A critical evaluation of Gene Ontology priors in biologically-informed neural networks

Abstract

Biologically-informed neural networks (BINNs) embed prior knowledge such as the Gene Ontology (GO) into their architecture to produce interpretable representations, yet whether and how this prior improves performance or interpretation remains unclear. We introduce GONNECT, a BINN incorporating GO into an autoencoder. We evaluate GO constraints in the encoder, decoder, or both on RNA-seq tumour samples from The Cancer Genome Atlas (TCGA), comparing against published BINNs (OntoVAE, VEGA), an unconstrained baseline, randomized priors, and untrained networks. Across various performance metrics, GO structure adds little to reconstruction or latent-space organization, frequently matched by randomized or unconstrained models. Node activations agree with a gene set enrichment analysis (GSEA)-derived reference, but this agreement largely arises from the architectural constraint: it is lost when GO connections are randomized, present even before training in the encoder, and learned in the decoder. Regularized soft links outside GO (GONNECT-SL) are unstable across runs and concentrate where the ontology is sparse, compensating for the prior's constraints rather than revealing new biology. Within this setting, we find no evidence that node activations capture biological information beyond that encoded by the predefined gene sets, suggesting that their apparent interpretability largely reflects the prior imposed by the architecture.