Page 106 - AJWEP-v22i3

Sonsare, et al.

6. Conclusion and future scope prediction, which makes them indispensable tools in
computational chemistry and drug discovery.
This study investigated the effectiveness of three GNN GINs are an effective methodology for predicting
architectures: GCN, GIN, and GAT, for molecular property environmental partition coefficients that use graph-based
prediction using the MUTAG dataset. GIN achieved the molecular representations. The model performs well for
best results, with an accuracy of 89.20% and an ROC- K but might improve for K with more vapor pressure-
aw
d
AUC of 0.94, confirming its superior expressiveness related properties. Future enhancements may include
in distinguishing molecular graph structures. GAT using transformer-based graph models (e.g., Graphormer,
demonstrated strong recall and interpretability through ChemBERTa) to increase prediction accuracy.
its attention mechanism, achieving 88.30% accuracy Our comparative study indicates that GCN, GAT, and
and an ROC-AUC of 0.93. GCN provided competitive GIN are all effective for molecular property prediction,
performance (accuracy of 87.50%) while being with performance variations that are relatively minor.
computationally efficient. The results align with or exceed Among them, GIN appeared to deliver slightly more
the performance of existing methods, including classical consistent and higher-quality predictions across different
graph kernel techniques and other GNN variants. GIN’s molecular properties, though not by a statistically
results were consistent with its established reputation for significant margin. Thus, while GIN shows promise
strong representational power, whereas GAT showcased for further development, all three models offer viable
robustness in complex graph scenarios. These findings approaches for QSAR modeling within environmental
reinforce the utility of GNNs for molecular property and pharmaceutical applications.
prediction, offering scalable, efficient, and accurate
alternatives to traditional methods in cheminformatics. Acknowledgments
The study demonstrates the effectiveness of GNN
architectures such as GIN, GAT, and GCN in molecular None.
property prediction. However, additional experiments
reveal that scalability and robustness remain critical Funding
challenges, especially for larger and noisier datasets.
Future work will focus on integrating scalable GNN None.
variants, such as GraphSAGE or Cluster-GCN, and
exploring domain-specific pretraining techniques to Conflict of interest
improve generalizability. Extending the evaluation
to datasets representing a wider range of molecular The authors declare no competing interests.
properties will further validate the applicability of these
models in real-world scenarios. Combining the strengths Author contributions
of multiple architectures, such as GIN’s expressiveness
and GAT’s attention mechanism, could lead to improved Conceptualization: Pravinkumar M. Sonsare
performance and interpretability. Developing methods Formal analysis: All authors
to interpret GNN predictions could enhance their Methodology: Pravinkumar M. Sonsare, Roshni
application in critical areas such as drug discovery Khedgaonkar
and toxicity prediction, where understanding decision- Writing – original draft: All authors
making processes is crucial. Leveraging pre-trained Writing – review & editing: Kavita Singh, Pratik
GNNs on large molecular datasets could improve Agrawal
performance on smaller datasets such as MUTAG
and accelerate training. Current models focus on two- Availability of data
dimensional molecular graphs. Incorporating three-
dimensional molecular geometry into GNNs could Data will be available on request from the corresponding
further improve predictions by capturing spatial features. author.
Extending these methods to real-world applications such
as virtual screening, material property prediction, and References
reaction prediction would validate their practical utility
and impact. By addressing these areas, future work can 1. Mayr A, Klambauer G, Unterthiner T, Hochreiter S.
further enhance the role of GNNs in molecular property DeepTox: Toxicity prediction using deep learning. Front

Volume 22 Issue 3 (2025) 100 doi: 10.36922/AJWEP025070041