The study was based on the UK Biobank, a prospective cohort study that recruited ~ 500,000 participants from 2006 to 2010 at 22 centers across the UK [23]. At recruitment, participants aged 37–73 years completed a touchscreen questionnaire [24], which collected extensive information on socioeconomic status, lifestyle, health conditions, etc. Research nurse interviews and physical examinations were also administered at recruitment. The UK Biobank study was approved by the North West Multi-Center Research Ethics Committee, and all participants provided written informed consent. The present study was conducted under application number 55005 of the UK Biobank resource.

From 2009 to 2012, ~ 42% of the UK Biobank participants repeatedly completed 24-h dietary recalls [25]. Among the 210,966 participants who completed at least one 24-h diet recall, we excluded 18,958 participants reporting extremely high total energy intake (> 20 MJ, i.e., 4785 kcal). Among the remaining participants, 158,447 completed the International Physical Activity Questionnaire (IPAQ), and we further excluded 4,585 participants who had diabetes at baseline or who were lost to follow-up with an unknown date of loss. The formal analyses included 153,862 participants (Figure S1).

In the UK Biobank, SB intake was assessed using the Oxford WebQ, a validated web-based questionnaire for 24-h diet recalls [25,26,27]. Eligible participants were invited to complete the online questionnaire on five occasions between April 2009 and June 2012, and the average number of repetitions in this study was 1.98. The specific time windows for the five events were April 2009 to September 2010, February 2011 to April 2011, June 2011 to September 2011, October 2011 to December 2011, and April 2012 to June 2012. In the questionnaire, participants were asked how many units (Glass/Carton/250 ml) of each type of beverage they consumed in the past 24 h. The options were 0, 0.5, 1, 2, 3, 4, 5, and 6 + . In the current study, we defined SSB as carbonated (fizzy) drinks and commercial fruit drinks (data fields 100,170 and 100,180), ASB as low-calorie drinks (data field 100,160), and NJ as pure orange juices, pure grape juices, and other pure juices (data fields 100,190, 100,200, and 100,210). For participants who completed the WebQ more than once (~ 50%), we took the average intake levels of the beverages. Consumption of SSB, ASB, and NJ was categorized as 0, > 0–1, > 1–2, or > 2 units/d according to a previous study [28].

Physical activity was assessed using the IPAQ [29], which asked how many minutes the participants spent on each type of physical activity. We calculated the Metabolic Equivalent Task Scores by summing weighted minutes per week for all activities, including walking, and moderate and vigorous activity, with weights of 3.3, 4.0, and 8.0, respectively. According to the strategy of a previous study and following the recommendation of the World Health Organization, we categorized participants into low, moderate, and high physical activity groups [30]. The corresponding variables were provided as data category 54 in the UK Biobank showcase.

The outcome of interest was the incidence of T2D. We incorporated data from self-reported diagnoses and medication and Hospital Episode Statistics (HES) data (ICD-9 250 and ICD-10 E10-E14) to identify prevalent and incident T2D cases, as described elsewhere [31]. Data on the linkage to HES was updated until December, 2022.

We included multiple covariates for confounding adjustments. All covariates, including sociodemographic, lifestyle, and health-related factors, were measured at recruitment. Sociodemographic factors included age, sex, race (White or non-White), education level (high school and below or college and above), and area-based Townsend deprivation index (TDI). Lifestyle factors (other than physical activity and beverage intake) included alternate healthy eating index (calculated according to a previous study [28]), smoking status (current, former, or never smokers), and drinking status (current, former, or never drinkers). Health-related factors included objectively measured body mass index (BMI) categories (calculated as weight in kilograms divided by height in meters squared, categorized into normal weight, overweight, and obesity with cutoffs of 25 and 30 [32]), hypertension, cardiovascular diseases, cancer, and dyslipidemia. History of chronic health conditions was self-reported and identified using linkages to the HES system, similar to T2D. Family history of diabetes was defined as the existence of diabetes among parents or other first-degree relatives.

We chose the date of completion of the first valid 24-h dietary recall as the study baseline. We calculated person-time from the study baseline to the date of the first record of incident T2D, date of death, or the end of follow-up (December, 2022), whichever occurred first. Baseline characteristics of participants were described by their physical activity categories. Continuous variables were presented as mean (standard deviation, SD), and categorical variables as number (percentage). We first assessed the independent relationships of intake of SSB, ASB, NJ, and physical activity with incident T2D with Cox proportional hazard models, with sequential adjustments for baseline socioeconomic status, health-related behaviors, and health conditions. Model 1 was adjusted for age, squared age, sex, and race. Model 2 was based on Model 1 and further adjusted for total energy intake, alternate healthy eating index, TDI, education level, physical activity (not when physical activity was the exposure of interest), smoking status, drinking status, and family history of diabetes. Model 3 was based on Model 2 and additionally adjusted for BMI categories.

To examine whether an individual’s level of physical activity modifies the effects of SB intake on T2D, we stratified the data according to physical activity level and used the above-mentioned Model 3 to explore whether each SB intake has different effects on T2D in different physical activity level groups. To evaluate the interactions between physical activity and SSB, ASB, and NJ, we performed Wald tests on the multiplicative terms of physical activity and each SB and reported P-interactions.

To assess their joint associations with T2D, we combined physical activity with each level of SB intake, respectively, forming three 12-level variables: physical activity × SSB, physical activity × ASB, and physical activity × NJ. With the above-mentioned Model 3, we calculated the hazard ratios (HRs) with 95% confidence intervals (CIs) for the levels of the combined variables. The reference groups were participants with high physical activity and zero intake of the corresponding SB.

We conducted several sensitivity analyses to test the robustness of our primary findings. First, we additionally adjusted the models for hypertension, cardiovascular diseases, cancer, and dyslipidemia to assess whether the observed associations were confounded by these chronic health conditions. Second, we additionally adjusted the models for the indicator of the first dietary assessment (e.g., participants who completed the first dietary assessment from April 2009 to September 2010 coded as 1, February 2011 to April 2011 coded as 2, etc.) to better control for the time lag between covariate set and dietary assessment. Third, we further adjusted the relations for total sugar intake. Fourth, we excluded participants who developed T2D within the first five years to reduce reverse causality and address the possibility that T2D cases may have occurred during the exposure measurements. Fifth, we mutually adjusted for three types of SB in the same model to test their independent associations with T2D. Finally, we evaluated the additive rather than multiplicative interactions using relative excess risk due to interaction (RERI), which reflected biological interactions [33].

Missing values for all continuous covariates were imputed by means, and categorical variables were imputed by the most populated categories. We reported two-sided P-values throughout and a P-value lower than 0.05 was considered an indicator of statistical significance. Statistical analyses were performed using R 4.1.0 from March to July, 2022.