Study population

During 2006–2010, the UK Biobank enrolled over 500,000 participants aged 37–73 years from the general population [21]. At baseline, participants completed a touch-screen questionnaire, provided biological samples, and had physical measurements taken in 22 assessment centers throughout England, Wales, and Scotland. The UK Biobank study was approved by the North West Multi-Centre Research Ethics Committee. All participants provided written informed consent.

In this analysis, we included 331,699 UK Biobank participants who had complete data on the 18 biomarkers used in our BA algorithms (Table 1). We also excluded n = 177 with outlier BA values (considered as ±5 standard deviations [SD] from mean) and n = 23,366 with any cancer diagnosis (except non-melanoma skin cancer) before baseline, yielding an analytical sample of n = 308,156.

Table 1 Biomarkers included in the biological age algorithms.Biological age measures

We quantified BA based on three composite measures of blood chemistry and other clinical data: Klemera-Doubal method (KDM) [22], PhenoAge [23], and homeostatic dysregulation (HD) [24]. Details on the calculations and interpretations of the three measures have been summarized previously [25, 26]. Briefly, KDM is calculated from a series of regressions of biomarkers on CA and can be interpreted as the age at which the average physiology in the US National Health and Nutrition Examination Surveys (NHANES) III (i.e., the training sample) matches the physiology of the person. PhenoAge is calculated based on a mortality prediction score of biomarkers and CA and can be interpreted as the age at which the average mortality risk in NHANES III matches the predicted mortality risk. Different from KDM and PhenoAge, HD does not include CA in the calculation, but it is calculated based on the Mahalanobis distance [27] for a set of biomarkers relative to a reference sample and can be interpreted as the deviation of the person’s physiology from a healthy sample of NHANES III participants aged 20–30. All three measures have previously been validated for their abilities to predict diseases, disability, and mortality [25, 26].

In general, any age-related biomarkers can be used for constructing the BA algorithms. To facilitate comparison, we selected the same set of biomarkers in our KDM, PhenoAge, and HD algorithms, which were computed using the R package BioAge [25] in three steps:

1.

Training in NHANES III. We first identified 19 potential biomarkers covering a range of organ systems (e.g., cardiometabolic, inflammatory, kidney, lung functions) that are routinely collected in clinical practice and were available in NHANES III and UK Biobank (Supplementary Fig. 1). Only those with ≤20% missing data and correlated with CA ( |r | > 0.1, in accordance with prior work [19, 28]) in NHANES III were considered. Pulse was excluded due to its high correlation with systolic blood pressure (r = 0.84). Therefore, our new KDM, PhenoAge, and HD algorithms included 18 biomarkers (Table 1): FEV1, systolic blood pressure, blood urea nitrogen, glycated hemoglobin (HbA1c), total cholesterol, creatinine, serum glucose, waist circumference, red cell distribution width, albumin, alkaline phosphatase, triglyceride, mean cell volume, uric acid, lymphocyte percent, red blood cell count, C-reactive protein, and diastolic blood pressure. Following previous work [25], we selected non-pregnant participants who aged 30–75 years and had complete biomarker data as the reference population for KDM (n = 7694). The reference population for PhenoAge included those aged 20–84 years and with complete biomarker data (n = 12,998). The reference population for HD included participants aged 20–30 years who were not obese and whose biomarker values were within the age- and sex-specific normal range (n = 258). Only one measurement occasion was available per person in the training set.

2.

Testing the new BA algorithms in comparison to the published algorithms for their ability to predict mortality in an independent cohort of NHANES IV participants (n = 3849), who were recruited during 1999–2014 and followed up to 2015. Similar to the original version of KDM [28] and PhenoAge [23] (constructed using an alternative list of biomarkers as shown in Table 1), our new BA algorithms were statistically significantly associated with mortality during a median follow-up time of 7.4 years (interquartile range 4.1–11.5) (Supplementary Table 1). Besides, the new KDM and PhenoAge were strongly correlated with CA (r > 0.9), and all BA measures were moderately correlated with each other in NHANES IV (Supplementary Figs. 2 and 3).

3.

Projecting the newly trained algorithms onto UK Biobank. Correlations among the 18 biomarkers are shown in Supplementary Table 2, and the distributions of the BA measures in UK Biobank are presented in Supplementary Fig. 4.

