PROTAC-Splitter: a machine learning framework for automated identification of PROTAC substructures.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Proteolysis-targeting chimeras (PROTACs) are heterobifunctional molecules with therapeutic potential.
• Manual annotation of PROTAC components (E3 ligase ligand, linker, warhead) is challenging and time-consuming.
• Automated methods are needed to accurately identify and annotate PROTAC substructures.

Purpose of the Study:

• To develop and validate PROTAC-Splitter, a machine learning framework for automated PROTAC substructure annotation.
• To address data scarcity by generating and releasing a large synthetic dataset of annotated PROTACs.
• To compare different machine learning models for PROTAC annotation.

Main Methods:

• Development of PROTAC-Splitter, a machine learning framework.
• Generation of a synthetic dataset of ~1.3 million annotated PROTAC structures.
• Implementation of Transformer-based and XGBoost-based models for substructure annotation.
• Evaluation on public and proprietary PROTAC datasets, including structurally novel compounds.
• Development of a Transformer wrapper (Transformer-Δ) to correct prediction errors.

Main Results:

• The Transformer model achieved 86% exact-match accuracy on public data but struggled with novel structures.
• The XGBoost model ensured chemical validity and perfect reassembly but had lower exact-match accuracy.
• Transformer-Δ improved reassembly accuracy to 96% (public) and 70% (internal datasets).
• A hybrid approach combining Transformer-Δ and XGBoost demonstrated robust annotation across diverse chemical spaces.

Conclusions:

• PROTAC-Splitter offers a reliable and scalable solution for automated PROTAC analysis.
• The hybrid approach effectively annotates PROTACs, overcoming limitations of individual models.
• The open-source availability of PROTAC-Splitter facilitates broader adoption in drug discovery.