Machine Learning for Structural Predictions of PROTACs - Predicting protein and molecular structures with AlphaFold and graph neural networks - Master’s thesis

Aug 4, 2026

Upload by Sage

PROteolysis TArgeting Chimeras (PROTACs) are molecules that induce targeted protein degradation by hijacking the ubiquitin–proteasome system. A PROTAC simultaneously binds an E3 ligase and a protein of interest to form a ternary complex, and productive ternary complex formation is critical for ubiquitination and subsequent degradation. Accurate structural modeling is valuable but limited by scarce PROTAC ternary complex data and by existing computational difficulties in simulating both protein partners together. This thesis applies AlphaFold for ternary modeling and introduces a graph neural network tool, the PROTAC Splitter, for predicting PROTAC substructures.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages118

Reading Time4 h 57 min

AccessFree to download

Abstract

Keywords

Acknowledgements

Machine Learning for Structural Predictions of PROTACs Predicting protein and molecular structures with AlphaFold and graph neural networks Master’s thesis in Biotechnology ANDERS KÄLLBERG DEPARTMENT OF LIFE SCIENCES CHALMERS UNIVERSITY OF TECHNOLOGY Gothenburg, Sweden 2024 [www.chalmers.se](www.chalmers.se) Master’s thesis 2024 Machine Learning for Structural Predictions of PROTACs Predicting protein and molecular structures with AlphaFold and graph neural networks Anders Källberg Department of Life Sciences Division of Data Science and AI AI Laboratory for Biomolecular Engineering Chalmers University of Technology Gothenburg, Sweden 2024 Machine Learning for Structural Predictions of PROTACs Predicting protein and molecular structures with AlphaFold and graph neural networks Anders Källberg © Anders Källberg, 2024 . Supervisor: Rocío Mercado, CSE Department at Chalmers Supervisor: Eva Nittinger, AstraZeneca Supervisor: Christian Tyrchan, AstraZeneca Examiner: Pernilla Wittung Stafshede, LIFE Department at Chalmers Master’s Thesis 2024 Department of Life Sciences Division of Data Science and AI AI Laboratory for Biomolecular Engineering Chalmers University of Technology SE-412 96 Göteborg Telephone: +46 31 772 1000 Cover: Crystallized PROTAC ternary structure (PDB: 7JTO) visualized in Molecular Operating Environment. Typeset in LATEX Gothenburg, Sweden 2024 Machine Learning for Structural Predictions of PROTACs Predicting protein and molecular structures with AlphaFold and graph neural networks Anders Källberg Department of Life Sciences Chalmers University of Technology Abstract PROteolysis TArgeting Chimeras (PROTACs) are molecules that induce the degradation of targeted proteins by hijacking the ubiquitin–proteasome system in the cell. A PROTAC binds simultaneously to an E3 ligase and a protein of interest (POI), forming a ternary complex. The ubiquitin–proteasome system tags the POI with ubiquitin, marking it for degradation by the proteasome. The formation of a good ternary complex is essential for the ubiquitination and subsequent degradation of the POI. Being able to accurately model ternary complexes thus provides critical advantages in the development of PROTACs; however, data on PROTACs and their crystallized ternary complexes are limited. Accurate predictions of these structures are desirable, but current computational methods struggle to simulate the interactions between the PROTAC and both proteins simultaneously. AlphaFold, a machine learning tool, has been shown to accurately predict protein complexes. Yet, research on applying AlphaFold to predict ternary complexes is scarce. In the first part of this thesis, the ternary complex was modeled using AlphaFold by utilizing the sequences of both natural and artificially linked POIs and E3 ligase. Nevertheless, it was determined that AlphaFold was unable to accurately predict these complexes, reasonably because it was notable to take the PROTAC into account in the predictions. The second part of this thesis focused on generating data on PROTAC substructures, essential for the development of these molecules. Despite the availability of such data, obtaining high-quality data on substructures of specific PROTACs can be challenging and time-consuming. To address this, the PROTAC Splitter, a novel machine learning tool based on graph neural networks, was developed to predict these substructures. The PROTAC Splitter predicts 99.7% of PROTACs, with known substructures, to a maximal error of 6 atoms wrong between the boundaries of the ligands and linker. It generalizes to PROTACs with three unknown substructures, where 23 . 1% of these predictions satisfy the same criteria. The code for the PROTAC splitter is available at [https://github.com/AndersKallberg/PROTAC_splitter. Although](https://github.com/AndersKallberg/PROTAC_splitter. Although) accurate predictions of ternary complexes remain challenging, the

FAQs

What is the role of a ternary complex in PROTAC function?

A PROTAC forms a ternary complex by binding an E3 ligase and a protein of interest at the same time. A good ternary complex is essential for ubiquitination and the subsequent degradation of the protein of interest.

How did the thesis use AlphaFold for structural predictions of PROTAC-related complexes?

The first part modeled ternary complexes using AlphaFold by using sequences of both the natural and artificially linked POIs and E3 ligase. Results showed AlphaFold could not accurately predict these complexes.

What is the PROTAC Splitter and what does it predict?

The second part developed the PROTAC Splitter, a graph neural network tool, to generate data on PROTAC substructures. It predicts most PROTAC substructures with low atomic boundary error and also generalizes to cases with unknown substructures.

Aug 4, 2026

Upload by Sage