Chemprop v2: An Efficient, Modular Machine Learning Package for Chemical Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Deep Learning

Background:

• Accurate molecular property prediction is crucial for computational chemistry and molecular design.
• Deep learning models, such as directed message-passing neural networks (D-MPNNs), are effective for predicting molecular properties directly from molecular graphs.
• Existing tools like the original chemprop facilitate these predictions but lack Python API integration and modularity.

Purpose of the Study:

• To address the need for improved usability and modularity in computational chemistry workflows.
• To rewrite the chemprop software, enhancing its speed, extensibility, and user experience.
• To provide researchers with a more effective tool for computational molecular design.

Main Methods:

• A ground-up rewrite of the chemprop software was performed, focusing on Python API integration and enhanced modularity.
• Directed message-passing neural network (D-MPNN) architecture was maintained for end-to-end learning of molecular properties.
• Extensive benchmarking was conducted to compare performance against the original chemprop release.

Main Results:

• The rewritten chemprop (v2) demonstrates algorithmic parity with the original version.
• Significant improvements were observed in execution time (approximately 2x faster) and memory usage (approximately 3x lower).
• The new version offers enhanced scalability to multiple GPUs, enabling larger and more complex model training.

Conclusions:

• chemprop v2 preserves the predictive accuracy of its predecessor while significantly enhancing speed, modularity, and usability.
• The updated software empowers researchers with more effective tools for computational molecular design.
• New features, documentation, and tutorials improve the accessibility and application of deep learning in chemistry.