Analyzing Learned Molecular Representations for Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Neural networks offer algorithmic solutions for molecular property prediction.
• Two main approaches include neural networks with fixed descriptors and graph convolutional neural networks (GCNNs).
• The superiority of these methods for generalizing to new chemical spaces and their industrial applicability remain unclear.

Purpose of the Study:

• To benchmark different molecular property prediction models.
• To compare fixed descriptor-based neural networks with GCNNs.
• To evaluate model performance on diverse public and proprietary industrial datasets.

Main Methods:

• Extensive benchmarking of models on 19 public and 16 proprietary datasets.
• Development and evaluation of a novel graph convolutional neural network architecture.
• Comparison against models using fixed molecular descriptors and prior graph neural architectures.

Main Results:

• The proposed GCNN model consistently matches or outperforms models using fixed molecular descriptors.
• The GCNN model also surpasses previous graph neural architectures on both public and proprietary datasets.
• While not yet at experimental reproducibility levels, significant improvements over current industrial models were observed.

Conclusions:

• The developed graph convolutional model shows strong performance for molecular property prediction.
• This model offers a significant advancement over existing methods in industrial research settings.
• Further research may bridge the gap between computational predictions and experimental reproducibility.