Exploratory bidirectional Mendelian randomisation analysis of cerebrospinal fluid metabolites associated with cerebral palsy and epilepsy

Abstract

Background

Cerebral palsy (CP) and epilepsy are prevalent neurodevelopmental disorders with shared aetiological complexities. This study employs bidirectional Mendelian randomisation (MR) to investigate associations consistent with potential causal effects between cerebrospinal fluid (CSF) metabolites and the pathogenesis of CP and epilepsy.

Methods

We conducted a two-sample MR analysis using genome-wide association study (GWAS) data from FinnGen, including 482 CP cases, 11 740 epilepsy cases and European ancestry controls. Instrumental variables for 338 CSF metabolites were selected using a relaxed threshold (p<1×10⁻⁵) due to the modest size of the metabolite GWAS. Nominal statistical significance was defined as p<0.05; to account for multiple comparisons, Benjamini-Hochberg false discovery rate (FDR) correction (Q<0.05) was applied across all five MR methods for each outcome.

Results

Six CSF metabolites showed nominal significant associations with CP risk in inverse variance weighted analyses, after FDR correction, three remained significant: 3-hydroxypyridine sulfate (OR=1.23, Q=0.040), isoleucine (OR=0.49, Q=0.046) and sphingomyelin (OR=1.12, Q=0.040). 17 metabolites were linked to epilepsy, of these, five remained significant after FDR correction, including 2,3-dihydroxy-5-methylthio-4-pentenoate (OR=0.81, Q=0.009) and 3-ureidopropionate (OR=1.09, Q=0.024). γ-Glutamylvaline showed nominally significant associations with both disorders but in opposite directions (CP: OR=1.25, p=0.040; epilepsy: OR=0.95, p=0.031); neither survived FDR correction. Reverse MR analyses did not support reverse associations for CP (all p>0.05), although genetic liability to epilepsy was associated with two CSF metabolites. Sensitivity analyses did not detect evidence of directional pleiotropy.

Conclusions

These findings provide hypothesis-generating evidence regarding metabolic pathways potentially involved in CP and epilepsy and may inform future experimental and clinical investigations. Independent validation and functional studies are required before clinical application can be considered.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Cerebral palsy (CP) and epilepsy are major neurodevelopmental disorders, and metabolic dysregulation may contribute to their pathogenesis. However, the role of cerebrospinal fluid metabolites remains poorly understood.

WHAT THIS STUDY ADDS

• Using bidirectional Mendelian randomisation, we identified several cerebrospinal fluid metabolites associated with CP and epilepsy risk, with three CP-related and five epilepsy-related associations remaining significant after false discovery rate correction. These findings provide genetically informed evidence that specific metabolic pathways may be associated with susceptibility to these neurological disorders.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• For clinicians, these findings do not currently support changes in clinical management or the use of the identified metabolites as diagnostic or therapeutic biomarkers. However, they highlight potentially relevant metabolic pathways that may inform future studies of risk stratification and therapeutic development, pending validation in independent and clinically relevant populations.

Introduction

Cerebral palsy (CP) and epilepsy represent two predominant neurodevelopmental disorders with profound lifelong impacts, yet their etiological complexities continue to challenge clinical management and mechanistic understanding. CP, affecting approximately 3.4 per 1000 live births in low- and middle-income countries,¹ manifests as non-progressive motor and postural impairments originating from prenatal or perinatal brain insults,² frequently accompanied by multisystem comorbidities including epilepsy (35%), sensory deficits (15%), cognitive impairment (49%) and chronic pain (75%).³ This clinical heterogeneity reflects diverse pathogenic pathways involving hypoxia-ischaemia (40% of cases), intrauterine infection (12%) and genetic predisposition (14%), though 30% of cases remain etiologically undefined.^4–6 Parallelly, epilepsy—occurring in 1/1200 infants annually—shares overlapping risk factors such as prematurity and cortical dysgenesis, with 35%–45% of CP patients developing comorbid epilepsy.⁷ This epidemiological convergence suggests shared neurobiological substrates, potentially mediated through oxidative stress cascades or synaptic plasticity dysregulation, though traditional observational studies have been limited in disentangling causation from confounding.

