ChemTorch: A Deep Learning Framework for Benchmarking and Developing Chemical Reaction Property Prediction Models.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Chemical reaction modeling

Background:

• Accurate chemical reaction modeling is crucial but computationally expensive with quantum mechanics.
• Deep learning offers a faster alternative, but a fragmented software ecosystem impedes progress.
• Reproducibility and fair comparison are hindered by the lack of standardized tools.

Purpose of the Study:

• To introduce ChemTorch, an open-source framework for streamlining deep learning in chemical reaction modeling.
• To provide modular pipelines, standardized configuration, and data splitters for efficient development and evaluation.
• To establish a foundation for community-driven method development and reproducible benchmarking.

Main Methods:

• Development of ChemTorch, an open-source deep learning framework.
• Implementation of modular pipelines and standardized configurations.
• Comparison of four modalities (fingerprint, sequence, graph, 3D) for barrier-height prediction using RDB7 dataset.

Main Results:

• ChemTorch facilitates streamlined model development, hyperparameter tuning, and benchmarking.
• Structurally informed models (graph and 3D) demonstrated clear advantages in barrier-height prediction.
• Significant performance degradation was observed for models under out-of-distribution conditions.

Conclusions:

• ChemTorch enhances reproducibility and facilitates fair comparison in chemical reaction modeling research.
• Rigorous benchmarking is essential, particularly highlighting the need for out-of-distribution evaluation.
• Future work includes expanding ChemTorch for broader community use and developing unified benchmarks.