HybridLinker: Topology-Guided Posterior Sampling for Enhanced Diversity and Validity in 3D Molecular Linker Generation
Minyeong Hwang, Ziseok Lee, Kwang-Soo Kim, Kyungsu Kim, Eunho Yang
TL;DR
HybridLinker introduces a novel diffusion posterior sampling framework that combines the strengths of point cloud-free and point cloud-aware models to generate diverse and valid 3D molecular linkers, advancing drug discovery methods.
Contribution
The paper presents HybridLinker, a new framework that improves molecular linker generation by integrating diverse topologies from point cloud-free models into point cloud-aware inference without extra training.
Findings
Significantly improves validity and diversity in linker generation.
Outperforms baseline models in drug optimization tasks.
Establishes a new diffusion posterior sampling framework for molecular design.
Abstract
Linker generation is critical in drug discovery applications such as lead optimization and PROTAC design, where molecular fragments are assembled into diverse drug candidates via molecular linker. Existing methods fall into point cloud-free and point cloud-aware categories based on their use of fragments' 3D poses alongside their topologies in sampling the linker's topology. Point cloud-free models prioritize sample diversity but suffer from lower validity due to overlooking fragments' spatial constraints, while point cloud-aware models ensure higher validity but restrict diversity by enforcing strict spatial constraints. To overcome these trade-offs without additional training, we propose HybridLinker, a framework that enhances point cloud-aware inference by providing diverse bonding topologies from a pretrained point cloud-free model as guidance. At its core, we propose LinkerDPS, the…
Peer Reviews
Decision·Submitted to ICLR 2026
The work’s main strength is conceptual clarity around the diversity–validity trade-off and a fair framing of where existing families succeed and fail. The generation-time hybridization is simple, modular, and persuasive: the paper explains, with figures and equations, how a high-entropy surrogate topology can be refined by a validity-focused prior, and it formalizes this with LinkerDPS as a posterior over conformations that is tractable to sample with the pretrained score network. Results back t
The weaknesses stem from reliance and scope. The approach inherits capabilities and biases from both the surrogate topology model and the diffusion prior; its success will vary with those choices, and the paper’s experiments are limited to ZINC-derived benchmarks, leaving questions about transfer to protein-contexted tasks like PROTACs or strictly pocket-aware objectives. The cross-domain likelihood that powers LinkerDPS uses a simple bond-length energy and assumes conditional independence betwe
The paper discusses a tension between point-cloud-free and point-cloud-aware models for small-molecule linker generation, and proposes a hybrid framework to address it. The presentation of the method is clear, and the topic addresses an important challenge in fragment-based drug discovery. The motivation for the work is clear and it connects directly to practical workflows that could leverage existing components.
Although the background is mostly clear, the introduction does not explain why Nref is given and does not discuss cases where this value (or even Rcond) are not specified precisely but might cover a range. The empirical evaluation of this work lacks full transparency on the data splits and the methodological details, and there are no codes provided for reproduction of the results. A key ambiguity concerns the exact identity of the test set of the 400 fragment-linker pairs out of the 250k molecu
* The motivation and illustration of tasks and methods are clear. * The paper presents thorough experimental results and efficiency analyses. * The related work is well covered, and the core idea of LinkerDPS is explained clearly.
* Generating 3D geometric structures based on an initially generated 2D topology is already a known strategy in molecular modeling, which limits the novelty of the proposed framework. * The metric definitions are somewhat misleading. Typically, Uniqueness refers to the ratio of unique molecules among the valid ones. However, the paper defines Uniqueness and V+U differently—computing the unique ratio over all generated molecules—which may invalidate the claimed diversity–validity trade-off betwee
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMonoclonal and Polyclonal Antibodies Research · Chemical Synthesis and Analysis · Wikis in Education and Collaboration
MethodsDiffusion