MaskMol: knowledge-guided molecular image pre-training framework for activity cliffs with pixel masking.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Activity cliffs, pairs of similar molecules with differing potency, challenge current machine learning models.
• High molecular similarity can cause model representation collapse, hindering accurate predictions.

Purpose of the Study:

• To develop a novel self-supervised learning framework for molecular image representation.
• To improve the accurate prediction of compound potency and activity cliff identification.
• To enhance virtual screening and drug discovery processes.

Main Methods:

• Developed MaskMol, a knowledge-guided molecular image self-supervised learning framework.
• Employed pixel masking tasks to extract fine-grained information from molecular images.
• Incorporated multi-level molecular knowledge (atoms, bonds, substructures) into the learning process.

Main Results:

• MaskMol accurately learns molecular image representations, outperforming 25 state-of-the-art methods in activity cliff estimation and potency prediction across 20 targets.
• Image-based approaches, like MaskMol, effectively capture distinctions missed by graph-based methods.
• Identified candidate EP4 inhibitors for tumor treatment with high biological interpretability.

Conclusions:

• MaskMol advances molecular image representation learning and virtual screening for drug discovery.
• The study highlights the importance of addressing activity cliffs in structure-activity relationship (SAR) analysis.
• Provides new insights into identifying subtle structural changes impacting molecular potency.