The UK Biobank is a comprehensive prospective cohort that enrolled ≥500,000 participants aged 40‐69 from 22 centers over the United Kingdom between 2006 and 2010. Neuropsychological tests were performed to assess cognitive function assessment in a subset of participants at baseline as well as 3 subsequent visits via a touch screen. Approximately 170,000 individuals participated in the baseline visit (2006‐2010), and ~20,000, ~60,000, and ~8000 individuals participated in the first (2012‐2013), second (2014+), and third (2019+) follow-up visits, respectively. For each follow-up visit dataset, consistent diagnostic criteria were applied for both sarcopenia and cognitive function. In our analysis, we included 35,000 participants who underwent cognitive function tests at baseline and had at least one subsequent follow-up visit (see Figure S1 in Multimedia Appendix 1).

Sarcopenia was defined according to the EWGSOP2, which uses the detection of low grip strength and low muscle mass to confirm sarcopenia, with poor physical performance indicating severe sarcopenia [23]. ALM was evaluated by bioelectrical impedance analysis with a Tanita BC418MA body composition analyzer and expressed in kilograms. We used the cut points as recommended in the EWGSOP2 definition of <7 kg/m² in men and 5.5 kg/m² in women. Grip strength was measured with a hydraulic handheld dynamometer (Jamar J00105). Specifically, 3 measurements were recorded for the maximum strength of both hands, and the highest values were used for analysis, expressed in kilograms. The cut-points for low grip strength recommended by EWGSOP2 were <27 kg in men and <16 kg in women. Gait speed was acquired directly from the participants through the question “How would you characterize your usual walking speed?” (options: “slow” [<3 miles per hour], “steady/average” [3‐4 miles per hour], “fast” [>4 miles per hour], or “prefer not to answer” [regarded as missing data]). Participants who reported that they were unable to walk or walked at a slow pace had low physical performance.

Cognitive function was assessed through 2 neuropsychological tests, namely fluid intelligence and prospective memory, by using a touch screen during visits to the UK Biobank assessment center. The fluid intelligence test was designed as a 13-item problem-solving task aiming at logic and reasoning abilities. Prospective memory, which is a form of episodic memory, was used as an indicator to measure the participants’ capacity to remember future tasks. The participants were instructed to reproduce a figure on a touch screen following a single instruction to be recalled later in the session. The responses were judged as either “correct on the first attempt” or otherwise, indicating a lapse in prospective memory (eg, “instruction not remembered, either skipped or incorrect” or “correctly recalled on the second attempt”).

The selection of potential confounders was based on previous literature [12,13], which was assessed through a touch-screen questionnaire at baseline. These variables included: age (continuous), sex (male or female), education level (college or university; A levels (Advanced Level), AS level (Advanced Subsidiary Level) or equivalent; O level, GCSEs (General Certificate of Secondary Education) or equivalent; CSEs or equivalent; NVQ (National Vocational Qualification), HND (Higher National Diploma), HNC (Higher National Certificate) or equivalent; other professional qualifications), self-reported race (White, Asian, Black, Mixed or other), assessment center, Townsend Deprivation Index at recruitment (continuous), BMI (continuous), presence of long-standing illness (no or yes), overall health rating (excellent, good, fair or poor), smoking status (never, former, or current), frequency of alcohol intake (daily or almost daily; 3 or 4 times a week; once or twice a week; 1 to 3 times a month or special occasions only), sleep duration (<7 h/day, 7‐9 h/day, or >9 h/day), and TV viewing duration (continuous). Specifically, the presence of long-standing illness was assessed using the following question, “Do you have any long-standing illness, disability, or infirmity”. Overall health rating was collected by asking “In general how would you rate your overall health?” All categorical variables strictly adhered to UK Biobank classification standards, with no additional category merging or processing.

To address the repeated measurements of the 2 cognitive tests, we used mixed-effects linear and logistic regression models to evaluate the associations of sarcopenia and its indices with fluid intelligence scores and with the risk of prospective memory impairment, respectively. All the mixed-effects models featured a random individual-specific intercept with a fixed slope. In regression models, we adjusted for sociodemographic, socioeconomic factors, and the number of cognitive assessments, and then further added health-related factors, including BMI, long-standing illness, and health rating. Finally, lifestyle factors, including smoking, alcohol intake frequency, sleep duration, and duration of TV viewing were added. We used a missing indicator approach in the primary analysis and conducted a sensitivity analysis by excluding missing data to evaluate its impact.

Given the lack of genome-wide association studies (GWASs) specifically focused on sarcopenia, we used the publicly available summary data on sarcopenic indices and cognitive function from the UK Biobank [24,25] and the COGENT Consortium [26]. The GWAS summary data for ALM (n=450,243), handgrip strength (n=461,089), and gait speed (n=461,089) were derived from the largest public GWASs of individuals with European ancestry in the UK Biobank [24,25]. The summary genetic statistics for the associations of general cognitive function were extracted from a comprehensive GWAS conducted through the UK Biobank and COGENT Consortium [26]. The summary phenotype source and outcome descriptions are provided in Table S1 in Multimedia Appendix 1.

