Flow-matching conformer generator that predicts 3D molecular shapes conditioned on environment: vacuum, protein pocket, solvent or crystal.
No providers recorded yet. Browse all providers
A drug molecule does not hold one shape. The same compound twists one way when it is isolated in vacuum, another when dissolved in water, another when packed into a crystal lattice, and another when it binds a protein pocket. Most neural conformer generators ignore this: they model the molecule in vacuum, or they train a dedicated model for one setting such as protein-ligand docking. That treats each environment as an unrelated task, even though the bonds, rings and torsional preferences underneath are the same chemistry everywhere. It also hurts most where data are thinnest, since crystal structures, solvated ensembles and ligand-bound poses are far scarcer than vacuum conformers.
EVA-Flow, from FAIR at Meta AI and New York University, treats conformer generation as one problem: learn the shared chemistry once, then let the environment bend it. The surroundings are encoded as an environment graph (protein-pocket atoms, a water proxy carrying solvation descriptors, or crystal space-group and lattice parameters) and compressed into an embedding. That embedding conditions a flow-matching decoder, which moves atoms from a graph-derived starting distribution to a final 3D conformation. Because the decoder is shared, what the model learns from abundant vacuum data carries over to the data-poor environments.
The work was posted to arXiv in October 2026 and accepted at Transactions on Machine Learning Research (TMLR) the same year.
EVA-Flow is a 12.1M-parameter model. A three-layer graph convolutional environment network feeds a 128-dimensional embedding to the encoder and to a 20-layer equivariant flow-matching decoder (160 hidden channels, 8 attention heads). It builds on the ETFlow equivariant flow-matching framework and adds environment conditioning, a variational latent and a multi-dataset training pipeline. Training jointly minimizes a flow-matching loss, two KL regularizers and a latent consistency loss. The pretraining corpus combines GEOM-DRUGS, GEOM-QM9, SPICE, CREST-relaxed SPICE and OE62, using randomly masked molecular subgraphs as context. Finetuning uses GEOM-DRUGS (vacuum, 272,641 entries), the Crystallography Open Database (packing, 20,765), PDBBind v2020 (docking, 9,940) and the Aquamarine implicit-solvent set (solvation, 1,276). Each 800-epoch stage takes roughly 530 A100 GPU-hours on eight A100s.
In the paper's four-conformer protocol, the pretrained and jointly finetuned model reaches Recall coverage of 0.684 on docking (2.0 Å), 0.949 on solvation (0.5 Å) and 0.644 on crystal packing (0.75 Å). Environment-specific models trained without pretraining score 0.553, 0.414 and 0.512, and a retrained ETFlow scores 0.391, 0.605 and 0.474. On vacuum GEOM-DRUGS the gains are small: EVA-Flow is comparable to retrained ETFlow and below the 242M-parameter MCF-L on recall, which the authors state plainly. For docking, EVA-Flow predicts the ligand's pocket-conditioned internal conformation, not its position or orientation in the protein, so it is not an end-to-end docking tool like DiffDock.
EVA-Flow targets the step before simulation and screening, where computational chemists need realistic 3D starting geometries. Pocket-conditioned ligand conformers can seed docking refinement and binding-free-energy workflows. Solvated conformer ensembles support solution-phase property calculations. Crystal-conditioned conformers bear on polymorph and crystal-structure prediction for pharmaceutical solids. Because one network covers all four settings, groups can generate environment-appropriate geometries without curating and maintaining a separate model for each. The authors recommend checking generated structures with physics-based calculations before downstream use.
EVA-Flow shows that conformer generation can be framed as transfer across physical environments rather than as a set of isolated benchmarks. The pay-off is concentrated in the data-poor docking, solvation and crystal settings. It is a small model, and its vacuum results only match existing generators. Its contribution is the conditioning design: one shared decoder steered by structured environment graphs, a pattern other generative chemistry models can adopt. The release is training, evaluation and data-preparation code under Meta's FAIR Noncommercial Research License, behind a gated request form; no trained checkpoint ships with it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.