Metabolomic, Proteomic, and Transcriptomic Changes in Adults with Epilepsy on Modified Atkins Diet

Abstract

Objective:

High fat and low carbohydrate diets can reduce seizure frequency in some treatment-resistant epilepsy patients, including the more flexible Modified Atkins Diet (MAD) that is more palatable, mimicking fasting and inducing high ketone body levels. Low carbohydrate diets may shift brain energy production, particularly impacting neuron and astrocyte linked metabolism.

Methods:

We evaluated the effect of short-term MAD on molecular mechanisms in adult epilepsy patients from surgical brain tissue and plasma compared to Control participants consuming a non-modified higher carbohydrate diet (n = 6 MAD, mean age 43.7 years, range 21–53, diet average 10 days; n = 10 Control, mean age 41.9 years, range 28–64).

Results:

By metabolomics, there were 13 increased metabolites in plasma (n = 15 participants with available specimens) that included 4.10-fold increased ketone body 3-hydroxybutyric acid, decreased palmitic acid in cortex (n = 16), and 11 decreased metabolites in hippocampus (n = 6) that had top associations with mitochondrial functions. Cortex and plasma 3-hydroxybutyric acid levels had a positive correlation (p = 0.0088, R² = 0.48). Brain proteomics and RNAseq identified few differences, including 2.75-fold increased hippocampal MT-ND3 and trends (p < 0.01, FDR > 5%) in hippocampal NADH related signaling pathways (activated oxidative phosphorylation and inhibited sirtuin signaling).

Significance:

Short-term MAD was associated with metabolic differences in plasma and resected epilepsy brain tissue when compared to Control participants, in combination with trending expression changes observed in hippocampal NADH related signaling pathways. Future studies should evaluate how brain molecular mechanisms are altered with long-term MAD in a larger cohort of epilepsy patients, with correlations to seizure frequency, epilepsy syndrome, and other clinical variables. [Clinicaltrials.gov NCT02565966]

Introduction

Brain metabolism is altered in epilepsies, playing a role in seizure generation and also altered as a result of seizure activity. ¹ Specifically, glucose metabolism can be impaired in epilepsy² and contribute to altered neuronal membrane potentials that foster a feed-forward cycle in seizure generation.³ Our proteomics studies across a broad range of epilepsy syndromes identified altered brain metabolism in several brain regions, including the hippocampus, frontal cortex, and brainstem.^4,5 Altered metabolic pathways have also been identified by RNAseq in the hippocampus of mesial temporal lobe epilepsy (MTLE) patients by our group.⁶ There is evidence that modifying brain metabolism with high fat and low carbohydrate diets may result in reduced seizure frequency in some treatment-resistant epilepsy patients.^1,7,8 These diets include variations of the ketogenic diet, including the more flexible Modified Atkins Diet (MAD) that is more palatable, mimicking fasting with high ketone body levels.^1,8 Low carbohydrate diets shift brain energy production, particularly impacting neuron and astrocyte linked metabolism.^9–12 Neurons primarily rely on glucose as an energy source, as glycolysis requires less oxygen, minimizes oxidative stress, and quickly generates ATP when compared to fatty acid beta-oxidation.¹⁰

Molecular studies of ketogenic diets implicate several potential mechanisms that impact seizure frequency, including activation of K_ATP channels by mitochondrial metabolism related to ketone bodies, elevated adenosine levels that may reduce neuronal excitability through adenosine A1 receptors, altered NAD/NADH levels that may impact redox and enzymatic reactions associated with sirtuins and poly(ADP-ribose) polymerases (PARP) that mediate transcription, and mammalian target of rapamycin (mTOR) inhibition.^8,13–15 In human studies, meta-analysis indicated some evidence for ketogenic diets reducing seizure frequency, particularly in pediatric epilepsy.⁷ A prospective study of MAD in 30 adult treatment-resistant epilepsy patients with various epilepsy syndromes indicated seizure reduction with a median time of 2 weeks on the diet, with a range from 1 to 8 weeks; overall with 47% of participants having >50% seizure reduction after 1 month.¹⁶ Diet induced brain molecular mechanisms have been explored more widely in animal models. In an epilepsy rat model, mTOR signaling pathway activation, observed in some epilepsy patients,^17–20 was prevented in hippocampus and cortex with a ketogenic diet as early as 7 days after seizure induction.²¹ Hippocampal proteomics in a young epilepsy rat model after a 28 day ketogenic diet showed top pathway enrichment for vitamin digestion and absorption.²² Although ketogenic diet therapies have been used to modulate seizure frequency since the 1920s,⁷ the mechanisms of action in human brain tissue are not well understood.

In this study, we evaluated molecular mechanisms in brain tissue and plasma from adult epilepsy patients after an average of 10 days of MAD prior to brain surgical resection compared to Control participants that consumed a non-modified, higher carbohydrate diet.

