Astronauts as a Model for Human Aging: How Space Travel May Change Biological Aging

Axiom Mission 2 (Ax‐2) was a 9‐day, 5‐h, 26‐min private astronaut flight to the International Space Station (ISS) operated by Axiom Space in partnership with NASA and SpaceX. The mission launched aboard SpaceX's Crew Dragon Freedom on a Falcon 9 from Launch Complex 39A, Kennedy Space Center, at 21:37 UTC on 21 May 2023, marking the second all‐commercial crewed visit to the ISS. Freedom docked to the Harmony zenith port on 22 May and remained attached for 8 days. Biospecimens were collected under standard ISS protocols during these studies form the basis of the analyses reported here. The spacecraft undocked on 30 May and splashed down in the Gulf of Mexico off Panama City, Florida, in the early hours of 31 May 2023, successfully concluding Axiom Space's second pathfinder mission.

We collected venous blood from the 4 astronauts of the Axiom Mission 2 before, during, and after a short‐duration mission onboard the ISS. Baseline (preflight) sampling occurred 45 days before launch (Launch (L) L‐45). In‐flight sampling occurred on (Flight Day (FD) +4 and +7 days). Postflight sampling occurred on (Return (R) R+1, R+7). Trained and certified astronauts performed in‐flight blood draws using standard phlebotomy techniques (venipuncture). We collected non‐fasted venous blood into plasma preparation tubes (PPTs) (BD Biosciences, cat # 362788). At R+1 only, we used K2 EDTA tubes in substitution of PPTs. The amount of blood drawn was 5 mL per astronaut at each timepoint. After collection, we centrifuged whole blood at room temperature at 1256 g for 25 min and then frozen at −80°C within 60 min of sampling. Then, we shipped the samples to Weill Cornell Medicine. The remaining samples now reside in the Cornell Aerospace Biobank (CAMbank) and all metadata is securely stored at the HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine (Weill Cornell Medicine, New York, United States).

We extracted high molecular weight (HMW) genomic DNA (gDNA) from the cell pellet below the gel barrier of the BD Vacutainer PPT tube (cat # 362788) following centrifugation, using the Monarch HMW DNA Extraction Kit for Cells & Blood (cat # T3050L). We sheared the extracted HMW gDNA using Covaris G‐Tubes (cat # 520079) down to a target size of 20 kb. We measured fragment length using Genomic ScreenTape (Agilent, cat # 5067 5365) on the Agilent 4200 TapeStation System (cat # G2991BA). We quantified DNA using the Qubit 1X dsDNA High Sensitivity Assay Kit (cat # Q33230) on the Qubit 4 Fluorometer (cat # Q33238). We measured the quantity and fragment length before and after shearing.

We assessed DNA methylation using the TruAge epigenetic age platform (TruDiagnostic Inc., Lexington, KY). We collected and lysed peripheral whole blood in a stabilization buffer. We extracted genomic DNA and bisulfite converted 500 ng using the EZ DNA Methylation Kit (Zymo Research) following the manufacturer's instructions. We hybridized the converted DNA to the Illumina Infinium MethylationEPIC v2.0 BeadChip (850 k) with the custom aging add‐on available from TruDiagnostic. We randomly assigned samples to BeadChip wells to reduce technical batch effects. We scanned arrays using the Illumina iScan SQ platform, generating IDAT files for each sample. We processed raw intensity data using the minfi R package (Aryee et al. 2014). We performed quality control using ENmix (Xu et al. 2021), identifying outliers based on internal control probe variance (> 3 SD from the mean); no outliers were detected. We applied single‐sample Noob (ssNoob) normalization to correct for background and dye bias. We calculated detection p values and removed all samples with a mean detection p value > 0.05. Additionally, we excluded CpGs failing detection in more than 5% of samples (p > 0.05). After QC, we collapsed probes unique to EPIC v2 arrays and imputed missing probes (relative to EPIC v1) using blood‐specific medians from the sesameData package (Zhou et al. 2025 ). The final normalized beta matrix was used for downstream analysis.