The advent of multiomics integration now offers unprecedented opportunities to decode these neurodevelopmental enigmas. Cerebrospinal fluid (CSF), the dynamic interface between cerebral metabolism and systemic circulation, encapsulates real-time pathophysiological signatures through its metabolomic profile—downstream products of genetic, enzymatic and environmental interactions.^8–10 Mendelian randomisation (MR) emerges as a valuable epidemiological approach that uses genetic variants as instrumental variables (IVs) to provide evidence consistent with potential causal associations while circumventing reverse causation biases inherent to conventional designs.^11–13 By synthesising CSF metabolomics with genome-wide association data, this approach enables systematic identification of metabolite–disease associations potentially consistent with causal effects across neurodevelopmental trajectories, convergent metabolic pathways bridging CP and epilepsy pathogenesis, and biologically relevant pathways and candidate targets for future investigation anchored in blood–brain barrier permeability.

Our study pioneers the application of bidirectional MR to elucidate CSF metabolomic determinants of CP and epilepsy, addressing two critical knowledge gaps: First, limited evidence regarding the potential role of metabolic dysregulation to neurodevelopmental sequelae; Second, the paucity of biomarkers potentially associated with disease susceptibility. Through this approach, we aim to establish a metabolite-centric paradigm for understanding shared etiopathogenesis while informing precision diagnostics and neuroprotective interventions.

Materials and methods

Study design

We conducted a two-sample MR study to estimate associations consistent with potential causal effects between CSF metabolites and CP and epilepsy. MR uses genetic variations as proxies for risk factors; hence, effective IVs must satisfy three critical criteria for causal inference: (1) genetic variation is directly associated with the exposure; (2) genetic variation is unaffected by potential confounders between the exposure and the outcome and (3) genetic variation does not influence the outcome through mechanisms other than the exposure.¹⁴ We employed various MR approaches in the research design to mitigate bias and yield precise estimates of the pertinent modifiable exposures and their relationships with the outcomes.

Data sources

We examined 338 CSF metabolites as the exposure factor, drawing from a metabolome-wide association analysis conducted by Panyard et al.¹⁵ This study retained 291 independent samples of persons of European descent following quality control and further data cleaning procedures. At the time of CSF extraction, all participants demonstrated normal cognitive function. Out of the 338 metabolites, 39 remain chemically unclear, whereas 299 have been chemically validated and are categorised into eight principal metabolic groups: amino acid, carbohydrate, cofactor and vitamin, energy, lipid, nucleotide, peptide and xenobiotic metabolism.

The CP genome-wide association study (GWAS) summary dataset was obtained from FinnGen (https://www.finngen.fi/endpoints/G6_CP) and includes 482 cases and 373 780 controls of European ancestry, encompassing 20 170 172 single nucleotide polymorphisms (SNPs). The epilepsy GWAS summary dataset was obtained from FinnGen (https://www.finngen.fi/endpoints/G6_STATUSEPI) and includes 11 740 cases and 287 837 controls of European ancestry, encompassing 20 168 098 SNPs.

No known sample overlap was identified between the CSF metabolite GWAS and the FinnGen outcome datasets. Therefore, bias arising from participant overlap is expected to be minimal.

It should be noted that the CSF metabolite GWAS was derived from individuals with normal cognitive function and was not specifically designed to represent fetal, neonatal or paediatric metabolic states.

Selection of IVs

SNPs are often employed genetic variants in MR analysis, signifying DNA sequence variation resulting from a single nucleotide change at the genomic level. In our study, we meticulously established a unique set of independent variables for each of the 338 CSF metabolites. This process was used to ensure that the genetic links relevant to each metabolite were accurately represented. Initially, recognising that genome-wide significance might be excessively stringent for CSF metabolites, we derived IVs from these metabolites employing a more lenient criteria (p<1×10⁻⁵).¹⁶ Subsequently, linkage disequilibrium (LD) criteria were set at r² <0.001 with a clumping distance of 10 000 kb to ensure independence of included SNPs. Then, we synchronised SNP-exposures and SNP-outcomes, removing ambiguous palindromic SNPs during harmonisation to avert distortion in strand orientation and allele coding throughout this procedure.¹⁷ Furthermore, to mitigate bias from weak IVs, we employed the F-statistic to evaluate the statistical strength of the association between each SNP and the exposures. IVs with an F-statistic greater than 10 were classified as strong instruments, while those with an F-statistic less than 10 were regarded as having a poor association between SNPs and the exposures. The F-statistic was computed using the formula F=R²/(1−R²) * (N−K−1)/K, where N denotes the sample size of the exposure GWAS, K signifies the number of SNPs, and R² indicates the proportion of variance elucidated by the SNPs in the exposure database.¹⁸