Materials and Methods

Biospecimens

This study (Clinicaltrials.gov NCT02565966) was approved by NYU Grossman School of Medicine Institutional Review Board. Participants were enrolled from 2015–2019. Written informed consent was provided by each participant. Adults (≥ 18 years) were included that were undergoing routine epilepsy surgical resection. Participants were excluded if following an Atkins or low glycemic index diet or if had systemic corticosteroid treatment five days prior to enrollment. The study was designed for participants to be randomized to MAD or no intervention, adhere to MAD for 1–4 weeks prior to surgery, followed by surgical resection, and biospecimen collection. Due to most participants coming from long distances, MAD was adhered to for an average of 10 days prior to surgical resection. Control participants adhered to a non-modified, higher carbohydrate diet. At surgical resection, both peripheral blood and brain tissue were collected. A portion of resected brain tissue was immediately frozen and stored at −80°C; another portion was processed by formalin fixed paraffin embedding (FFPE) for neuropathology. Nineteen participants were enrolled; 16 participants with biospecimens. Participants were age (MAD: 43.7 years mean, range 21–53, Control: 41.9 years, range 28–64; p = 0.80) and sex matched (p > 0.99; Table 1, Supplemental Table 1). Changes to seizure frequency were not evaluated, given brief study duration and high intra-individual variability in seizure frequency.

Table 1.

Case History

Diet Group	Case ID	Age at Surgery	Sex	Age of Seizure Onset	Brain Regions	Neuropathology
Control	1	55	M	25	TL, HP	TL: cortical dysplasia (mild to moderate), white matter hypercellularity and heterotopia, gliosis, microglial activation, perivascular dilitation and fibrosis, leptomeningeal histiocytic infiltrate;
						Amygdala/HP: hypercellular, cortical dysplasia (mild), gliosis, neuronal loss, histiocytic infiltrate, microglial activation, vascular hyalinization
Control	2	39	M	27	TL/cavernoma	Cavernoma
Control	3	63	F	44	TL, HP	TP: cavernomas with microcalcifications;
						HP: reactive changes (focal DG bilamination, focal neuronal loss)
Control	4	33	F	30	TL, HP	TL and HP: reactive gliosis
Control	5	64	F	51	TL	TL: reactive gliosis;
						HP: hippocampal sclerosis
Control	6	29	F	16	TL, HP	TL: FCD IB, reactive gliosis, numerous white matter neurons;
						HP: FCD IIA
Control	7	33	M	26	TL/cavernoma	Cavernoma
Control	8	28	F	25	TL	Mild architectural abnormalities, reactive changes (focal neuronal loss and reactive gliosis), epidermal inclusion cyst
Control	9	40	F	21	OC	FCD IIA
Control	10	35	M	28	TL	Cavernoma with evidence of previous hemorrhage; reactive changes (mild hypercellularlity, gliosis)
MAD	11	52	M	7	TL	DNT, WHO grade I
MAD	12	44	F	42	TL	Cavernoma
MAD	13	21	F	19	FR	Low-grade DNT
MAD	14	49	M	26	TL, HP	TL: FCD IC;
						HP: unremarkable
MAD	15	43	F	43	TL	TL: gliosis;
						Amygdala/HP: gliosis and focal neuronal loss
MAD	16	53	M	50	TL, HP	TL: unremarkable;
						HP: cavernoma and hippocampal sclerosis

TL = temporal lobe, FR = frontal lobe, OC = occipital lobe, HP = hippocampus, DG = dentate gyrus, FCD = focal cortical dysplasia, DNT = dysembryoplastic neuroepithelial tumor

Note: Although neuropathology performed in all resected brain regions, not all brain regions available for additional analyses (i.e. hippocampus for cases Control-5, MAD-15). Cortical tissue n = 16 (10 control, 6 MAD) and hippocampus n = 6 (n = 4 control, n = 2 MAD).

Plasma Metabolomics

Peripheral whole blood was collected into EDTA tubes (n = 15) and stored at −80°C. Whole blood was thawed on ice and processed to remove hemoglobin by NuGel-HemogloBind with Spin-X column (Biotech Support Group) according to manufacturer protocol, as plasma was not available. For each participant 20ul whole blood was diluted in 200ul Hemoglobin Binding Buffer, 175ul of diluted blood was processed on Spin-X column, eluate collected, and dried down for processing with NYU Metabolomics Core. Briefly, samples were analyzed with the hybrid LCMS assay after scaling metabolite extraction to 100ul/sample. Overall, library coverage was moderate with 95/147 metabolites detected in ≥3 samples, 65 detected in n=15 after background threshold correction. Data were analyzed by principal components analysis (PCA) and differential expression (significance at p<0.05). Pathways related to significant metabolites were evaluated in Metaboanalyst 5.0 (metaboanalyst.ca) and by Human Metabolome Database (HMDB).²³ Receiver operating characteristic (ROC) was evaluated in GraphPad Prism 9.2.0.

