Hyperpolarized 13C metabolic imaging detects long-lasting metabolic alterations following mild repetitive traumatic brain injury

Career athletes, active military, and head trauma victims are at increased risk for mild repetitive traumatic brain injury (rTBI), a condition that contributes to the development of epilepsy and neurodegenerative diseases. Standard clinical imaging fails to identify rTBI-induced lesions, and novel non-invasive methods are needed. Here, we evaluated if hyperpolarized 13C magnetic resonance spectroscopic imaging (HP 13C MRSI) could detect long-lasting changes in brain metabolism 3.5 months post-injury in a rTBI mouse model. Our results show that this metabolic imaging approach can detect changes in cortical metabolism at that timepoint, whereas multimodal MR imaging did not detect any structural or contrast alterations. Using Machine Learning, we further show that HP 13C MRSI parameters can help classify rTBI vs. Sham and predict long-term rTBI-induced behavioral outcomes. Altogether, our study demonstrates the potential of metabolic imaging to improve detection, classification and outcome prediction of previously undetected rTBI.


INTRODUCTION
Individuals subject to frequent concussions such as career athletes (e.g. football players, boxers), accidental head trauma victims, domestic abuse victims, or active military are some of the population at risk for mild repetitive traumatic brain injury (rTBI). Indeed, rTBI is being steadily recognized as a risk factor for the development of epilepsy and neurodegenerative diseases, particularly chronic traumatic encephalopathy (CTE) 1,2,3 . However, to date, non-invasive diagnostic biomarkers of rTBI are lacking. In particular, clinical computed tomography (CT) and magnetic resonance imaging (MRI), the standard imaging methods for trauma patients, are unable to detect rTBI-induced pathology 4,5,6 . This lack of imaging techniques hampers proper diagnosis and appropriate clinical care, as a result new approaches are critically needed.
To further our understanding of rTBI, several models were developed over the past years 7,8 . The closedhead impact model of engineered rotational acceleration (CHIMERA) device was designed to deliver multiple subconcussive mild TBI in a controlled and reproducible manner 9 . CHIMERA-induced rTBI has been shown to lead to reproducible pathological and behavioral changes up to several months following impacts 10 , mirroring the long-term effects of rTBI seen in the clinic (review by McNamara et al. 11 ). Only a few studies have investigated the use of magnetic resonance imaging (MRI) to detect CHIMERA-induced rTBI. Diffusion MRI, which is sensitive to water diffusion in tissue and changes in tissue microstructure, detected subtle differences in rTBI animals at 7 days following injury and in the optic tract, brachium of the superior colliculus, corpus callosum and hippocampus regions 12,13 ; however, long-lasting changes were not studied. T 2 -weighted MRI, which is sensitive to changes in tissue microstructure, edema, and myelination, did not detect signs of brain injury at 7 days and 40 days post-injury 10,12 . All these studies are in line with clinical ndings, and further highlight the need for more sensitive approaches to detect and monitor long-term brain changes after rTBI.
Metabolic impairment following TBI has been well documented in patients and animal models in the hours following TBI using 13 C-labelled substrates infusion and metabolomics approaches (reviews by Jalloh, Dermers-Marcil, and Carpenter 14,15,16 ). Notably, cerebral microdialysis studies have identi ed that the lactate / pyruvate ratio (Lac/Pyr) parameter is associated with poor outcome 17,18 . However, as cerebral microdialysis is an invasive method, its use at chronic timepoints following TBI, or in closed head injury and concussion is not feasible. Hyperpolarized 13 C magnetic resonance spectroscopic imaging (HP 13 C MRSI) is a unique technology that allows to measure metabolic uxes in vivo, and to compute lactate / pyruvate ratio values as well. HP 13 C MRSI applications have been extensively described in the oncology eld 19,20 , and its use is emerging to probe brain metabolism in health and diseases 21 . HP 13 C MRSI of [1-13 C]pyruvate enables to monitor the conversion of this key metabolite into its product(s) [1-13 C]lactate and/or [ 13 C]bicarbonate in the brain, which provides unprecedented metabolic information 22,23 . Prior studies of moderate TBI have shown changes in the HP 13 C lactate / pyruvate ratio (HP 13 C Lac/Pyr), and HP 13 C bicarbonate / lactate ratio at early timepoints (4 hours up to 7 days) following injury in preclinical models and in patients 24,25,26  Here, we questioned if advanced imaging approaches that have never been applied to the study of rTBI could detect long-lasting alterations following rTBI in the CHIMERA model. In addition to the above described HP 13 C MRSI of HP [1-13 C]pyruvate and HP [ 13 C]urea, we also evaluated Susceptibility-weighted imaging (SWI) and T 1 mapping. SWI is an MRI method particularly sensitive to iron that can inform on venous deoxygenated blood and iron deposition in tissue, and which has proven very valuable to identify microbleeds in TBI, but has not yet been investigated in the CHIMERA model 32,33 . Recent reports have highlighted the potential of T 1 mapping to detect oxidative stress in the rodent brain 34 , and thus this technique holds great potential to detect the production of reactive oxygen species that may occur following diffuse axonal injury and in ammatory processes observed in the CHIMERA model 11 .
