Microbiome-based prediction of allogeneic hematopoietic stem cell transplantation outcome.

Background: Allogeneic hematopoietic stem cell transplantation (HSCT) is potentially curative for hematologic malignancies but is frequently complicated by relapse and immune-mediated complications, such as graft-versus-host disease (GVHD). Emerging evidence suggests a role for the intestinal and oral microbiome in modulating HSCT outcomes, yet predictive models incorporating microbiome data remain limited.

Methods: We applied the RATIO (suRvival Analysis lefT barrIer lOss) model to longitudinal stool and saliva microbiome data from 204 adult HSCT recipients to predict the timing of seven outcomes: overall survival (OS), non-relapse mortality (NRM), relapse, acute GVHD (grades II-IV and III-IV), chronic GVHD, and oral chronic GVHD. A total of 514 stool and 1291 saliva samples were collected over 70 weeks post-HSCT. Model performance was evaluated using the concordance index (CI) and Spearman correlation coefficient (SCC), with SHAP (SHapley Additive exPlanations) analysis used for model interpretability.

Results: Oral and stool microbial dysbiosis peaked within the first 2 weeks post-HSCT, followed by partial recovery. Using the RATIO model, we found that microbiome features from early time points (weeks 1-2) were most predictive of short-term complications such as acute GVHD, while later samples (weeks 36-70) were more informative for long-term outcomes, including overall survival. RATIO outperformed traditional survival models (Cox and Random Survival Forest) across most outcomes (median CI > 0.65), with stool microbiota showing greater predictive power than saliva. SHAP analysis identified specific stool genera, including Collinsella and Eggerthella, associated with shorter time to various complications. External validation using a pediatric GVHD cohort confirmed the model's generalizability and reproducibility. External validation using a pediatric HSCT cohort (n = 90) confirmed the reproducibility and generalizability of these microbiome-based predictions.

Conclusions: Microbiome profiling of stool and saliva samples offers robust, time-sensitive prediction of post-HSCT complications. The RATIO model enables interpretable, time-to-event prediction across multiple outcomes and may inform microbiome-guided interventions to improve transplant success.