TreeGazer: Prospecting Protein Sequence-Function Landscapes via Phylogenetic Structure

Abstract

Building diverse and informative protein sequence datasets is critical for understanding how function varies across sequence space. Because only a small fraction of sequences in a dataset can typically be experimentally characterised, strategies for selecting what sequences to characterise should maximise the information gained from each experiment. Here, we present TreeGazer, a phylogeny-informed framework that combines Bayesian optimisation with the topology of a tree to guide sequence selection. TreeGazer balances exploitation of sequences predicted to exhibit favourable properties against exploration of regions higher model uncertainty. Unlike existing approaches that apply Bayesian optimisation for sequence selection, TreeGazer does not rely on black-box models and instead uses latent representations of property distributions that are directly tied to phylogenetic structure. Modelling properties in this way enables biologically interpretable predictions and uncertainty estimates. Across two simulated selection campaigns, TreeGazer consistently selected sequences that produced datasets more representative of the underlying property distribution than alternative strategies that used protein language models. TreeGazer also performed effectively in low-data settings, where tree-guided selection enabled accurate identification of functional transitions across clades. TreeGazer can be run on conventional laptop computers while still providing equivalent or superior performance to embedding-based approaches. These results demonstrate that phylogenetic structure is a powerful and underutilised prior for guiding informative sequence selection.