Microbiome-based disease prediction with multimodal variational information bottlenecks
Fig 2
Ablation study results. Comparisons between single-modality and bimodal MVIB.
The shown values are ROC AUC on test sets, error bars represents standard error computed by repeating each experiment five times on different random train/test splits. We leveraged the JMVIB objective (Eq 5) for optimisation. MVIB-A indicates the model performance only on species-relative abundances at test time. MVIB-M indicates the model performance only on strain-level markers at test time. MVIB-A+M indicates the model performance on both species-relative abundances and strain-level markers at test time.