espaloma

espaloma generates molecular mechanics (MM) force field parameters using graph neural networks in an end-to-end differentiable framework to enable parameter optimization against quantum chemical and physical property data for biopolymers and small molecules.

Key Features:

• Graph Neural Networks: Uses graph neural networks to perceive chemical environments and generate continuous atom embeddings.
• End-to-End Differentiability: Implements a modular, end-to-end differentiable parameter assignment pipeline with gradients with respect to model parameters.
• Reproduction and Extension of Force Fields: Can reproduce legacy atom types and extend or construct molecular mechanics force fields directly from quantum chemical calculations.
• Arbitrary Loss Functions: Supports training with arbitrary loss functions to fit quantum chemical and physical property data.
• Parameter Prediction: Predicts valence and nonbonded parameters applicable to both biopolymers and small molecules.
• Partial Charge Modeling: Fits partial charge models simultaneously, producing high-quality atomic charges with faster performance than existing best practices.

Scientific Applications:

• Biomolecular Modeling and Drug Discovery: Provides parameters for applications ranging from rapid virtual screening to detailed free energy calculations.
• High-Fidelity Force Fields: Enables construction of self-consistent, high-fidelity force fields applicable to biopolymers and small molecules by training against target data.
• Efficient Charge Model Adaptation: Supports efficient adaptation and generation of partial charge models with minimal inaccuracy.

Methodology:

Uses graph neural networks to produce continuous atom embeddings and smooth neural functions across stages, applies end-to-end differentiable parameter assignment, and trains models with arbitrary loss functions including simultaneous fitting of partial charge models.