A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks

Alexander Kensert; Gert Desmet; Deirdre Cabooter

doi:10.1016/j.aca.2024.343046

A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks

Alexander Kensert
, Gert Desmet
, Deirdre Cabooter

Research output: Contribution to journal › Article › peer-review

4 Citations (Scopus)

8 Downloads (Pure)

Abstract

This hands-on tutorial offers a practical guide to implementing Graph Neural Networks (GNNs) within the scope of Quantitative Structure-Activity Relationship (QSAR) modeling. Readers will be guided through a detailed, step-by-step process of applying a state-of-the-art GNN algorithm for activity predictions. It is anticipated that this practical tutorial will provide readers with sufficient knowledge to apply GNNs to their problems as well as to gain a good intuition of the underlying concepts. Specifically, the first part of this tutorial introduces the basic theory of GNNs for QSAR, including molecular graphs and their transformation in the GNN. The second part translates this theory into practical code, culminating in a working implementation of a QSAR model using a GNN. Basic knowledge of QSAR modeling and basic proficiency in the Python programming language and Keras deep learning framework is required. The code for the implementation detailed in this tutorial is available at https://github.com/akensert/molgraph/.

Original language	English
Article number	343046
Number of pages	12
Journal	Analytica Chimica Acta
Volume	1331
DOIs	https://doi.org/10.1016/j.aca.2024.343046
Publication status	Published - 1 Dec 2024

Bibliographical note

Funding Information:
The importance of QSAR modeling for toxicity predictions was expressed in 2014 when the organizers of Tox21 (Toxicology in the 21st Century) challenged invited participants to develop computational methods for predicting the toxicity of chemical compounds [1,2]. More specifically, the challenge was aimed at discovering modern methods capable of accurately predicting the activity of chemical compounds against specific types of receptors in the body (Fig. 1), including Steroid Receptors (SR) and Non-Steroid Receptors (NR), which when activated, may disrupt biological pathways in ways that may result in toxic effects. Among the various computational methods tested, including Random Forests and Support Vector Machines (often operating on molecular descriptors [3]), a Deep Neural Network (DNN), known as DeepTox, demonstrated superior performance for this specific challenge [4]. DeepTox, with its fully-connected neural network (FCNN) architecture, operated on a specific type of numerical representation of molecules, known as extended-connectivity fingerprints (ECFP) [5], which directly leveraged the graph structure of molecules to capture the presence or absence of molecular substructures of different radii.Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \u2013 research project \u201CChimic\u201D (EOS ID: 30897864)).

Funding Information:
Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \u2013 research project \u201CChimic\u201D (EOS ID: 30897864)).

Publisher Copyright:
© 2024

Keywords

Bioinformatics
Computational chemistry
Deep learning
Neural networks
Predictive modeling
Toxicity predictions

Access to Document

10.1016/j.aca.2024.343046

graph convolutional networks as submitted 2021Accepted author manuscript, 1.32 MBLicence: CC BY-NC-ND

Handle.net

Persistent link

Cite this

@article{e39b0dc76a2448dd8060dec49ac4e3cc,

title = "A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks",

abstract = "This hands-on tutorial offers a practical guide to implementing Graph Neural Networks (GNNs) within the scope of Quantitative Structure-Activity Relationship (QSAR) modeling. Readers will be guided through a detailed, step-by-step process of applying a state-of-the-art GNN algorithm for activity predictions. It is anticipated that this practical tutorial will provide readers with sufficient knowledge to apply GNNs to their problems as well as to gain a good intuition of the underlying concepts. Specifically, the first part of this tutorial introduces the basic theory of GNNs for QSAR, including molecular graphs and their transformation in the GNN. The second part translates this theory into practical code, culminating in a working implementation of a QSAR model using a GNN. Basic knowledge of QSAR modeling and basic proficiency in the Python programming language and Keras deep learning framework is required. The code for the implementation detailed in this tutorial is available at https://github.com/akensert/molgraph/.",

keywords = "Bioinformatics, Computational chemistry, Deep learning, Neural networks, Predictive modeling, Toxicity predictions",

author = "Alexander Kensert and Gert Desmet and Deirdre Cabooter",

note = "Funding Information: The importance of QSAR modeling for toxicity predictions was expressed in 2014 when the organizers of Tox21 (Toxicology in the 21st Century) challenged invited participants to develop computational methods for predicting the toxicity of chemical compounds [1,2]. More specifically, the challenge was aimed at discovering modern methods capable of accurately predicting the activity of chemical compounds against specific types of receptors in the body (Fig. 1), including Steroid Receptors (SR) and Non-Steroid Receptors (NR), which when activated, may disrupt biological pathways in ways that may result in toxic effects. Among the various computational methods tested, including Random Forests and Support Vector Machines (often operating on molecular descriptors [3]), a Deep Neural Network (DNN), known as DeepTox, demonstrated superior performance for this specific challenge [4]. DeepTox, with its fully-connected neural network (FCNN) architecture, operated on a specific type of numerical representation of molecules, known as extended-connectivity fingerprints (ECFP) [5], which directly leveraged the graph structure of molecules to capture the presence or absence of molecular substructures of different radii.Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \textbackslash{}u2013 research project \textbackslash{}u201CChimic\textbackslash{}u201D (EOS ID: 30897864)). Funding Information: Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \textbackslash{}u2013 research project \textbackslash{}u201CChimic\textbackslash{}u201D (EOS ID: 30897864)). Publisher Copyright: {\textcopyright} 2024",

