What is it about?
Chemistry studies of reaction mechanisms rely on a calculation called the nudged elastic band (NEB). The NEB connects a starting structure to a product and pins down the transition state: the highest-energy configuration on the path. In practice, running an NEB means several setup steps first -- minimizing the endpoints, aligning the structures, and generating an initial guess for the path. Researchers usually glue these steps together with one-off scripts, and the scripts drift between groups, machines, and versions of the same tool. Results are hard to reproduce. We built a Snakemake workflow that handles every step end to end. It pulls the PET-MAD machine-learned potential for force evaluations, hands off to the eOn saddle point search code for the NEB itself, and resolves every software dependency from conda-forge so the same workflow runs identically on a laptop or a cluster. The dependency graph is explicit: each stage has defined inputs and outputs, and Snakemake re-runs only what changed. We validated the workflow on the HCN to HNC isomerization and recovered the correct single-barrier profile and product energy without manual intervention. Code and dependency lock are public.
Featured Image
Why is it important?
Computational chemistry benchmarks often report numbers without the workflow that produced them. Two groups running the "same" NEB rarely start from the same endpoint minima, the same path initialization, or the same version of the underlying force code. The gap between "we used NEB" and "here is the pipeline that produces these numbers" is exactly where reproducibility fails. A Snakemake pipeline closes that gap. The graph is the specification: anyone can run it, get the same outputs, and inspect what changed. Coupling to a machine-learned potential (PET-MAD) gives the pipeline small-molecule throughput without sacrificing accuracy for the regimes it targets. The structure extends to other ML potentials and other saddle search codes, so the approach is not tied to any single tool.
Perspectives
I kept finding myself re-running the same four or five commands before every NEB I wanted to study. Over the past year, glue scripts for endpoint minimization and path generation had drifted into different shapes across different projects, and I could not reproduce my own earlier results on a new machine. Snakemake forces you to be explicit: every input, every output, every version pinned. Coupling this to eOn (which I maintain) and to PET-MAD gave me a setup where the whole pipeline is one =snakemake --configfile= away. The HCN to HNC validation was the first system I ran end to end without touching intermediate files. Anyone who wants to benchmark saddle search methods against an ML potential, or regenerate NEB data for a paper, should be able to clone and run.
Rohit Goswami
University of Iceland
Read the Original
This page is a summary of: Reproducible orchestration of best practices for reaction path optimization with the nudged elastic band, MethodsX, June 2026, Elsevier,
DOI: 10.1016/j.mex.2026.103899.
You can read the full text:
Contributors
The following have contributed to this page