The single nucleotide polymorphism (SNPs) meeting the following criteria were selected as instrumental variables: (1) significantly associated with ALM, handgrip strength, or usual gait speed (P<5×10⁻⁸), (2) independent (linkage disequilibrium r²<.001 within 10,000 kb), (3) with appropriate effect allele frequencies (≥1%), and (4) not palindromic (adenine/thymine or cytosine/guanine). To correct for multiple comparisons (3 exposures), the Bonferroni method was used. Associations with P<.016 (0.05/3) were deemed statistically significant.

As shown in Figure 2, for an increase of each one unit in genetically predicted ALM, handgrip strength, and gait speed, the general cognitive function score was increased by 0.10 (95% CI 0.07-0.12; P<.001), 0.18 (95% CI 0.08-0.29; P<.001), and 0.78 (95% CI 0.53-1.02; P<.001), respectively. To assess the consistency of directional causation, the effect estimates of 3 different sensitivity methods, that is, weighted median, MR-Egger, and MR-PRESSO were examined and plotted, confirming the directional causation between sarcopenic indices to general cognitive function (see Figure 2 and Figure S2 in Multimedia Appendix 1). There was some evidence of heterogeneity in the SNP effects although the MR-Egger intercepts indicated limited evidence of directional pleiotropy (see Table S3 in Multimedia Appendix 1). In radialMR analyses, outliers were detected (see Figure S3 in Multimedia Appendix 1). MR results remained consistent, showing slightly smaller effects after removing outliers (see Table S4 in Multimedia Appendix 1). For all sarcopenic indices and cognitive function phenotypes, the type 1 error rate was robustly controlled below 0.05 and bias estimates were confined to a narrow range of −0.01 to 0.01 (see Table S5 in Multimedia Appendix 1). Validation through the MRlap method further confirmed that sample overlap did not substantially influence the causal inferences (see Table S6 in Multimedia Appendix 1).

Figure 3 illustrates the effects of sarcopenic indices on 11 potential mediators. Univariable IVW MR analysis revealed that genetically predicted ALM exhibited protective effects on MVPA (β=0.07, 95% CI 0.05-0.09; P<.001) and negative associations with omega-3 fatty acids (β=−0.08, 95% CI −0.12 to −0.05; P<.001), falls (β=−0.03, 95% CI −0.05 to -0.003; P=.03), and frailty (β=−0.05, 95% CI −0.07 to −0.03; P<.001). Genetically predicted handgrip strength showed a positive association with MVPA (β=0.10, 95% CI 0.02-0.19; P=.02) and negative associations with falls (β=−0.15, 95% CI −0.24 to −0.07; P<.001), and frailty (β=−0.22, 95% CI −0.30 to −0.14; P<.001), depression (β=−0.09, 95% CI −0.16 to −0.01; P=.02), and stroke (β=−0.41, 95% CI −0.69 to −0.14; P=.003). Genetically predicted gait speed was positively associated with MVPA (β=0.83, 95% CI 0.65-1.01; P<.001). It also demonstrated significant protective effects against sleep disorders, falls, frailty, anxiety, depression, metabolic syndrome, and type 2 diabetes (with βs ranging from −0.01 to −2.32, all P values <.05). Further mediation analysis revealed that the total effect of genetically predicted ALM on general cognitive function decreased from 0.10 (95% CI 0.07-0.12) to 0.09 (95% CI 0.06-0.11) after adjusting for MVPA in MVMR analysis (Table 3). Similarly, the total effect of genetically predicted handgrip strength on general cognitive function attenuated from 0.18 (95% CI 0.08-0.29) to 0.16 (95% CI 0.05-0.27) with adjustment in MVMR analysis. MVPA mediated 8.2% of the total direct effect of ALM on cognitive function, and 10.6% of the total direct effect of handgrip strength on cognitive function. No apparent mediation effect was observed through omega-3 fatty acids, vitamin D level, falls, frailty, sleep disorders, anxiety, depression, stroke, metabolic syndrome, or type 2 diabetes. The type 1 error rate due to sample overlap between sarcopenic indices and the mediators remained below 0.05 (with bias estimates under 0.01) for all phenotypes (see Table S5 in Multimedia Appendix 1).

We conducted bidirectional MR analysis to investigate the potential reverse causality between sarcopenia and cognitive function. Table S7 in Multimedia Appendix 1 shows evidence supporting a causal effect of genetically predicted cognitive function on all sarcopenic indices. The effect estimates demonstrated a general consistency across different sensitivity methods (see Figure S2 in Multimedia Appendix 1). Some heterogeneity in the SNP effects was observed, although the MR-Egger intercepts suggested no evidence of directional pleiotropy (see Table S7 in Multimedia Appendix 1). In radialMR analyses, outliers were detected (see Figure S4 in Multimedia Appendix 1). The MR results remained consistent both before and after outlier correction and are presented in Tables S7 and S8 in Multimedia Appendix 1.