Brain Metabolomics

NYU Metabolomics Core analyzed brain tissue with the hybrid LCMS assay after scaling metabolite extraction to 10mg brain tissue/1ml extra buffer (n=22). Cocktail of isotope labeled amino acid standards was spiked into metabolite extraction solvent cocktail. Peak intensities were extracted by library of m/z values and retention times for doubly labeled (13C and 15N uniform) amino acids. Intensities were extracted with an in-house script with 10ppm tolerance for theoretical m/z of each isotope labeled standard, and maximum 30sec retention time window. The retention time range and mass error was evaluated as well as percent coefficient of variance (CV) in peak intensity across 16 labeled standards. Overall, library coverage was moderate with 113/147 metabolites detected in ≥3 samples, and 38 detected in n=22 after background threshold correction. Data were analyzed similar to plasma metabolomics above.

Brain Proteomics

Protein extraction and digestion.

With NYU Proteomics Laboratory, samples (10mg) were solubilized in 100ul of 100% TFA according to “SPEED” sample prep workflow.²⁴ Briefly, samples were in TFA for 15min at room temperature (RT) following 5min (73°C). Lysates were cooled (RT) and spun 16,500g 5min, supernatants were neutralized by 10 times excess (v/v) 2M TRIS containing 10mM TCEP and 20mM chloroacetamide. After 30min (90°C), samples were centrifuged and supernatants transferred into clean tubes. Protein concentrations were measured by 280nm absorbance. Samples were diluted 6 times in sequencing-grade modified trypsin at 50:1 mass ratio (protein to trypsin) in water overnight (37°C). Enzymatic digestions were stopped by acidification with TFA (to final 2%) and peptide concentrations measured by 280nm absorbance.

TMT-Pro labeling.

TMT-Pro isobaric labeling (ThermoFisher) was combined with peptide desalting to minimize sample handling and associated variability. Peptides (20ug/sample) were loaded into Waters tC18 u-elution 96 plate (5mg sorbent/well), washed 2 × 200ul of 2% ACN 0.2% formic acid (FA) and eluted in mixture of 18ul of 50 mM HEPES (pH 8) and 10ul of TMT-Pro label ACN stock (12.5mg/ml). Peptides were eluted from C18 directly into TMT labeling buffer to limit Speedvac steps. TMT labeling was performed for 30min (RT). Excess TMT reagent was neutralized with 500mM ammonium bicarbonate (to final 50mM, 30min, 37°C). Small sample aliquots were analyzed by LC-MS/MS to confirm labeling completion. Samples were pooled (16 cortex and 6 hippocampus split in separate TMT batches) and desalted on SepPak C18 cartridges. Eluates were dried by Speedvac and stored (−20°C).

Offline basic pH reverse phase fractionation.

TMT labeled peptides were resolubilized in 50ul of 5% ACN 10mM ammonium bicarbonate and fractionated by high-pH reverse-phase chromatography on Waters XBridge BEH 130A C18 3.5um 4.63mm ID x 250mm column coupled to Agilent 1260 Infinity series HPLC system at 1ml/min flow rate with three buffer lines: Buffer A water, buffer B ACN, and Buffer C 100mM ammonium bicarbonate. Peptides were separated by linear gradient from 5% B to 35% B (62min) followed by linear increase to 60% B (5min), and ramped to 70% B (3min). Buffer C was constantly introduced throughout gradient at 10%. Fractions were collected at 60s, combined in nonconsecutive pooling scheme for 24 final samples. Fractions were acidified with FA to 0.5% final concentration and vacuum centrifugation concentrated.

LC-MS/MS.

Peptides resuspended by 0.1% FA were loaded on Evosep tips according to manufacturer and separated on C18 analytical column (15cm x 150um ID, packed with ReproSil-Pur C19 1,9A beads, Evosep cat# EV1106) over 88min gradient using extended Evosep One method (15SPD) on Evosep One LC system (www.evosep.com). Peptides were gradient eluted from column directly to Orbitrap HFX mass spectrometer. High resolution full MS spectra were acquired with resolution of 120,000, AGC target of 3e6, with maximum ion injection time of 100ms, and scan range of 400–1600 m/z. Following each full MS scan, 20 data-dependent HCD MS/MS scans were acquired at resolution of 60,000, AGC target of 5e5, maximum ion time of 100 ms, one microscan, 0.4 m/z isolation window, NCE of 35, fixed first mass 100 m/z, and dynamic exclusion for 45s. Both MS and MS/MS spectra were recorded in profile mode.

Data Analysis.