As shown in Fig. 1, we induced rTBI in two-month old male mice using the CHIMERA apparatus, tested risk-taking behavior 3 months post rTBI, and performed four MRI-based scans. We investigated the potential of 13 C MRSI of HP [1-13 C]pyruvate and [ 13 C]urea to detect metabolic and tissue perfusion impairment, T 2 -weighted MRI to assess structural changes (as clinical standard of MR imaging), T 1mapping to evaluate tissue microstructure alterations and oxidative stress, and SWI MRI to detect changes in tissue microstructure, microbleed and tissue oxygenation following rTBI. Last, brain tissue was collected to evaluate changes in enzymes activity and transporter protein expression. Given the multidimensional nature of the data, we used a Machine Learning (ML) approach to identify how measured parameters could best predict changes in risk-taking behavior and HP 13 C MRSI.

HP 13 C MRSI detects long-lasting metabolic alterations following rTBI
We investigated whether HP 13 C MRSI could be used as a non-invasive tool to detect rTBI-induced longlasting changes, and speci cally to question whether metabolic alterations are present at chronic time points following injury.
Altogether, our results indicate that HP 13 C MRSI can detect region-speci c long-lasting metabolic changes following mild rTBI.

Multimodal MRI does not detect long-lasting effect of injury in rTBI
We evaluated whether a comprehensive multimodal MRI approach could detect changes between rTBI and Sham mice at 3.5 months post-injury, when metabolic alterations where detected by HP 13 C MRSI.
We rst used T 2 -weighted MRI, the clinical standard of MRI approach, that is sensitive to in ammation and/or changes in myelin content. We found that T 2 signal intensities were not different between rTBI and Sham mice in any of the regions studied (prefrontal cortex, cortex, hippocampus and thalamus (subcortex)) ( Fig. 4.a-b). In addition, we did not detect any differences in brain region volumes between groups ( Supplementary Fig. 1). Next, we used a T 1 mapping sequence that was shown to be sensitive to changes in microstructure or alterations related to oxidative stress. Similar to the T 2 intensities, we did not observed any differences in the T 1 values in any of the region studied between rTBI and Sham mice ( Fig. 4.c-d). Lastly, a SWI sequence was used, as it is capable of detecting microbleeds as well as potential changes in oxygenation following injury. Once again, no differences in SWI values were observed between rTBI and Sham mice in any region (Fig. 4.e-f). We did not detect any microbleed lesions in any of the studied animals.
Altogether, our results indicate that a comprehensive multimodal MRI approach combining T 2 -weighted MRI, T 1 mapping and SWI was not able to detect any signs of injury in rTBI mice at 3.5 months postinjury, unlike HP 13 C MRSI.
Disrupted enzymatic activity, but not transporter expression, is observed at chronic time points after rTBI To further investigate potential underlying mechanisms responsible for the observed changes in HP 13 C MRSI readouts, we evaluated the activity of enzymes responsible for pyruvate conversion into its downstream metabolites and the expression of transporters that control the entry of pyruvate into cells and the e ux of metabolites outside of the cells.