We removed SNPs having an F-statistic value below 10, as an F-statistic value beyond 10 signifies adequate strength to validate the SNPs.¹⁹ To minimise potential confounding in MR analyses, we used the LDtrait tool for LD pruning and confounder screening of IVs.^20–22 The workflow included: (1) Confounder screening: SNPs significantly associated with known confounders (eg, body mass index, smoking) were removed using GWAS summary statistics from the IEU OpenGWAS database (p<5×10⁻⁸). (2) Sensitivity validation: Horizontal pleiotropy was assessed via MR-Egger intercept and MR Pleiotropy Residual Sum and Outlier (MR-PRESSO) tests. Ultimately, in the leave-one-out strategy, we removed SNPs with significant outliers, leading to the identification of IVs used for analysis.

Statistical analysis

In this study, we used the inverse variance weighted (IVW), MR-Egger, weighted median, weighted mode and simple mode methods to estimate causal effect estimates. Given the IVW method’s superior test efficacy compared with other MR methods, we selected it as the primary method. The IVW method is a highly successful approach for estimating causal effects, particularly in contexts including many genetic variants as IVs. It employs the correlations among genetic variations, exposure and outcome to estimate causal effects, achievable by calculating a weighted average of the ratio estimates for each genetic variant.²³ Nominal statistical significance was defined as p<0.05. To account for multiple comparisons across all five MR methods and all metabolites for each outcome, Benjamini-Hochberg false discovery rate (FDR) correction was applied globally to all p values from the five MR methods (MR Egger, weighted median, IVW, simple mode, weighted mode). Associations with FDR-adjusted p values (Q<0.05) were considered statistically significant after correction for multiple testing. Results with p<0.05 but Q≥0.05 are reported as nominally significant and interpreted cautiously. An OR value exceeding 1 indicates a positive correlation between the exposure and the result, implying that an increase in the exposure factor elevates the probability of the outcome, hence suggesting that the exposure may serve as a risk factor for the outcome. An OR value below 1 signifies a negative correlation between the exposure and the outcome, indicating that an increase in the exposure factor results in a reduction in the risk of the outcome, hence implying that the exposure may serve as a protective factor for the outcome.

The leave-one-out strategy was used to ascertain any IVs that may affect the estimation of causal effects. The Egger intercept was used to evaluate horizontal pleiotropy, and the MR-PRESSO test was conducted to identify and eliminate SNPs potentially affected by pleiotropy. Heterogeneity assessments were performed using the Cochran Q test to analyse the variability of the SNPs. All analyses were conducted using the TwoSampleMR R tools within R software V.4.4.1.

To further validate the assumed causal direction, we performed the MR Steiger directionality test using the directionality_test function in the TwoSampleMR R package. This test compares the proportion of variance explained by the IVs for the exposure versus the outcome. A greater proportion of variance explained in the exposure than in the outcome was considered directionally consistent with the hypothesised exposure-to-outcome direction. However, the Steiger test results were interpreted cautiously, particularly for binary outcomes such as CP and epilepsy, and were not considered definitive evidence of causal direction when the corresponding statistical test was not significant.

Due to file-size limitations, representative plots are presented in the manuscript and online supplemental figures, while complete visual outputs for all analysed metabolites are provided as online supplemental zip files.

Patient and public involvement

No patients or members of the public were involved in this study. This study is a two-sample MR analysis focusing on CSF metabolites and CP/epilepsy, using publicly available GWAS data (from FinnGen) and metabolomic data (from previously published studies). Since no direct patients were enrolled, and the analysis relied on secondary data, patient involvement was not feasible.

