MolPROP: Molecular Property prediction with multimodal language and graph fusion

Pretrained deep learning models self-supervised on large datasets of language, image, and graph representations are often fine-tuned on downstream tasks and have demonstrated remarkable adaptability in a variety of applications including chatbots, autonomous driving, and protein folding. Additional research aims to improve performance on downstream tasks by fusing high dimensional data representations across multiple modalities. In this work, we explore a novel fusion of a pretrained language model, ChemBERTa-2, with graph neural networks for the task of molecular property prediction. We benchmark the MolPROP suite of models on seven scaffold split MoleculeNet datasets and compare with state-of-the-art architectures. We find that (1) multimodal property prediction for small molecules can match or significantly outperform modern architectures on hydration free energy (FreeSolv), experimental water solubility (ESOL), lipophilicity (Lipo), and clinical toxicity tasks (ClinTox), (2) the MolPROP multimodal fusion is predominantly beneficial on regression tasks, (3) the ChemBERTa-2 masked language model pretraining task (MLM) outperformed multitask regression pretraining task (MTR) when fused with graph neural networks for multimodal property prediction, and (4) despite improvements from multimodal fusion on regression tasks MolPROP significantly underperforms on some classification tasks. MolPROP has been made available at https://github.com/merck/MolPROP . This work explores a novel multimodal fusion of learned language and graph representations of small molecules for the supervised task of molecular property prediction. The MolPROP suite of models demonstrates that language and graph fusion can significantly outperform modern architectures on several regression prediction tasks and also provides the opportunity to explore alternative fusion strategies on classification tasks for multimodal molecular property prediction.