Ligand Conformation Generation: from singleton to pairwise

Jiamin WU, He CAO, Jia Li, Yuan Yao

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: molecular conformation generation, conditional diffusion model, graph neural network

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Drug discovery is a time-consuming process, primarily due to the vast number of molecular structures that need to be explored. One of the challenges in drug design involves generating rational ligand conformations. For this task, most previous approaches fall into the singleton category, which solely rely on ligand molecular information to generate ligand conformations. In this work, we contend that the ligand-target interactions are also very important in providing crucial semantics for ligand generation. To address this, we introduce PsiDiff, a comprehensive diffusion model that incorporates target and ligand interactions, as well as ligand chemical properties. By transitioning from singleton to pairwise modeling, PsiDiff offers a more holistic approach. One challenge of the pairwise design is that the ligand-target binding site is not available in most cases and thus hinders the accurate message-passing between the ligand and target. To overcome this challenge, we employ graph prompt learning to bridge the gap between ligand and target graphs. The graph prompt learning of the insert patterns enables us to learn the hidden pairwise interaction at each diffusion step. Upon this, our model leverages the Target-Ligand Pairwise Graph Encoder (TLPE) and captures ligand prompt entity fusion and complex information. Experimental results demonstrate significant improvements in ligand conformation generation, with a remarkable 18\% enhancement in Aligned RMSD compared to the baseline approach.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1743