Results

Causal effects of CSF metabolites on CP

Our research results indicate that 6 types of CSF metabolites showed nominally significant associations with CP risk. All retained SNPs had F-statistics >10, with minimum F-statistics ranging from 19.30 to 19.44, median F-statistics ranging from 20.80 to 22.88, and mean F-statistics ranging from 21.52 to 24.90 across the six metabolites (online supplemental table 1). Among the 6 CSF metabolites associated with CP, there are five identified metabolites and one with unknown chemical properties. Six CSF metabolites demonstrated associations with CP risk in IVW analyses. All results are depicted in figure 1. The reverse MR analysis of the 6 CSF metabolites and CP reveals that all p values exceed 0.05, reverse MR analyses did not provide evidence that CP causally altered the evaluated CSF metabolites.

Figure 1. Forest plot of Mendelian randomisation estimates between CSF metabolites and CP. CP, cerebral palsy; CSF, cerebrospinal fluid.

The IVW exhibited statistical significance in the end outcomes (p<0.05). 3-hydroxypyridine sulfate, Sphingomyelin (d18:1/20:0, d16:1/22:0), and Gamma-glutamylvaline demonstrated similar orientations and magnitudes throughout the five methodologies: IVW, MR-Egger, weighted median, weighted mode and simple mode, producing predominantly consistent results (online supplemental table 2). Scatter plots for six identified CSF metabolites across various tests are displayed in figure 2.

Figure 2. Scatterplot of genetic association between CSF metabolites and CP, with lines for different MR methods showing the direction and magnitude of the estimated association. (A–F) Associations consistent with potential causal effect of the six CSF metabolites on CP. CP, cerebral palsy; CSF, cerebrospinal fluid; IVW, inverse variance weighted; MR, Mendelian randomisation; SNP, single-nucleotide polymorphism.

After Benjamini-Hochberg FDR correction across all five MR methods, three of these associations remained significant (Q<0.05): 3-hydroxypyridine sulfate (Q=0.040), isoleucine (Q=0.046) and sphingomyelin (Q=0.040). The remaining three metabolites (X-24452, gamma-glutamylvaline, theobromine) did not survive multiple-testing correction and should be interpreted as hypothesis-generating findings requiring replication.

MR-Egger intercept and MR-PRESSO analyses did not detect evidence of directional pleiotropy (table 1), Cochran’s Q test did not detect substantial heterogeneity (table 2).

Table 1. The pleiotropy test of CSF metabolites and CP in this study.

Exposure	Outcome	Egger-intercept	SE	P value	MR-PRESSO (p value)
3-hydroxypyridine sulfate	CP	−0.0369	0.0428	0.3989	0.497
Isoleucine	CP	0.0204	0.0475	0.6699	0.664
X-24452	CP	0.0819	0.0398	0.0505	0.16
Sphingomyelin(d18:1/20:0, d16:1/22:0)	CP	0.0466	0.0269	0.0845	0.441
Gamma-glutamylvaline	CP	0.0368	0.0359	0.3089	0.61
Theobromine	CP	−0.0096	0.0337	0.7763	0.174

CP, cerebral palsy; CSF, cerebrospinal fluid; MR-PRESSO, Mendelian Randomisation Pleiotropy Residual Sum and Outlier.

Table 2. The heterogeneity test of CSF metabolites and CP in this study.

Exposure	Outcome	Method	Q	Q_df	Q_p value
3-hydroxypyridine sulfate	CP	MR Egger	20.1022	20	0.4516
3-hydroxypyridine sulfate	CP	IVW	20.8489	21	0.4682
Isoleucine	CP	MR Egger	28.1599	31	0.6129
Isoleucine	CP	IVW	28.3451	32	0.6522
X-24452	CP	MR Egger	28.2567	25	0.2962
X-24452	CP	IVW	33.0291	26	0.1613
Sphingomyelin(d18:1/20:0, d16:1/22:0)	CP	MR Egger	219.6737	218	0.4555
Sphingomyelin(d18:1/20:0, d16:1/22:0)	CP	IVW	222.7005	219	0.4178
Gamma-glutamylvaline	CP	MR Egger	60.1857	63	0.5773
Gamma-glutamylvaline	CP	IVW	61.2382	64	0.5748
Theobromine	CP	MR Egger	68.1797	58	0.1695
Theobromine	CP	IVW	68.2755	59	0.1912

CP, cerebral palsy; CSF, cerebrospinal fluid; IVW, inverse variance weighted; MR, Mendelian randomisation; Q, Cochran Q test.

