Mapping Inflammation Dynamics Across Brain,CSF,and Serum after TBI: A Multi-Compartment and Machine Learning Analysis of Injury and Celastrol Effects

Abstract

Traumatic brain injury (TBI) involves a transition from acute to chronic neuroinflammation that contributes to lasting neurological deficits; however, the mechanisms by which brain-derived inflammation orchestrates brain-periphery interactions and how to modulate them remain incompletely understood. This multi-compartment study assessed cytokine concentrations across paired serum, cerebrospinal fluid, and ipsilateral hippocampal lysate samples 48 h after controlled cortical impact. Celastrol, an understudied anti-inflammatory compound in TBI, was used as a model therapeutic to evaluate the utility of multi-compartment sampling in pre-clinical therapy screening. Multi-compartment factor analysis (MCFA) was applied as a novel machine learning workflow to identify covariance patterns across fluid and tissue TBI biomarkers. Injury-induced responses were greatest in the hippocampus (IL-6, KC/GRO, and tumor necrosis factor-alpha [TNF-α] increases), with serum decreases in IL-6 and TNF-α. Celastrol restored serum TNF-α to control levels. Sensitivity analyses after outlier removal additionally revealed hippocampal IFN-γ suppression, CSF IL-1β and TNF-α elevation, and celastrol-associated increases in hippocampal IL-10 and IFN-γ. MCFA uncovered coordinated CSF shifts in IL-6, IL-1β, and KC/GRO not individually significant by standard methods and identified a serum inflammatory pattern tracking hippocampal inflammation. Cross-compartment correlations were limited, supporting local regulation. Multivariate dispersion analysis revealed that untreated injury doubled inter-animal inflammatory heterogeneity across the factor space, while celastrol restored dispersion to sham-like levels, indicating treatment re-consolidated the multi-compartment inflammatory response. These findings demonstrate a compartment-specific inflammatory response 48 h post-TBI and suggest celastrol exerts cytokine- and compartment-specific immunomodulatory effects. MCFA provided insights beyond traditional analyses and a standardizable framework for characterizing serum signatures of active brain injury with translational potential for broader TBI biomarker datasets.

Keywords

biomarker celastrol compartmental analysis cytokine neuroinflammation traumatic brain injury

Introduction

TBI is a significant contributor to trauma-related death and disability worldwide, increasingly recognized to initiate secondary injury cascades with long-term consequences that drive a substantial public health burden.¹ Despite advances in supportive and rehabilitative care, more than 30 clinical trials have failed to yield safe and effective neuroprotective therapies, highlighting the need for novel research strategies.^2–8

Developing effective TBI treatments is complicated by the heterogeneity of secondary injury. Primary mechanical damage initiates molecular and cellular cascades involving both protective and pathological mechanisms.^9–11 Blood–brain barrier (BBB) disruption induced by mechanical and inflammatory processes post-TBI permits bidirectional exchange of immune mediators between peripheral circulation and brain parenchyma, perturbing the baseline immune-privileged environment. The extent to which this exchange modulates the balance of protective and pathological factors across compartments remains poorly understood, complicating the design and interpretation of pre-clinical therapeutic trials.

Recent TBI research has exploited BBB permeability to detect brain-derived biomarkers such as GFAP and UCH-L1 in more accessible biological fluids, such as serum.¹² While these structural markers originating from damaged neural tissue provide clinical prognostic utility, their value for identifying therapeutic targets in pre-clinical study is not fully established.

In contrast to structural biomarkers, neuroinflammation has emerged as a pivotal mechanism linking primary and secondary injury phases post-TBI, with landmark studies demonstrating protracted innate and adaptive immune responses driving secondary pathologies.^13–16 Sustained elevation of neuroinflammatory biomarkers correlates with the development of traumatic encephalopathies and neurodegenerative disorders, including Alzheimer’s and Parkinson’s diseases.^14,17–19 The hippocampus is particularly vulnerable to TBI-induced inflammation, which strongly correlates with cognitive deficits and poor long-term neurological outcomes.^13,20–30 However, changes in inflammatory marker distributions across paired serum, CSF, and hippocampal samples from the same animal after TBI have not been characterized.

Extensive research has reported single-compartment cytokine dynamics and associated cellular or molecular changes post-TBI.^{13,14,16,18,19,23,26,29} However, inferring the interplay of inflammatory processes across serum, CSF, and brain-specific compartments such as the hippocampus from these studies is difficult due to heterogeneity of injury models and severities, detection and quantification methods, inflammatory markers measured, compartments studied, and whether samples originate from the same or different animals.^13,26,31,32

This variability across the existing literature emphasizes the need for standardized experimental approaches that permit direct intra-animal comparisons across multiple biological compartments. To address this gap, the present study examined pro- and anti-inflammatory cytokines IFN-γ, IL-1β, IL-4, IL-5, IL-6, IL-10, IL-13, KC/GRO, and TNF-α 48 h post-CCI across serum, CSF, and ipsilateral hippocampal tissue lysate. This timepoint was selected to capture a predominantly pro-inflammatory phase preceding the peak of endogenous anti-inflammatory responses, at which therapeutic intervention may be beneficial.^13,33

To assess the utility of this multi-compartment approach for pre-clinical therapeutic assessment, we investigated celastrol, a plant-derived triterpene with broad anti-inflammatory and antioxidant properties demonstrated across rheumatoid arthritis, stroke, and neurodegenerative disease models.^34–44 In a single TBI study, celastrol elevated heat shock protein 70 levels after CCI, improved cognitive and motor function, and eliminated post-CCI mortality.⁴⁵ Because these effects have not been characterized across multiple biological compartments simultaneously, celastrol served as a model therapeutic to probe compartment-specific immunomodulation rather than to evaluate its clinical development.

To address the complexity of multi-compartment cytokine pattern detection, Multi-Omics Factor Analysis (MOFA), originally developed by Argelaguet et al. for genomics research, was adapted for a novel application termed Multi-Compartment Factor Analysis (MCFA).^46,47 This unsupervised machine learning approach, typically used to integrate multiple molecular data types (e.g., genomics, transcriptomics, proteomics), was repurposed here as an exploratory complement to standard statistical analyses given the complex data structure. Unlike traditional principal component analysis (PCA), which operates on a single concatenated matrix and can be dominated by the highest-variance compartment, MCFA treats each compartment as a separate data view, isolating shared cross-compartment variation from compartment-specific noise while incorporating Bayesian sparsity priors that regularize the model for small sample sizes. MCFA identifies latent factors that explain covariation patterns between cytokines across anatomical compartments by weighing covariances of each cytokine measurement across all compartments simultaneously, while its probabilistic framework accommodates missingness, as cytokines are frequently below the lower limit of detection (LLOD) in particular experimental conditions.

Methods and Materials

Animals

All experimental procedures were approved by the University of Pittsburgh Institutional Animal Care and Use Committee in accordance with the guidelines established by the National Institutes of Health in the Guide for the Care and Use of Laboratory Animals. Animals were housed up to two rats per cage in the University of Pittsburgh vivarium with a 12:12 light/dark photoperiod (lights on at 7:00 a.m.) and provided food and water ad libitum. Animals were monitored daily by researchers and veterinary technicians. Forty-eight male Sprague–Dawley rats (Envigo, Indianapolis, IN) weighing 275–300 g were used in this study and randomly assigned to four experimental conditions (n = 12 per condition) to receive either a sham procedure or controlled cortical impact (CCI) injury and subsequent treatment with either celastrol or vehicle (dimethyl sulfoxide [DMSO]) (Fig. 1). Experimental conditions are hereafter referenced as control (sham-vehicle), drug (sham-celastrol), injury (CCI-vehicle), and treatment (CCI-celastrol). All surgical procedures and treatment administrations were performed in a blinded fashion.

FIG. 1.

Experimental design. Surgery—single CCI modeling moderate-severe TBI, or sham surgery. intraperitoneal (IP) Injection—Celastrol in DMSO, or DMSO (vehicle) alone, administered intraperitoneally 24 and 48 h after injury. Sample collection—brain, cerebrospinal fluid (from cisterna magna), and serum (from tail vein) terminally collected 48 h after CCI or sham surgery. Ipsilateral hippocampal tissue lysate was prepared from each brain sample prior to analysis. Sample analysis—Meso Scale Discovery (MSD) electro chemiluminescent sandwich enzyme-linked immunosorbent assay (ELISA) detection of neuroinflammatory biomarkers IFN-γ, IL-1Β, IL-4, IL-5, interleukin-6 (IL-6), IL-10, IL-13, keratinocyte chemoattractant/growth-related oncogene (KC/GRO), TNF-α ran in technical duplicate. Created in BioRender. Gjesdal, B. (2026) https://BioRender.com/628vwz6. CCI, controlled cortical impact; TBI, traumatic brain injury.

