MolPIF: A Parameter Interpolation Flow Model for Molecule Generation

2025-07-18

AI Summary

Existing Bayesian Flow Networks (BFNs) for molecular generation suffer from rigid parameter-space modeling, limiting flexible design of distribution transformation paths and suboptimal efficiency. This work proposes Interpolative Flow in Parameter Space (IFPS), the first flow-based framework to embed a learnable interpolation mechanism directly into the flow architecture, enabling concise, efficient, and customizable distribution transformations. By bypassing traditional Bayesian inference constraints, IFPS supports end-to-end molecular structure generation and optimization. On key drug design tasks—including property optimization and scaffold hopping—IFPS significantly outperforms state-of-the-art BFNs as well as VAE- and GAN-based baselines, achieving superior performance in generation quality, structural diversity, and sampling efficiency. These results validate the effectiveness and generalizability of the parameter-space flow modeling paradigm for molecular generative AI.

Technology Category

• Machine Learning: Deep Neural Architectures and Foundation Models
• Search and Optimization: Sampling/Simulation-based Search
• Natural Language Processing: Generation

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

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.

Problem

• Limitations in flexible distribution transformation pathways for molecule generation
• Challenges adapting to diverse data distributions and task requirements
• Unexplored potential for simpler parameter-space-based molecular models

Innovation

• Parameter Interpolation Flow model for molecules
• Simpler parameter-space-based generative modeling
• Superior performance in structure-based drug design