Loading...

Controllable De Novo Antibody Design Using Probabilistic Diffusion Models

Halvaei, Reihaneh | 2025

53 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58835 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sharifi Zarchi, Ali; Asgari, Ehsaneddin
  7. Abstract:
  8. This research addresses the challenge of controllability in computational antibody design using generative diffusion models, particularly RFantibody. Although these models have demonstrated success in generating diverse antibody structures, they still face limitations in steering the generative process toward specific design goals and directly optimizing binding affinity. The main innovation of this study lies in introducing two external guiding potentials during inference, which enhance the generation process in a modular manner without retraining the model. The first potential, focusing on interface hotspots, directs the antibody toward the desired epitope. The second potential, inspired by biophysical principles, reinforces shape complementarity at the antibody–antigen interface during the final refinement steps, resulting in more stable binding. This approach is implemented within the RFantibody framework, leveraging the RFdiffusion backbone and RoseTTAFold architecture for accurate 3D structure prediction. Experimental results demonstrate that the proposed method preserves structural fidelity while significantly reducing binding free energy and enhancing epitope engagement. Comparative evaluations against advanced models such as DiffAb and AbX indicate competitive or superior performance. From a computational perspective, the proposed method—despite its simplicity and independence from retraining—achieves a remarkable improvement in controllability and antibody generation efficiency. Overall, it represents a promising step toward accelerating in silico antibody design and advancing the development of targeted antibody-based therapeutics
  9. Keywords:
  10. Antibody Design ; Diffusion Model ; Binding Affinity ; Protein Engineering ; Deep Learning ; Guided Inference

 Digital Object List

 Bookmark

...see more