EVA-Flow: Environment-Aware Flow Matching for Unified 3D Molecular Conformation Generation

ICLR 2026 Conference Submission13268 Authors

Keywords: Flow matching, Conditional generative modeling, Geometric deep learning, 3D molecular conformation generation

TL;DR: EVA-Flow is a unified VAE + flow-matching model that generates environment-aware 3D molecular conformations across vacuum, protein–ligand docking, solvation, and crystal packing.

Abstract: Predicting the 3D geometry of molecules is central to applications in drug discovery, materials design, and molecular modeling. However, molecular geometry can change dramatically across environments (e.g., crystal lattice versus protein binding pocket). Existing generative approaches are typically environment-agnostic or require separate models for each environment, which limits generalization. We introduce EVA-Flow, a unified framework for environment-aware conformation generation. EVA-Flow combines a variational autoencoder with a flow matching decoder and incorporates environment information through a learned embedding. Across four environments including vacuum, protein-ligand docking, solvation, and crystal packing, EVA-Flow substantially improves generation accuracy through pretraining and unification. Analysis of shared molecules that appear in multiple environments further shows that EVA-Flow generates distinct, environment-specific conformations rather than memorizing a single geometry.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 13268