VisXAI: Structure-Mapped Explainable Artificial Intelligence for Molecular Machine Learning
Abstract
Background. Molecular machine-learning models use representations such as fingerprints, tokenized simplified molecular input line entry system (SMILES) strings, and molecular graphs. Explanation methods then return importance values for fingerprint bits, sequence tokens, graph nodes, or graph edges. These values must then be mapped to atoms and bonds. This mapping depends on the representation and is often implemented separately for each model. Results. We present VisXAI, a modular Python framework for molecular featurization, trainedmodel interfaces, attribution, structural mapping, and visualization. It supports TreeSHAP for Morgan and MACCS fingerprints. For tokenized SMILES models, it supports attention, Integrated Gradients, and Grad-CAM. For PyTorch Geometric graph neural networks, it supports Integrated Gradients and Grad-CAM. Pretrained Hugging Face sequence checkpoints are supported through a tokenizer adapter and model loader. Every pathway produces the same explanation object containing atom and optional bond scores in RDKit index space. Explanations can be rendered as static two-dimensional scalable vector graphics (SVG), compared in a notebook dashboard with shared or per-panel color scales, or explored using a linked molecule/SMILES widget inspired by XSMILES. Six executable notebooks cover the three representation families, pretrained sequence models, interactive comparison, and a logP case study with an exact Crippen reference decomposition. The current source tree defines 619 Python test cases and 44 JavaScript tests, with separate continuous-integration jobs for the Python and widget front ends. Conclusions. VisXAI does not introduce a predictive architecture or attribution algorithm. It provides a tested integration and mapping layer for expressing different types of molecular explanations in the same chemical coordinates. The framework supports model analysis, method comparison, and reproducible molecular-XAI workflows without requiring separate mapping and visualization code for each representation.
Related
- Artificial Intelligence-Assisted Engineering of Nanocomposite Corrosion Protection Materials: A Critical Review of Machine Learning, Molecular Simulation, and Experimental Verification
- Explainable Molecular Structure Inference from GC--MS with Diffusion Models and LLM Reranking
- Pushing the Limits of One-Dimensional NMR Spectroscopy for Automated Structure Elucidation Using Artificial Intelligence
- Integrating artificial intelligence and molecular simulation for peptide discovery: limitations, validation and decision-centered workflows
- Artificial Intelligence and Machine Learning Model Implementation for UC Irvine Dataset 322
BibTeX
@misc{yao2026visxai,
title = {{VisXAI: Structure-Mapped Explainable Artificial Intelligence for Molecular Machine Learning}},
author = {Jiaming Yao and Samanyu Kulkarni},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008138/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008138/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - VisXAI: Structure-Mapped Explainable Artificial Intelligence for Molecular Machine Learning
AU - Jiaming Yao
AU - Samanyu Kulkarni
PY - 2026
DA - 2026/09/01
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008138/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008138/v1
AB - Background. Molecular machine-learning models use representations such as fingerprints, tokenized simplified molecular input line entry system (SMILES) strings, and molecular graphs. Explanation methods then return importance values for fingerprint bits, sequence tokens, graph nodes, or graph edges. These values must then be mapped to atoms and bonds. This mapping depends on the representation and is often implemented separately for each model. Results. We present VisXAI, a modular Python framework for molecular featurization, trainedmodel interfaces, attribution, structural mapping, and visualization. It supports TreeSHAP for Morgan and MACCS fingerprints. For tokenized SMILES models, it supports attention, Integrated Gradients, and Grad-CAM. For PyTorch Geometric graph neural networks, it supports Integrated Gradients and Grad-CAM. Pretrained Hugging Face sequence checkpoints are supported through a tokenizer adapter and model loader. Every pathway produces the same explanation object containing atom and optional bond scores in RDKit index space. Explanations can be rendered as static two-dimensional scalable vector graphics (SVG), compared in a notebook dashboard with shared or per-panel color scales, or explored using a linked molecule/SMILES widget inspired by XSMILES. Six executable notebooks cover the three representation families, pretrained sequence models, interactive comparison, and a logP case study with an exact Crippen reference decomposition. The current source tree defines 619 Python test cases and 44 JavaScript tests, with separate continuous-integration jobs for the Python and widget front ends. Conclusions. VisXAI does not introduce a predictive architecture or attribution algorithm. It provides a tested integration and mapping layer for expressing different types of molecular explanations in the same chemical coordinates. The framework supports model analysis, method comparison, and reproducible molecular-XAI workflows without requiring separate mapping and visualization code for each representation.
ER -