In the experiment, confounding SNPs that influence the results and SNPs that are substantial outliers in the leave-one-out analysis were excluded and reanalysed. The leave-one-out study validated that the exclusion of any one SNP did not induce bias in the MR estimation (online supplemental figure 1). The funnel plots are displayed in online supplemental figure 2.

Causal effects of CSF metabolites on epilepsy

Our research findings suggest that 17 categories of CSF metabolites showed nominally significant associations with epilepsy risk. For the epilepsy analysis, the number of SNPs retained per metabolite, the variance explained (R² and the minimum, median, and maximum F-statistics across SNPs for each metabolite are reported in (online supplemental table 3). Across all metabolites nominally associated with epilepsy, the minimum F-statistics ranged from 19.35 to 19.71, median F-statistics ranged from 20.16 to 22.20, and mean F-statistics ranged from 21.20 to 27.69—all substantially exceeding the conventional threshold of 10. There are 17 CSF metabolites linked to epilepsy, of which 16 have been characterised and 1 has unidentified chemical characteristics. The association is consistent with a potential causal effect between the risk of epilepsy and 17 CSF metabolites (figure 3). Reverse MR analyses suggested associations between epilepsy liability and two CSF metabolites. They are 3-methylhistidine and N-acetyl-3-methylhistidine. Their IVW results are as follows: 3-methylhistidine levels (OR 1.2736, 95% CI 1.0771 to 1.5060, p=0.0047), N-acetyl-3-methylhistidine levels (OR 1.1999, 95% CI 1.0641 to 1.3531, p=0.0029). Other results of reverse MR are detailed in the online supplemental table 4.

Figure 3. Forest plot of Mendelian randomisation estimates between CSF metabolites and epilepsy. CSF, cerebrospinal fluid.

The IVW exhibited statistical significance in the end outcomes (p<0.05). 1-palmitoyl-2-stearoyl-gpc (16:0/18:0), 1-stearoyl-gpc (18:0), isovalerate (i5:0), tartarate, DMTPA, 3-methylhistidine, 3-ureidopropionate, gamma-glutamylleucine, gamma-glutamylvaline, methionine sulfone, MMA and X-24431 demonstrated similar orientations and magnitudes throughout the five methodologies: IVW, MR-Egger, weighted median, weighted mode and simple mode, producing predominantly consistent results (online supplemental table 5). Scatter plots for 17 identified CSF metabolites across various tests are displayed in figure 4.

Figure 4. Scatterplot of genetic association between CSF metabolites and epilepsy, with lines for different MR methods showing the causal relationship’s direction and strength. (A–F) Representative scatter plots illustrating associations consistent with potential causal effects of six selected CSF metabolites on epilepsy. Complete scatter plots for all 17 epilepsy-associated metabolites are provided in scatter plots for epilepsy.zip. CSF, cerebrospinal fluid; IVW, inverse variance weighted; MR, Mendelian randomisation; SNP, single-nucleotide polymorphism.

We performed Benjamini-Hochberg FDR correction across all P-values from five MR methods (MR Egger, weighted median, IVW, simple mode, weighted mode) for the association between CSF metabolites and epilepsy. After FDR correction, four IVW results (1-palmitoyl-2-stearoyl-gpc, DMTPA, 3-methylxanthine, 3-ureidopropionate) and one weighted median result (gamma-glutamylleucine) remained significant at FDR <0.05.

MR-Egger intercept and MR-PRESSO test did not detect evidence of directional pleiotropy (table 3). Cochran Q test no substantial heterogeneity was detected, except the IVW of hydroxy-cmpf and MR Egger of tartarate showing heterogeneity (table 4).

Table 3. The pleiotropy test of CSF metabolites and epilepsy in this study.

