MAPCliff-WMGR: Exploring Activity Cliffs in Molecular Activity Prediction Enhanced by Weighted Molecular Graph...

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Accurate molecular activity prediction is vital for drug discovery.
• Activity cliffs, where similar molecules have different activities, pose a significant challenge.

Purpose of the Study:

• Introduce MAPCliff-WMGR, a computational framework for predicting molecular activity, specifically addressing activity cliffs.
• Enhance the accuracy of molecular activity predictions in scenarios involving activity cliffs.

Main Methods:

• Utilize weighted molecular graphs and a core mGraphSNN_GAT module with model-specific adjustments.
• Employ an Independent Feature Mapping (IFM) module with sinusoidal transformations to address spectral bias in activity cliff data.
• Develop the MACE-R7 benchmark platform for evaluating prediction performance.

Main Results:

• MAPCliff-WMGR achieved an average RMSE of 0.677 for cliff molecules, outperforming baselines by 7.2%.
• The method showed an average improvement of 3.2% overall and 8.7% for cliff molecules on the MACE-R7 benchmark.
• Model interpretability revealed critical atoms contributing to activity cliffs via attention analysis and dimensionality reduction.

Conclusions:

• MAPCliff-WMGR effectively predicts molecular activity in the presence of activity cliffs.
• The framework demonstrates potential for virtual drug screening, as shown in a case study on ERα inhibitors for breast cancer.
• The study highlights the importance of specialized computational methods for tackling challenges like activity cliffs in drug discovery.