Controlled cortical impact

A schematic of the experimental design can be found in Figure 1. Rats were anesthetized using 4% isoflurane with a 2:1 N₂O/O₂ mixture in a ventilated anesthesia chamber. Following endotracheal intubation, the rats were mechanically ventilated with a 2% isoflurane mixture. Animals were placed in a stereotaxic frame, and body temperature was monitored by rectal thermistor probe and maintained at 37°C with a heating pad. Following a midline incision, the soft tissues were reflected, and a 7 mm craniectomy was performed over the right parietal cortex, between bregma and lambda, and centered 5 mm lateral of the sagittal suture to expose the dura mater. Control sham injury animals were subjected to anesthesia and surgical procedures but did not receive a TBI. The CCI injury device is a small bore (1.975 cm) double-acting stroked-constrained pneumatic cylinder with a 5.0 cm stroke. A 6 mm diameter, flat impactor tip was set to produce a tissue deflection of 2.5 mm at a velocity of 4 m/sec with a dwell time of 100 msec. After each sham or CCI injury, the scalp was sutured, gas anesthetics were turned off, and righting time was monitored. Once ambulatory, the animals were returned to their home cages.

Celastrol administration

Every 24 h after animals had undergone either CCI or sham surgery, animals received a 0.5 mg/kg dose of celastrol or vehicle (DMSO) administered intraperitoneally and returned to their home cages. Animals received two total doses.

Serum, CSF, hippocampal brain tissue lysate collection, and preparation

At 48 h post-CCI or sham surgery, animals received an overdose of sodium pentobarbital (intraperitoneally, 100 mg/kg Fatal-plus, Vortech Pharmaceuticals, Dearborn, MI). CSF samples were then immediately collected from the cisterna magna manually via needle aspiration and stored at −80°C. Following rapid decapitation, whole blood samples were immediately collected from the trunk of each animal, incubated at room temperature for 30 min, centrifuged at 1500g for 15 min, aliquoted, and stored at −80°C. The ipsilateral hippocampus was dissected on a chilled ice plate and immediately snap frozen in liquid nitrogen and stored at −80°C. Ipsilateral hippocampus samples were homogenized by sonication in lysis buffer (0.1 M NaCl, 0.01 M Tris-Cl, 0.001 M ethylenediaminetetraacetic acid, pH 7.6) with protease and phosphatase inhibitors (Sigma-Aldrich, St. Louis, MO). The homogenized whole cell samples were centrifuged at 13,000g at 4°C for 30 min, and the supernatants collected. Total protein concentration was determined by a BCA protein assay kit (Thermo Scientific, Pittsburgh, PA) using a 96-well microplate reader (Biotek, Winooski, VT).

Inflammatory biomarker analysis

All sample analyses were conducted in a blinded fashion. IFN-γ, IL-1β, IL-4, IL-5, IL-6, IL-10, IL-13, KC/GRO, and TNF-α concentrations in serum, CSF, and hippocampal tissue lysate samples were determined using electrochemiluminescence immunoassays on the MESO QuickPlex SQ 120MM system with V-PLEX Proinflammatory Panel 2 Rat Kits (Meso Scale Discovery [MSD], Maryland, USA). Serum, CSF, and hippocampal dilutions were tested at fourfold dilution per kit guidelines. Calibrator detection ranges spanned 0–30,000 pg/mL. No values were recorded above the upper limit of detection. Handling of values at or below the LLOD is described in Data Analysis below.

Data analysis

All primary statistical analyses were conducted using R (v4.5.0). p values < 0.05 were considered statistically significant. Benjamini-Hochberg false discovery rate (FDR) correction was applied separately to Kruskal–Wallis tests, pairwise tests, and correlation tests; To account for inter-plate variability, cytokine concentrations were normalized as a percentage of the plate control (sham vehicle) median for each cytokine and therefore do not represent absolute concentrations. Median and interquartile range (IQR) are reported where appropriate to summarize variably skewed distributions. FDR-corrected p values are reported in the text and tables; both uncorrected and FDR-adjusted p values are provided in Supplementary Table S1.

Cytokine concentrations were back-calculated from the standard curve by MSD Discovery Workbench software. The software reports a calculated concentration for all samples, including those below the manufacturer-specified LLOD; values for which the software could not generate a concentration estimate, where the electrochemiluminescence signal was indistinguishable from background, were recorded as not detected. All instrument-reported concentrations were retained in the primary analysis regardless of whether they fell above or below the manufacturer LLOD, because the statistical tests employed (Kruskal–Wallis, Mann–Whitney U, Wilcoxon signed-rank, Spearman correlation) are rank-based and rely on ordinal position rather than absolute concentration, making them robust to measurement imprecision in low-signal samples. Only not-detected values were treated as missing data and excluded from the primary analysis. IL-5 data across all compartments and IL-13 data in hippocampus are reported in tables and figures for completeness but not interpreted due to high rates of non-detection (most IL-5 readings not detected across compartments, IL-13 mostly not detected in hippocampal samples; per-group missingness rates are detailed in Supplementary Fig. S1).

Intra-compartment differences in normalized cytokine expression between experimental conditions were assessed using the Kruskal–Wallis test; when significant, pairwise Mann–Whitney U tests assessed distributions between conditions, with results depicted in Table 1 and visualized in Figure 2. Standardized effect sizes (Cohen’s d with pooled standard deviation) were calculated for all pairwise group comparisons per compartment and cytokine (Table 1). Paired Cohen’s d was calculated for within-group compartment comparisons (Table 2).

FIG. 2.

Cytokine expression by compartment and experimental condition at 48 h after controlled cortical impact. Male Sprague–Dawley rats were assigned to Sham Vehicle (control), Sham Celastrol (drug), CCI Vehicle (injury), or CCI Celastrol (treatment) (N samples = 12 per condition; per-cytokine N varies due to missing values, N per condition presented in Table 1). Bars show medians with interquartile ranges for ipsilateral hippocampal tissue lysate, CSF, and serum per condition. Values are normalized to the plate control median for Sham Vehicle and displayed as percent of control. Cytokines measured by electrochemiluminescence (MSD V-PLEX; 4× dilution). Statistical testing within each compartment and cytokine used Kruskal–Wallis; when significant, subsequent pairwise Mann–Whitney U tests determined condition differences within compartments per cytokine. Significance annotated as follows: NS = nonsignificant Kruskal–Wallis test; *=p < 0.05, **=p < 0.01, ***=p < 0.001. CCI, controlled cortical impact; CSF, cerebrospinal fluid.

Table 1.

