PAIRMAP: A Unified Geometry-Aware Pairwise-Map Framework for Molecular Representation Learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular informatics

Background:

• Existing molecular representation learning methods often overlook crucial pairwise interactions and lack adaptability and physical grounding.
• These limitations hinder their effectiveness in complex drug discovery tasks.

Purpose of the Study:

• To introduce PAIRMAP, a unified framework designed to overcome the limitations of current methods in molecular representation learning.
• To enhance the prediction of molecular properties and binding affinities by integrating geometry-informed atom pair embeddings and physical constraints.

Main Methods:

• Developed PAIRMAP, a framework featuring learnable geometric encoding with task-adaptive basis functions.
• Incorporated geometric triangular attention with physical constraints (triangle inequality, cosine law) as learnable biases.
• Utilized hierarchical pair pooling to aggregate interactions across chemically relevant distance scales within an E(3)-equivariant framework.

Main Results:

• Achieved state-of-the-art performance across 20 benchmarks, including high ROC-AUC on MoleculeNet and improved accuracy on QM8 quantum benchmarks.
• Demonstrated superior binding affinity prediction (low RMSE) on the ATOM3D LBA benchmark, even with low protein sequence identity.
• Showcased generalization to structurally novel proteins and outperformed large-scale pretrained models without extensive pretraining.

Conclusions:

• PAIRMAP offers a powerful and versatile approach to molecular representation learning, effectively capturing pairwise interactions and physical constraints.
• The framework's ability to generalize and achieve state-of-the-art results highlights its potential to accelerate drug discovery and development.
• PAIRMAP provides a more efficient and physically grounded alternative to existing methods, including large pretrained models.