In the brain, lactate dehydrogenase (LDH) converts pyruvate into lactate, and pyruvate dehydrogenase (PDH) controls pyruvate entry into the tricarboxylic cycle and its conversion into acetyl-coA. We observed that PDH was 1.6 fold lower in the prefrontal cortex and 1.7 fold lower in the cortex of rTBI compared to Sham mice (Table 1, p = 0.0044 and p = 0.0375, respectively). No differences in PDH were observed in subcortical areas that include the hippocampus and thalamus. The activity of LDH was not signi cantly different between rTBI and Sham mice for cortical and subcortical areas.  Machine learning identi es rTBI/Sham classi ers, and predictors of behavior and HP 13 C readouts The machine learning (ML) analysis included all the data described above, as well as behavioral data we previously reported in Krukowski et al. 35 , which showed that mild rTBI leads to increased risk-taking behavior in male mice at 100 days post-injury.
Given n = 20 (10 for rTBI and 10 for Sham) mice along with 44 measured variables (see Table 3 for list of variables and abbreviations), we used ML to perform two types of analyses. First, we wanted to identify the best classifying variables allowing for separation of the two groups (rTBI vs Sham). Second, we aimed to nd the best predictors of changes in risk-taking behavior, as it recapitulates a key behavioral component observed in rTBI patients, and of cortical HP 13 C Lac/Pyr, due to its potential to serve as a novel biomarker for long-lasting consequences of rTBI. Various classi cation and feature extraction methods were implemented to identify the best classifying variables between Sham and rTBI mice. Consequently, we identi ed ve triplets of variables that could accurately distinguish between either group and ranked them based on their feature importance scores computed using various feature extraction algorithms (see methods). The top two triplets with high feature scores are presented in Fig. 5.a and the others are shown in Supplementary Fig. 2. The top two triplets are: 1) PDH Pfc, LDH Thal, and LDH Hp, and 2) PDH Ctx, LDH Thal, and HP 13 C Lac/Pyr Ctx. These ndings suggest that although one single feature is not su cient to identify the difference between rTBI and Sham, the combined PDH and LDH activity, as well as HP 13 C imaging readouts can help distinguish differences between the two conditions. The three lower-tier triplets consisted of the subsequent variables: 3) PDH Pfc, EPM duration openandcenter, and LDH Thal, 4) PDH Pfc, EPM duration openandcenter, and HP 13 C Lac/Pyr Ctx, and 5) PDH Pfc, MCT1 Pfc, and HP 13 C Lac/Pyr Ctx. These three last triplets further idenfy the risk-taking behavior and MCT1 expression in the prefrontal cortex as important variables to classify Sham and rTBI.
Next, we identi ed the best predictors of the changes in risk-taking behavior and HP 13 C Lac/Pyr Ctx presented in Fig. 5.b and 5.c, respectively. Interestingly, four variables, namely EPM frequency openandcenter, MCT1 Pfc, nT 2 Pfc, and HP 13 C Lac/Pyr Ctx, were su cient to predict risk-taking behavior ( Fig. 5.b bottom panel) with similar accuracy than when all variables were used for prediction ( Fig. 5 24,25 . In contrast, in this study we observed a decreased HP 13 C Lac/Pyr at chronic timepoints, suggesting different underlying pathological changes between contusion injury and rTBI. Studies performed using positron emission tomography (PET) imaging with the glucose analogue 18 F-uorodeoxyglucose ( 18 F-FDG) have detected long-term brain hypometabolism following TBI 4,5,36 , which is in line with the lower HP 13 C Lac/Pyr measured here at chronic timepoints. To the best of our knowledge, 18 F-FDG PET imaging has never been applied to CHIMERA rTBI model. Furthermore, the use of this method for TBI is limited by ionizing radiations, and the high background of 18 F-FDG PET signal in the brain tissue, hampering the detection of small changes in glucose uptake.
