Age-adjusted machine learning identifies facial skin microbes associated with skin quality among Korean women

Recognizing specific microbes that significantly influence skin quality is becoming an essential aspect of personalized skincare. However, conventional large-scale cohort skin microbiome studies often overlook important confounders, such as age, leading to missing meaningful microbe-skin relationships. In this study, we developed an age-adjusted machine learning (AAML) framework to identify microbial candidates associated with skin quality by determining optimal age ranges that enhance age-independent signals of skin microbes. It allowed the identification of distinct age groups that clearly explain specific skin microbial effects, as well as potential microbes showing notable age-independent links to skin quality, which were not observed in analyses across the entire age spectrum. In particular, Corynebacterium propinquum (C. propinquum) was recognized as a key species that positively impacts the middle-aged group, especially regarding skin tone. We further validated its dermatological significance using functional assays in human skin cell lines, performed a gene-level functional analysis, and suggested a potential mechanism. Our AAML method can be adapted to other microbiome analyses to precisely measure factors unaffected by age-related confounding factors.IMPORTANCEAge is a crucial but often intractable confounder in microbiome studies, obscuring how specific microbes affect human traits. We developed an age-adjusted machine learning (AAML) framework that automatically finds age ranges where the microbiome best predicts skin quality, rather than relying on arbitrary age groups. In a Korean facial skin cohort, AAML revealed three biologically meaningful age windows and uncovered microbial effects that are invisible in whole-age analyses. AAML identified Corynebacterium propinquum as a previously unrecognized commensal microbe that improves skin tone in the middle-aged group, and we mechanistically linked this effect to resveratrol production. Our framework provides a general, confounder-aware strategy for discovering age-independent microbiome-host relationships.