Feature selection-guided graph neural networks for compound activity prediction

Molecular graphs serve as input for graph neural networks (GNNs) for compound classification and regression tasks. While chemical information captured by molecular graphs is often sufficient for compound classification and molecular property prediction tasks, it is often attempted to further increase the chemical information content of graph representations by augmenting them with various molecular descriptors. For this purpose, several approaches have been introduced that operate at different stages of GNN training, either before, during, or after message passing. Despite their methodological differences, these approaches commonly challenge GNNs to reduce feature noise, redundancy, or biases during learning, which often negatively affects predictions. As a conceptual alternative, we introduce a two-stage methodology in which a feature ranking GNN is used first to automatically select features important for compound activity prediction, without requiring the calculation of any descriptor values at this stage, followed by the derivation of a classification GNN based on the preselected feature set. For different compound activity classes and other datasets, we show that small preselected descriptor subsets achieve GNN predictive performance comparable to or better than original descriptor pools used for graph augmentation and also comparable to reference methods. Control calculations reveal that small preselected descriptor subsets add task-specific information for molecular graph classification, consistently outperforming randomly chosen subsets of equal size. Furthermore, we demonstrate that selected features aid in explaining predictions, are chemically interpretable, and represent activity determinants. Scientific contribution We present a new methodological concept for increasing the chemical information content of molecular graphs as input for GNNs through feature augmentation. The approach separates the augmentation and learning phases by using a feature ranking GNN in combination with a classification GNN. This approach essentially maintains or increases predictive performance and supports the explanation of predictions.

Authors

Institutions

Publication Details

Journal
Journal of Cheminformatics
Published
2026-11-18
DOI
https://doi.org/10.1186/s13321-026-01328-w
Primary Topic
Computational Drug Discovery Methods
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Feature selection-guided graph neural networks for compound activity prediction

Fouad Alkhoury, Andrea Mastropietro, Jürgen Bajorath, Tiago Janela et al.
1 links
Journal of Cheminformatics
Computational Drug Discovery Methods
article

Feature selection-guided graph neural networks for compound activity prediction

Fouad Alkhoury, Andrea Mastropietro, Jürgen Bajorath, Tiago Janela, Tamás Horváth
article en

Abstract

Molecular graphs serve as input for graph neural networks (GNNs) for compound classification and regression tasks. While chemical information captured by molecular graphs is often sufficient for compound classification and molecular property prediction tasks, it is often attempted to further increase the chemical information content of graph representations by augmenting them with various molecular descriptors. For this purpose, several approaches have been introduced that operate at different stages of GNN training, either before, during, or after message passing. Despite their methodological differences, these approaches commonly challenge GNNs to reduce feature noise, redundancy, or biases during learning, which often negatively affects predictions. As a conceptual alternative, we introduce a two-stage methodology in which a feature ranking GNN is used first to automatically select features important for compound activity prediction, without requiring the calculation of any descriptor values at this stage, followed by the derivation of a classification GNN based on the preselected feature set. For different compound activity classes and other datasets, we show that small preselected descriptor subsets achieve GNN predictive performance comparable to or better than original descriptor pools used for graph augmentation and also comparable to reference methods. Control calculations reveal that small preselected descriptor subsets add task-specific information for molecular graph classification, consistently outperforming randomly chosen subsets of equal size. Furthermore, we demonstrate that selected features aid in explaining predictions, are chemically interpretable, and represent activity determinants. Scientific contribution We present a new methodological concept for increasing the chemical information content of molecular graphs as input for GNNs through feature augmentation. The approach separates the augmentation and learning phases by using a feature ranking GNN in combination with a classification GNN. This approach essentially maintains or increases predictive performance and supports the explanation of predictions.

Journal of Cheminformatics
University of Bonn (DE), Fraunhofer Institute for Intelligent Analysis and Information Systems (DE), Bonn Aachen International Center for Information Technology (DE), Lamarr Institute for Machine Learning and Artificial Intelligence (DE), Nara Institute of Science and Technology (JP)
Rheinische Friedrich-Wilhelms-Universität Bonn
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon