thermo Fisher, Q33211).
The 15 mostintense peptide ions were chosen for fragmentation by higherenergy collisional dissociation (resolution, 15,000; isolation width, 1.6 m/z; automatic gain control target, 1 × 10 5 ; normalized colli sion energy, 26%). A dynamic exclusion of 120 s was applied for fragment ion spectra acquisition. For EAE samples, two technical replicates were measured per sample.

Liquid chromatography-tandem mass spectrometry data analysis
The LC-MS/MS raw data were analyzed with the software MaxQuant (v.1.6.3.3 or v.1.6.10.43) 56 . The data was searched against a one gene per protein canonical database of Mus musculus downloaded from UniProt (downloaded 14 March 2019, 22,294 entries). Carbamidomethylation of cysteines was defined as fixed modification, whereas oxidation of methionines and acetylation of protein N termini were set as variable modifications. Peptide mass recalibration using the first search option with 20 ppm mass tolerance was enabled. For the main search, pep tide and peptide fragment mass tolerances were set to 7 and 20 ppm, respectively. Peptide spectrum match and protein false discovery rates were set to 1% and false discovery rates were controlled using a forward and reverse concatenated database search approach. The option match between runs was enabled with a matching time window of 1.5 s. Protein labelfree quantification (LFQ) was performed based on at least two ratio counts of unique peptides per protein. For EAE, LFQ intensities of technical replicates were averaged. If a protein was detected in <50% of the samples within one experimental group (EAE or control), all values were imputed with values drawn from a standard distribution with a mean value which was 2 × s.d. below the mean LFQ value of the entire dataset and a 0.3fold s.d. of the overall s.d. of the dataset (to control for zero inflation). The remaining missing values (of proteins detected in >50% of the samples within one experimental group) were imputed by multiple imputations by chained equations (MICE, R pack age 'mice'). LFQ values were log 2 transformed. For subsequent analysis, LFQ values were treated as interval scale values. To test the similarity of samples within the respective experimental group and to identify outliers, principalcomponent analysis was performed. Subsequently, outliers were removed from further analyses. The first two principal components with their explained variance of each sample were visual ized. To test whether a protein was differentially expressed between the EAE and control group, we calculated a Student's ttest and a fold change of the LFQ values between the experimental groups. To control type I error inflation, P values were corrected according to Bonferroni. MitoCarta 2.0 (ref. 57) was used to annotate mitochondrial pro teins as well as their submitochondrial location. Gene set enrichment analysis was performed with the Python package nezzworker (https:// github.com/engelsdaniel/nezzworker) with REACTOME (v.7.4) as a reference gene set library. Proteins which are members of gene sets that showed the most extreme NESs were further analyzed. To test whether mitochondrial, nonmitochondrial proteins and proteins from different mitochondrial respiratory chain complexes were equally abundant in EAE and control samples, oneway ANOVA was calculated to test for statistically significant differences between the mean LFQ values. To assess transcriptional changes of motor neurons in EAE, we reanalyzed previously published data from Schattling et al. 27 , using a custom RDre script with the DESeq2 library. To assess the correla tion of protein lifetime and the alterations of neuronal mitochondrial proteome, we extracted the protein halflives dataset from Fornasiero et al. 41 . All analysis was performed in R studio (v.4.2.2) or Python (v.3, https://github.com/engelsdaniel/mitoproteomics).
Heatinduced antigen retrieval was performed on PFAfixed tis sue following deparaffination for human tissue sections. After 10% FCS/PBS unspecific binding blockade, sections were incubated with primary antibodies in Dako Diluent (Dako, 52022) overnight at 4 °C. After washing with Wash Buffer (Dako, 53006) and autofluorescence removal treatment (Merck, 2160), secondary antibodies and DAPI (Life, D3571) at a dilution of 1:200 were incubated at room temperature for 1 h and further mounted with Fluoromount Mounting Medium (Sigma, F4680). The use of human samples followed institutional ethical guide lines and was approved by the ethics committee of the University of Geneva. Written informed consent to use autopsy samples for research purposes was obtained for all samples, with the exceptions of autopsies that were performed more than 20 years ago. In all cases, no samples were used from patients who refused involvement in research projects. Patient information is provided in Extended Data Table 1. Data from the human cortex are representative of two experiments performed on seven different cases.

