Resource type
Thesis type
(Thesis) Ph.D.
Date created
2026-04-10
Authors/Contributors
Author: Shen, Tongzhou (Tony)
Abstract
Practical structure-based drug design requires generating molecules that are optimised for favourable properties, are easily synthesizable, and exhibit three-dimensional interactions with a protein target structure. A central limitation is that existing generative models typically address only one of these aspects, while real-world drug discovery requires jointly addressing all three aspects simultaneously. This thesis develops a unified research program based on generative flow models for addressing three these interacting desiderata that arise in practical drug design. First, the challenge of efficient reward-guided exploration is addressed by TacoGFN, which reframes structure-based drug design as a pocket-conditioned reinforcement learning problem. Second, the requirement of chemical synthesizability grounded in explicit reaction pathways is incorporated by RxnFlow, which models molecular generation directly over feasible reaction pathways. Lastly CGFlow enables the co-design of synthesis pathways and 3D molecular conformations within a single generative framework. Together, these contributions establish compositional generative flows as a unified approach to practical drug design: reward-guided exploration is performed over objects that are reachable by explicit reaction pathways and compatible with three-dimensional binding geometry. This perspective opens a path toward future systems that incorporate both structure modelling, synthesis planning and RL in increasingly autonomous discovery pipelines.
File
Extent
143 pages.
Identifier
etd24253
Copyright statement
Copyright is held by the author(s).
Academic Supervisor
Thesis advisor: Ester, Martin
Language
English
Member of collection
| Download file | Size |
|---|---|
| etd24253.pdf | 13.07 MB |