A comparison of probabilistic generative frameworks for molecular simulations

Area of Science:

• Molecular Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Generative artificial intelligence (AI) is increasingly utilized in molecular science.
• A lack of numerical experiments benchmarks the performance of probabilistic generative models on molecular data.
• Understanding model performance is crucial for selecting appropriate tools for molecular tasks.

Purpose of the Study:

• To introduce and categorize probabilistic generative models for molecular science.
• To benchmark the performance of representative flow-based and diffusion models.
• To guide the selection of generative models for diverse molecular applications.

Main Methods:

• Categorization of generative models into flow-based and diffusion models.
• Selection and implementation of three representative models: neural spline flows, conditional flow matching, and denoising diffusion probabilistic models.
• Evaluation of model accuracy, computational cost, and generation speed using tunable datasets, including a Gaussian mixture model and peptide dihedral angle distributions.

Main Results:

• No single generative model framework is universally superior across all molecular data types and complexities.
• Neural spline flows demonstrate superior performance in capturing mode asymmetry in low-dimensional data.
• Conditional flow matching excels in high-dimensional, low-complexity data scenarios.
• Denoising diffusion probabilistic models show the best performance for low-dimensional, high-complexity data.

Conclusions:

• The choice of generative model in molecular science depends heavily on data characteristics such as dimensionality, complexity, and modal asymmetry.
• The study provides a valuable taxonomy and empirical evidence to aid researchers in selecting appropriate probabilistic generative models for their specific molecular tasks.
• Further research can expand this benchmarking to a wider array of generative models and molecular datasets.