A Biologically Informed Machine Learning Pipeline Uncovers Metabolic Features of Intestinal Barrier Dysfunction

Intestinal
barrier dysfunction is a pivotal factor in
diverse pathologies,
yet clear diagnostic criteria remain elusive, and biomarker discovery
is often hindered by the high costs of large-scale clinical cohorts.
Here, we developed a biologically informed machine learning (ML) pipeline
that integrates function-guided network biology with ensemble feature
selection (EFS) to identify key metabolic features associated with
this condition. As an integrative framework, the pipeline combines
knowledge-based screening and computational prioritization with biological
validation, reconciling statistical rigor with biological interpretability.
We first constructed a heterogeneous data set from multiphenotypic
murine models and derived a continuous intestinal barrier index as
a predicted variable. This approach reframes the conventional classification
paradigm into a supervised regression task, enabling the discovery
of disease-relevant clues without clear diagnostic end points. By
leveraging network biology-informed EFS, we identified a functional
feature subset of 10 core metabolites. The subset demonstrated remarkable
stability across five distinct regression architectures, maintaining
consistent robust performance (R2: 0.604–0.654;
mean absolute error: 0.319–0.352). Furthermore, independent
clinical validation coupled with in vitro biotransformation assays
confirmed the applicability of these prioritized metabolic features,
highlighting serotonin, N-acetylputrescine, d-phenylalanine, and hippuric acid as functional metabolic features
of barrier dysfunction. In this work, ML was repurposed as a tool
for feature prioritization, robustness evaluation, and hypothesis
generation, reducing reliance on large cohorts. This integrated pipeline
facilitates the identification of intestinal barrier dysfunction associated
metabolic features and enables high-efficiency evaluation of multicomponent
therapeutic responses, offering new insights into the management of
complex multifactorial diseases.