MS raw data were analyzed using MaxQuant software (1.6.3.4) and searched against SwissProt human uniprot database (http://www.uniprot.org/) containing 20,430 entries. Database search was performed in Andromeda integrated into MaxQuant. A list of 248 common laboratory contaminants were also added to database as well as reversed versions of all sequences. For searching, enzyme specificity was set to trypsin with 2 maximum missed cleavages. Precursor mass tolerance was set to 20ppm for first search used for non-linear mass re-calibration,²⁵ and then to 6ppm for main search. Oxidation of methionine was searched as variable modification; carbamidomethylation of cysteines was searched as fixed modification. TMT labeling was set to lysine residues and N-terminal amino groups, corresponding batch-specific isotopic correction factors were accounted for. The false discovery rate (FDR) for peptide, protein, and site identification was set to 1%, minimum peptide length was set to 6. Mass spectrometry raw files are accessible under MassIVE ID: MSV000089863.

The protein expression matrix was filtered to include proteins identified by ≥two peptides/brain region (n=3462 cortex, n=3530 hippocampus). All TMT reporter ion intensities were log-transformed and normalized to obtain same median value across all TMT channels. For PCA, missing values were imputed from normal distribution (0.3 width, 1.8 downshift, relative to measured protein intensity distribution). Two sample t-tests were performed on non-imputed data in Perseus (http://www.perseus-framework.org/) to detect differences at FDR<5% (permutation-based with 250 data randomizations). Ingenuity Pathway Analysis (IPA, Qiagen) identified signaling pathways associated with trending differentially expressed proteins (p<0.01). All detected proteins were included in IPA dataset for each brain region, including UniProtID, fold change, and p value. An IPA Core Analysis was performed in each brain region for proteins (p<0.01, FDR>5%). Pathways were considered enriched at p value of overlap<0.05 and activated/inhibited as result of combined protein fold changes in pathway as reflected by |z score|≥2. Brain cell type annotation was performed as reported previously.^4,5,26–28

Brain RNAseq

RNA was isolated from brain tissue (10mg/sample) with miRNeasy Purification of Total RNA (Qiagen). Homogenization was performed in Qiazol with handheld device equipped with pestle, followed by 22-gauge needle. RNA quality and concentration was determined by Agilent Bioanalyzer. Libraries were created by NYU Genome Technology Center with Illumina Stranded Total RNA Prep with Ribo-Zero Plus (Cat.20040529), and sequenced on Illumina NovaSeq6000 with S2 100 cycle flow cell, 50bp paired-end. PCA and differential gene expression was evaluated with NYU Applied Bioinformatics Laboratory. Data was analyzed by sns rna-star pipeline (https://igordot.github.io/sns/routes/rna-star.html). Adapters and low quality bases were trimmed using Trimmomatic (v.0.36). Sequencing reads were mapped to the reference genome (hg38) using STAR aligner (v.2.7.3). Alignments were guided by Gene Transfer Format (GTF). Mean read insert sizes and standard deviations were calculated using Picard tools (v.2.18.20). Genes-samples counts matrix was generated using featureCounts (v.1.6.3), normalized by library size factors using DEseq2, and differential expression analysis was performed. The Read Per Million (RPM) normalized BigWig files were generated using deepTools (v.3.1.0). Statistical analyses were performed in the R environment (4.0.3). Signaling pathway analysis was performed as above with trending transcripts (p<0.01, adjusted p>0.05) in IPA (Qiagen). Brain cell type annotation was evaluated as noted above.

Results

Case History

There were 6 adult epilepsy MAD participants (n = 2 participants with both hippocampus and cortex) with surgical resected brain tissue and whole blood available for analysis, after an average of 10 days MAD. There were 10 adult epilepsy Control participants (n = 4 participants with both hippocampus and cortex) with surgical resected brain tissue available for comparison, as well as 9 participants with whole blood. Clinical information is summarized in Table 1 and detailed in Supplemental Table 1.

Plasma and Brain Metabolomics

In plasma, PCA indicated no segregation of MAD and Control participants (p = 0.15; Fig 1A). Differential expression analysis identified 13 increased metabolites in MAD participants at p<0.05 (Fig 1B, Table 2, Supplemental Table 2), including 4.10-fold increased ketone body 3-hydroxybutyric acid. Related pathways associated with the altered metabolites are noted in Table 2.

Fig 1. Metabolomics identified altered metabolites in modified Atkins diet (MAD) versus control plasma and brain.

(A) Principal components analysis (PCA) indicated no segregation of the control and MAD participants in PCA1 (p = .15). (B) Differential expression analysis in plasma identified 13 increased metabolites in MAD participants. Dashed line indicates significance at p < .05. (C) PCA in cortex indicated no segregation of the control and MAD participants in PCA1 (p = .24). (D) Differential expression analysis in cortex identified one decreased metabolite in MAD participants. Dashed line indicates significance at p < .05. (E) PCA in hippocampus indicated segregation of the control and MAD participants in PCA1 (p = .032). (F) Differential expression analysis in hippocampus identified 11 decreased metabolites in MAD participants. Dashed line indicates significance at p < .05. AMP, adenosine monophosphate; CMP, cytidine monophosphate; GMP, guanosine monophosphate.

