MolPIF: A Parameter Interpolation Flow Model for Molecule Generation

TL;DR

MolPIF introduces a novel parameter interpolation flow model for molecule generation, demonstrating superior performance and flexibility in drug design tasks, advancing the use of parameter-space-based generative models in chemistry.

Contribution

This work presents a new parameter interpolation flow model (MolPIF) with theoretical foundations, enabling more flexible and efficient molecule generation compared to existing Bayesian flow networks.

Findings

MolPIF outperforms baselines across multiple metrics.

The model demonstrates superior flexibility in diverse chemical tasks.

Parameter-space-based modeling shows promise for molecular generation.

Abstract

Advances in deep learning for molecular generation show promise in accelerating drug discovery. Bayesian Flow Networks (BFNs) have recently shown impressive performance across diverse chemical tasks, with their success often ascribed to the paradigm of modeling in a low-variance parameter space. However, the Bayesian inference-based strategy imposes limitations on designing more flexible distribution transformation pathways, making it challenging to adapt to diverse data distributions and varied task requirements. Furthermore, the potential for simpler, more efficient parameter-space-based models is unexplored. To address this, we propose a novel Parameter Interpolation Flow model (named PIF) with detailed theoretical foundation, training, and inference procedures. We then develop MolPIF for structure-based drug design, demonstrating its superior performance across diverse metrics compared to baselines. This work validates the effectiveness of parameter-space-based generative modeling paradigm for molecules and offers new perspectives for model design.