Decreased HP 13 C bicarbonate / lactate ratio was found in moderate TBI models 25 , and decreased HP 13 C bicarbonate levels were found in TBI patients 26 , highlighting possible changes in mitochondrial function and aerobic versus anaerobic respiration following trauma. In agreement with these ndings, a decreased of PDH activity after injury has been previously described 24,37 , including in our current ndings in rTBI. In this study, we were not able to detect 13  where increased MCT1 expression leads to increased HP 13 C Lac/Pyr in selected cell lines 41 . In this study, we did not nd any signi cant differences in MCT1 and MCT4 between rTBI and Sham mice, suggesting that they do not play a prominent role in the changes observed in the HP 13 C Lac/Pyr at this late time post injury.
It has previously been shown that the blood-brain-barrier (BBB) limits the entry of HP probes, which could in turn in uence the measured HP 13  Conventional and advanced anatomical MRI did not detect any differences between Sham and rTBI at 3.5 months post-injury. In agreement with these ndings, Haber et al. and our group previously reported no differences using T 2 MRI at 7 days and 40 days post-injury, respectively 10, 12 . These results suggest that conventional T 2 MRI may not be able to detect rTBI patholology-induced using the CHIMERA device, either at early or late timepoints following injury. Using T 1 mapping we investigated whether we could detect changes in tissue microstructure and reactive oxygen species production, but found no differences between Sham and rTBI 34 . Oxidative stress and reactive oxygen species have been shown to play an important role in TBI pathogenesis and in mediating axonal degeneration 46, 47, 48 . However it remains unclear if these events may predominantly occur at early timepoints following injury and would have resolved by the time we performed our imaging study (3.5 months injury), or whether T 1 mapping was not able to detect these events in this rTBI model. HP [1-13 C]dehydroascorbic acid (DHA) and HP [1-13 C]Nacetyl cysteine (NAC) have been shown to be sensitive probes to investigate redox changes in vivo 49,50 , and therefore represent attractive probes to further interrogate the involvement of oxidative stress using HP 13 C MRSI. We also included SWI MRI exams as this method has been shown to improve the detection of microbleeds and hemorrhagic diffusive axonal injury after TBI, which was associated with neurologic de cits and long-term outcome in human TBI 51,52 . However, we did not detect any differences in SWI MRI between Sham and rTBI. As for T 1 mapping, it remains to be determined whether no microbleeds or oxygenation changes occurs in this model at early timepoints, or whether potential changes have resolved by the time we conducted our imaging exams.
We took a systems approach to measure behavior outcomes using ML analyses. We found that sets of variables were able to classify rTBI and Sham mice. These variables are linked to cognitive abilities (risktaking behavior), metabolism and molecule transport (PDH and LDH activity, HP 13 C Lac/Pyr, and MCT1), highlighting the importance of long-term metabolic impairment in rTBI and suggesting their potential as injury biomarkers. Interestingly, ML was able to identify that enzymatic changes in the thalamus and hippocampus regions, which were not statistically signi cant using conventional unpaired t-test in isolation, became important variables to classify rTBI and Sham when considered together. This is in agreement with changes in hippocampal function that have been previously reported in rTBI up to 6 months post-injury 53 . Future HP 13 C imaging studies will aim to also include the thalamus and hippocampus to determine whether in vivo metabolic changes can be detected in these regions. ML was also used to determine which variables are best predictors of the risk-taking behavior. We found that the changes in risk-taking behavior were best predicted by variables from the cortical areas, including HP 13 C Lac/Pyr, MCT1 expression, structural MRI, and behavioral parameters. Similarly, cerebral microdialysis studies have shown that the Lac/Pyr is an important variable associated to clinical outcome. High Lac/Pyr within the rst days after injury was associated to poor clinical outcome 6 months later 18 , while here we show that a lower HP 13 C Lac/Pyr is associated to higher risk-taking behavior 3.5 months after injury. This discrepancy might be explained by different timing of the Lac/Pyr measurement (days versus months after injury), and the injury severity (severe TBI versus mild rTBI). Nonetheless, our study provides further evidence that the Lac/Pyr is a useful marker to predict behavioral outcome, which can now be measured in a non-invasive manner using HP 13 C MRSI, thus opening new avenues to evaluate metabolic alterations months after trauma. Last, ML was used to determine the best predictors of the HP 13 C Lac/Pyr in the cortex, and identi ed subcortical variables, including PDH activity, SWI MRI and metabolic MRSI, as well as risk-taking behavior. Altogether, these results demonstrate the importance of multimodal approaches to detect rTBI pathology and associated long-lasting changes.
