Back to Catalog

Explainability in Graph Neural Networks: Molecular Insights

AdvancedGuided Project

Explaining how Graph Neural Networks reason is essential for validating their structural understanding. This project examines GNNExplainer for revealing which graph components drive model behavior, analyzing influential nodes, edges, and functional motifs, and evaluating explanation faithfulness through principled sparsification and substructure tests. The explainability method is applied to molecular graphs in the MUTAG dataset to uncover which atomic interactions and functional groups most strongly drive mutagenicity predictions, linking model reasoning to meaningful chemical insights.

Language

  • English

Topic

  • Artificial Intelligence

Skills You Will Learn

  • Artificial Intelligence, Graph Neural Networks (GNNs), Molecular Modeling, Explainable Artificial Intelligence (XAI), Cheminformatics, PyTorch Geometric (PyG)

Offered By

  • IBMSkillsNetwork

Estimated Effort

  • 60 minutes

Platform

  • SkillsNetwork

Last Update

  • January 28, 2026
About this Guided Project
In this project, you’ll learn how explainability techniques reveal the internal reasoning of graph-based models. You will examine how GNNExplainer highlights influential nodes, edges, and functional motifs, and how faithfulness metrics help validate whether explanations reflect true model behavior. These methods are then applied to the MUTAG molecular graph dataset, where you will investigate which atomic interactions and functional groups most strongly drive mutagenicity predictions, connecting explainability methods to real scientific use cases.

Who Is It For

This project is designed for learners who have basic familiarity with graph machine learning and want a practical, hands-on introduction to explaining Graph Neural Networks (GNNs). It is suitable for developers, data scientists, and machine learning enthusiasts who are interested in understanding how GNNs make decisions, especially in molecular prediction tasks. The project is particularly valuable for learners who want to move beyond treating GNNs as black boxes and gain insight into how graph-based models reason over nodes, edges, and substructures.

What You’ll Learn

By the end, you’ll be able to interpret GNN decisions in a principled and scientifically informed way.
  • Understand the core principles behind explaining Graph Neural Networks.
  • Use GNNExplainer to identify influential substructures and graph components.
  • Evaluate explanation quality using faithfulness-based tests.

What You'll Need

This project requires basic familiarity with Python, graph-structured data, and machine learning concepts. An understanding of molecules as graphs (atoms as nodes, bonds as edges) will be helpful. The libraries used in this project can be installed directly within the IBM Skills Network Labs environment, allowing you to set up and run the workflow without any external configuration. The project works best on modern browsers such as Chrome, Edge, Firefox, or Safari.

Instructors

Zikai Dou

Data Scientist at IBM

Ph.D. Candidate in Computer Science at McMaster University, specializing in Federated Learning (FL), Graph Neural Networks (GNNs), and Computer Vision (CV). I develop privacy-preserving, distributed AI systems that tackle real-world challenges in healthcare, finance, and enterprise applications. Passionate about bridging academic research with industry impact to advance scalable and trustworthy AI.

Read more

Contributors

Joseph Santarcangelo

Senior Data Scientist at IBM

Joseph has a Ph.D. in Electrical Engineering, his research focused on using machine learning, signal processing, and computer vision to determine how videos impact human cognition. Joseph has been working for IBM since he completed his PhD.

Read more

Wojciech "Victor" Fulmyk

Data Scientist at IBM

Wojciech "Victor" Fulmyk is a Data Scientist and AI Engineer on IBM’s Skills Network team, where he focuses on helping learners build expertise in data science, artificial intelligence, and machine learning. He is also a Kaggle competition expert, currently ranked in the top 3% globally among competition participants. An economist by training, he applies his knowledge of statistics and econometrics to bring a distinctive perspective to AI and ML—one that considers both technical depth and broader socioeconomic implications.

Read more