year = "2024",

month = dec,

day = "1",

doi = "10.1016/j.aca.2024.343046",

language = "English",

volume = "1331",

journal = "Analytica Chimica Acta",

issn = "0003-2670",

publisher = "Elsevier",

}

TY - JOUR

T1 - A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks

AU - Kensert, Alexander

AU - Desmet, Gert

AU - Cabooter, Deirdre

N1 - Funding Information: The importance of QSAR modeling for toxicity predictions was expressed in 2014 when the organizers of Tox21 (Toxicology in the 21st Century) challenged invited participants to develop computational methods for predicting the toxicity of chemical compounds [1,2]. More specifically, the challenge was aimed at discovering modern methods capable of accurately predicting the activity of chemical compounds against specific types of receptors in the body (Fig. 1), including Steroid Receptors (SR) and Non-Steroid Receptors (NR), which when activated, may disrupt biological pathways in ways that may result in toxic effects. Among the various computational methods tested, including Random Forests and Support Vector Machines (often operating on molecular descriptors [3]), a Deep Neural Network (DNN), known as DeepTox, demonstrated superior performance for this specific challenge [4]. DeepTox, with its fully-connected neural network (FCNN) architecture, operated on a specific type of numerical representation of molecules, known as extended-connectivity fingerprints (ECFP) [5], which directly leveraged the graph structure of molecules to capture the presence or absence of molecular substructures of different radii.Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \u2013 research project \u201CChimic\u201D (EOS ID: 30897864)). Funding Information: Alexander Kensert is funded by a joint initiative of the Research Foundation Flanders (FWO) and the Walloon Fund for Scientific Research (FNRS) (EOS \u2013 research project \u201CChimic\u201D (EOS ID: 30897864)). Publisher Copyright: © 2024

PY - 2024/12/1

Y1 - 2024/12/1

N2 - This hands-on tutorial offers a practical guide to implementing Graph Neural Networks (GNNs) within the scope of Quantitative Structure-Activity Relationship (QSAR) modeling. Readers will be guided through a detailed, step-by-step process of applying a state-of-the-art GNN algorithm for activity predictions. It is anticipated that this practical tutorial will provide readers with sufficient knowledge to apply GNNs to their problems as well as to gain a good intuition of the underlying concepts. Specifically, the first part of this tutorial introduces the basic theory of GNNs for QSAR, including molecular graphs and their transformation in the GNN. The second part translates this theory into practical code, culminating in a working implementation of a QSAR model using a GNN. Basic knowledge of QSAR modeling and basic proficiency in the Python programming language and Keras deep learning framework is required. The code for the implementation detailed in this tutorial is available at https://github.com/akensert/molgraph/.

AB - This hands-on tutorial offers a practical guide to implementing Graph Neural Networks (GNNs) within the scope of Quantitative Structure-Activity Relationship (QSAR) modeling. Readers will be guided through a detailed, step-by-step process of applying a state-of-the-art GNN algorithm for activity predictions. It is anticipated that this practical tutorial will provide readers with sufficient knowledge to apply GNNs to their problems as well as to gain a good intuition of the underlying concepts. Specifically, the first part of this tutorial introduces the basic theory of GNNs for QSAR, including molecular graphs and their transformation in the GNN. The second part translates this theory into practical code, culminating in a working implementation of a QSAR model using a GNN. Basic knowledge of QSAR modeling and basic proficiency in the Python programming language and Keras deep learning framework is required. The code for the implementation detailed in this tutorial is available at https://github.com/akensert/molgraph/.

KW - Bioinformatics

KW - Computational chemistry

KW - Deep learning

KW - Neural networks

KW - Predictive modeling

KW - Toxicity predictions

UR - https://www.scopus.com/pages/publications/85203554601

U2 - 10.1016/j.aca.2024.343046

DO - 10.1016/j.aca.2024.343046

M3 - Article

AN - SCOPUS:85203554601

SN - 0003-2670

VL - 1331

JO - Analytica Chimica Acta

JF - Analytica Chimica Acta

M1 - 343046

ER -

A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks

Abstract

Bibliographical note

Keywords

Access to Document

Handle.net

Other files and links

Fingerprint

OZRMETH7: BruChrom: Breaking Boundaries in Chromatography

Cite this

A hands-on tutorial on quantitative structure-activity relationships using fully expressive graph neural networks

Abstract

Bibliographical note

Keywords

Access to Document

Handle.net

Other files and links

Fingerprint

Projects

OZRMETH7: BruChrom: Breaking Boundaries in Chromatography

Cite this