Cade Gordon

GENERATIVE HUMANIZATION FOR THERAPEUTIC ANTIBODIES 1. Introduced a novel framework reframing antibody humanization as a conditional generative modeling task, leveraging masked language models (MLMs) trained on human antibody data to produce diverse, humanized sequences with reduced immunogenicity risk. 2. Demonstrated that the generative humanization method enriches therapeutic attributes like binding affinity and thermostability through guided sampling, outperforming traditional and deterministic approaches. 3. Validated the approach in silico and in lab settings, showing that the model generates multiple candidates with superior antigen binding compared to baseline methods, such as Sapiens and manual humanization. 4. Employed guided sampling with a “product of experts” approach, integrating humanness metrics and oracle-predicted properties like binding affinity, resulting in highly humanized antibodies that retain or improve therapeutic properties. 5. Highlighted the flexibility of the method, enabling iterative optimization campaigns where therapeutic attributes are progressively enhanced while maintaining high humanness. 6. Provided experimental results showing guided sampling surpasses unguided approaches in generating candidates with both high humanness and improved binding affinity, validated through high-throughput lab measurements. 7. Introduced structure-guided sampling for scenarios lacking specific attribute oracles, leveraging structural predictors to maintain antigen binding properties while humanizing antibody sequences. 8. Positioned the approach as a transformative method for accelerating the development of antibody therapies, reducing immunogenicity risks while optimizing therapeutic efficacy. @peytongreenside @CadeGordonML @BigHatBio 📜Paper: arxiv.org/abs/2412.04737 #TherapeuticAntibodies #AntibodyHumanization #GenerativeModeling #Bioinformatics #AI #DrugDevelopment