![[2602.20573] Benchmarking GNN Models on Molecular Regression Tasks with CKA-Based Representation Analysis](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20573] Benchmarking GNN Models on Molecular Regression Tasks with CKA-Based Representation Analysis
Summary
This paper benchmarks various Graph Neural Network (GNN) models for molecular regression tasks, highlighting the effectiveness of a hierarchical fusion framework over standalone GNNs.
Why It Matters
Understanding the performance of GNNs in molecular property prediction is crucial for advancements in computational chemistry and drug discovery. This study provides insights into the architectures' effectiveness and their representational similarities, which can guide future research and applications in the field.
Key Takeaways
- The study benchmarks four GNN architectures across diverse datasets.
- A hierarchical fusion framework (GNN+FP) shows over 7% RMSE improvement compared to standalone GNNs.
- CKA analysis reveals independent latent spaces for GNN and fingerprint embeddings.
- Isotopic models like GCN, GraphSAGE, and GIN show high convergence in representation.
- Understanding GNN performance on smaller datasets is essential for practical applications.
Computer Science > Machine Learning arXiv:2602.20573 (cs) [Submitted on 24 Feb 2026] Title:Benchmarking GNN Models on Molecular Regression Tasks with CKA-Based Representation Analysis Authors:Rajan, Ishaan Gupta View a PDF of the paper titled Benchmarking GNN Models on Molecular Regression Tasks with CKA-Based Representation Analysis, by Rajan and 1 other authors View PDF HTML (experimental) Abstract:Molecules are commonly represented as SMILES strings, which can be readily converted to fixed-size molecular fingerprints. These fingerprints serve as feature vectors to train ML/DL models for molecular property prediction tasks in the field of computational chemistry, drug discovery, biochemistry, and materials science. Recent research has demonstrated that SMILES can be used to construct molecular graphs where atoms are nodes ($V$) and bonds are edges ($E$). These graphs can subsequently be used to train geometric DL models like GNN. GNN learns the inherent structural relationships within a molecule rather than depending on fixed-size fingerprints. Although GNN are powerful aggregators, their efficacy on smaller datasets and inductive biases across different architectures is less studied. In our present study, we performed a systematic benchmarking of four different GNN architectures across a diverse domain of datasets (physical chemistry, biological, and analytical). Additionally, we have also implemented a hierarchical fusion (GNN+FP) framework for target prediction. We ob...
