AUTHORS
Zeyang Li, PhD, Keimya Sadeghi, MS2, Tyler Funnell, PhD3, Natalie Grier Smith, BA4, Jennifer Angelica Funes, MS5, Amina Lazrak, MS6, Yuqing Chen, MS7, Nathan Duong, BS8, Yi-Tzu Lin, PhD9, Arnab Chowdhury, MS10, Ryotaro Nakamura, MD11, Joycelynne Palmer, PhD12, Robert Jenq, MD, PhD13, Antonio Luiz LC Cruz Gomes, PhD14, Jenny Paredes, PhD15, Marcel R.M. van den Brink Jr, M.D., Ph.D.16
AFFILIATIONS
1 City of Hope, Duarte, California, United States
2 City of Hope, Duarte, California, United States
3 City of Hope, Duarte, California, United States
4 City of Hope, Duarte, California, United States
5 City of Hope, Duarte, California, United States
6 City of Hope, Duarte, California, United States
7 City of Hope, Duarte, California, United States
8 City of Hope, Duarte, California, United States
9 City of Hope, Duarte, California, United States
10 City of Hope, Duarte, California, United States
11 City of Hope, Duarte, California, United States
12 City of Hope, Duarte, California, United States
13 City of Hope, Duarte, California, United States
14 City of Hope, Duarte, California, United States
15 City of Hope, Duarte, California, United States
16 City of Hope, Duarte, California, United States
ABSTRACT
Background
The intestinal microbiome regulates immune homeostasis and clinical outcomes following allogeneic hematopoietic cell transplantation (allo-HCT). We previously developed PARADIGM (PARameters Associated with DynamIcs of Gut Microbiota), a machine learning model integrating longitudinal drug exposures with microbiome sequencing to quantify drug-microbiome interactions as bacteria-drug response scores associated with allo-HCT outcomes (Nguyen et al., Cell 2023). We hypothesized that PARADIGM-derived bacteria-drug response scores could retain predictive power without microbiome data and identify microbiome-mediated therapeutics for graft-vs-host disease (GVHD).
Methods
We analyzed an independent cohort of 3,913 adult allo-HCT patients (2012 to 2025) at City of Hope (COH). In the absence of microbiome profiling, we integrated COH longitudinal drug exposures with the original Memorial Sloan Kettering (MSK) bacteria-drug response scores to infer COH patient-specific taxa response scores. We used multivariable Cox models to evaluate overall survival (OS) and transplant-related mortality (TRM). The therapeutic candidate ursodiol (UDCA) was evaluated using a preclinical major MHC-mismatched (C57BL/6J to BALB/c) model of GVHD by dietary supplementation at 13, 26, and 52 mg/kg starting day −7 relative to transplantation until the end of study.
Results
PARADIGM retained predictive power in our independent COH cohort. α-diversity patient response scores were associated with improved OS and TRM, whereas Enterococcus patient response scores were associated with worse OS and TRM, reproducing the original PARADIGM findings (adjusted P < 0.05). We further expanded the PARADIGM analysis from 4 to 36 taxa associated with OS and TRM, with most associations consistent with published literature. Correlation analysis of all taxa scores revealed a pathogen cluster and a commensal cluster without predefined groupings. Next, we screened bacteria-drug response profiles of 62 standard of care drugs from the original MSK PARADIGM analysis to identify candidates that suppress the pathogen cluster. UDCA emerged as a top-ranked therapeutic predicted to suppress the pathogen cluster and was selected for downstream clinical and preclinical validation.
We classified patients as UDCA-exposed (n=1,554) or UDCA-unexposed (n=2,293) using an exposure window from day −100 to day 30, to capture prophylactic UDCA administration before the cohort median onset of acute GVHD (day 30). Consistent with PARADIGM predictions, UDCA-exposed patients exhibited significantly lower pathogen cluster response scores than UDCA-unexposed patients (P < 0.0001). Multiple pathogen-associated taxa, including Enterococcus and Enterobacter, exhibited significantly higher predicted drug response scores among UDCA-unexposed patients (adjusted P < 0.0001). Peri-transplant UDCA exposure was independently associated with improved OS in the COH cohort (adjusted P < 0.0001). Notably, higher-than-median pathogen cluster scores predicted worse OS among UDCA-unexposed patients (adjusted P = 0.043), whereas no significant association was observed among UDCA-exposed patients.
UDCA was further evaluated in a murine GVHD model. Consistent with the clinical findings, dietary UDCA supplementation significantly improved post-transplant survival across treatment groups (overall log-rank P = 0.0006) and reduced intestinal Enterococcus dominance (P = 0.008). Mice treated with 52 mg/kg UDCA exhibited reduced Th1 cells (P = 0.0004) and increased regulatory T cells (P = 0.002) in the large intestinal lamina propria, together with decreased MHCII expression on large intestinal epithelial cells (P = 0.04). Furthermore, fecal microbiota transplantation from 26 mg/kg and 52 mg/kg UDCA-treated donors for 7 days recapitulated the survival and immunoregulatory benefits of dietary UDCA (overall log-rank P < 0.05), providing preliminary support for a microbiome-dependent mechanism of UDCA.
Conclusions
Together, these findings demonstrated the predictive power of PARADIGM-derived taxa–drug response scores in cohorts lacking microbiome profiling, expanded the analysis to 36 clinically relevant taxa, and identified UDCA as a microbiome-mediated therapeutic predicted to suppress a pathogen-associated microbial cluster. Clinical and preclinical studies validated this prediction, showing improved overall survival with peri-transplant UDCA prophylaxis and supporting a microbiome-dependent mechanism of GVHD protection.
DISCLOSURES
Zeyang Li, PhD: No relevant conflicts of interest to declare.