Thanks to the recent efforts of the community to enable easy data sharing though data repository 54 , future studies with higher sample size will become possible, which in turn can lead to improvement of our understanding of biological and functional pathway involved in rTBI, and help identify novel biomarkers.
In summary, our ndings demonstrate the potential of HP [1-13 C]pyruvate to detect long-lasting metabolic alterations in a mouse model of rTBI. In addition, ML identi ed HP 13 C MRSI as a key parameter to predict long-term rTBI-induced behavioral outcomes. Over the past few years, the use of HP 13 C MRSI in clinical trials worldwide has been rapidly expanding, and the injection of HP [1-13 C]pyruvate has proven feasible and safe, with no reported side effects 55 . In this study, we were able to measure changes in HP 13 C MRS parameters from two regions, the cortex which is closest to the impact, and the subcortex, which is more remote. We were not able to differentiate between smaller brain regions (e.g. prefrontal cortex, hippocampus, thalamus) due to the large voxel size used in this study relative to the size of the mouse brain. Current sequences available on clinical scanners can achieve up to 1 cm 3 spatial resolution and cover the entire human brain, thus providing metabolic information from brain areas close and remote to the site of injury. With the growing availability of the HP 13 C MRS technology, our ndings provide a strong rationale to translate its use in patients suffering from rTBI, with the aim to improve the detection of rTBI-induced damages, help in understanding metabolic pathways involved in rTBI pathogenesis, and eventually aid the development of treatment strategies.

Animals and rTBI model induction
All animal research was approved by the Institutional Animal Care and Use Committee of the University of California, San Francisco. Mice were given one week of acclimation and housed with a reversed 12-h light/12-h dark cycle and provided food and water ad libitum. At 8 weeks of age, mice were randomly assigned to the rTBI or sham control group. Animals were anesthetized using iso urane (2-3%) in oxygen 1 L/min during the procedure. rTBI animals were subjected to multiple, mild, closed-head injuries using the CHIMERA device as previously reported 10,35 . Brie y, rTBI animals were placed supinely into an angled holding platform without any shaving of the head or incision into the skin so that the head was level with the piston target hole while aligning the eyes, ears, and nose such that the impact was centered on the dorsal convexity of the skull, targeting a 5-mm area surrounding bregma. A nose cone delivering iso urane was removed just prior to the impact. Impact was initiated using RealTerm software, which was connected to a system including air tank, pressure regulator, digital pressure gauge, two-way solenoid valve, and piston. The impact was administered with a velocity range of 3.9-4.5 m/sec, resulting in an impact energy of 0.5 J from the 5 mm, 50 g piston 10,35 . Animals were moved to an incubator immediately after the impact and monitored until fully recovered. rTBI animals received an injury once per day for 5 days with a 24 h interval in between impacts. Five repeated hits were chosen to speci cally focus on the effects of repeated exposure to TBI, as athletes, veterans and sometimes trauma victims are exposed to constant and repeated blows, even without experiencing concussive symptoms.
Sham mice were exposed to the same iso urane anesthesia paradigm without sustaining an impact. Skull fractures, seizures, apnea, or mortality were not observed in any animals, and no animals were excluded from the study due to injury parameters.

Risk-taking behavioral test
For all behavioral assays, the experimenters were blinded to surgery. Before behavioral analysis, animals were inspected for gross motor impairments. Animals were inspected for whisker loss, limb immobility (including grip strength), and eye occlusions. If animals displayed any of these impairments, they were eliminated from the study. Behavioral tests were recorded and scored using a video tracking and analysis setup (Ethovision XT 8.5, Noldus Information Technology). If tracking was unsuccessful, videos were scored by two individuals blinded to surgery. Risk-taking behavioral phenotype was evaluated using the Elevated Plus Maze (EPM) at ~ 100 days (3 months) post-injury (counted from the day of the rst injury) as described previously 10,35 . The EPM consists of Magnetic resonance imaging data analysis Brain regions were manually delineated on T 1 maps, T 2 -weighted and SWI magnitude images for each mouse based on the Allen Adult Mouse Brain atlas (Allen Institute) using the Aedes region of interest package for MATLAB (Mathworks). For each region, brain volumes were calculated with the T 2 -weighted data, as well as the mean T 2 -weighted values which were normalized to the mean of the cerebrospinal uid signal from the ventricles as signal value standard. The mean T 1 relaxation times was calculated from T 1 maps generated in VNMRj by pixel-wise tting according to (Eq. 1).