Table 2.

Altered Plasma and Brain Metabolites in MAD vs. Control

Metabolite	p Value	Fold Change	Related Pathway
Plasma
Increased
Uric acid	5.44E-03	1.72	Purine metabolism
Fructose 6-phosphate	1.11E-02	1.79	Starch and sucrose metabolism
D-Ribose 5-phosphate	1.43E-02	2.19	Purine metabolism
Ureidopropionic acid	1.46E-02	1.53	Pantothenate and CoA biosynthesis
3-Hydroxybutyric acid	2.06E-02	4.10	Synthesis and degradation of ketone bodies
5-Phosphoribosylamine	2.79E-02	1.14	Purine metabolism
L-Asparagine	2.84E-02	1.24	Aminoacyl-tRNA biosynthesis
L-Isoleucine	3.11E-02	1.95	Valine, leucine, isoleucine biosynthesis
Dihydroxyacetone phosphate	3.22E-02	3.89	Glycerolipid metabolism
L-Valine	3.37E-02	1.41	Valine, leucine, isoleucine biosynthesis
L-Acetylcarnitine	4.07E-02	1.62	Biosynthesis and oxidation of fatty acids
L-Threonine	4.43E-02	1.68	Valine, leucine, isoleucine biosynthesis
Cortex
Decreased
Palmitic acid	2.72E-02	2.26	Fatty acid biosynthesis
Hippocampus
Decreased
Fumaric acid	2.86E-03	3.00	TCA cycle
L-Malic acid	3.07E-03	2.90	TCA cycle
L-Tyrosine	8.06E-03	2.39	Aminoacyl-tRNA biosynthesis
L-Valine	8.60E-03	2.41	Aminoacyl-tRNA biosynthesis
NADH	1.31E-02	3.90	Oxidative phosphorylation
Taurine	1.74E-02	3.11	Taurine and hypotaurine metabolism
Guanosine monophosphate	1.91E-02	2.98	Purine metabolism
Adenosine monophosphate	2.12E-02	2.62	Purine metabolism
L-Tryptophan	3.34E-02	2.89	Aminoacyl-tRNA biosynthesis
Cytidine monophosphate	3.84E-02	2.59	Pyrmidine metabolism
L-Kynurenine	4.31E-02	2.88	Tryptophan metabolism
L-Leucine	4.97E-02	2.45	Valine, leucine, isoleucine biosynthesis

In brain, PCA did not indicate segregation of MAD and Control participants in PCA1 of the cortex (p = 0.24), but there was segregation in PCA1 of the hippocampus (p = 0.032; Fig 1). There was one decreased metabolite in cortex of MAD participants and 11 decreased metabolites in hippocampus at p<0.05 (Fig 1, Table 2, Supplemental Tables 3-4). The highest fold change observed in either brain region was 3.90-fold decreased NADH in hippocampus.

Correlations of significant metabolites in at least plasma or brain (n = 24 metabolites) indicated that the majority of significant plasma metabolites were changing in the same direction in the cortex of MAD participants when compared to Control, although with no overall correlation (p = 0.32, R² = 0.045; Fig 2A). In hippocampus, fewer (n = 2) metabolites were changing in the same direction when compared to plasma and overall there was a positive correlation (p = 0.037, R² = 0.18; Fig 2B). For hippocampus and cortex, there were 5 lower metabolites in both regions but with no overall correlation (p = 0.71, R² = 0.0066; Fig 2C). The ketone body, 3-hydroxybutyric acid, increased in plasma positively correlated to cortex levels (p = 0.0088, R² = 0.48; Fig 2D). There were too few hippocampal samples (n = 2) with corresponding plasma 3-hydroxybutyric acid values to perform a correlation analysis. To evaluate the potential of these metabolites as biomarkers of successful MAD induction, we performed a ROC curve analysis for 3-hydroxybutyric acid with respect to participant diet. There was an area under the curve (AUC) of 0.80 for both plasma (p = 0.059) and cortex (p = 0.079), with each approaching significance (Fig 2E), indicating promising biomarker potential given the small sample size in the current study.

Fig 2. Significant metabolomics correlations in plasma and brain.