We calculated epigenetic age difference (EAD) as the difference between epigenetic age and chronological age. We quantified epigenetic age acceleration (EAA) as the residual deviation of each DNA methylation‐based age estimate from its expected value given chronological age and sex. For every clock—including principal component–based clocks (PCHorvath1, PCHorvath2, PCHannum, PCPhenoAge, and PCGrimAge) (Higgins‐Chen et al. 2022), first‐generation clocks (Horvath (Horvath 2013), Hannum (Hannum et al. 2013)), PhenoAge (Levine et al. 2018), OMICmAge (Chen et al. 2023), DNAmFitAge (McGreevy et al. 2023), causal clocks (AdaptAge, CausAge, and DamAge) (Ying et al. 2024), the intrinsic clock (IntrinClock) (Tomusiak et al. 2024), stochastic clocks (Stochastic.Zhang, Stochastic.Horvath, and Stochastic.PhenoAge) (Tong et al. 2024), retroelement‐based clocks (Retroclock and Retroclockv2) (Ndhlovu et al. 2024), and 12 organ‐system clocks (Blood, Brain, Inflammation, Heart, Hormone, Immune, Kidney, Liver, Metabolic, Lung, MusculoSkeletal, and SystemsAge) (Sehgal et al. 2025). We fitted a separate linear regression model of the form:EpigeneticAge=β₀+β₁·ChronologicalAge_i+β₂·Sex_i+ε_iwhere Sex was encoded as Female = 1 and Male = 0. The residual ε_i, extracted via R's resid() function, represents the EAA for sample i, (i.e., the residual epigenetic age after adjusting for age and sex). We predicted cell composition using a DNA methylation reference matrix (Luo et al. 2023), which deconvolutes 12 immune cell subtypes based on Illumina 850 k DNAm profiles of cell‐sorted samples from Salas et al. (2022). We calculated intrinsic epigenetic age acceleration (IEAA) by fitting a linear model of epigenetic age on chronological age, sex, and the predicted fraction of the 12 cell types, and extracting the residual (ε_i).

Given the limited number of samples, we evaluated the changes in biological age (EAD, EAA, and IEAA) for each crew member across timepoints. We treated the set of 31 epigenetic clocks expressed in units of years as repeated measures. We employed two statistical approaches to compare timepoints within each subject: (1) We initially assessed differences using a two‐sided paired Wilcoxon signed‐rank test. This nonparametric test compares the paired distributions of biological estimates across clocks to determine if the median difference between timepoints differs from zero. (2) To account for the high correlation between epigenetic clocks, which violates the independence assumption of standard tests, we implemented a permutation‐based test on the mean differences (i.e., p_perm). For each astronaut and timepoint, we calculated the vector of paired differences for all clocks. We then constructed a null distribution by randomly assigning a positive or negative sign to each difference and calculating the mean of these sign‐flipped values in 10,000 permutations. Finally, we calculated the two‐sided p values as the proportion of permuted means with an absolute value greater or equal to the absolute value of the observed mean difference.

In addition, we calculated overall statistically significant changes between timepoints across astronauts and epigenetic clocks using linear mixed models (i.e., p_LMM). We included timepoint as a fixed effect while considering random intercepts for both the crew member and the epigenetic clock to account for repeated measures and between‐clock variability.EpigeneticAge=β₀+β₁·Timepoint_k+b_0i+c_0j+ε_ijkwhere b_0i represents the random intercept for Astronaut i (accounts for 1|crew) and c_0j represents the random intercept for Clock j (accounts for 1|clock). We calculated this using the function lmer from the lme4 R package.

Considering that the use of different epigenetic clocks as repeated measures might be sensitive to the influence of specific clocks with large variation, we performed a sensitivity analysis where we removed one clock and a time and repeated the calculations for mean, Wilcoxon test and permutation‐based test. Overall, when p values were below 0.05 in the test with all the clocks included, they remained below 0.05 when an individual clock was excluded (Figure S7). The only exception was the removal of AdaptAge in A1 for the age difference using a permutation‐based approach which resulted in a p = 5.04e‐2.