GONNECT: A Gene Ontology-guided Neural Network for Explainable Cancer Typing

Abstract

Biologically-informed neural networks (BINNs) embed prior knowledge such as the Gene Ontology (GO) into their architecture to produce structurally interpretable representations, yet whether and how this prior improves performance or interpretation remains unclear. Here, 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 and VEGA), randomized-prior controls, and an unconstrained baseline. Across metrics, GO structure adds little to reconstruction or latent-space organization, frequently matched by randomized or unconstrained models. Its value lies in node activations, particularly in the encoder, where they correlate with a gene set enrichment analysis (GSEA)-derived reference. GONNECT-SL introduces regularized connections outside GO, but these soft links are unstable across seeds and concentrate where the ontology is sparse, appearing to compensate for the priors constraints rather than reveal new biology. They recover near-unconstrained reconstruction, keeping encoder activations interpretable. We identify the soft-link encoder as most promising. Our results clarify what biological priors contribute: their value lies not in the identity of the imposed connections or in improved performance, but in organizing activations into biologically meaningful units that can be interrogated directly.