A new AI framework called MOZAIC can help doctors more accurately match fecal transplant donors and recipients and increase treatment success by predicting how the gut microbiome will converge after treatment.
Research: Matching donor and recipient gut microbiota with artificial intelligence for optimized fecal microbiota transplantation. Image credit: Life Sciences/Shutterstock.com
recent cell reports In this study, we investigated whether donor and recipient gut microbiomes utilizing AI consistent with the MOZAIC framework can improve the clinical efficacy of fecal microbiota transplantation (FMT) by optimizing microbiome convergence after FMT and predicting patient outcomes.
Challenges and determinants of FMT effectiveness
Fecal microbiota transplantation (FMT) is an established treatment for relapse clostridioides difficile infection (CDI) and is also being evaluated for other gastrointestinal and metabolic disorders. FMT restores intestinal microbial diversity and metabolic function, effectively reverses dysbiosis, and supports intestinal homeostasis.
Despite the effectiveness of FMT, the results of FMT vary among recipients. Although most optimization has focused on donor selection, recipient-specific factors are increasingly recognized as key determinants of engraftment and treatment response. Differences in outcomes between recipients transplanted from the same donor highlight the importance of incorporating recipient resilience into FMT strategies.
Although donor and recipient microbial interactions critically determine FMT outcomes, current computational models lack the ability to fully capture the complex multidimensional microbial dynamics and interindividual response variability. Applications of machine learning (ML) have been attempted to predict recipient microbiome profiles and clinical responses after FMT, but model limitations have prevented comprehensive integration of the two-way donor-recipient interactions. An enhanced computational framework is needed to achieve accurate donor-recipient matching and improve the effectiveness of FMT.
Multidimensional FMT evaluation using MOZAIC
In the current study, we systematically analyzed 515 FMT events obtained from 30 diverse datasets (consisting of 24 public datasets and 6 in-house datasets, spanning 3 healthy volunteers and 12 disease indications). Of these, 94 metagenomes from 44 FMTs were newly collected in-house.
The researchers conducted extensive taxonomic profiling of bacterial, fungal, viral, and archaeal communities, as well as functional analysis of metabolic pathways and gene families.
We used an advanced bioinformatics pipeline to interpret metagenomic data to ensure a multidimensional view of the gut microbiome before and after FMT. This analysis took into account confounding variables, particularly adjusting for disease type, patient age, sex, and previous antibiotic treatment.
Given the heterogeneity and complexity of microbial alterations observed across different diseases and patient backgrounds, in this study we developed MOZAIC, an advanced deep learning framework specifically for FMT donor-recipient matching. Unlike traditional approaches that rely on simple ecological indicators or individual traits, MOZAIC processes a wide range of taxonomic and functional data from both donor and recipient.
Its architecture consists of five tightly interconnected neural computational blocks, each designed to extract and process compositional data such as microbial species and pathway abundances and functional gene family information in parallel. Downstream layers of the network then integrate these features to identify potential patterns of compatibility and complementarity that are unique to each donor-recipient pair.
The model incorporates advanced ML strategies such as regularization, dropout, and dynamic learning rate adjustment to ensure robust and generalizable predictions. Using this sophisticated design, MOZAIC can more accurately predict which donor-recipient pairs will achieve microbiome convergence after FMT. This is a result that is closely associated with clinical success and outperforms traditional machine learning models in predictive performance.
However, the authors point out that MOZAIC is still a relatively “black box” deep learning system, and its decision-making processes are not yet easily interpretable in terms of specific microbial taxa or pathways.
Microbiome convergence and predictive modeling shapes FMT outcomes
Recipients who improved clinically after FMT showed significant changes toward a donor-like profile in their microbiome, particularly in bacterial composition and metabolic function. However, non-responders showed minimal convergence and retained distinct microbiome signatures. Therefore, the success of FMT is strongly associated with the recipient’s microbiome becoming more similar to that of the donor, both taxonomically and functionally.
The greater the ecological distance between recipient and donor microbiomes, the greater the likelihood of convergence after FMT. This wide gap may increase the opportunity for donor-derived microorganisms to colonize.
Notably, donor microbiome diversity did not predict convergence success. Instead, recipients with lower baseline microbial diversity were more susceptible to colonization and remodeling by donor microorganisms, reflecting a more dysbiotic or less resilient gut environment. However, this association was attenuated after adjusting for disease type and other confounding variables. This effect was strongest in the CDI, ulcerative colitis, and irritable bowel syndrome cohorts.
These findings highlight the importance of recipient baseline ecology and donor-recipient complementarity in successful microbiome integration after FMT.
Traditional ML models based on standard ecological metrics achieved only moderate accuracy in predicting convergence after FMT. This indicates that these metrics do not fully capture complex donor-recipient dynamics and highly heterogeneous disease-specific microbial migration patterns. In contrast, MOZAIC consistently outperformed traditional models, with an average area under the curve (AUC) for predicting microbiome convergence of approximately 0.88, and precision and recall greater than 0.80.
In a retrospective analysis of an independent testing dataset, MOZAIC’s donor-recipient matching prediction achieved 78.7% accuracy in predicting clinical outcomes. Its robust performance persisted even when the definition of microbiome convergence changed, highlighting its adaptability.
Feature analysis showed that integrating both donor and recipient microbiome data is essential for optimal predictions, as models using only one source are much less effective. These findings highlight the need to consider multidimensional interactions between donor and recipient microbiomes to accurately predict FMT outcomes.
A retrospectively simulated clinical utility analysis showed that applying MOZAIC to donor-recipient matching could increase FMT success rate by 1.44 times compared to baseline. This improved efficacy persisted even after excluding cases with inherently high response rates, such as those with CDI. These findings highlight the potential of MOZAIC to significantly optimize clinical outcomes across a wide range of diseases and patient populations by systematically identifying the most compatible donor-recipient pairs.
AI-guided microbiome matching advances highly accurate FMT strategies
The present study demonstrated that the success of FMT is dependent on donor and recipient compatibility, as measured by AI analysis of microbiome characteristics. MOZAIC helps optimize donor selection and addresses important barriers in microbiome therapy. This study guides precision engineering of the gut ecosystem by linking microbiome convergence to clinical outcomes.
Next steps include testing MOZAIC in clinical trials and uncovering how its predictions work to better connect microbial ecology and personalized medicine. The authors also emphasized that the findings were based on a retrospective analysis and that prospective validation and improved interpretability of the AI framework is required before routine clinical implementation.
Click here to download your PDF copy.
