Collaborative research · Shared first author

Coarsening molecular graphs around functional groups.

FunQG uses a quotient-graph perspective to build smaller molecular representations from functional groups, then evaluates those representations with graph neural networks for molecular property prediction.

Contribution: Co-author · shared first author
Research context: Machine Learning and Graph Mining Lab
Published: Journal of Chemical Information and Modeling · 2023

Read the published FunQG paper ↗

Research question

Can a molecular graph be simplified into a smaller, more informative graph that supports molecular-property prediction while reducing the depth and computational burden associated with message passing on the original graph?

The published work motivates this through known difficulties in deeper GNNs, including over-smoothing and over-squashing, and the cost of larger parameterized models.

Method

FunQG uses functional groups as meaningful molecular building blocks. The repository describes a pipeline that extracts functional groups, partitions molecular graphs, generates quotient/coarsened graphs, and trains GNN models for prediction. The code includes graph generation, training/evaluation, and hyperparameter-tuning entrypoints, with molecular-property datasets and scaffold/random splitting options documented.

I developed, trained, benchmarked, and evaluated graph-based ML models, including work on FunQG. This is collaborative research; shared first authorship does not imply sole ownership of the code or results.

The published FunQG framework: a molecule is decomposed into functional groups, represented through quotient-graph coarsening, then processed by a graph neural network for molecular-property prediction. — Framework figure distributed with the collaborative FunQG repository. It illustrates the molecular-graph coarsening and GNN prediction flow; it is not a portfolio-created benchmark graphic.

Published findings

The article reports that the resulting molecular quotient graphs are substantially smaller than the original molecular graphs, and that experiments across molecular-property prediction benchmarks produce notable prediction results while reducing model parameters and computational costs. Results vary across the reported tasks and experimental settings; see the paper for benchmark-specific results.

The publication compares GNN baselines on the coarsened datasets with state-of-the-art baselines on the original datasets. The implementation README describes evaluation on datasets including Tox21, BBBP, BACE, SIDER, ESOL, FreeSolv, HIV, QM7/QM8, and PDBbind tasks.

Research scope

The evidence describes molecular-property prediction benchmarks and a functional-group-based graph-coarsening method. The portfolio does not generalize these findings to unrelated graph tasks or production systems. Exact benchmark outcomes depend on the paper’s dataset, split, baseline, and training setup; consult the article and supporting information for details.

This was collaborative research. My CV records shared first authorship and direct work on model development, training, benchmarking, and evaluation.

Publication & code

Hajiabolhassan, H.; Taheri, Z.; Hojatnia, A.; Yeganeh, Y. T. “FunQG: Molecular Representation Learning via Quotient Graphs.” Journal of Chemical Information and Modeling 2023, 63(11), 3275–3287. DOI: 10.1021/acs.jcim.3c00445 ↗

Collaborative code repository ↗ · ACS article record ↗ · 2022 preprint ↗

The arXiv link is the earlier preprint; the citation above refers to the final 2023 journal article.

See research overview