A) Of the 24 metabolites significant in at least plasma, cortex, or hippocampus of MAD participants when compared to Control, there were more metabolites changing in the same direction when comparing plasma and cortex (increased). However, there was no overall correlation (p = 0.32, R² = 0.045). Purple points indicate metabolites changing in the same direction and yellow points indicate metabolites changing in opposite directions (i.e. up in cortex and down in plasma of MAD participants). B) When comparing plasma and hippocampus, there were 2 metabolites changing in the same direction (increased). There was an overall positive correlation (p = 0.037, R² = 0.18). C) For hippocampus and cortex, there were 5 metabolites changing in the same direction (decreased). There was no overall correlation (p = 0.71, R² = 0.0066). D) For the ketone body 3-hydroxybutyric acid, there was a positive correlation of plasma and cortex levels (p = 0.0088, R² = 0.48). There were two detectable values available for both hippocampus and plasma in the same participants, thus no correlation analysis was performed. (E) The receiver operating characteristic (ROC) for 3-hydroxybutyric acid indicated an area under the curve = .80 for both plasma (p = .059) and cortex (p = .079). AMP = adenosine monophosphate, CMP = cytidine monophosphate, GMP = guanosine monophosphate.

Brain Proteomics

There were 3462 proteins detected in cortex and 3530 proteins detected in hippocampus (Supplemental Tables 5-6). PCA did not indicate segregation of MAD and Control participants in PCA1 of the cortex (p = 0.92) or hippocampus (p = 0.15; Fig 3A-B). Differential expression analyses identified 11 proteins in cortex and 48 proteins in hippocampus at p < 0.01, but no altered proteins in either brain region at FDR < 5–15% (Fig 3C-D; Supplemental Tables 7-8). The mitochondrial protein TTC19 (tetratricopeptide repeat domain 19; Q6DKK2; p = 5.24 × 10⁻⁴) had the strongest trend for difference in cortex with a 1.51-fold increase in MAD participants. APOC2 (apolipoprotein C2; P02655; p = 3.02 × 10⁻⁴) had the strongest trend for difference in hippocampus with a 1.26-fold increase. The 11 trending cortical proteins (p < 0.01, FDR > 5%) were not associated with signaling pathways (Supplemental Table 9). Among the hippocampal proteins that showed a trend for altered expression in MAD participants (p < 0.01, FDR > 5%; Supplemental Table 8), there were five increased NADH dehydrogenases (NDUFS2, NDUFS3, NDUFA10, NDUFC2, NDUFA13). Thus the trending hippocampal proteins (p < 0.01, 48 proteins) were associated with oxidative phosphorylation activation (p = 7.41 × 10⁻⁵, z = 2.00) and sirtuin signaling pathway inhibition (p = 3.02 × 10⁻⁴, z = −2.00), although no signaling pathways were significant at FDR < 5% (Supplemental Table 10).

Fig 3. Differential proteomic expression analyses in MAD vs. control brain tissue.

A) PCA in cortex indicated no segregation of the Control and MAD participants in PCA1 (p = 0.92). B) PCA in hippocampus indicated no segregation of the Control and MAD participants in PCA1 (p = 0.15). C) Differential expression analysis in cortex did not identify altered proteins at FDR < 5%. The protein with the strongest trend for difference is annotated, TTC19. D) Differential expression analysis in hippocampus did not identify altered proteins at FDR < 5%. The protein with the strongest trend for difference is annotated, APOC2. Cell type annotations are indicated for each protein, with “Neuron” indicating a protein with both excitatory and inhibitory neuron annotations.

Brain cell type annotation identified 11.5% (397/3462) proteins in cortex and 11.4% (401/3530) proteins in hippocampus with annotations. The majority of proteins were not annotated (“Undefined”), as the proteins are expressed ubiquitously among multiple cell types or it is unknown. The majority of annotated proteins (46%) in both regions were generally neuronal (“Neuron” annotation).

Brain RNAseq

PCA did not indicate segregation of MAD and Control participants in PCA1 of the cortex (p = 0.26) or hippocampus (p = 0.51; Fig 4A-B). Differential expression analyses identified 3 increased transcripts in cortex at adj. p < 0.05 (Fig 4C, Supplemental Table 11). AP001025.1 and RNU4–1 (RNA, U4 Small Nuclear 1) are non-protein coding transcripts. RAB3IP (RAB3A Interacting Protein) is a protein coding transcript and was increased 7.06-fold. There were 3 altered transcripts in hippocampus (2 increased, 1 decreased; Fig 5D; Supplemental Table 12). MT-ND3 (mitochondrially encoded NADH:ubiquinone oxidoreductase core subunit 3) was increased 2.75-fold. GRB14 (growth factor receptor bound protein 14) was increased 3.46-fold. Lower hippocampal XIST (1448.15-fold) is likely related to the comparison of a male MAD group and a predominantly female Control group ²⁹. There was no signaling pathway enrichment associated with the altered transcripts at adj. p < 0.05. Differential expression of trending transcripts (p < 0.01, adj. p > 0.05) identified 829 transcripts in cortex (Supplemental Table 13) and 102 transcripts in hippocampus (Supplemental Table 14). The top signaling pathway associated with trending cortex transcripts was EIF2 signaling activation (p = 6.66 × 10⁻²⁸, z = 4.81; Supplemental Table 15). In the hippocampus, trending transcripts were associated with oxidative phosphorylation activation (p = 7.45 × 10⁻⁵, z = 2.00) and sirtuin signaling inhibition (p = 2.99 × 10⁻⁴, z = −2.00; Supplemental Table 16).

