Auto3D_pkg

What Auto3D does

Auto3D generates low-energy conformers from SMILES or SDF input. The pipeline enumerates tautomers and stereoisomers, embeds and optimizes them with a neural network potential such as AIMNet2, ANI2x, or ANI2xt, removes duplicates, and ranks what remains by energy. The result is the lowest-energy conformer per input molecule, each carrying its energy value, written to an SDF file in a timestamped directory. The tool is aimed at computational chemistry workflows where you need a reasonable 3D structure before further calculation. The README emphasizes that this happens in one command or one function call, which keeps it approachable for scripting. Because the engines are neural network potentials rather than classical force fields, the optimization can be faster while still producing physically meaningful geometries. For batches of up to 150 molecules you can skip the job directory and get RDKit molecules back directly, which is convenient inside a larger Python pipeline. The enumeration of tautomers and stereoisomers is worth calling out because it means the input string is expanded into the set of chemically plausible forms before geometry is built, so you are less likely to miss a low-energy variant the flat SMILES implied. That expansion step is what separates Auto3D from a bare embedder that takes the input at face value.

Installation and requirements

Auto3D requires Python 3.11 or newer and PyTorch 2.8 or newer. For GPU acceleration you install a CUDA-enabled PyTorch build first. AIMNet2 weights download to a cache directory under your home folder on first use, so the first run needs network access even though later runs can work offline. The README notes a conda wrinkle: installing via conda-forge pulls version 2.3.0 rather than 3.1.0, because a core dependency named aimnet since 3.0.0 is not yet a conda package, nor is its own dependency. Auto3D works fine inside a conda environment, but you should install the package with pip to get the current release, and an installation file sets up the supported combination. Documentation is hosted on Read the Docs, and the project links a guide for building the conda package if you need it. The Python 3.11 and PyTorch 2.8 floors mean you should check your environment before installing, since older interpreters will not import the package. The network requirement on first run is easy to forget in an air-gapped lab, so plan to prime the cache on a connected machine if your compute nodes are offline. Those two constraints, the version floor and the first-run download, are the parts most likely to trip up a fresh install.

Command line and API

On the command line, exactly one of two window flags is required, and the tool uses the GPU by default. Requesting the GPU with no visible CUDA device is a fatal error rather than a fallback, so on a CPU-only machine you pass a no-gpu flag. In Python, the main function returns a WorkflowResult that subclasses string and also carries the molecule count, conformer count, and any failures, so you can use the return value directly as the output path. The quick start shows both paths, and the what-you-get section describes the run output in more detail. Beyond conformer generation, the README points to engines and key parameters sections plus an upgrading note for users moving from the 2.x series. The project is MIT licensed, which lets you use it in academic or commercial pipelines. For cheminformatics teams, Auto3D fills the gap between a flat SMILES string and a geometry ready for downstream modeling without hand-tuning a force field. The fatal-on-missing-GPU behavior is intentional: the authors prefer a loud failure over a silent CPU fallback that would quietly change your run time and possibly your results. That strictness is a small but welcome signal that the tool treats reproducibility seriously enough to refuse to guess what you meant.

Editorial conclusion

Auto3D is published under the MIT license, requires Python 3.11 or newer, and depends on PyTorch 2.8 or newer.