MolPROP: Molecular Property prediction with multimodal language and graph fusion

Area of Science:

• Computational chemistry and cheminformatics
• Deep learning and artificial intelligence
• Drug discovery and development

Background:

• Pretrained deep learning models excel in various domains by fine-tuning on downstream tasks.
• Multimodal data fusion aims to enhance performance by integrating diverse data representations.
• Molecular property prediction is crucial for drug discovery and materials science.

Purpose of the Study:

• To introduce and evaluate MolPROP, a novel multimodal fusion model for molecular property prediction.
• To benchmark MolPROP against state-of-the-art architectures on diverse molecular datasets.
• To investigate the impact of different pretraining strategies on multimodal fusion performance.

Main Methods:

• Fusion of a pretrained language model (ChemBERTa-2) with graph neural networks.
• Benchmarking on the MolPROP suite across seven scaffold-split MoleculeNet datasets.
• Comparison of masked language model (MLM) and multitask regression (MTR) pretraining tasks.

Main Results:

• MolPROP matches or outperforms modern architectures on regression tasks like hydration free energy, solubility, lipophilicity, and toxicity.
• Multimodal fusion primarily benefits regression tasks.
• ChemBERTa-2's MLM pretraining superior to MTR when fused with graph neural networks.
• MolPROP underperforms on certain classification tasks despite regression improvements.

Conclusions:

• Novel multimodal fusion of language and graph representations enhances small molecule property prediction, particularly for regression tasks.
• The choice of pretraining strategy (MLM vs. MTR) impacts multimodal fusion effectiveness.
• Further exploration of fusion strategies is warranted for classification tasks in multimodal molecular property prediction.