Intra-Compartment Cytokine Responses Between Experimental Conditions

	Control (SV)		Injury (CV)		Treatment (CC)		Drug (SC)		All	Sham vs. CCI		CV vs. CC		SV vs. SC
Cytokine	N	Median (IQR)	N	Median (IQR)	N	Median (IQR)	N	Median (IQR)	KW p	MWU p	Cohen d	MWU p	Cohen d	MWU p	Cohen d
Ipsilateral hippocampal tissue lysate
KC/GRO	10	94.86 (86.87–126.01)	12	3854.01 (270.94–12933.35)	12	313.33 (225.07–503.96)	12	56.90 (45.30–70.08)	1.2E-5*	1.2E-8*	−0.54	0.223	0.77	0.0956	0.69
IL-6	10	105.48 (91.83–116.03)	12	2342.63 (182.69–4412.07)	12	293.08 (175.40–446.65)	11	74.82 (57.93–91.45)	1.9E-5*	6.2E-8*	−0.43	0.325	0.61	0.203	0.91
IL-1β	10	98.60 (88.22–110.24)	12	965.83 (97.56–3830.21)	12	114.29 (102.89–141.17)	12	102.18 (90.06–127.19)	0.281	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
TNF-α	9	84.81 (75.51–150.83)	11	1371.62 (613.90–11070.30)	12	625.81 (505.68–863.57)	10	73.00 (56.43–138.11)	1.2E-5*	1.8E-9*	−0.79	0.244	1.10	0.698	0.04
IFN-γ	10	100.24 (82.19–123.62)	9	61.58 (51.46–1952.52)	10	91.78 (70.80–120.75)	12	98.87 (75.17–121.82)	0.78	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-10	9	95.76 (88.07–113.55)	9	76.15 (74.09–1452.33)	9	129.18 (102.10–134.74)	11	98.19 (79.04–118.19)	0.627	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-4	9	87.07 (83.44–116.14)	8	80.02 (67.44–669.95)	9	105.73 (93.71–153.79)	11	87.48 (75.76–121.33)	0.557	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-13	4	97.44 (86.66–117.49)	3	1726.49 (1444.68–1899.97)	4	209.29 (173.19–479.34)	5	159.45 (68.76–241.06)	0.114	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-5	5	86.13 (82.68–116.11)	2	1951.77 (1283.04–2620.50)	3	153.85 (146.79–225.88)	5	131.05 (122.25–134.27)	0.13	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
CSF
KC/GRO	11	107.49 (59.93–130.23)	10	115.86 (66.79–131.38)	11	98.78 (49.44–258.85)	11	68.83 (34.08–90.15)	0.505	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-6	9	65.67 (48.49–206.22)	10	136.97 (93.26–207.34)	11	87.52 (70.86–430.41)	8	35.30 (28.13–55.18)	0.0539	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-1β	7	99.33 (77.66–100.00)	9	184.47 (100.26–217.78)	11	154.28 (31.72–321.99)	8	60.27 (4.29–93.55)	0.227	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
TNF-α	8	78.81 (65.49–172.46)	9	222.68 (131.62–831.51)	9	258.90 (139.67–357.34)	8	40.16 (31.07–127.49)	0.235	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IFN-γ	8	93.49 (88.40–130.74)	7	131.04 (60.68–133.57)	8	94.19 (62.86–159.00)	8	76.44 (72.37–82.45)	0.626	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-10	8	90.33 (79.68–110.84)	9	57.84 (50.87–87.85)	9	56.48 (36.81–124.85)	8	28.36 (7.52–45.37)	0.0364*	0.921	−0.35	0.987	−0.34	0.0044*	2.17
IL-4	6	102.44 (98.42–110.23)	4	83.39 (72.78–93.56)	7	38.85 (8.76–107.67)	4	30.41 (21.15–62.40)	0.412	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-13	7	138.93 (71.58–154.23)	8	177.90 (46.79–309.83)	7	148.24 (123.64–189.73)	7	51.74 (46.19–105.47)	0.29	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-5	4	106.64 (68.89–145.20)	4	87.68 (62.44–119.17)	2	68.51 (62.69–74.32)	2	50.89 (44.63–57.15)	0.412	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
Serum
KC/GRO	11	100.37 (76.41–133.84)	11	39.15 (30.79–54.37)	11	111.85 (46.94–164.34)	11	74.04 (41.00–250.26)	0.164	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-6	11	99.24 (83.31–108.67)	11	20.88 (7.23–25.86)	11	103.88 (49.36–132.67)	11	68.37 (57.64–171.66)	0.0133*	0.0644	0.55	0.0611	−1.17	0.839	−0.35
IL-1β	10	104.21 (59.15–139.87)	7	48.11 (42.48–64.92)	10	54.93 (22.31–103.89)	10	92.07 (66.33–127.86)	0.174	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
TNF-α	11	96.14 (92.82–110.05)	11	49.49 (23.54–65.92)	11	101.37 (68.74–144.17)	11	104.61 (65.52–121.10)	0.0134*	0.0956	0.56	0.0403*	−1.33	1.0	−0.21
IFN-γ	11	102.65 (86.49–105.85)	11	97.30 (76.81–100.59)	10	67.47 (62.43–108.49)	11	100.39 (88.96–120.54)	0.412	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-10	11	101.12 (96.30–105.26)	11	83.77 (77.45–96.59)	11	92.61 (80.96–98.70)	11	101.75 (88.91–106.24)	0.0745	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-4	11	98.64 (93.53–103.80)	11	94.98 (91.73–96.91)	11	96.07 (91.68–103.51)	11	98.78 (93.83–105.58)	0.412	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-13	11	102.35 (91.58–112.17)	11	80.94 (70.58–96.14)	11	90.17 (64.11–99.06)	11	90.89 (88.22–114.29)	0.281	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW
IL-5	2	134.38 (89.50–179.26)	0	NA	1	81.82 (81.82-81.82)	2	122.31 (87.10–157.53)	1.0	ns KW	ns KW	ns KW	ns KW	ns KW	ns KW

Median (IQR) cytokine expression per condition within each compartment, normalized to plate control median for each cytokine and expressed as percent of control.

N indicates non-missing samples.

Global group differences were assessed using the Kruskal–Wallis test (KW p); where KW p was significant (p < 0.05), pairwise Mann–Whitney U tests were performed (MWU p) and standardized effect sizes were calculated (Cohen d; pooled standard deviation).

All reported p values are Benjamini-Hochberg FDR-corrected p values; uncorrected p values are reported in Supplementary Table S1.

Positive Cohen d indicates group 1 > group 2; negative indicates group 1 < group 2.

Conditions: Control (SV) = sham vehicle; Drug (SC) = sham celastrol; Injury (CV) = CCI vehicle; Treatment (CC) = CCI celastrol. Pairwise comparisons: Sham vs. CCI = pooled injury effect (SV + SC) vs. (CV + CC); CV vs. CC = treatment effect in injured animals; SV vs. SC = drug effect in sham animals. Significance (p < 0.05) denoted by *. NA = too few samples above detectable limit.

CCI, controlled cortical impact; CSF, cerebrospinal fluid.

Table 2.

Within-Group Compartment Differences (Effect-Size Ratios Relative to Sham Vehicle)

Group	Cytokine	Hipp:CSF:Serumeffect size (Δ vs. sham)	SM p	Hipp vs. CSF		Hipp vs. serum		CSF vs. serum
Group	Cytokine	Hipp:CSF:Serumeffect size (Δ vs. sham)	SM p	WSR p	Paired d	WSR p	Paired d	WSR p	Paired d
Sham vehicle	KC/GRO	1 : −1.46 : −0.07	0.902	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-6	1 : −6.26 : −0.14	0.44	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-1β	1 : 0.48 : −3.01	0.605	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	TNF-α	1 : 1.39 : 0.25	0.627	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IFN-γ	1 : −27.10 : 11.03	0.753	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-10	1 : 2.28 : −0.27	0.326	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-4	1 : −0.19 : 0.11	0.745	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-13	1 : −15.20 : −0.92	0.605	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-5	1 : −0.48 : −2.48	0.902	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
CCI vehicle	KC/GRO	1 : 0.00 : −0.02	0.0039*	0.0338*	0.57	0.0275*	0.59	0.197	0.11
	IL-6	1 : 0.02 : −0.04	0.0057*	0.0417*	0.47	0.0275*	0.47	0.0888	0.60
	IL-1β	1 : 0.10 : −0.06	0.168	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	TNF-α	1 : 0.10 : −0.04	0.0095*	0.159	0.70	0.0338*	0.92	0.0677	0.75
	IFN-γ	1 : −0.81 : 0.07	0.96	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-10	1 : 1.77 : 0.68	0.336	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-4	1 : 0.83 : 0.25	0.431	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-13	1 : 0.05 : −0.01	0.197	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-5	NA	0.304	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
CCI celastrol	KC/GRO	1 : −0.01 : 0.06	0.295	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-6	1 : −0.06 : 0.02	0.0972	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-1β	1 : 3.80 : −3.15	0.605	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	TNF-α	1 : 0.30 : 0.00	0.0228*	0.326	0.48	0.0275*	1.64	0.159	0.60
	IFN-γ	1 : 0.71 : 3.96	0.431	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-10	1 : −1.49 : −0.25	0.326	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-4	1 : −10.67 : −0.69	0.437	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-13	1 : 0.44 : −0.09	0.104	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-5	1 : −0.58 : −0.34	0.336	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
Sham celastrol	KC/GRO	1 : 0.72 : 0.60	0.336	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-6	1 : 2.57 : 1.26	0.434	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-1β	1 : −18.20 : −3.63	0.52	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	TNF-α	1 : 2.22 : −0.17	0.326	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IFN-γ	1 : 20.81 : −0.35	0.738	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-10	1 : 39.67 : −0.97	0.0275*	0.059	2.25	0.935	−0.01	0.059	−3.30
	IL-4	1 : 5.56 : 0.10	0.274	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-13	1 : −0.81 : −0.15	0.326	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM
	IL-5	1 : −1.58 : 0.72	0.295	ns SM	ns SM	ns SM	ns SM	ns SM	ns SM

Effect-size ratios express each compartment’s median group difference (relative to Sham Vehicle) as a proportion of the hippocampal effect (hippocampus = 1). Global compartment differences assessed using the Skillings–Mack test (SM p); where SM p < 0.05, pairwise Wilcoxon signed-rank tests (WSR p) and paired Cohen’s d were computed on matched samples. All reported p values are Benjamini-Hochberg FDR-corrected p values; uncorrected p values in Supplementary Table S1. Significance (p < 0.05) denoted by *. CCI, controlled cortical impact; CSF, cerebrospinal fluid.

Two prespecified sensitivity analyses were conducted to evaluate the robustness of findings. First, all intra-compartment comparisons were repeated after excluding within-group outliers defined as values exceeding three times the median absolute deviation (MAD) from the group median; when the MAD was near zero, a fallback threshold of three times the IQR was used (outlier removal sensitivity analysis; Supplementary Table S2). Second, not-detected values were imputed at half the minimum observed instrument-reported concentration for each compartment–cytokine pair, ensuring all imputed values ranked below all observed measurements, and Kruskal–Wallis tests were repeated to evaluate robustness to missing data handling (imputation sensitivity analysis; Supplementary Table S3).

Inter-compartment, within-condition differences were summarized as hippocampus: CSF: serum effect-size ratios. Effect size was defined as normalized % cytokine expression median − 100 such that ratios represent the distribution of the condition-induced directional change across compartments proportional to the control condition. Ratios are reported relative to the hippocampal effect size as 1: (CSF−100)/(Hippocampus−100): (Serum−100)/(Hippocampus−100). Differences across compartments were assessed using the Skillings–Mack test; when significant, paired Wilcoxon signed-rank tests were applied to within-sample effect-size differences between compartments: Hippocampus—CSF, Hippocampus—Serum, and CSF—Serum, with results depicted in Table 2.

Within-compartment cytokine co-expression patterns and single-cytokine associations between compartments were assessed using Spearman rank correlation, depicted in Tables 3 and 4.

