Interpretable machine learning for low-sample multi-omics: a case study of ferret vaccine response.

Motivation: Machine Learning approaches continue to be critical to the modelling of complex biological pathways, but many of the most-used models suffer from being uninterpretable, providing little insight into underlying mechanisms. Interpretable machine learning (IML) offers a pathway to bridge predictive modeling and biological understanding. Results: Here, we apply an IML framework combining TreeFARMS and Rashomon Set analysis to multi-omics data from a controlled ferret vaccination study as a case study. The dataset comprised transcriptomic, proteomic, lipidomic, and metabolomic profiles collected across ten time points from animals receiving different Inovio vaccine formulations. Using TreeFARMS, we generated sparse, interpretable decision trees optimized for accuracy and compactness, and explored their Rashomon Sets to identify stable and alternative molecular rules predictive of vaccination status. The resulting models outperformed the ensemble methods while producing concise if-then rules that directly linked measurable molecular features, such as AZGP1 expression and ketoleucine levels, to immune response patterns. This proof-of-concept study demonstrates that interpretable ML can capture biologically meaningful signatures of vaccine response without sacrificing predictive accuracy, providing a transparent and reproducible alternative to black-box approaches in multi-omics analysis. Availability and implementation: The datasets generated and/or analysed during the current study are available at: https://github.com/nehlehk/iml-ferret-vaccine/tree/main/data  .