Exposure	Outcome	Egger-intercept	SE	P value	MR-PRESSO (p value)
1-palmitoyl-2-stearoyl-gpc (16:0/18:0)	epilepsy	−0.0014	0.0074	0.8498	0.771
1-stearoyl-gpc (18:0)	epilepsy	0.0023	0.0070	0.7388	0.647
Argininosuccinate	epilepsy	−0.0015	0.0087	0.8654	0.718
Hydroxy-cmpf	epilepsy	0.0156	0.0098	0.1261	0.056
Isovalerate (i5:0)	epilepsy	−0.0072	0.0078	0.3605	0.524
Tartarate	epilepsy	−0.0018	0.0137	0.8971	0.056
Uracil	epilepsy	0.0164	0.0136	0.2382	0.167
2,3-dihydroxy-5-methylthio-4-pentenoate (dmtpa)	epilepsy	−0.0056	0.0086	0.5231	0.653
3-methylhistidine	epilepsy	0.0137	0.0093	0.1512	0.671
3-methylxanthine	epilepsy	−0.0010	0.0106	0.9251	0.778
3-ureidopropionate	epilepsy	0.0130	0.0086	0.1433	0.358
Gamma-glutamylleucine	epilepsy	0.0070	0.0069	0.3174	0.895
Gamma-glutamylvaline	epilepsy	0.0055	0.0074	0.4567	0.941
Methionine sulfone	epilepsy	−0.0070	0.0067	0.3044	0.818
Methylmalonate (mma)	epilepsy	0.0023	0.0054	0.6750	0.225
N-acetyl-3-methylhistidine	epilepsy	−0.0024	0.0065	0.7078	0.728
X-24431	epilepsy	−0.0016	0.0070	0.8246	0.481
epilepsy	3-methylhistidine	−0.0065	0.0154	0.6727	0.532
epilepsy	N-acetyl-3-methylhistidine	0.0048	0.0111	0.6647	0.986

CSF, cerebrospinal fluid; MR-PRESSO, Mendelian Randomisation Pleiotropy Residual Sum and Outlier.

Table 4. The heterogeneity test of CSF metabolites and epilepsy in this study.

Exposure	Outcome	Method	Q	Q_df	Q_p value
1-palmitoyl-2-stearoyl-gpc (16:0/18:0)	epilepsy	MR Egger	39.4402	45	0.7057
1-palmitoyl-2-stearoyl-gpc (16:0/18:0)	epilepsy	IVW	39.4765	46	0.7405
1-stearoyl-gpc (18:0)	epilepsy	MR Egger	73.3085	77	0.5981
1-stearoyl-gpc (18:0)	epilepsy	IVW	73.4204	78	0.6257
Argininosuccinate	epilepsy	MR Egger	28.1736	32	0.6607
Argininosuccinate	epilepsy	IVW	28.2028	33	0.7049
Hydroxy-cmpf	epilepsy	MR Egger	35.9448	25	0.0724
Hydroxy-cmpf	epilepsy	IVW	39.5452	26	0.0432
Isovalerate (i5:0)	epilepsy	MR Egger	58.9279	60	0.5149
Isovalerate (i5:0)	epilepsy	IVW	59.7770	61	0.5203
Tartarate	epilepsy	MR Egger	29.2298	18	0.0456
Tartarate	epilepsy	IVW	29.2577	19	0.0620
Uracil	epilepsy	MR Egger	31.1044	25	0.1856
Uracil	epilepsy	IVW	32.9209	26	0.1644
2,3-dihydroxy-5-methylthio-4-pentenoate (dmtpa)	epilepsy	MR Egger	19.7281	22	0.6000
2,3-dihydroxy-5-methylthio-4-pentenoate (dmtpa)	epilepsy	IVW	20.1492	23	0.6329
3-methylhistidine	epilepsy	MR Egger	28.7553	34	0.7224
3-methylhistidine	epilepsy	IVW	30.9115	35	0.6659
3-methylxanthine	epilepsy	MR Egger	20.1615	25	0.7383
3-methylxanthine	epilepsy	IVW	20.1705	26	0.7834
3-ureidopropionate	epilepsy	MR Egger	23.5598	23	0.4285
3-ureidopropionate	epilepsy	IVW	25.9119	24	0.3577
Gamma-glutamylleucine	epilepsy	MR Egger	50.2626	64	0.8950
Gamma-glutamylleucine	epilepsy	IVW	51.2782	65	0.8927
Gamma-glutamylvaline	epilepsy	MR Egger	46.9649	63	0.9346
Gamma-glutamylvaline	epilepsy	IVW	47.5258	64	0.9387
Methionine sulfone	epilepsy	MR Egger	52.6511	63	0.8206
Methionine sulfone	epilepsy	IVW	53.7232	64	0.8166
Methylmalonate (mma)	epilepsy	MR Egger	337.8991	316	0.1899
Methylmalonate (mma)	epilepsy	IVW	338.0874	317	0.1988
N-acetyl-3-methylhistidine	epilepsy	MR Egger	115.7907	125	0.7105
N-acetyl-3-methylhistidine	epilepsy	IVW	115.9318	126	0.7290
X-24431	epilepsy	MR Egger	52.1005	51	0.4309
X-24431	epilepsy	IVW	52.1512	52	0.4680
epilepsy	3-methylhistidine	MR Egger	69.2092	69	0.4703
epilepsy	3-methylhistidine	IVW	69.3898	70	0.4981
epilepsy	N-acetyl-3-methylhistidine	MR Egger	46.5385	69	0.9827
epilepsy	N-acetyl-3-methylhistidine	IVW	46.7280	70	0.9854

