Presented by: Venkata Suhas Maranganti
View Abstract
Introduction
Understanding the relationship between microbiome dynamics and host outcomes is critical for defining predictive models and ultimately understanding the mechanisms through which the microbiome causes disease. Although off-the-shelf “black box” machine learning methods are widely deployed in the field, many methods do not consider biologically relevant structure in the data and are not interpretable.
Methods
We developed a new model, MDITRE: Microbial Differentiable and Interpretable Temporal Rule Engine that accelerates our earlier fully-Bayesian method, for learning human-interpretable rules to predict host outcomes from longitudinal microbial data. MDITRE uses continuous relaxations of discrete variables that capture relevant phylogenetic and temporal features, using novel domain-specific attention mechanisms, which enables highly efficient gradient-based optimization inference algorithms on GPUs.
Results
MDITRE achieves similar predictive performance as our original method on a suite of longitudinal microbiome datasets, while running 30X-70X faster. Moreover our model learns biologically meaningful relationships that our prior model did not.
Conclusion
We developed MDITRE, a highly scalable and accurate supervised machine learning model that learns human interpretable rules from longitudinal microbiome data. We demonstrate that our model can provide new insights into the complex and dynamic host-microbial ecosystems.
Venkata Suhas Maranganti – Poster Description
If you have any questions regarding the poster, feel free to reach out to Venkata Suhas Maranganti here.