Confocal imaging and image processing
Sections stained by immunofluorescence labeling to quantify mito chondrial proteins in neuronal somata, were imaged with an upright Olympus FV1000 confocal system equipped with ×10/0.4 air, ×20/0.85 and ×60/1.42 oil immersion objectives or a Leica SP8 equipped with ×20/0.75 HC PL APO CS2 and ×40/1.30 oil immersion HC PL APO CS2 objectives. Images were obtained using standard filter sets and pro cessed with Fiji. For representative figure panels, different channels of image series were pseudocolorcoded in Fiji or Adobe Photoshop; schematics were created using BioRender.com. Contrast and bright ness were equally adjusted across the entire image. For the panels dis played in Fig. 4 for intensity comparison, immunofluorescence images of both control and EAE Thy1mitoRFP tissues were acquired with the same settings and were adjusted with the same processing parameters. In panels that do not primarily illustrate quantitative differences, γ was adjusted nonlinearly to enhance the visibility of lowintensity objects. Figures were assembled in Adobe Illustrator.
Human tissue sections were scanned using a whole slide scanner (PANNORAMIC P250 II, 3DHistech). Regions of interest were manually selected using SlideViewer software (v.2.3, 3DHistech) and exported as individual images for further processing using a custom rulebased script in Definiens Developer XD (v.2.7, Definiens). In short, specific signals for each marker were detected based on their respective inten sity and ratio to tissue background intensity. NF70 (neurofilament) was detected first and Idh3a detection was restricted to the inside of NF70positive structures. The total area was quantified for each marker. Additionally, mean intensities were exported for each object individually. For quantification of TCA cycle enzymes in mouse axons, we followed the same approach both for reasons of consistency, but also because mitochondria in axons can be difficult to stain in nonparaffinembedded tissue. For this, slides of Thy1mitoRFP mice stained with NF70 together with the mitochondrial markers Mhd2, Idh2 or Idh3a were imaged on Leica SP8 with ×40/1.30 oil immersion using standard filter sets and processed with Fiji. Notably, RFP fluorescence is abolished in this protocol, so independent masking of mitochondria based on a mitochondrial transgene or staging of FAD were not possible in this dataset. Image analysis was further performed in Visiopharm (v.2021.09). In summary, NF70positive structures were detected and taken as the reference regions. Mitochondrial markerpositive puncta were then detected by a 120% intensity cutoff and a minimum size 4 pixel cutoff and quantified strictly within NF70positive structures. The total area was quantified for NFpositive structures and markerpositive puncta, with the ratio representing 'occupancy'. Within the thus defined mitochondrial voxel, the mean intensity value for each individual object was quantified and the integrated density represents the sum of the values of the pixels in the image or selection, equivalent to the product of occupancy and mean intensities. Data were processed using R (v.4.2.2, Rproject.org).

Sequential cytochrome c oxidase histochemistry and immunofluorescence histochemistry
To assess COX activity in single axons, COX histochemistry was com bined with immunofluorescence histochemistry with a primary antibody against OFP followed by a directly conjugated secondary antibody. This sequential technique was performed in the same tis sue section and has already been described and validated in previous studies 5,24,58 . The COX medium consisted of 100 µM cytochrome c, 4 mM diaminobenzidine tetrahydrochloride and 20 µg ml −1 catalase in 0.1 M phosphate buffer at pH 7.0. Cryosections were incubated at 37 °C for 30 min and washed in PBS. The cryosections were processed through the immunofluorescent histochemistry steps as mentioned above. The sequentially stained sections were mounted in glycerol with Hoechst nuclear stain and stored at −20 °C until required for imaging by the Zeiss Imager Z1 Apotome 2 microscope.
To assess COX activity in axons, a mask of the individual axon identified by the fluorescent labeling of OFP was generated in ImageJ and superimposed onto the brightfield image of COX histochemistry. The total area occupied by COX active elements within a single axon was determined as a percentage of the axonal area. Twenty axons (at least 25 µm in length) per region were randomly selected from each animal for quantitation. Assessors were blinded by coding the axons in ascending numerical order.