Table 3.

Intra-Compartment Cytokine-Cytokine Correlations

	IL-6			IL-1β			TNF-α			IFN-γ			IL-10			IL-4			IL-13			IL-5
	N	ρ	p	N	ρ	p	N	ρ	p	N	ρ	p	N	ρ	p	N	ρ	p	N	ρ	p	N	ρ	p
Ipsilateral hippocampus tissue lysate
KC/GRO	45	0.95	2.3E−21*	46	0.56	2.0E−4*	42	0.88	3.2E−13*	41	0.10	0.58	38	0.29	0.115	37	0.23	0.231	16	0.59	0.0336*	15	0.58	0.0445*
IL-6				45	0.62	2.9E−5*	41	0.89	3.2E−13*	40	0.18	0.331	37	0.29	0.123	36	0.27	0.159	15	0.70	0.0094*	15	0.62	0.0275*
IL-1β							42	0.49	0.0032*	41	0.68	9.6E−6*	38	0.72	3.3E−6*	37	0.61	2.7E−4*	16	0.85	1.5E−4*	15	0.75	0.0034*
TNF-α										38	−0.04	0.842	36	0.32	0.0978	34	0.37	0.0618	16	0.59	0.0358*	15	0.83	4.1E−4*
IFN-γ													38	0.77	3.0E−7*	37	0.61	2.4E−4*	16	0.57	0.0436*	15	0.84	3.0E−4*
IL-10																36	0.78	3.0E−7*	15	0.67	0.0135*	14	0.85	3.5E−4*
IL-4																			14	0.35	0.28	12	0.66	0.0411*
IL-13																						13	0.81	0.0022*
CSF
KC/GRO	38	0.37	0.0445*	35	0.42	0.0258*	34	0.32	0.108	31	0.26	0.211	34	0.55	0.0022*	21	0.56	0.0181*	29	0.24	0.267	12	−0.01	0.992
IL-6				34	0.71	1.4E−5*	34	0.72	1.1E−5*	30	0.41	0.0445*	34	0.48	0.0094*	21	0.21	0.433	29	0.57	0.0035*	12	0.62	0.0571
IL-1β							32	0.33	0.108	28	0.09	0.715	31	0.54	0.0046*	21	0.27	0.286	27	0.33	0.142	10	0.53	0.167
TNF-α										29	0.24	0.267	32	0.14	0.515	21	0.04	0.898	28	0.50	0.0141*	11	0.04	0.933
IFN-γ													28	0.58	0.0035*	21	0.45	0.0754	28	0.34	0.121	12	0.53	0.118
IL-10																21	0.77	2.0E−4*	27	0.34	0.129	11	0.51	0.159
IL-4																			21	−0.08	0.783	9	0.23	0.606
IL-13																						12	0.44	0.204
Serum
KC/GRO	44	0.87	3.2E−13*	37	0.70	1.1E−5*	44	0.74	1.9E−7*	43	0.13	0.458	44	0.51	0.0016*	44	0.31	0.0754	44	0.21	0.236	5	0.60	0.338
IL-6				37	0.70	1.1E−5*	44	0.84	2.9E−11*	43	0.21	0.236	44	0.58	1.5E−4*	44	0.43	0.0094*	44	0.29	0.0974	5	0.60	0.338
IL-1β							37	0.66	4.5E−5*	36	0.10	0.606	37	0.53	0.0022*	37	0.31	0.0974	37	0.19	0.331	5	0.60	0.338
TNF-α										43	0.31	0.08	44	0.72	5.8E−7*	44	0.44	0.0071*	44	0.24	0.172	5	0.50	0.454
IFN-γ													43	0.57	2.5E−4*	43	0.68	5.9E−6*	43	0.65	1.4E−5*	5	0.20	0.783
IL-10																44	0.61	5.1E−5*	44	0.47	0.0035*	5	0.20	0.783
IL-4																			44	0.60	8.0E−5*	5	0.00	1.0
IL-13																						5	0.20	0.783

Spearman rank correlations between all cytokine pairs within each compartment, presented as a lower-triangular matrix. N = number of paired observations; ρ = Spearman rank correlation coefficient; p = Benjamini-Hochberg FDR-corrected p value (corrected across all 108 within-compartment correlation tests); uncorrected p values in Supplementary Table S1. Significance (p < 0.05) denoted by *.

Table 4.

Cross-Compartment Spearman Rank Correlations

Compartment pair	Cytokine	N	Spearman ρ	p
Hippocampus vs. CSF	IFN-γ	12	0.03	0.93
	IL-1β	15	0.36	0.54
	IL-10	10	0.41	0.54
	IL-13	10	0.42	0.54
	IL-4	5	0.56	0.54
	IL-5	8	0.12	0.864
	IL-6	18	0.4	0.446
	KC/GRO	23	0.28	0.54
	TNF-α	12	0.19	0.657
Hippocampus vs. serum	IFN-γ	19	0.39	0.446
	IL-1β	17	−0.21	0.594
	IL-10	16	−0.19	0.601
	IL-13	16	0.07	0.864
	IL-4	15	0.21	0.594
	IL-5	5	0.6	0.54
	IL-6	23	−0.38	0.446
	KC/GRO	24	−0.23	0.54
	TNF-α	22	−0.49	0.283
CSF vs. serum	IFN-γ	30	−0.13	0.601
	IL-1β	32	0.15	0.594
	IL-10	34	0.15	0.594
	IL-13	29	0.04	0.883
	IL-4	21	0.26	0.54
	IL-5	3	1.0	<0.001*
	IL-6	38	−0.17	0.54
	KC/GRO	43	0.19	0.54
	TNF-α	34	−0.34	0.441

Spearman rank correlations for each cytokine across compartment pairs. All reported p values are Benjamini-Hochberg FDR-corrected p values (corrected across all 27 cross-compartment correlation tests); uncorrected p values in Supplementary Table S1. Significance (p < 0.05) denoted by *. CSF, cerebrospinal fluid; FDR, false discovery rate.

MCFA was performed by adapting the MOFA + v2.0 package (https://biofam.github.io/MOFA2/) in Python (v3.9.6) using log-transformed plate-normalized data (Fig. 3).^46,47 IL-5 was removed from input matrices due to missingness. MCFA was trained without further filtering. Model parameter tuning is outlined in Supplementary Table S4.

FIG. 3.

Multi-compartment factor analysis (MCFA) structural adaptation from multi-omics factor analysis (MOFA). Schematic representation of data integration across biological compartments. Cytokine concentrations measured from serum (red), cerebrospinal fluid (blue), and hippocampal tissue (pink) are organized into compartment-specific data matrices (Y¹, Y², Y³). MCFA decomposes these matrices into compartment-specific weight matrices (W¹, W², W³) capturing cytokine loadings and a shared factor matrix (Z) containing sample scores, revealing coordinated inflammatory programs across compartments. Created in BioRender. Gjesdal, B. (2026) https://BioRender.com/628vwz6.

MCFA decomposed multi-compartment cytokine data into latent factors capturing coordinated patterns of variation across biological compartments. Each factor represents a distinct inflammatory axis defined by specific combinations of cytokines from hippocampus, CSF, and serum that covary together. Factor loadings quantify how strongly each cytokine-compartment pair contributes to each latent factor, with positive loadings indicating cytokines that increase when the factor is elevated and negative loadings indicating inverse relationships. Factor scores represent where each animal falls along these inflammatory axes, allowing assessment of how experimental conditions shift the overall multi-compartment inflammatory state.

Experimental condition labels were not provided to the algorithm; factor score distributions were therefore interpreted post hoc relative to the extent to which they separated animals by condition (Fig. 4A), the contribution of each cytokine-compartment pairing to each factor (Fig. 4B), and the mean group separation by factor (Fig. 4C). To formally assess group separation in factor space, Kruskal–Wallis tests were applied to per-factor scores across all four groups, with pairwise Mann–Whitney U tests for key comparisons. Permutational Multivariate Analysis of Variance (PERMANOVA; 9,999 permutations, Euclidean distance) was applied to the joint three-factor space to test overall multivariate group separation. Permutational Multivariate Analysis of Dispersion (PERMDISP; 9,999 permutations) was used to evaluate whether groups differed in within-group spread in factor space.

FIG. 4.

Latent factor patterns learned by multi-compartment factor analysis (MCFA). (A) 3D plots of sample scores on Factors 1–3 separated by experimental condition, depicting compartment-specific inflammatory changes. Each panel shows one condition’s distribution in factor space, where Factor 1 captures hippocampus-dominant inflammation, Factor 2 represents CSF-specific patterns, and Factor 3 reflects serum responses, with PERMANOVA confirming significant multivariate group separation (pseudo-F = 5.90, p < 0.001). Axes and viewing angle are identical across panels to facilitate direct comparison of how each condition shifts the multi-compartment inflammatory profiles by learned factors. Each point represents an individual animal; gray convex hulls encompass all samples within each group. (B) Factor loading weights for ipsilateral hippocampus tissue lysate, CSF, and serum. Positive loadings indicate higher feature contribution to the factor. (C) Group means ± standard error of the mean of factor scores for Factors 1–3. MCFA was trained on log(x + 1)-transformed normalized cytokine expressions arranged as three views (hippocampus, CSF, serum); IL-5 was excluded from model training due to missingness; K = 3 factors. CSF, cerebrospinal fluid.

