Reproduce “Implicit Neural Representations of Molecular Vector-Valued Functions” on OpenResearch

About this paper

Jirka Lhotka, Daniel Probst

Researchers at EPFL and Wageningen University & Research developed "molecular neural fields," a framework that represents molecules as continuous, vector-valued functions using implicit neural networks. This method provides resolution-independent representations that capture both external and internal molecular features, enabling super-resolution reconstruction and interpolable latent spaces for molecular conformers and structures.

Key insights

• Implicit Neural Representations (INRs), a technique from computer vision, can be successfully adapted to create continuous and resolution-independent representations of 3D molecular structures.
• Representing molecules as vector-valued functions allows for the capture of both external and internal molecular features, addressing limitations of surface-only methods.
• The latent space learned from these molecular neural fields encodes chemically relevant information, enabling smooth interpolation between molecular conformers and distinct molecules.

Summary by alphaXiv.