Active learning enables evolutionary discovery and characterization of fungal transcriptional activators

Abstract

Biological discovery and design are increasingly guided by predictive models trained on data from high-throughput technologies rather than costly experiments. However, existing datasets are often biased by overrepresentation of model organisms, causing models to fail in evolutionary studies of non-model species. We focus on transcriptional activators, which contain activation domains (ADs) that promote gene expression. ADs are intrinsically disordered and poorly conserved, limiting their study using comparative genomics. We present a hybrid framework that leverages high-throughput molecular assays and active learning to quantify biological properties across evolutionary space. We develop ADhunter, a high-capacity regression model that outperforms state-of-the-art algorithms in identifying transcriptional activators and quantifying their strength. We use model-based uncertainty to guide evolutionary sampling across 7,842,516 proteins from 2,400 fungal genomes. We functionally characterize 9,836 ADs from 1,071 fungal genomes, providing a 15.5-fold expansion in genome representation compared with existing datasets. Comprehensive sampling improves model generalizability and provides the first functional annotation for 3,416 proteins in non-model fungi. Interpretability analysis of ADhunter aligns with biophysical models and reveals novel, underrepresented protein codes. These results highlight the importance of sampling from non-model organisms to build evolutionarily robust functional genomics models. Our framework provides a general strategy for building predictive models that better capture the diversity of natural sequence-to-function relationships.