We demonstrate how NetworkVI identifies immune evasion mechanisms via GO programs in a CRISPR-perturbed melanoma dataset not detectable with standard methods. This highlights the medical utility of interpretable multimodal models.

NetworkVI also incorporates a GO-specific covariate attention mechanism, allowing the model to adjust for sample-level metadata (e.g., donor or condition), improving both performance and interpretability in heterogeneous datasets.

These scores reveal associations between genes and GO terms, helping to identify active biological processes and their regulatory drivers in a given condition or perturbation.
This opens a path toward mechanistic interpretation beyond enrichment tests.

NetworkVI’s architecture incorporates prior biological knowledge:

1. Gene-gene interactions from TADs
2. The structure of the Gene Ontology
This enables inference of gene and GO term importance scores at modality- and cell-specific resolution.