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

Gene Ontology priors in biologically-informed neural networks mainly organize activations into meaningful units rather than improving performance, with the most promising approach being soft-link encoders.

• Why it matters: Understanding how biological priors influence neural network interpretability and performance is crucial for advancing biologically relevant machine learning models, yet their actual contribution remains unclear.
• What they did: The study introduced GONNECT, an autoencoder incorporating GO constraints into different network parts, and compared it against other models and controls using RNA-seq tumor data from TCGA, analyzing activation patterns and stability.
• The result: Findings show GO structure adds little to reconstruction accuracy but helps organize activations, especially in the encoder, enabling biological interpretation; soft links outside GO improve interpretability without compromising reconstruction.