Results

Intra-compartment cytokine responses between experimental conditions

Condition-specific cytokine responses and between-condition comparisons from ipsilateral hippocampal tissue lysate, CSF, and serum samples are presented in Table 1 and visualized in Figure 2.

In the hippocampus, injury increased IL-6, KC/GRO, and TNF-α (Kruskal–Wallis p < 0.001 for each) while IFN-γ, IL-10, IL-4, and IL-1β remained unchanged. In CSF, no individual cytokine reached significance after FDR correction; IL-6 showed a near-significant injury-associated increase (Kruskal–Wallis p = 0.054). In serum, injury decreased IL-6 and TNF-α (Kruskal–Wallis p = 0.013 for each); IL-10 showed a near-significant decrease (p = 0.074), while IFN-γ, IL-4, KC/GRO, IL-13, and IL-1β remained unchanged. Treatment in injured animals did not significantly alter any hippocampal or CSF cytokines. In serum, treatment restored TNF-α to sham vehicle control levels (p = 0.040); IL-6 showed a similar direction of effect (p = 0.061, not significant). All other cytokines were unchanged. Drug administration in sham animals decreased CSF IL-10 (p = 0.0044); hippocampal KC/GRO showed a near-significant decrease (p = 0.096) with all other cytokines unchanged across either compartment. Serum cytokines were not significantly altered by drug administration. All p values reported in the text are FDR-corrected; uncorrected values are provided in Supplementary Table S1. Effect sizes for injury comparisons were moderate to large for hippocampal TNF-α (|d| = 0.79), serum IL-6 (|d| = 0.55), serum TNF-α (|d| = 0.56), and others (Table 1).

The outlier removal sensitivity analysis revealed additional condition effects not significant in the primary analysis (Supplementary Table S2). For injury: hippocampus showed decreased IFN-γ (p = 0.021); CSF showed increased IL-1β and TNF-α (p = 0.003 for each); serum showed decreased KC/GRO (p = 0.014). For treatment: hippocampus showed restored IFN-γ to sham vehicle control levels (p = 0.016) and increased IL-10 (p = 0.003); serum showed increased KC/GRO (p = 0.036). For drug: CSF additionally showed decreased TNF-α and IFN-γ (p = 0.021–0.030).

To assess the stability of primary findings to missing data handling, not-detected values were imputed at half the minimum observed concentration. All Kruskal–Wallis findings significant under exclusion (hippocampal IL-6, KC/GRO, and TNF-α; CSF IL-6; serum IL-6, TNF-α, and IL-10) remained significant (p < 0.05) under imputation, with both approaches agreeing on the direction of all tested comparisons. Five additional tests reached significance only under imputation (CSF IL-1β and TNF-α; hippocampal IFN-γ and IL-10; serum KC/GRO), consistent with increased power from larger sample sizes; no finding significant under exclusion was lost (Supplementary Table S3).

Cross-compartment cytokine response effect size ratios per experimental condition

Effect size ratios comparing the relative magnitude of cytokine changes across compartments within each experimental condition are presented in Table 2. Following injury, IL-6, KC/GRO, and TNF-α showed hippocampal increases with minimal CSF changes and opposing serum decreases (hippocampus > CSF, p = 0.006–0.008; hippocampus > serum, p = 0.004–0.006); IL-6 and TNF-α also showed CSF > serum differences (p = 0.018–0.025). Following treatment in injured animals, differences in IL-6 and TNF-α hippocampus-serum effect size remained significant, while much smaller in magnitude (p = 0.004–0.014). Following drug administration in sham animals, only IL-10 showed compartment-specific differences: CSF suppression exceeded hippocampal suppression (p = 0.014) and differed from serum (p = 0.014). All other cytokines showed no significant pairwise compartment differences in effect sizes. Paired Cohen’s d values for significant compartment comparisons are reported in Table 2.

Intra-compartment cytokine–cytokine correlations and multi-compartment correlations by cytokine

Within-compartment cytokine co-expression patterns and cross-compartment cytokine associations are presented in Tables 3 and 4. Within compartments, moderate and strong correlations (ρ ≥ 0.60) demonstrated distinct recurring cytokine correlation networks.

In hippocampus, KC/GRO, IL-6, and TNF-α formed a pro-inflammatory network (ρ = 0.88–0.95, p < 0.001) with IL-1β correlating with IL-6 (ρ = 0.62, p < 0.001). An immunomodulatory co-expression cluster comprised IL-10, IFN-γ, and IL-4 (ρ = 0.61–0.78, p < 0.001). IL-1β bridged these networks through correlations with IL-10 (ρ = 0.72, p < 0.001), IFN-γ (ρ = 0.68, p < 0.001), and IL-4 (ρ = 0.61, p < 0.001). In CSF, IL-6 and TNF-α remained correlated (ρ = 0.72, p < 0.001), while KC/GRO dissociated from this pairing. IL-1β correlated with IL-6 (ρ = 0.71, p < 0.001). IL-10 and IL-4 remained correlated (ρ = 0.77, p < 0.001), and IFN-γ showed moderate associations with IL-10 (ρ = 0.58, p = 0.0035) and IL-4 (ρ = 0.45, p = 0.075), indicating partial preservation of the regulatory cluster observed in hippocampus. In serum, the pro-inflammatory network was preserved with KC/GRO, IL-6, and TNF-α (ρ = 0.74–0.87, p < 0.001), IL-1β with IL-6 (ρ = 0.70, p < 0.001), and TNF-α (ρ = 0.66, p < 0.001). The immunomodulatory cluster remained intact (ρ = 0.57–0.68, p < 0.001) with additional correlations between IL-13 and IFN-γ (ρ = 0.65, p < 0.001) and IL-4 (ρ = 0.60, p < 0.001).

Cross-compartment analyses revealed nearly uniform compartment independence. Of all cytokines measured, only TNF-α demonstrated nominally significant cross-compartment correlations, showing an inverse hippocampus-serum relationship (ρ = −0.49, p = 0.28) and a weaker CSF-serum trend (ρ = −0.34, p = 0.44); neither was significant after FDR correction (Supplementary Table S1), consistent with predominantly local regulation.

Multi-compartment factor analysis

MCFA identified three latent factors explaining 89.5%, 59.0%, and 74.9% of variance in ipsilateral hippocampus tissue lysate, CSF, and serum, respectively, capturing 74.5% of total multi-compartment variance (Fig. 4A). Projection into the three-dimensional factor space separated each experimental condition into a distinct multi-compartment inflammatory trajectory, weighted by compartment-cytokine covariances.

To interpret the axes, we considered factor feature loadings (Fig. 4B) together with condition separation by latent factor (Fig. 4C). Factor 1 reflected a hippocampus-dominant inflammatory axis (KC/GRO 0.864; IL-6 0.756; IL-1β 0.718; IFN-γ 0.522; TNF-α 0.400). Factor scores were elevated in injury (+1.854 ± 0.743) and negative in both sham groups (control −0.488 ± 0.068; drug −0.488 ± 0.085); celastrol nearly reversed the injury-induced shift in Factor 1 scores (treatment −0.256 ± 0.075). Factor 2 reflected a CSF-dominant inflammatory axis (IL-6 1.390; IL-1β 0.913; TNF-α 0.540; KC/GRO 0.323). Factor scores were positive with injury (+0.179 ± 0.250) versus negative in shams (control −0.322 ± 0.464; drug −0.642 ± 0.182). To a lesser extent than Factor 1, celastrol reduced injury-induced Factor 2 scores (treatment + 0.023 ± 0.275). Factor 3 reflected a serum-dominated inflammatory axis (IL-6 1.320; KC/GRO 1.237; IL-1β 0.649; TNF-α 0.599). Factor scores increased with celastrol (drug + 0.576 ± 0.317; treatment + 0.370 ± 0.202) and were negative in vehicle groups (control −0.421 ± 0.129; injury −0.558 ± 0.266).

Among serum features loading on the hippocampal Factor 1, the largest magnitude was serum IL-6 (−0.291), followed by IL-10 (+0.137) and IL-1β (−0.113), with smaller KC/GRO (+0.080). These serum signatures represent the inflammatory pattern most strongly associated with hippocampal inflammation at 48 h after CCI.

Beyond group-level shifts in factor scores, PERMDISP testing revealed that groups differed substantially in multivariate dispersion within the three-factor space (F = 14.73, permutation p = 0.001). CCI Vehicle animals exhibited roughly twice the mean distance to their group centroid (2.66) compared with Sham Vehicle (1.31), Sham Celastrol (1.04), and CCI Celastrol (1.07). That is, untreated injured animals did not converge on a single shifted inflammatory profile; rather, different CCI Vehicle animals occupied widely divergent positions across the hippocampal, CSF, and serum inflammatory axes, reflecting pronounced inter-animal heterogeneity in the multi-compartment response to injury. By contrast, CCI Celastrol animals exhibited dispersion statistically indistinguishable from both sham groups, indicating that celastrol treatment not only shifted the mean inflammatory profile but also re-consolidated the multi-compartment inflammatory response to sham-like homogeneity. This pattern suggests that untreated TBI produces a destabilized inflammatory state in which individual animals follow disparate multi-compartment trajectories and that celastrol restores coherent regulation across compartments.