CSF, cerebrospinal fluid; IVW, inverse variance weighted; MR, Mendelian randomisation; Q, Cochran Q test.

In the experiment, confounding SNPs that influence the results and SNPs that are substantial outliers in the leave-one-out analysis were excluded and reanalysed. The leave-one-out study validated that the exclusion of any one SNP did not induce bias in the MR estimation (online supplemental figure 3). The funnel plots are displayed in online supplemental figure 4.

For brevity, only six representative metabolites are displayed in the manuscript and online supplemental figures, whereas the complete scatter plots, leave-one-out analyses, and funnel plots for all 17 epilepsy-associated CSF metabolites are provided in the online supplemental zip files (Scatter plots for epilepsy.zip, Leave-one-out for epilepsy.zip, and Funnel plots for epilepsy.zip).

To further evaluate the assumed causal direction, we performed Steiger directionality tests for all exposure–outcome pairs. The proportion of variance explained by the genetic instruments was generally greater for CSF metabolites than for the corresponding disease outcomes, which was directionally consistent with the hypothesised exposure-to-outcome relationship. However, Steiger p values ranged from 0.12 to 0.82 (all >0.05), indicating that the statistical evidence for directionality was inconclusive. Therefore, the Steiger analysis did not provide statistically significant support for the proposed causal direction and could not definitively exclude the possibility of reverse causation (online supplemental table 6).

X-24452 and X-24431 remain chemically unidentified and should therefore be interpreted cautiously.

Discussion

This exploratory MR analysis identified several genetically predicted CSF metabolites with nominally significant associations for CP and epilepsy risk, of which a subset remained significant after FDR correction. These findings provide hypothesis-generating evidence that metabolic pathways related to amino acid metabolism, lipid metabolism, oxidative stress and xenobiotic processing may warrant further investigation in relation to CP and epilepsy susceptibility.

γ-Glutamylvaline showed nominally significant associations with both CP and epilepsy, although the directions differed (higher levels associated with increased CP risk but decreased epilepsy risk). Given its involvement in glutathione metabolism, this observation generates the hypothesis that redox-related pathways may differ between these disorders. However, neither association survived FDR correction (both Q ≥0.05); therefore, these findings are exploratory and require validation in larger datasets before further interpretation.

3-Hydroxypyridine sulfate was one of the metabolites associated with CP risk in our analysis. This metabolite has previously been recognised as a biomarker of dietary and environmental exposures, particularly those related to coffee and tea consumption.^24–26 Previous observational studies have reported associations between 3-hydroxypyridine sulfate and several health outcomes, although the direction and magnitude of these associations have varied across different disease contexts.^27–32 Therefore, the biological significance of the present finding remains uncertain.

