Interpretable Graph-Network-Based Machine Learning Models via Molecular Fragmentation.

Area of Science:

• Computational Chemistry
• Machine Learning
• Deep Learning

Background:

• Interpreting complex deep learning models in computational chemistry is challenging.
• Existing models often lack transparency in their predictions.
• Accurate prediction of molecular properties like thermochemistry is crucial.

Purpose of the Study:

• To develop an interpretable graph network for computational chemistry.
• To provide decomposed, fragment-wise predictions for enhanced model understanding.
• To improve the accuracy of thermochemical predictions using deep learning.

Main Methods:

• Proposed an interpretable graph network named FragGraph.
• Utilized Δ-learning to predict corrections to Density Functional Theory (DFT) calculations.
• Applied the model to predict atomization energies for the GDB9 dataset.

Main Results:

• FragGraph achieved G4(MP2)-quality thermochemistry predictions with <1 kJ mol^-1 accuracy.
• Fragment corrections quantitatively revealed deficiencies in B3LYP DFT.
• Node-wise predictions significantly outperformed previous global state vector models.

Conclusions:

• Interpretable graph networks like FragGraph enhance the utility of deep learning in computational chemistry.
• Fragment-wise contributions offer deeper insights into model behavior and chemical properties.
• Node-wise predictions demonstrate improved accuracy and generality for larger molecules.