AvgFlow is now available on Vecura

What is AvgFlow?

AvgFlow is a JAX-based diffusion-transformer model that generates 3D atomic coordinates (conformers) for small organic molecules directly from SMILES strings. The model uses a novel SO(3)-Averaged Flow-Matching objective that analytically computes probability flow from noise to all rotations of training data, eliminating the need for rotational data augmentation and improving both training efficiency and sample quality.

It helps users generate physically plausible 3D geometries for drug-like molecules, which is a critical step in computational drug discovery pipelines. It is especially useful for molecular docking, scoring, molecular dynamics seeding, and property prediction tasks that require realistic 3D structures rather than flat 2D graphs.

What can users do with AvgFlow on Vecura?

With AvgFlow on Vecura, users can:

• Generate multiple 3D conformers (up to any number) for drug-like molecules from simple SMILES input

• Choose between different model variants optimized for speed (single-step distillation) or quality (multi-step full models)

• Obtain ready-to-use SDF files with explicit hydrogens and corrected stereochemistry for immediate use in downstream applications

• Process molecules with up to 200 atoms for high-throughput virtual screening campaigns

What the output means

The output provides a multi-record SDF file containing all generated 3D conformers, with each conformer as a separate record including explicit hydrogens and post-hoc stereochemistry correction applied. The file also returns the count of conformers generated and echoes back the input SMILES for traceability.

This output should be used to support scientific decision making. It does not replace experimental validation. The model outputs raw 3D geometries without energy rankings or stability predictions, so users should apply appropriate filtering or energy minimization for their specific use cases.

Why this matters

Accurate 3D conformer generation is fundamental to structure-based drug design, yet traditional methods like distance geometry or molecular dynamics can be computationally expensive or produce unrealistic geometries. AvgFlow's rotation-aware flow-matching approach represents a significant advance in generative modeling for molecular science, providing high-quality conformer ensembles that capture the true conformational landscape of drug-like molecules.

The availability of multiple checkpoint variants—from full multi-step models for maximum fidelity to single-step distilled versions for high-throughput screening—makes this tool adaptable to diverse workflows, from detailed lead optimization studies to large-scale virtual screening campaigns across millions of compounds.

• Developed by: NVIDIA Corporation (BioNeMo team)

• Source: NVIDIA NGC Catalog, GitHub repository (NVIDIA-BioNeMo/avgflow)

• Reference: Original paper (ICML 2025, arXiv:2507.09785)