TeBaAb: Text-Based Antigen-Conditioned Antibody Redesign via Directed Evolution

The design of antibodies with high affinity and specificity for target antigens is a cornerstone of therapeutic and diagnostic innovation. Traditional optimization strategies, such as phage or yeast display and directed evolution, remain resource-intensive and limited in their ability to integrate contextual information. Recent AI-driven approaches have accelerated protein engineering, but most rely exclusively on structured inputs, overlooking the potential of natural language as a flexible design interface. In this work, we introduce TeBaAb, a novel text-based antigen-conditioned framework for antibody redesign that combines generative modeling with iterative optimization inspired by directed evolution. TeBaAb integrates a Conditional Variational Autoencoder (CVAE) jointly conditioned on antigen sequences and textual descriptions of antibody properties, coupled with a two-stage binding affinity predictor and an iterative enrichment loop. To support this approach, we curated AbDes, a new dataset of 7,800 text–antibody–antigen pairs with accompanying structural and binding information. Experimental evaluations demonstrate that TeBaAb improves the predicted binding affinity by an average of 15.5% compared to the original antibodies, while preserving structural confidence (RMSPE < 1 angstrom) and generating sequences that are diverse and novel. By enabling text-conditioned antigen-specific antibody design, TeBaAb provides a promising new paradigm for accelerating therapeutic antibody discovery and expanding the antibody design space beyond traditional methods.