Protuna: a modular pipeline for proposing the next protein engineering batch

Protuna is an optimization framework for iterative protein engineering campaigns. Starting from historical mutagenesis data or a single hit, it proposes the batch of variants to test next — the decision that usually sits between a model score and a lab plate. The intended user is a team already running design-build-test cycles, not someone looking for a single predictor. Protuna does not replace the assay or the protein language model. It organizes the loop around them: define where you are allowed to mutate, generate candidates there, score them, and select a bounded set for the next round. Four interchangeable stages The core abstraction is a Pipeline. The README describes four steps that can be swapped independently, which is the main engineering value of the project. DesignSpace: holds starting sequences, allowed substitutions, and rules such as fixing sites, denying residues, or protecting conserved positions from an MSA. It can also sample smaller subspaces for independent exploration. Sampler: proposes sequences within each subspace. The quickstart uses a random sampler, but the framework also supports reward-guided generation with Feynman–Kac steering, classifier guidance, and beam search. Predictor: estimates the objective property. Hard constraints reject candidates outright; soft constraints return feasibility probabilities instead. Acquisition: selects candidates within a budget using criteria such as predicted fitness, uncertainty, or diversity. Selection is limited to what the sampler actually proposed. That last limit is worth noting. Protuna explicitly separates generation from selection because a protein language model provides a prior over plausible sequences, and that prior need not align with a campaign’s objective. Reward-guided generation is presented as a way to steer exploration toward desired properties at inference time, through externally aligned generators, or both. Multiple goals are handled with a primary objective