An innovative AI-based platform for antibody stability improvement and affinity optimization

Background: Bispecific antibodies (bsAbs), which simultaneously target two distinct antigens, offer advantages in specificity, efficacy, and resistance management, making them an increasingly important modality in therapeutic antibody development. However, their complex formats impose stringent requirements on affinity and stability. Traditional affinity maturation methods, such as saturation mutagenesis and phage display, are costly, time-consuming, and limited in their ability to improve stability. To overcome these challenges, we developed an AI-based antibody engineering approach that uses deep learning to predict key mutations based on engineering objectives and integrates these predictions with high-throughput expression, enabling greatly reduced screening efforts and rapid identification of optimized antibody variants. Methods: We developed a deep learning-based AI model to support antibody engineering by simulating antigen-antibody docking and predicting affinity changes, enabling targeted mutation design according to defined optimization goals. For antibodies with poor stability, the model proposes engineered disulfide bonds or CDR/framework mutations to adjust surface hydrophobicity while maintaining affinity. For functional enhancement, it identifies key CDR residues and generates combinatorial multi-site mutations, allowing selection of variants with preserved affinity but improved blocking or functional performance. Results: As an example using a symmetric scFv bispecific antibody, we applied AI-driven design to generate 50 candidate variants, followed by binding-based screening to eliminate molecules with altered affinity. Then candidate molecules were expressed, purified, and subjected to stability evaluation. This approach effectively find a new mutant reduced aggregation under one-week accelerated thermal stress from 100% to less than 5%, while increasing the melting temperature (Tm) by up to ~10 °C. For functional enhancement of nanobodies, we applied single-point mutagenesis followed by three rounds of combinatorial design, generating a total of 260 variants-representing a ~1000-fold reduction compared with traditional multi-site saturation libraries (10³-10⁵). This process yielded a nine-site mutant (with at least one mutation per CDR) that demonstrated a two-fold improvement in reporter cells blocking tests. Conclusions: We developed an AI-guided approach to design targeted antibody variants, accelerating the discovery of molecules for bispecific antibody assembly and drug development.