Where y is the measured signal from fast spin echo with multiple inversion recovery, and the three t parameters: relaxation time T 1 , equilibrium longitudinal magnetization M 0 , and pre-inversion recovery longitudinal relaxation M(0). The TI list is used as input for time t.
SWI data were processed as previously described 32 with phase images unwrapped by PRELUDE (FSL), High pass Gaussian ltered with pixel size 32 x 32, and positive phase map scaling used (Eq. 2).
Normalized positive phase map φ pos (t), where φ(t) is the ltered, unwrapped phase at location t, and φ max is the maximum phase of the slice of interest. The positive phase map is a spatial map varying between zero and one, with higher phase approaching zero and thus increasing contrast on the nal merged SWI data. The phase map is multiplied with the magnitude image four times to create nal SWI data 56 . The mean SWI intensity was calculated for each mouse for each brain region and normalized to the mean of the Sham for each region, which corresponds to 1.

Machine Learning analyses
The ML pipeline is summarized in Supplementary Fig. 3. Preprocessing of the raw experimental data was performed before ML analyses to (i) scale the measurements for fair comparison and (ii) predict the missing measurements. The raw measurements were rescaled using Scikit-learn python library's standard scaler 57 (see Supplementary Fig. 4 for an example of original versus scaled data distributions). For a few mice, we were not able to measure some of the variables due to tissue isolation (n = 2 rTBI and n = 2

Statistical analysis
Results are expressed as mean ± standard deviation (SD). Statistical analyses of MRI, behavioral, and ex vivo parameters was performed using unpaired t-test (GraphPad Prism (v 9.1.2), (*p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, ****p ≤ 0.0001).  Experimental timeline of the study. Two-month old male mice received a rTBI using the CHIMERA device or underwent a Sham procedure (no impact). Risk-taking behavior was evaluated at 3 months post-injury using the Elevated Plus Maze. MR imaging was performed 3.5 months after Sham or rTBI, and included HP 13 C MRSI, T 2 -weighted MRI, T 1 mapping MRI, and SWI MRI. Tissue was collected 4 months after Sham or rTBI procedures to evaluate PDH and LDH activities, and expression of MCT1 and MCT4. ML analyses methods were used to identify the best classi ers between rTBI and Sham, and the best predictors of the risk-taking behavior and HP 13 C Lac/Pyr in the cortex.  Lac/Pyr in cortical areas in rTBI mice. N = 9 rTBI and 10 Sham mice. Unpaired t-test (**p ≤ 0.01); data are expressed as means ± SD.

Figure 4
Multimodal MRI does not detect long-lasting effect of injury in rTBI.
(a) Representative T 2 -weighted MRI data and corresponding manual brain masking. (b) Quantitative analyses of T 2 -weighted signal intensity revealed no signi cant differences for brain subregions between Sham and rTBI. (c) Representative T 1 map and corresponding manual brain masking. (d) Quantitative analyses of T 1 maps revealed no signi cant differences for brain subregions between Sham and rTBI. (e) Representative SWI data and corresponding manual brain masking. (f) Quantitative analyses of SWI intensity revealed no signi cant differences for brain subregions between Sham and rTBI. Brain masking color code: yellow: cortex, green: light blue: prefrontal cortex; hippocampus; dark blue: thalamus. N = 10 rTBI and 10 Sham mice. Unpaired t-test; data are expressed as means ± SD. ML analyses identify best rTBI and Sham classi ers and best predictors of changes in risk-taking behavior and HP 13 C Lac/Pyr Ctx.
(a) Top two triplets that can classify rTBI (red) and Sham (black) mice. Here, circles represent the mice for which all three variables are measured whereas triangles represent mice for which at least one of the