Further testing of factor score separation confirmed group differences. PERMANOVA across the joint three-factor space was highly significant (pseudo-F = 5.90, p < 0.001; 9,999 permutations), with pairwise separation between Sham Vehicle and CCI Vehicle (p = 0.010), CCI Vehicle and CCI Celastrol (p = 0.009), and Sham Vehicle and Sham Celastrol (p = 0.049). Per-factor testing showed Factor 1 (hippocampal axis) separated Sham from CCI groups (Kruskal–Wallis p = 0.045; pooled Sham vs. CCI Mann–Whitney p = 0.012) and Factor 3 (serum axis) separated vehicle from celastrol groups (Kruskal–Wallis p = 0.010; CCI Vehicle vs. CCI Celastrol p = 0.032; Sham Vehicle vs. Sham Celastrol p = 0.021). Factor 2 (CSF axis) did not reach significance across all four groups (Kruskal–Wallis p = 0.24) but showed a pooled Sham vs. CCI difference (p = 0.066).

Discussion

This paired, within-animal multi-compartment study of ipsilateral hippocampus tissue lysate, CSF, and serum characterized key cytokine markers of neuroinflammation 48 h after CCI. Injury produced an inflammatory profile with the greatest detectable changes in the hippocampus, smaller effects in CSF, and minimal serum changes. Summarizing within-condition changes as effect-size ratios across compartments provided distributional insight beyond absolute concentrations and revealed a modest positive CSF component alongside a negative serum component for several pro-inflammatory cytokines, contrary to the uniformly pro-inflammatory gradient across compartments that spatial proximity to the injury site might predict. Celastrol was associated with reallocation of these responses in a cytokine- and compartment-specific manner. MCFA defined coherent inflammatory and immunomodulatory axes and identified serum patterns that tracked hippocampal signals, reinforcing a compartment-specific interpretation of post-TBI inflammation.

Injury produced robust hippocampal IL-6, TNF-α, and KC/GRO increases; serum reductions in IL-6 and TNF-α, with a near-significant IL-10 decrease; and minimal CSF changes. The outlier removal sensitivity analysis additionally showed hippocampal IFN-γ suppression. Effect-size ratios for IL-6, KC/GRO, and TNF-α were dominated by hippocampal changes, with CSF changes generally exceeding serum, suggesting that most of the pro-inflammatory response is concentrated in the hippocampus with a smaller CSF component and a lesser, opposing serum contribution. Several mechanisms may account for this compartmental divergence: high local production within parenchyma, rapid dilution and turnover in CSF, and active clearance or counter-regulation in blood at this acute timepoint. Prior studies support early hippocampal pro-inflammatory elevation and mechanisms for compartmental divergence but, to our knowledge, do not provide paired within-animal multi-compartment data for direct comparison. The coordinated changes in IL-6, TNF-α, and KC/GRO we report align with acute activation of canonical glial innate pathways and chemokine-driven recruitment after experimental TBI.^13,26

In uninjured animals, celastrol reduced CSF IL-10 with no serum effects. The IL-10 effect-size ratio was disproportionately CSF-driven (1:39.67:−0.97), suggesting a CSF-centric regulatory signal in the absence of injury. In injured animals, the outlier removal sensitivity analysis revealed celastrol increased hippocampal IL-10 and restored IFN-γ, effects not significant in the primary analysis. For IL-6, the effect-size ratio developed a small serum component while remaining hippocampus-dominated (1:−0.06:0.02), whereas TNF-α remained hippocampus-dominated with CSF greater than serum (1:0.30:0.00). This recalibration of injury effects by celastrol aligns with expected microglia/macrophage targets (e.g., suppression of NF-κB/JNK/STAT1 pathways, modulation of HMGB1, promotion of reparative phenotypes in other CNS injury models), but given that cell-level phenotyping and migration data are beyond the scope of this study, conclusions about the immune cell trafficking and activation patterns that may underlie these compartmental shifts remain inferential.^48–50 Nonetheless, the multi-compartment approach revealed compartment-specific therapeutic effects that would not have been apparent from single-compartment analyses alone.

Pooled data across conditions demonstrated KC/GRO–IL-6–TNF-α coupling within hippocampus and serum; in CSF this triad was only partially preserved, with weaker KC/GRO relationships and stronger IL-6–TNF-α and IL-6–IL-1β associations. IL-10–IL-4 coordination was present in all compartments, with IFN-γ strongly co-varying with these cytokines in hippocampus and serum but showing weaker associations in CSF. Although IFN-γ is traditionally classified as a pro-inflammatory (type 1 helper T cell) cytokine, its empirical co-expression with IL-10 and IL-4 in our data is consistent with its documented context-dependent immunomodulatory roles, including regulatory feedback during CNS injury and its capacity to promote both classically activated and alternatively activated microglial phenotypes depending on the local cytokine milieu.^13,17 We therefore interpret this cluster as immunomodulatory rather than strictly anti-inflammatory. Cross-compartment associations were limited, consistent with predominantly local regulation. The tight KC/GRO–IL-6–TNF-α coupling within compartments is characteristic of coordinated innate/glial programs and chemokine-driven recruitment. Similar co-varying clusters and temporal co-peaks among TNF-α, KC/GRO, and IL-6 have been demonstrated within brain extracellular fluid after severe TBI, with different timing and magnitude than in blood, consistent with the compartment-specific coordination we observed.^11,29,51,52 Partial preservation of this triad in CSF, with weaker KC/GRO links and stronger IL-6–TNF-α and IL-6–IL-1β associations, is consistent with CSF acting as a diluted, region-averaged compartment with distinct kinetics and BBB-dependent dynamics.¹¹

MCFA identified a hippocampus-dominant inflammatory axis, an immunomodulatory axis, and a serum-dominated axis elevated by celastrol. Factor score shifts clarified how serum tracked hippocampal changes even when direct CSF signals were subtle. For the hippocampal axis (Factor 1), MCFA learned a covariant serum fingerprint: IL-6 and IL-1β negative, IL-10 and KC/GRO positive (−0.291, −0.113, + 0.137, and +0.080, respectively). In practice, when hippocampal inflammation is high, serum IL-6 and IL-1β are relatively lower, while serum IL-10 and KC/GRO are relatively higher. This exploratory serum signature of hippocampal inflammation, while derived from a limited sample, warrants prospective validation. The ability of MCFA to learn this graded serum fingerprint associated with active hippocampal pro-inflammatory signals suggests the utility of an integrative framework for multi-compartment TBI biomarker studies. Future studies can leverage longitudinal datasets, which MCFA was designed to accommodate, incorporating paired neurological outcomes, injury markers, genotypes, and cell-specific changes across multiple timepoints alongside inflammatory biomarkers.

The multivariate dispersion analysis of factor scores revealed a finding with potentially broad implications for TBI research: injury did not simply shift the multi-compartment inflammatory profile to a new location in factor space; it destabilized it. CCI Vehicle animals exhibited roughly twice the within-group spread of any other group across the three-factor space, meaning that individual untreated injured animals followed widely divergent multi-compartment inflammatory trajectories at 48 h post-CCI. This heterogeneity is consistent with the well-documented clinical and pre-clinical variability of TBI pathophysiology, in which the extent and localization of secondary injury cascades, BBB disruption, and immune cell infiltration differ substantially between individuals even under controlled experimental conditions.^26,53 Notably, this heterogeneity was captured simultaneously across hippocampal, CSF, and serum compartments, revealing that the variability is not confined to the injury site but extends across the brain-periphery axis. Celastrol treatment reconsolidated the multi-compartment inflammatory response: CCI Celastrol animals showed dispersion indistinguishable from both sham groups (1.07 vs. 1.04 and 1.31), suggesting that the therapeutic effect extends beyond shifting mean cytokine levels and includes restoration of coordinated, predictable regulation across compartments. If replicated, the ability to quantify multi-compartment inflammatory destabilization and its reversal by treatment could serve as an integrative end-point for pre-clinical therapeutic screening, complementing traditional single-analyte comparisons with a measure of how coherently the inflammatory system is regulated across biological compartments.

MCFA factor scores and loadings may also support precision medicine approaches in future pre-clinical studies by identifying individual animals with outlier inflammatory trajectories and defining intervention thresholds based on temporally sampled molecular markers and neurological assessments.

MCFA, adapted from the MOFA+ framework, differs from traditional PCA in several respects relevant to multi-compartment biomarker studies. PCA, whether supervised or unsupervised, operates on a single concatenated data matrix and maximizes total variance, making it susceptible to compartments or cytokines with the largest absolute variances dominating the solution. MCFA treats each compartment as a separate data view and identifies latent factors that explain shared variation across views while allowing view-specific weights, thereby distinguishing coordinated multi-compartment patterns from compartment-specific noise. Unlike supervised methods that require group labels during decomposition, MCFA is fully unsupervised, reducing the risk of overfitting to known group structure in small samples. The Bayesian sparsity priors in MOFA+ further regularize the model, automatically driving negligible loadings toward zero, which is advantageous when the number of features approaches or exceeds the sample size. In our application with 24 features across 48 animals, MOFA+’s built-in regularization provides a principled framework that standard PCA lacks to better handle sample-to-feature ratios observed in small series multiplex studies.