Metabolic modeling
Metabolic modeling was performed using the QSM data analysis platform provided by Doppelganger Biosystem, using the mass spec trometry analysis described above for MitoTagderived neuronal mitochondria from acute (2-3 d following clinical onset) EAE versus controls (n = 6 and 5 mice, respectively). A detailed description of the approach used can be found in Berndt et al. 46 . In brief, the QSM kinetic model includes the major cellular metabolic pathways of mitochon drial energy metabolism and glycolysis, as well as key electrophysi ological processes at the inner mitochondrial membrane (membrane transport of various ions, mitochondrial membrane potential, gen eration and utilization of the proton motive force). concentrations of a health neuronal tissue 46 . To establish individual metabolic models, the approach uses the protein intensity profiles of quantitative shotgun proteomics of MitoTagderived neuronal mitochondria. The maximal activities of enzymes and transporters are scaled, exploiting the fact that the maximal activity of an enzyme is proportional to the abundance of the enzyme protein via v max (EAE) = v max (normal) × (E(EAE) / E(mean control)). v max (normal) denotes the maximal activity for a given enzyme as used by Berndt et al. 46 , E(EAE) denotes the protein abundance measured of enzyme E in a given EAE sample and E(mean control) represents the mean protein abundance of an enzyme E averaged across the control sample measurements. Applying the model involves tailored quality control (QC), which evaluates both the number of proteins of interest found (QC score) and the number of metabolic processes associated with the enzymes found (QSM score), which indicated good suitability for applying the QSM kinetic model (QC > 90%; QSM > 100%, with 75% being the standard cutoff). To evaluate energetic capacities, the model calcu lates the changes of metabolic state due to increasing the rate of ATP consumption above the resting value, with the ATP consumption rate being modeled by a generic hyperbolic rate law of the form v ATP = k load × (ATP / ATP + K m ). To model increased 'metabolic load', the parameter k load was increased in steps until convergence of the ATP production rate to its maximal value.

Statistics and reproducibility
All experiments in this study include at least three biological replicates, except in Extended Data Fig. 2e. The number of biological replicates (n) is mentioned for each experiment in the figure legend. No statisti cal methods were used to predetermine sample size, but our sample sizes are similar to those reported in previous publications 6,14,17,23 . Data collection and analysis were performed blind to the conditions of the experiment, unless this was not possible due to an obvious dis ease or labeling phenotype. Statistical analysis was performed using Microsoft Excel software and GraphPad Prism (GraphPad Software, v.7.0). Sample sizes were chosen according to previous in vivo imag ing studies of spinal axons 6,14 . Normality of distribution was assessed with the Shapiro-Wilk test. When normal distribution was confirmed, a twotailed Student's ttest was applied for two groups comparison and oneway ANOVA followed by Bonferroni's or Turkey's post hoc comparison was used for more than two groups. The Kruskal-Wallis test followed by Dunn's multiple comparisons test or Mann-Whitney Utest was used where normal distribution could not be confirmed.
Obtained P values were stated as significance levels in the figure legends (****P < 0.001, ***P < 0.005; **P < 0.01; *P < 0.05; NS, not significant). Data are expressed as the mean and the error bars indicate s.e.m. unless specified otherwise in the legends.

Reporting summary
Further information on research design is available in the Nature Port folio Reporting Summary linked to this article.

Data availability
All data are available in the manuscript or the supplementary materials. Raw data are available upon reasonable request to the correspond ing authors. The transgenic mouse lines (Thy1mitoTwitch2b and Thy1PercevalHR) and the plasmids (pAAV.hSyn:Idh3a.P2A.tdTomato, pAAV.hSyn:mMdh2.P2A.tdTomato, pAAV.hSyn:Ateam1.03 and pAAV. hSyn:SypHer3s) are available on request. All proteomic datasets gener ated within this study are deposited online to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD032363 and at the GitHub repository (https://github.com/engels daniel/mitoproteomics). The codes used to reanalyze the singlecell data (Fig. 3f) from Schattling et al. are publicly available on the GitHub repository (https://github.com/engelsdaniel/mitoproteomics). Source data are provided with this paper.

Extended Data Fig. 4 | Expression of ETC components in EAE neurons.
(a) Top: Schematic of the experiment, analysis of same data sets as shown in Fig. 3. Bottom: Relative abundance of individual ETC complex components in neuronal mitochondria. Average shown as colorcoded log 2 (EAE/Control) for acute and chronic EAE compared to respective controls. (b) In situ analysis of axonal COX IV activity using in situ histochemical assay in Thy1OFP EAE spinal cords. Top: Schematic of the experiment. Bottom: Confocal image of control and EAE stage 1 axons, OFP (red) and COX IV (arrow heads indicate mitochondria as dark areas, as fluorescence of OFP is quenched by reaction product of COX IV assay). (c) Quantification of COX IV activity signal's occupancy of OFP axon area on axonal level. Mean ± s.e.m.; n = 180 axons from nine mice for control and 208 axons from nine mice for EAE using a twotailed, MannWhitney test, p = 5.2 × 10 12 . Scale bar: 25 µm in b. ****, p < 0.001. See source data for individual data points and further statistical parameters. Illustration created with BioRender.