Bringing clarity to how machine learning models “see” molecules 

Bringing clarity to how machine learning models “see” molecules 

New Publication in the Journal of Computational Science

Our researchers Adam Sułek and Tomasz Kościołek, together with their co-authors, have published a new paper in the Journal of Computational Science, addressing one of the key challenges in molecular machine learning: evaluating atom-level explanations.

“Benchmarking atom-level explainability against pharmacophore-computed labels in molecular machine learning”

Authors: Adam Sułek, Jakub Klimczak, Jakub Jończyk, Tomasz Kościołek, Tomasz Danel and Barbara Pucelik

Machine learning models can predict molecular properties with impressive accuracy—but do their explanations identify atoms that are genuinely relevant to molecular structure and function?

The study introduces a controlled benchmark based on pharmacophores and geometric rules describing the spatial arrangement of functional elements within molecules. At the core of the framework is PharmacoScore, a metric that quantifies how closely atom-level model explanations align with pharmacophore-derived reference annotations.

Using this framework, the authors compared explanation methods across several molecular machine learning architectures. Distance-aware transformer models, which explicitly encode interatomic geometry, consistently achieved higher PharmacoScore values than models that do not explicitly represent molecular geometry.

High predictive accuracy does not necessarily mean that a model relies on chemically meaningful features. With PharmacoScore, we created a controlled way to test whether atom-level explanations recover the spatial relationships defined by pharmacophores. This gives us a common reference point for comparing explainability methods across different molecular machine learning architectures.

— Adam Sułek

The study provides a step towards molecular machine learning models that are evaluated not only by their predictive performance, but also by the chemical relevance of their explanations.

Congratulations to all the authors!

This publication is one of the outcomes of the FIRST TEAM FENG project FNP Foundation for Polish Science, under which Łukasiewicz – Krakow Institute of Technology is developing new approaches to the identification and evaluation of therapeutic candidates in hormone-dependent breast cancer.