Fig 4. RNAseq differential expression analysis in MAD vs. control brain tissue.

A) PCA in cortex indicated no segregation of the Control and MAD participants in PCA1 (p = 0.26). B) PCA in hippocampus indicated no segregation of the Control and MAD participants in PCA1 (p = 0.51). C) Differential expression analysis in cortex identified 3 increased transcripts in MAD participants. Dashed line indicates significance at adjusted p < 0.05. D) Differential expression analysis in hippocampus identified 3 altered transcripts (1 down, 2 up) in MAD participants. Dashed line indicates significance at adjusted p < 0.05. The altered transcripts did not have cell type annotations (“Undefined”).

The altered transcripts (adj. p < 0.05) were not associated with cell type annotations, rather were “Undefined” as the transcripts are expressed ubiquitously among multiple cell types or it is unknown.

Discussion

We identified altered metabolites in the brain and plasma after short-term MAD in epilepsy patients, including increased plasma ketone body 3-hydroxybutyric acid (β-hydroxybutyrate) that correlated to cortex levels and altered brain metabolites associated with mitochondrial functions. Mitochondrial related changes were reflected in RNAseq and proteomics of the hippocampus, with an increased mitochondrially encoded NADH transcript (MT-ND3) and a trend in other increased MT-ND transcripts as well as NADH dehydrogenase proteins. Further, these mitochondrial changes were represented in trends in both RNAseq and proteomics associated with oxidative phosphorylation activation and sirtuin signaling pathway inhibition, related to MT-ND transcript and NDUF protein levels in mitochondrial complex I.

In plasma, altered metabolites were also associated with various cellular energetic functions. We observed elevated plasma 3-hydroxybutyric acid, as in previous ketogenic diet studies, which can be synthesized by the liver from fatty acid beta-oxidation-derived acetyl-CoA with transport into the brain across the blood brain barrier. We found a tight correlation between plasma and brain levels of 3-hydroxybutyric acid, in line with expectations of efficient uptake by the brain, but we did not observe changes in other brain ketone bodies. Synthesis of 3-hydroxybutyric acid can also be carried out locally in the brain,^30,31 but all tissues can utilize 3-hydroxybutyric acid where it functions as an energy source and as a signaling molecule, influencing cellular energy levels, epigenetics, and inflammatory response.³¹ Some studies indicate minimal changes in brain ketone bodies and inconsistent correlation of plasma ketone bodies to seizure frequency.^30,32 Other altered plasma metabolites included increased branched chain amino acids, as well as metabolites associated with lipid and purine metabolism. It will be interesting in future long-term MAD studies to evaluate brain ketone body levels, as well as whether the altered plasma amino acids have implications for a shift in transport across the blood brain barrier along with the glutamate precursor glutamine.^33,34

In cortex, we identified decreased palmitic acid in MAD participants. Palmitic acid is a long chain fatty acid that is one of the most common saturated fatty acids endogenously, can be synthesized intracellularly, or come from dietary sources.³⁵ Palmitic acid can function as a precursor for ketone body synthesis and has been shown to increase ketone body synthesis more prominently in astrocytes when compared to neurons in vitro, which may contribute to a shift in energy production through the astrocyte-neuron lactate and ketone body shuttle systems.^11,36,37 In a non-epilepsy ketogenic diet rat model (10 days), fatty acid metabolism was altered but palmitic acid was not altered when analyzing the whole brain.³⁸ Future studies should evaluate whether decreased palmitic acid is sustained with long-term MAD in this brain region with corresponding changes in downstream metabolites, whether this decrease is specific to astrocytes, neurons, or intravascular levels, and whether levels are associated with different epilepsy syndromes as we see a wide distribution of palmitic acid levels in the Control group and altered levels are reported in some disease groups.³⁵

In hippocampus, we identified several altered metabolites associated with mitochondrial functions, including decreased NADH. In a non-epilepsy ketogenic diet rat model, NAD metabolism is altered early after two days in hippocampus, higher NAD levels are sustained for at least 3 weeks, but not altered in cortex.^39,40 NAD is a rate limiting factor in multiple mitochondrial ATP generating pathways, involved in sirtuin mediated deacetylation, PARP mediated DNA damage repair, and is also secreted influencing multiple extracellular functions.¹⁵ We did not observe significantly altered NAD, NAD/NADH ratio, nor a change in the NAD degradation product¹⁵ adenosine. Changes in NAD and the other hippocampal metabolites should be evaluated further in additional patients, as well as in long-term MAD, particularly because this brain region can be more vulnerable to hyperexcitability and damage due to metabolically demanding functions like plasticity and seizure activity.^2,3,41

