Taking a Respite from Representation Learning for Molecular Property Prediction

Artificial intelligence (AI) has been widely applied in drug discovery with a major task as molecular property prediction. Despite booming techniques in molecular representation learning, fundamentals underlying molecular property prediction haven't been carefully examined yet. In this study, we conducted a systematic evaluation on a collection of representative models using various molecular representations. In addition to the commonly used MoleculeNet benchmark datasets, we also assembled a suite of opioids-related datasets from ChEMBL and two additional activity datasets from literature. To interrogate the basic predictive power, we also assembled a series of descriptors datasets with varying sizes to evaluate the models' performance. In total, we trained 62,820 models, including 50,220 models on fixed representations, 4,200 models on SMILES sequences and 8,400 models on molecular graphs. We first conducted dataset profiling and highlighted the activity-cliffs issue in the opioids-related datasets. We then conducted rigorous model evaluation and addressed key questions therein. Furthermore, we examined inter-/intra-scaffold chemical space generalization and found that activity cliffs significantly can impact prediction performance. Based on extensive experimentation and rigorous comparison, representation learning models still show limited performance in molecular property prediction in most datasets. Finally, we explored into potential causes why representation learning models fail and highlighted the importance of dataset size. By taking this respite, we reflected on the fundamentals underlying molecular property prediction, the awareness of which can, hopefully, bring better AI techniques in this field.