Limitations

This study examined a single acute timepoint in male rats, limiting inference on temporal trajectories, pharmacodynamics, and sex differences. Overall, 19.5% of cytokine measurements were not detected; rates were highest in CSF (26.9%), IL-5 (67–86%), and IL-13 (30–54%) (Supplementary Fig. S1). IL-5 and IL-13 were excluded from interpretation due to nonrandom missingness. An additional 30.4% fell below the manufacturer LLOD but were retained given rank-based testing. CSF power was limited by low analyte concentrations and reduced post-injury volumes.^53,54 Exclusion of not-detected values in primary analysis may reduce power and introduce bias; however, all findings significant under exclusion remained significant under half-minimum imputation, and five additional tests gained significance under imputation, suggesting exclusion was the more conservative approach (Supplementary Table S3).

Plate-level normalization produces non-absolute values; effect-size ratios reflect within-condition apportionment relative to hippocampus, not absolute concentration differences. Our paired design reduces between-animal variance but does not identify cellular sources or transport mechanisms; we did not measure BBB permeability, pharmacokinetics, or immune cell trafficking, so mechanistic inferences about compartment shifts remain provisional. MCFA models are observational and require prospective validation. Future studies will incorporate longitudinal sampling, female cohorts, analyte-specific dilution optimization, barrier and pharmacokinetic measurements, and cell-resolved assays paired with MCFA to test the mechanistic hypotheses raised.

Conclusions

Multi-compartment profiling identified a hippocampus-dominant IL-6, TNF-α, and KC/GRO inflammatory response 48 h after CCI, with preserved immunomodulatory coordination within hippocampus, CSF, and serum. Celastrol reversed injury-induced serum TNF-α changes; sensitivity analyses additionally showed restoration of hippocampal IFN-γ, indicating cytokine- and compartment-specific regulation rather than uniform suppression. Untreated injury doubled multi-compartment inflammatory heterogeneity between animals, while celastrol restored this dispersion to sham-like levels, suggesting that treatment re-consolidated a destabilized inflammatory state across compartments. MCFA defined coherent inflammatory and immunomodulatory axes, clarified serum-hippocampus coupling, and provided a method for multi-compartment TBI biomarker analysis that warrants validation in larger, longitudinal datasets incorporating paired outcomes, genotypes, and cell-specific data.

Authors’ Contributions

Conceptualization (B.G., S.R., S.C., and E.D.); Formal analysis (B.G., Y.C., S.C., and E.D.); Investigation (B.G., Z.R., J.H., and S.C.); Methodology (B.G., S.C., and E.D.); Software (B.G.); Visualization (B.G.); Writing—original draft (B.G. and S.R.); Writing—review and editing (B.G., E.P., S.C., and E.D.); Project administration (J.H., S.C., and E.D.); Validation (Y.C., S.C., and E.D.); Funding acquisition (E.D.); Resources (S.C. and E.D.); and Supervision (S.C. and E.D.).

Footnotes

Author Disclosure Statement

The authors declare no competing, personal financial, funding, employment, or other competing related to the experiments described in this article.

Funding Information

No funding was received for this article.

Supplemental Material

References

1. Gu

, Zhang

, Li

, et al. Global, regional, and national burden of traumatic brain injury, 1990–2021: A systematic analysis for the global burden of disease study 2021. J Neurotrauma 2025;42(19–20):1805–1815; doi: 10.1089/neu.2025.0039

2. Narayan

, Michel

, Ansell

, et al. Clinical trials in head injury. J Neurotrauma 2002;19(5):503–557; doi: 10.1089/089771502753754037

3. Maas

AIR

, Marmarou

, Murray

, et al. Clinical trials in traumatic brain injury: Current problems and future solutions. In: Mechanisms of Secondary Brain Damage from Trauma and Ischemia. ( Baethmann

, Eriskat

, Lehmberg

, et al. eds.) Springer Vienna: Vienna; 2004; pp. 113–118; doi: 10.1007/978-3-7091-0603-7_16

4. Loane

, Faden

. Neuroprotection for traumatic brain injury: Translational challenges and emerging therapeutic strategies. Trends Pharmacol Sci 2010;31(12):596–604; doi: 10.1016/j.tips.2010.09.005

5. Lei

, Gao

, Jiang

. Acute traumatic brain injury: Is Current Management Evidence Based? An empirical analysis of systematic reviews. J Neurotrauma 2013;30(7):529–537; doi: 10.1089/neu.2012.2548

6. Stein

. Embracing failure: What the Phase III progesterone studies can teach about TBI clinical trials. Brain Inj 2015;29(11):1259–1272; doi: 10.3109/02699052.2015.1065344

7. Chakraborty

, Skolnick

, Narayan

. Neuroprotection trials in traumatic brain injury. Curr Neurol Neurosci Rep 2016;16(4):29; doi: 10.1007/s11910-016-0625-x

8. Hawryluk

GWJ

, Bullock

. Past, present, and future of traumatic brain injury research. Neurosurg Clin N Am 2016;27(4):375–396; doi: 10.1016/j.nec.2016.05.002

9. Orr

, Lesha

, Kramer

, et al. Traumatic Brain Injury: A comprehensive review of biomechanics and molecular pathophysiology. World Neurosurg 2024;185:74–88; doi: 10.1016/j.wneu.2024.01.084

10.

10. Thapa

, Khan

, Singh

, et al. Traumatic brain injury: Mechanistic insight on pathophysiology and potential therapeutic targets. J Mol Neurosci 2021;71(9):1725–1742; doi: 10.1007/s12031-021-01841-7

11.

11. Crupi

, Cordaro

, Cuzzocrea

, et al. Management of traumatic brain injury: From present to future. Antioxidants (Basel) 2020;9(4):297; doi: 10.3390/antiox9040297

12.

12. Diaz-Arrastia

, Wang

KKW

, Papa

, TRACK-TBI Investigators. et al.; Acute biomarkers of traumatic brain injury: Relationship between plasma levels of ubiquitin C-Terminal Hydrolase-L1 and glial fibrillary acidic protein. J Neurotrauma 2014;31(1):19–25; doi: 10.1089/neu.2013.3040

13.

13. Witcher

, Bray

, Chunchai

, et al. Traumatic brain injury causes chronic cortical inflammation and neuronal dysfunction mediated by microglia. J Neurosci 2021;41(7):1597–1616; doi: 10.1523/JNEUROSCI.2469-20.2020

14.

14. Helmy

, Carpenter

, Menon

, et al. The cytokine response to human traumatic brain injury: Temporal profiles and evidence for cerebral parenchymal production. J Cereb Blood Flow Metab 2011;31(2):658–670; doi: 10.1038/jcbfm.2010.142

15.

15. Frugier

, Morganti-Kossmann

, O’Reilly

, et al. In Situ detection of inflammatory mediators in Post Mortem human brain tissue after traumatic injury. J Neurotrauma 2010;27(3):497–507; doi: 10.1089/neu.2009.1120

16.

16. Clausen

, Marklund

, Hillered

. Acute inflammatory biomarker responses to diffuse traumatic brain injury in the rat monitored by a novel microdialysis technique. J Neurotrauma 2019;36(2):201–211; doi: 10.1089/neu.2018.5636

17.

17. Ansari

. Temporal profile of M1 and M2 responses in the hippocampus following early 24 h of neurotrauma. J Neurol Sci 2015;357(1–2):41–49; doi: 10.1016/j.jns.2015.06.062

18.

18. Loane

, Kumar

, Stoica

, et al. Progressive neurodegeneration after experimental brain trauma: Association with chronic microglial activation. J Neuropathol Exp Neurol 2014;73(1):14–29; doi: 10.1097/NEN.0000000000000021

19.

19. Drieu

, Lanquetin

, Prunotto

, et al. Persistent neuroinflammation and behavioural deficits after single mild traumatic brain injury. J Cereb Blood Flow Metab 2022;42(12):2216–2229; doi: 10.1177/0271678X221119288

20.

20. Boone

, Weisz

, Willey

, et al. Traumatic brain injury induces long-lasting changes in immune and regenerative signaling. Plos One 2019;14(4):e0214741; doi: 10.1371/journal.pone.0214741

21.

21. Aungst

, Kabadi

, Thompson

, et al. Repeated mild traumatic brain injury causes chronic neuroinflammation, changes in hippocampal synaptic plasticity, and associated cognitive deficits. J Cereb Blood Flow Metab 2014;34(7):1223–1232; doi: 10.1038/jcbfm.2014.75

22.

22. Komoltsev

, Frankevich

, Shirobokova

, et al. Neuroinflammation and neuronal loss in the hippocampus are associated with immediate posttraumatic seizures and corticosterone elevation in rats. Int J Mol Sci 2021;22(11):5883; doi: 10.3390/ijms22115883

23.

