When Do Models Win? A Learning Curve Benchmark for Molecular Property Prediction in Low-Data Regimes
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
In many molecular property prediction tasks, such as drug discovery and OLED material design, labeled data are often limited to only tens to thousands of samples due to high experimental costs. Despite the rapid development of machine learning models, it remains unclear which model classes are most effective under such low-data regimes and how their performance evolves with increasing data. In this work, we present a comprehensive learning curve benchmark across traditional machine learning methods, graph neural networks (GNNs), transformer-based molecular language models, and hybrid architectures. We systematically vary the training data size from 50 to 3,000 samples and evaluate performance on QM9, ESOL, Lipophilicity, and BACE datasets using scaffold splits with repeated trials for statistical robustness. We further propose two hybrid models combining GCN and ChemBERTa with different fusion strategies-concatenation (GTCA-Cat) and cross-attention (GTCA-CA)-and analyze their data efficiency through ablation studies. Our results reveal that (1) hybrid models substantially improve over single-modality GCN in low-data regimes for physicochemical property targets (e.g., 32% RMSE reduction on ESOL at N=50), (2) simple concatenation fusion generally outperforms or matches cross-attention due to reduced overfitting and improved stability, and (3) clear crossover points exist where traditional kernel-based methods outperform deep learning models depending on the task type. Importantly, we identify regimedependent model optimality: deep learning models dominate continuous physicochemical property prediction, while kernel-based methods can be more effective for bioactivity prediction across most data scales. Our findings provide actionable guidelines for model selection under data scarcity and highlight the importance of learning curve analysis as a principled tool for understanding model behavior.
Fields of Study & Subjects
Computational Drug Discovery Methods
Machine Learning in Materials Science
Advanced Chemical Physics Studies
Publication details
- Journal / source
- ChemRxiv
- Publication date
- 25 March 2026
- Language
- EN
- Document type
- Preprint