Journal Article Ahead of print
PubMed: 42270769 DOI: 10.1038/s41598-026-53215-x
Structural diversity and chemical space analysis of a PROTAC database using unsupervised machine learning.
Sci Rep
Authors: Ashutosh Kharwar, Alberto Marbán-González, José L Medina-Franco, Carlos A Velázquez-Martínez
Affiliations
- Faculty of Pharmacy and Pharmaceutical Sciences, University of Alberta, Edmonton, AB, T6G 2E1, Canada. Department of Chemistry, Babasaheb Bhimrao Ambedkar University, Lucknow, 226025, Uttar Pradesh, India.
- DIFACQUIM Research Group, Department of Pharmacy, Faculty of Chemistry, Universidad Nacional Autónoma de México, Mexico City, 04510, Mexico.
- Faculty of Pharmacy and Pharmaceutical Sciences, University of Alberta, Edmonton, AB, T6G 2E1, Canada.
Abstract
Targeted protein degradation (TPD) mediated by proteolysis-targeting chimeras (PROTACs) has emerged as a powerful therapeutic strategy, enabling the catalytic and selective elimination of disease-relevant proteins via the ubiquitin-proteasome system. Despite their ability to overcome drug resistance and address traditionally undruggable targets, the structural complexity and vast chemical diversity of PROTACs present challenges for systematic analysis and rational design. Here, we present a systematic unsupervised machine-learning framework that represents the first large-scale similarity-driven clustering and scaffold-centric analysis of the PROTAC chemical space, to comprehensively characterize the structural, functional, and physicochemical landscape of PROTAC molecules and to support data-driven lead optimization and next-generation degrader design. An initial dataset of 9,380 compounds was curated from the publicly available PROTACs Database (PROTAC-DB 3.0), followed by rigorous standardization and filtering, resulting in 6,113 unique, chemically valid compounds. The chemical space was explored using a multi-step computational pipeline involving dimensionality reduction and a comparative evaluation of diverse clustering algorithms. Among the evaluated approaches, a refined clustering strategy demonstrated superior performance in partitioning the dataset into structurally coherent groups. Structural analysis revealed a pronounced convergence around canonical PROTAC architectures, characterized by conserved E3 ligase-binding motifs, diverse target-binding frameworks, and heterogeneous linker designs. Functional group profiling and physicochemical analysis further demonstrated that these compounds predominantly occupy a specialized chemical space beyond traditional drug-like limits, marked by high molecular weight and substantial conformational flexibility. Collectively, these findings provide data-driven design guidance for PROTAC optimization by highlighting frequent scaffold architectures and preferred property ranges, thereby informing the development of next-generation degraders.
MeSH
Pharmaceutical Preparations, Machine Learning, Cluster Analysis, Learning, Unsupervised Machine Learning