23. Komoltsev

, Tret’yakova

, Frankevich

, et al. Neuroinflammatory cytokine response, neuronal death, and microglial proliferation in the hippocampus of rats during the early period after lateral fluid percussion- Induced traumatic injury of the neocortex. Mol Neurobiol 2022;59(2):1151–1167; doi: 10.1007/s12035-021-02668-4

24.

24. Lopez-Caperuchipi

, Kürzinger

, Hopp-Krämer

, et al. Posttraumatic learning deficits correlate with initial trauma severity and chronic cellular reactions after closed head injury in male mice. Exp Neurol 2021;341:113721; doi: 10.1016/j.expneurol.2021.113721

25.

25. Almeida-Suhett

, Prager

, Pidoplichko

, et al. GABAergic interneuronal loss and reduced inhibitory synaptic transmission in the hippocampal CA1 region after mild traumatic brain injury. Exp Neurol 2015;273:11–23; doi: 10.1016/j.expneurol.2015.07.028

26.

26. Almeida-Suhett

, Li

, Marini

, et al. Temporal course of changes in gene expression suggests a cytokine-related mechanism for long-term hippocampal alteration after controlled cortical impact. J Neurotrauma 2014;31(7):683–690; doi: 10.1089/neu.2013.3029

27.

27. Fortier-Lebel

, Jobin

, Lécuyer-Giguère

, et al. Verbal episodic memory alterations and hippocampal atrophy in acute mild traumatic brain injury. J Neurotrauma 2021;38(11):1506–1514; doi: 10.1089/neu.2020.7475

28.

28. Broussard

, Redell

, Zhao

, et al. Mild traumatic brain injury decreases spatial information content and reduces place field stability of hippocampal CA1 neurons. J Neurotrauma 2020;37(2):227–235; doi: 10.1089/neu.2019.6766

29.

29. Witgen

, Lifshitz

, Smith

, et al. Regional hippocampal alteration associated with cognitive deficit following experimental brain injury: A systems, network and cellular evaluation. Neuroscience 2005;133(1):1–15; doi: 10.1016/j.neuroscience.2005.01.052

30.

30. De Simoni

, Grover

, Jenkins

, et al. Disconnection between the default mode network and medial temporal lobes in post-traumatic amnesia. Brain 2016;139(Pt 12):3137–3150; doi: 10.1093/brain/aww241

31.

31. Dyhrfort

, Shen

, Clausen

, et al. Monitoring of protein biomarkers of inflammation in human traumatic brain injury using microdialysis and proximity extension assay technology in neurointensive care. J Neurotrauma 2019;36(20):2872–2885; doi: 10.1089/neu.2018.6320

32.

32. Lassarén

, Lindblad

, Frostell

, et al. Systemic inflammation alters the neuroinflammatory response: A prospective clinical trial in traumatic brain injury. J Neuroinflammation 2021;18(1):221; doi: 10.1186/s12974-021-02264-2

33.

33. Lindblad

, Pin

, Just

, et al. Fluid proteomics of CSF and serum reveal important neuroinflammatory proteins in blood–brain barrier disruption and outcome prediction following severe traumatic brain injury: A prospective, observational study. Crit Care 2021;25(1):103; doi: 10.1186/s13054-021-03503-x

34.

34. Dalgard

, Cole

, Kean

, et al. The cytokine temporal profile in rat cortex after controlled cortical impact. Front Mol Neurosci 2012;5:6; doi: 10.3389/fnmol.2012.00006

35.

35. Allison

, Cacabelos

, Lombardi

VRM

, et al. Celastrol, a potent antioxidant and anti-inflammatory drug, as a possible treatment for Alzheimer’s disease. Prog Neuropsychopharmacol Biol Psychiatry 2001;25(7):1341–1357; doi: 10.1016/S0278-5846(01)00192-0

36.

36. Yang

, Chen

, Cui

, et al. Celastrol, a triterpene extracted from the Chinese “Thunder of God Vine,” Is a potent proteasome inhibitor and suppresses human prostate cancer growth in nude mice. Cancer Res 2006;66(9):4758–4765; doi: 10.1158/0008-5472.CAN-05-4529

37.

37. Li

, He

, Zhang

, et al. Protective effect of celastrol in rat cerebral ischemia model: Down-regulating p-JNK, p-c-Jun and NF-κB. Brain Res 2012;1464:8–13; doi: 10.1016/j.brainres.2012.04.054

38.

38. Venkatesha

, Yu

, Rajaiah

, et al. Celastrus-derived celastrol suppresses autoimmune arthritis by modulating antigen-induced cellular and humoral effector responses. J Biol Chem 2011;286(17):15138–15146; doi: 10.1074/jbc.M111.226365

39.

39. Zhang

, Zhong

, Chen

, et al. Corrigendum: Neuroprotective effects of celastrol on transient global cerebral ischemia rats via regulating HMGB1/NF-κB signaling pathway. Front Neurosci 2021;15:637004; doi: 10.3389/fnins.2021.637004

40.

40. Paris

, Ganey

, Laporte

, et al. Reduction of β-amyloid pathology by celastrol in a transgenic mouse model of Alzheimer’s disease. J Neuroinflammation 2010;7(1):17; doi: 10.1186/1742-2094-7-17

41.

41. Cleren

, Calingasan

, Chen

, et al. Celastrol protects against MPTP‐ and 3‐nitropropionic acid‐induced neurotoxicity. J Neurochem 2005;94(4):995–1004; doi: 10.1111/j.1471-4159.2005.03253.x

42.

42. Sethi

, Ahn

, Pandey

, et al. Celastrol, a novel triterpene, potentiates TNF-induced apoptosis and suppresses invasion of tumor cells by inhibiting NF-κB–regulated gene products and TAK1-mediated NF-κB activation. Blood 2007;109(7):2727–2735; doi: 10.1182/blood-2006-10-050807

43.

43. Westerheide

, Bosman

, Mbadugha

BNA

, et al. Celastrols as Inducers of the Heat Shock Response and Cytoprotection. J Biol Chem 2004;279(53):56053–56060; doi: 10.1074/jbc.M409267200

44.

44. Eroglu

, Kimbler

, Pang

, et al. Therapeutic inducers of the HSP70/HSP110 protect mice against traumatic brain injury. J Neurochem 2014;130(5):626–641; doi: 10.1111/jnc.12781

45.

45. Divya

, Dineshbabu

, Soumyakrishnan

, et al. Celastrol enhances Nrf2 mediated antioxidant enzymes and exhibits anti-fibrotic effect through regulation of collagen production against bleomycin-induced pulmonary fibrosis. Chem Biol Interact 2016;246:52–62; doi: 10.1016/j.cbi.2016.01.006

46.

46. Argelaguet

, Velten

, Arnol

, et al. Multi‐OMICS Factor Analysis–-a framework for unsupervised integration of multi‐omics data sets. Mol Syst Biol 2018;14(6):e8124; doi: 10.15252/msb.20178124

47.

47. Argelaguet

, Arnol

, Bredikhin

, et al. MOFA+: A statistical framework for comprehensive integration of multi-modal single-cell data. Genome Biol 2020;21(1):111; doi: 10.1186/s13059-020-02015-1

48.

48. An

, Youn

, Kim

, et al. Celastrol suppresses expression of adhesion molecules and chemokines by inhibiting JNK-STAT1/NF-κB activation in poly(I:C)-stimulated astrocytes. BMB Rep 2017;50(1):25–30; doi: 10.5483/BMBRep.2017.50.1.114

49.

49. Liu

D-D

, Luo

, Gu

, et al. Celastrol exerts a neuroprotective effect by directly binding to HMGB1 protein in cerebral ischemia–reperfusion. J Neuroinflammation 2021;18(1):174; doi: 10.1186/s12974-021-02216-w

50.

50. Jiang

, Liu

, Zhang

, et al. Celastrol treatment protects against acute ischemic stroke-induced brain injury by promoting an IL-33/ST2 axis-mediated microglia/macrophage M2 polarization. J Neuroinflammation 2018;15(1):78; doi: 10.1186/s12974-018-1124-6

51.

51. Samal

, Waites

, Almeida-Suhett

, et al. Acute response of the hippocampal transcriptome following mild traumatic brain injury after controlled cortical impact in the rat. J Mol Neurosci 2015;57(2):282–303; doi: 10.1007/s12031-015-0626-2

52.

52. Helmy

, Antoniades

, Guilfoyle

, et al. Principal component analysis of the cytokine and chemokine response to human traumatic brain injury. PLoS ONE 2012;7(6):e39677; doi: 10.1371/journal.pone.0039677

53.

53. Santacruz

, Vincent

J-L

, Bader

, et al. Association of cerebrospinal fluid protein biomarkers with outcomes in patients with traumatic and non- traumatic acute brain injury: Systematic review of the literature. Crit Care 2021;25(1):278; doi: 10.1186/s13054-021-03698-z

54.

54. Agoston

, Shutes-David

, Peskind

. Biofluid biomarkers of traumatic brain injury. Brain Inj 2017;31(9):1195–1203; doi: 10.1080/02699052.2017.1357836