Proteomics and RNAseq showed few differences with short-term MAD in both brain regions evaluated, but did indicate trends in activated oxidative phosphorylation and inhibited sirtuin signaling in the hippocampus associated with mitochondrial complex I. By proteomics, there was a trend in increased NADH dehydrogenases (NDUFS2, NDUFS3, NDUFA10, NDUFC2, NDUFA13). RNAseq identified increased MT-ND3 and other trending increased MT-ND transcripts. Mutations in NDUF and MT-ND3 genes are associated with mitochondrial complex I deficiency, resulting in Leigh syndrome that can cause neurodegeneration and neurodevelopmental delay, as well as seizures.^42,43 Our previous proteomics study of hippocampus in epilepsy compared to control cases identified inhibited oxidative phosphorylation, including decreased expression of multiple NADH dehydrogenases, as well as sirtuin signaling activation.⁴ Similar findings of decreased NDUF and MT-ND expression in the hippocampus of MTLE patients compared to controls have been observed with RNAseq by our group.⁶ In combination with the altered changes observed by metabolomics in short-term MAD, there were trends in hippocampal NADH protein and related pathway alterations that contrasted with our previous findings in epilepsy cases compared to control cases^4,6 and may indicate that MAD influences these signaling pathways altered in epilepsy. Future studies should evaluate long-term altered brain molecular mechanisms with correlations to seizure frequency, whether increased mitochondrial complex I protein expression is associated with mitochondrial biogenesis that is reportedly induced by ketogenic diets,^44–46 whether MAD induces a reversal of mitochondrial complex I protein and transcript expression dyshomeostasis that we have observed in epilepsy,^4,6 whether mitochondrial function changes differ by brain subregion or cell type,^4,47, and whether other clinical factors are related to expression changes like epilepsy syndrome, anti-seizure medication mitochondrial toxicity,⁴⁸ and cognitive deficits that can be influenced by ketogenic diets.^49,50

There were several limitations in our study. There was a low number of participants and the study was short-term due to recruitment difficulties, associated with long traveling distances for surgical resection. Several clinical variables were heterogeneous, including seizure etiology, neuropathology, medications, and disease duration.

In summary, we identified metabolic differences in MAD participant plasma and resected epilepsy brain tissue. There are trends in the hippocampus that may indicate a shift in cellular energy production that should be evaluated further in future studies with long-term MAD, with correlations to specific epilepsy syndromes and seizure frequency.

Supplementary Material

supinfo

Supplemental Table 1. Participant Clinical History

Supplemental Table 2. Metabolomics in plasma of MAD vs. Control

Supplemental Table 3. Metabolomics in cortex of MAD vs. Control

Supplemental Table 4. Metabolomics in hippocampus of MAD vs. Control

Supplemental Table 5. LC-MS/MS Proteomics in cortex of MAD vs. Control

Supplemental Table 6. LC-MS/MS Proteomics in hippocampus of MAD vs. Control

Supplemental Table 7. Proteins at p < 0.01 in cortex of MAD vs. Control (FDR > 5%)

Supplemental Table 8. Proteins at p < 0.01 in hippocampus of MAD vs. Control (FDR > 5%)

Supplemental Table 9. Proteomics IPA Pathways in cortex of MAD vs. Control at p < 0.01

Supplemental Table 10. Proteomics IPA Pathways in hippocampus of MAD vs. Control at p < 0.01

Supplemental Table 11. RNAseq in cortex of MAD vs. Control

Supplemental Table 12. RNAseq in hippocampus of MAD vs. Control

Supplemental Table 13. Transcripts at p < 0.01 in cortex of MAD vs. Control (adj. p > 0.05)

Supplemental Table 14. Transcripts at p < 0.01 in hippocampus of MAD vs. Control (adj. p > 0.05)

Supplemental Table 15. RNAseq IPA Pathways in cortex of MAD vs. Control at p < 0.01

Supplemental Table 16. RNAseq IPA Pathways in hippocampus of MAD vs. Control at p < 0.01

Key Points.

1. Short-term MAD was associated with 13 increased plasma metabolites, 11 decreased in hippocampus, and decreased cortical palmitic acid.

2. Ketone body 3-hydroxybutyric acid was increased in plasma and had a positive correlation to cortex levels.

3. Short-term MAD differences in plasma and resected epilepsy brain tissue had top associations with mitochondrial functions.

Data Availability

The data that support the findings of this study are available in the public repositories listed in the methods, supplemental files, and from the authors upon reasonable request.