Structural Resolution in Feature Attribution Analysis of Molecular Graph Neural Networks
Abstract
Machine learning models based on molecular graphs can evaluate the contribution of each atom to a prediction using feature attribution methods. However, in chemistry, molecular properties are frequently discussed in terms of both individual atoms and structural units comprising multiple atoms, e.g., functional groups and substituents. In this study, we defined the granularity of the structural units used to present and evaluate feature attributions as their structural resolution, and we investigated how the relationship between feature attributions obtained from different molecular graph models varied with structural resolution. These attributions were evaluated at two structural resolutions, i.e., individual atoms and chemically defined structural groups. The group-level feature attributions exhibited higher intermodel agreement than the corresponding atom-level attributions for every target property and model pair. These results indicate that although the allocation of contributions to individual atoms can vary depending on the model architecture, the contribution patterns shared across multiple models can be retained at the level of entire structural groups.
Related
- Cover Feature: Stress‐Programmed Microcrack Networks in Gallium–Carbon Nanotube Thin Films for Plasma‐Active and Electroresponsive Antibacterial Interfaces (Small Struct. 10/2026)
- LM2KM: feature extraction from pretrained protein and molecular language models combined with quantum mechanics (QM) features for KM predictions
- SpectralTCN: A multi-resolution wavelet gating network for vibration-based structural damage detection
- How Molecular Weight Shapes the Relative Importance of Chemical and Structural Descriptors for Aqueous Solubility
- Polyoxometalate-Assisted Crystallization: A General Strategy Enabling Structural Characterization of Molecular Radium Complexes
BibTeX
@misc{sadakane2026structural,
title = {{Structural Resolution in Feature Attribution Analysis of Molecular Graph Neural Networks}},
author = {Teruhisa Sadakane and Daisuke Yokogawa},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010435/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010435/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Structural Resolution in Feature Attribution Analysis of Molecular Graph Neural Networks
AU - Teruhisa Sadakane
AU - Daisuke Yokogawa
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010435/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010435/v1
AB - Machine learning models based on molecular graphs can evaluate the contribution of each atom to a prediction using feature attribution methods. However, in chemistry, molecular properties are frequently discussed in terms of both individual atoms and structural units comprising multiple atoms, e.g., functional groups and substituents. In this study, we defined the granularity of the structural units used to present and evaluate feature attributions as their structural resolution, and we investigated how the relationship between feature attributions obtained from different molecular graph models varied with structural resolution. These attributions were evaluated at two structural resolutions, i.e., individual atoms and chemically defined structural groups. The group-level feature attributions exhibited higher intermodel agreement than the corresponding atom-level attributions for every target property and model pair. These results indicate that although the allocation of contributions to individual atoms can vary depending on the model architecture, the contribution patterns shared across multiple models can be retained at the level of entire structural groups.
ER -