Sub-clinical acute rejection (subAR) in kidney transplant recipients (KTR) leads to chronic rejection and graft loss. Non-invasive biomarkers are needed to detect subAR. 307 KTR were enrolled into a multi-center observational study. Precise clinical phenotypes (CP) were used to define subAR. Differential gene expression (DGE) data from peripheral blood samples paired with surveillance biopsies were used to train a Random Forests (RF) model to develop a gene expression profile (GEP) for subAR. A separate cohort of paired samples was used to validate the GEP. Clinical endpoints and gene pathway mapping were used to assess clinical validity and biologic relevance. DGE data from 530 samples (130 subAR) collected from 250 KTR yielded a RF model: AUC 0.85; 0.84 after internal validation with bootstrap resampling. We selected a predicted probability threshold favoring specificity and NPV (87% and 88%) over sensitivity and PPV (64% and 61%, respectively). We tested the locked model/threshold on a separate cohort of 138 KTR undergoing surveillance biopsies at our institution (rejection 42; no rejection 96): NPV 78%; PPV 51%; AUC 0.66. Both the CP and GEP of subAR within the first 12 months following transplantation were independently associated with worse graft outcomes at 24 months, including de novo donor-specific antibody (DSA). Serial GEP tracked with response to treatment of subAR. DGE data from both cohorts mapped to gene pathways indicative of allograft rejection.