To calculate the deviation between BA and CA, we regressed out CA (as 3 degrees-of-freedom natural spline) from KDM and PhenoAge in each cohort and considered them as “age residuals” (also known as “age acceleration”) [29]. Residuals were not calculated for HD as it was not an age measure by definition and it already quantifies deviation from a reference population [25, 26], but it was log-transformed due to the skewed distribution. Higher values of KDM residual, PhenoAge residual, and HD represent advanced BA. The KDM residual, PhenoAge residual, and HD were then standardized with mean = 0 and SD = 1 to allow comparison of effect sizes in subsequent analyses.

Cancer ascertainment

We studied five common cancers in Europe, including breast (for women), prostate (for men), and for both sexes, lung, colorectal, and melanoma skin cancer [30]. Incident cancers were ascertained from the cancer registries in England, Wales, and Scotland, where complete follow-up was available through February 29, 2020. We defined cancer diagnoses using the International Classification of Diseases, 10th revision (ICD-10) codes: any cancer (C00-97, excluding non-melanoma skin cancer C44), breast (C50), prostate (C61), lung (including trachea, C33-34), colorectal (C18-20), and melanoma (C43).

Statistical analyses

Participants were followed from the date of baseline assessment to the date of cancer diagnosis, death, or end of follow-up, whichever occurred first. Hazard ratios (HRs) for cancer risks per 1-SD increase in each BA measure were estimated using multivariable Cox proportional-hazards models, where attained age was used as the underlying timescale. The models were first adjusted for birth year and sex, and were further adjusted for baseline assessment center, ethnic background, body mass index (BMI), smoking, alcohol, physical activity, education, and Townsend deprivation index [31] in analyses of all cancer sites. Additionally, we adjusted for cancer-specific covariates such as family history of cancers, women-specific factors (menopausal status, hormone replacement therapy use, oral contraceptive use, parity), cancer screening, diet, and sun exposure variables, as relevant for each cancer site based on the literature. Covariates used in each model are listed in the footnote of the corresponding tables, and their definitions and descriptive statistics are provided in Supplementary Table 3. Missing data on covariates were coded as indicator variables in the models. To assess the associations of individual clinical biomarkers—in comparison to composite BA measures—with cancer risks, we also calculated the HRs per 1-SD increase in each clinical biomarker from the fully-adjusted models.

We performed subgroup analyses to test whether the associations may differ by age at baseline (<60 vs. ≥60 years), sex (women vs. men), and ethnicity (white vs. non-white). For breast cancer and lung cancer, we additionally stratified the analyses by menopausal status (premenopausal vs. postmenopausal) and smoking (never-smokers vs. ever-smokers), respectively. The proportional-hazards assumption was formally tested using Schoenfeld residuals. When the proportional-hazards assumption did not hold in the exposure of interest (P < 0.05), we fitted a time-varying model by including interaction terms between the BA measure and age (split into 5-year intervals) to calculate HRs over different periods of follow-up. Besides, to examine potential non-linear relationships between BA measures and cancer risks, we compared model fit of a restricted cubic spline model with a linear model using likelihood ratio tests and plotted the models with evidence of non-linearity (P < 0.05).

Several sensitivity analyses were performed. First, we assessed the association between the original KDM [28] and PhenoAge [23] algorithms (Table 1) and cancer risks to analyze if the biomarker composition would affect the results. Second, as we found an unexpected protective effect of BA measures for prostate cancer and that HbA1c and serum glucose may be associated with reduced prostate cancer risk [32, 33], we repeated the analysis using modified versions of KDM, PhenoAge, and HD computed from 16 biomarkers (i.e., removing HbA1c and serum glucose) to further examine whether the observed inverse relationship may be influenced by these two biomarkers. Third, instead of using indicator variables for missing values, we performed a complete-case analysis using available data (i.e., excluding individuals with missing data on any covariates). Finally, as individuals with a cancer diagnosis during the first 2 years of follow-up might have undiagnosed or subclinical disease at baseline, we performed a sensitivity analysis by excluding the first 2 years of follow-up to minimize reverse causation.

All analyses were performed in R 4.1.3 and Stata 16. To account for multiple testing (3 BA measures × 5 cancers), we applied the Bonferroni correction and considered a two-sided P < 0.0033 (i.e., 0.05/15) as statistically significant.