Structural Bias in Three-Dimensional Autoregressive Generative Machine Learning of Organic Molecules

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Generative machine learning models are increasingly used for designing novel molecules and materials.
• Three-dimensional structure generation is crucial for quantum chemistry workflows and property prediction.
• Model performance evaluation traditionally focuses on novelty, validity, and uniqueness, but chemical space reproduction is also vital.

Purpose of the Study:

• To investigate the G-SchNet autoregressive model's ability to reproduce chemical space and property distributions of training data.
• To assess the impact of G-SchNet's generation bias on molecular properties and chemical space coverage.
• To explore methods for mitigating generation bias, such as functional group constraints and composite datasets.

Main Methods:

• Analysis of elemental composition, size, bond-length, functional group, and chemical space distributions for training and generated molecules.
• Principal Component Analysis (PCA) of chemical space to identify generation biases.
• Application of decision tree models to discriminate between training and generated data and reveal chemical differences.

Main Results:

• G-SchNet exhibits a generation bias, producing molecules that are less saturated and contain more heteroatoms, irrespective of hyperparameters or training data distribution.
• Purely aliphatic molecules are largely absent in the generated set.
• Decision tree models successfully identified the generation bias and highlighted key chemical differences affecting electronic properties like the HOMO-LUMO gap.

Conclusions:

• The G-SchNet model demonstrates a significant bias in generating molecules, leading to an inaccurate representation of the training data's chemical space and properties.
• Functional group constraints and composite datasets can partially alleviate the observed generation bias.
• Understanding and addressing generation bias is critical for accurately designing functional molecules with desired electronic properties using generative models.