Importantly, because 3-hydroxypyridine sulfate is influenced by dietary intake (eg, coffee, tea) and other environmental factors, genetically predicted levels may reflect broader metabolic handling of exogenous compounds rather than a direct neurodevelopmental mechanism. This metabolite passed FDR correction (Q=0.040); however, its strong dependence on dietary exposures limits its interpretability as a neurodevelopmental biomarker. Future studies incorporating maternal exposure data are needed.

Sphingomyelin is an important structural component of neuronal membranes and myelin sheaths and has been implicated in several neurological disorders. Previous studies have reported altered sphingomyelin profiles in neurodegenerative diseases, including Alzheimer’s disease, and have suggested potential associations between sphingolipid metabolism and cognitive decline.^33–35 These findings indicate that sphingomyelin homeostasis may play a broader role in neurological health and disease.

In the present study, genetically predicted CSF sphingomyelin levels were associated with an increased risk of CP. Reverse MR analyses did not provide evidence supporting a reverse association from CP to sphingomyelin levels. Although these findings suggest a potential role for sphingomyelin metabolism in CP, the biological significance of this association remains unclear. Given the importance of sphingolipids in neural development and cellular signalling, further experimental studies are warranted to investigate the underlying mechanisms. Moreover, replication in independent cohorts and targeted lipidomic analyses will be necessary to determine whether specific sphingomyelin species contribute to CP susceptibility or merely represent correlated metabolic markers.

Several limitations of this study should be acknowledged. First, the CSF metabolite GWAS included a relatively small sample size, which may limit instrument strength and increase susceptibility to weak-instrument bias despite the application of standard quality-control procedures and F-statistic filtering. Second, the CP GWAS dataset contained a limited number of cases, potentially reducing statistical power and increasing uncertainty around modest effect estimates. Therefore, replication in larger and independent datasets is warranted. Third, both CP and epilepsy are clinically heterogeneous disorders comprising multiple etiological and phenotypic subtypes. Subtype-specific analyses were not possible using the available summary-level GWAS data, which may limit biological interpretation of the observed associations. Future studies investigating specific CP and epilepsy subgroups may provide additional mechanistic insights. Fourth, the CSF metabolite GWAS was derived from individuals with normal cognitive function and was not specifically designed to represent fetal, neonatal or paediatric metabolic states. Consequently, genetically predicted metabolite levels may not fully capture the metabolic environment during critical periods of neurodevelopment, particularly for CP. The Steiger directionality analyses showed that the genetic instruments generally explained more variance in the CSF metabolites than in the corresponding binary disease outcomes, which was directionally consistent with the hypothesised exposure-to-outcome direction. However, all Steiger P values were >0.05, and therefore the directionality evidence was statistically inconclusive. These results should not be interpreted as definitive evidence supporting causal direction or as evidence that excludes reverse causation.

Finally, although MR is less susceptible to confounding and reverse causation than conventional observational studies, residual horizontal pleiotropy cannot be completely excluded. In addition, genetically predicted metabolite levels represent lifelong genetic predisposition rather than dynamic metabolite fluctuations influenced by environmental exposures, diet, medication use or other time-varying factors. Therefore, the translational relevance of these findings should be interpreted with appropriate caution.

In this exploratory bidirectional MR analysis, several genetically predicted CSF metabolites showed nominally significant associations with CP and epilepsy risk; a subset (three for CP, five for epilepsy) remained significant after FDR correction. γ-Glutamylvaline showed nominally significant associations with both disorders but in opposite directions; however, neither survived multiple-testing correction. These findings provide hypothesis-generating evidence regarding metabolic pathways that may warrant further investigation in CP and epilepsy. Independent validation in larger, paediatric-derived or perinatal-derived metabolomics datasets and functional studies are required before any clinical translation can be considered. The modest sample sizes of the exposure and outcome GWAS and the developmental timing mismatch are important limitations that temper the strength of our conclusions.

Despite some authors' affiliations with the funder, the funder had no influence on the study results and had no role in the study design; collection, analysis, or interpretation of data; writing of the manuscript; or the decision to submit the manuscript for publication.

Data availability statement

Data are available in a public, open access repository. All data relevant to the study are included in the article or uploaded as supplementary information.