Steering Protein Generative Models at Test-Time for Guided AAV2 Capsid Design

Recent advances in protein generative models have created new opportunities for protein engineering. However, a significant challenge remains in effectively steering these models to generate sequences with specific, desired functionalities, especially when these properties are defined by "black-box" or non-differentiable fitness functions. To address this, we present ProVADA+, a model-agnostic framework that guides pretrained generative models at testtime without costly retraining. Our approach introduces a reinforcement learning-based adaptive masking technique (MADA-DUCB) that significantly accelerates convergence. We demonstrate this framework on the challenging task of designing novel Adeno-Associated Virus 2 (AAV2) capsids. By coupling a ProteinMPNN generative prior with a fine-tuned AAV viability oracle, our method successfully navigates the rugged fitness landscape where unguided random mutagenesis is ineffective-with prior experiments showing as few as 0.3% of variants with six or more mutations are viable. In its final iterations, ProVADA generated a pool of novel candidates with a mean viral selection score of 2.72, consistently scoring highly viable variants while maintaining a diverse range of sequence similarity to the wildtype sequence. Our results show that ProVADA provides a powerful and efficient framework for accelerating the design of proteins with complex, user-defined properties.