Active Probabilistic Drug Discovery

Xun Deng, Junlong Liu, Jiansheng Wu, Fuli Feng, Xiangnan He, Zheng Wang, Jieping Ye

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

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: Drug Discovery, Active Learning, Molecule Clustering

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

Abstract:

Early drug discovery plays a crucial role in the development of new medications
by focusing on the identification and optimization of lead molecules that specif-
ically bind to target proteins. However, this process is accompanied by various
challenges, such as the vastness of molecule libraries, high attrition rate, and the
intricate nature of molecular interactions. To overcome these challenges, there is
a paradigm shift towards integrating intelligence and automation into end-to-end
operations. Intelligent computing aids in the discovery and recommendation of
molecules, while automated experiments offer data validation and feedback. This
innovative approach can be viewed as an active probabilistic learning problem,
assuming that active molecules binding to a specific target are typically a small
proportion and exhibit cluster-distributed characteristics. Based on this formu-
lation, we propose a novel active probabilistic drug discovery (APDD) method,
which iteratively updates the binding probabilities of molecules to progressively
enhance drug discovery performance with three consecutive steps of probabilistic
clustering, selective docking, and active wet-experiment. We conduct extensive
experiments on two benchmark datasets of DUD-E and LIT-PCBA and a simu-
lated virtual library. The results demonstrate the feasibility and efficiency of our
approach, showcasing substantial cost savings with an average reduction of 80% in
computational docking expenses and 70% in wet experimental costs, while main-
taining high accuracy in lead molecule discovery.

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: 2696