---
name: molecular-geometry-neural-network-potential-evaluation
description: Use when after RDKit has generated multiple conformations for a molecule in an SDF or XYZ format, and you need to reduce the conformational ensemble to a tractable size (by energy-based ranking) before submitting to expensive quantum-chemical calculations (e.g., QUICK).
license: CC-BY-4.0
metadata:
  edam_operation: http://edamontology.org/operation_2996
  edam_topics:
  - http://edamontology.org/topic_3314
  - http://edamontology.org/topic_0097
  tools:
  - ASE-ANI
  - RDKit
  - QUICK
  - Snakemake
  techniques:
  - ion-mobility-MS
  - NMR
derived_from:
- doi: 10.1021/jasms.1c00315
  title: POMICS
evidence_spans:
- 'ASE-ANI: For conformation filtering'
claims: []
provenance:
  collection: https://w3id.org/holobiomicslab/asb-skill/collection/metabolomics/v2
  assembled_by: scripts/collect_metabolomics_collection.py
  sources:
  - build: coll_pomics_cq
    doi: 10.1021/jasms.1c00315
    title: POMICS
  dedup_kept_from: coll_pomics_cq
schema_version: 0.2.0
attribution:
  generator: AgenticScienceBuilder
  original_doi: 10.1021/jasms.1c00315
  all_source_dois:
  - 10.1021/jasms.1c00315
  zenodo_doi: 10.5281/zenodo.20794027
  curators: []
  promoter: Louis-Félix Nothias
  sponsor: CNRS & Université Côte d'Azur
---

# molecular-geometry-neural-network-potential-evaluation

## Summary

Rank and filter molecular conformers generated by RDKit using ASE-ANI neural network potentials to compute single-point energies, retaining low-energy geometries for downstream quantum-chemical calculations. This step bridges classical conformation sampling with quantum-accurate geometry selection in automated CCS prediction workflows.

## When to use

After RDKit has generated multiple conformations for a molecule in an SDF or XYZ format, and you need to reduce the conformational ensemble to a tractable size (by energy-based ranking) before submitting to expensive quantum-chemical calculations (e.g., QUICK). Use this skill when your workflow must process many molecules and adducts with high automation and parallelization on HPC systems, where computational cost of evaluating all conformers is prohibitive.

## When NOT to use

- Input conformer set contains elements outside CHNO (ASE-ANI potentials ANI-1x and ANI-1ccx support only C, H, N, O; will fail or give nonsensical predictions on S, F, Cl, etc.)
- Input is already a pre-filtered or experimentally validated conformer set; re-ranking may discard valid geometries
- Workflow does not have NVIDIA GPU access with CUDA 9.2 and modern compute capability (ASE-ANI requires GPU and is tied to specific CUDA/Python 3.6 versions per README; now deprecated in favor of TorchANI)

## Inputs

- RDKit-generated conformer ensemble (SDF or XYZ format)
- Molecule specification (e.g., SMILES or structure file containing C, H, N, O atoms only)

## Outputs

- Ranked conformer subset (lowest-energy geometries in XYZ or format compatible with QUICK)
- Energy ranking metadata (per-conformer ANI energies in eV)

## How to apply

Load the conformer structures generated by RDKit in SDF or XYZ format into ASE-ANI. Initialize the ASE-ANI potential (ANI-1x or ANI-1ccx, which support CHNO elements) and compute single-point energy for each conformer using the neural network model. Rank all conformers by ascending energy. Apply a cutoff strategy: either retain the top N lowest-energy conformers (e.g., best 5–10) or select all conformers within a specified energy threshold relative to the global minimum (e.g., within 5 kcal/mol). Export the filtered conformer set in a format compatible with downstream quantum-chemical software (QUICK uses standard xyz or similar input). The rationale is that ASE-ANI provides fast, DFT-level accuracy geometry assessment at force-field computational cost, allowing rapid elimination of high-energy non-representative conformations before expensive quantum calculations.

## Related tools

- **RDKit** (generates initial conformer ensemble prior to energy-based filtering) — https://www.rdkit.org
- **ASE-ANI** (computes single-point neural network potential energies for each conformer and ranks them) — https://github.com/isayev/ASE_ANI
- **QUICK** (receives filtered conformer set for downstream quantum-chemical CCS calculations)
- **Snakemake** (orchestrates the conformer filtering step within the automated HPC-parallelized workflow) — github.com/DasSusanta/snakemake_ccs

## Examples

```
python examples/ani_quicktest.py  # to validate ASE-ANI installation; then iterate over RDKit conformer SDF and call ASE-ANI energy evaluation for ranking
```

## Evaluation signals

- Output conformer count is ≤ input conformer count and matches the specified cutoff strategy (top N or energy threshold)
- All output conformer energies are ranked in ascending order and all are lower than discarded conformers (if any retained)
- Output geometries are in the expected file format (XYZ or compatible with QUICK) and are parseable by downstream quantum software
- Energy values reported are in eV and are consistent with ANI-1x/ANI-1ccx model predictions (spot-check against known reference molecules or literature ANI energies)
- Downstream QUICK quantum calculations run without geometry parsing errors on the filtered conformer set

## Limitations

- ASE-ANI repository is deprecated and no longer supported; users should migrate to TorchANI for ongoing maintenance and broader element support
- ASE-ANI is restricted to CHNO elements; molecules containing S, F, Cl, Br, I, or other heteroatoms cannot be evaluated and will fail or yield invalid predictions
- Requires NVIDIA GPU with compute capability ≥5.0, CUDA 9.2, and Python 3.6; hardware and software stack constraints limit portability
- ANI neural network potentials, while DFT-accurate for equilibrium geometries, may not reliably rank highly strained or non-standard conformations; ranking is best-effort for chemically reasonable structures
- Energy-based filtering alone does not guarantee that retained conformers are biochemically relevant; post-hoc validation (e.g., against experimental NMR or crystallographic data) may be needed for critical applications

## Evidence

- [other] Initialize ASE-ANI potential and compute single-point energy for each conformer. Rank conformers by energy and select the lowest-energy subset (retain top N conformers or apply an energy threshold).: "Initialize ASE-ANI potential and compute single-point energy for each conformer. Rank conformers by energy and select the lowest-energy subset (retain top N conformers or apply an energy threshold)."
- [other] Load conformer structures generated by RDKit in a standard molecular format (e.g., SDF or XYZ). Export filtered conformer set in a format compatible with downstream quantum-chemical software (QUICK).: "Load conformer structures generated by RDKit in a standard molecular format (e.g., SDF or XYZ). Export filtered conformer set in a format compatible with downstream quantum-chemical software (QUICK)."
- [readme] Current ANI-1x and ANI-1ccx potentials provide predictions for the CHNO elements.: "Current ANI-1x and ANI-1ccx potentials provide predictions for the CHNO elements."
- [readme] This is a prototype interface for ANI-1x and ANI-1ccx neural network potentials for The Atomic Simulation Environment (ASE).: "This is a prototype interface for ANI-1x and ANI-1ccx neural network potentials for The Atomic Simulation Environment (ASE)."
- [readme] DEPRECATED and no longer supported, please use TorchANI implementation: "DEPRECATED and no longer supported, please use TorchANI implementation"