A systematic study of key elements underlying molecular property prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Artificial intelligence (AI) is increasingly used in drug discovery for molecular property prediction.
• Molecular representation learning techniques have advanced, but their underlying mechanisms in prediction remain unclear.
• This hinders progress in AI-driven drug discovery.

Purpose of the Study:

• To extensively evaluate AI models for molecular property prediction using diverse representations.
• To investigate model performance across different dataset sizes, including low-data scenarios.
• To identify key factors influencing prediction accuracy and model limitations.

Main Methods:

• Trained 62,820 models across fixed representations, SMILES sequences, and molecular graphs.
• Evaluated models on MoleculeNet, opioid-related, and additional activity datasets.
• Assessed predictive power in both low-data and high-data regimes using varying dataset sizes.

Main Results:

• Representation learning models demonstrated limited performance on most molecular property prediction tasks.
• Several key factors were identified as significantly impacting prediction evaluation results.
• Activity cliffs were found to substantially influence model predictions.
• Model performance was strongly correlated with dataset size, with larger datasets being essential for representation learning.

Conclusions:

• Current representation learning models show limited efficacy in molecular property prediction.
• Dataset size is a critical determinant for the success of representation learning in this domain.
• Further research is needed to understand and overcome the limitations of AI models in drug discovery.