Engineering protein variants to function in exogenous environments remains a significant challenge due to the complexity of sequence and fitness landscapes. Experimental strategies often require extensive labor and domain expertise. While recent advances in protein generative modeling offer a promising in silico alternative, many of these methods rely on differentiable fitness predictors, which limits their applicability. To this end, we introduce Protein Variant ADAptation (ProVADA), an ensemble-guided, test-time steering framework that combines implicit generative priors with fitness oracles via a composite functional objective. ProVADA leverages Mixture-Adaptation Directed Annealing (MADA), a novel sampler integrating population-annealing, adaptive mixture proposals, and directed local mutations. Furthermore, ProVADA requires no gradients or explicit likelihoods, yet efficiently concentrates sampling on high-fitness, low-divergence variants. We demonstrate its effectiveness by in silico redesigning human renin for cytosolic functionality. Our results achieve significant gains in predicted localization fitness while preserving structural integrity.
Lu et al. (Thu,) studied this question.