NPstereo

Assigning the Stereochemistry of Natural Products by Machine Learning

NPstereo is a transformer-based language model that assigns stereochemistry (tetrahedral R/S, double bond E/Z) to natural products (NPs) from their SMILES representations. It was trained on referenced data from the open-access COCONUT database and can complete missing stereocenters in partially assigned or fully unassigned NP structures.

This repository contains:

• Data extraction and augmentation scripts
• Model training and evaluation code
• Utilities for running predictions on new molecules

Overview

Nature shows strong stereochemical regularities — L-amino acids in proteins, D-sugars in nucleotides. The idea of NPstereo is to investigate whether similar patterns exist in other NPs and whether they can be machine-learned.

Key features:

• Predicts stereochemistry directly from SMILES
• Supports full assignment and completion of partially assigned stereochemistry
• Achieves 80.2% per-stereocenter accuracy for full assignment and 85.9% for partial assignment on canonical SMILES
• Works across diverse NP classes (alkaloids, polyketides, lipids, terpenes, peptides, glycosides, etc.)

For details, see the manuscript!

Installation

Clone the repository:

git clone https://github.com/reymond-group/NPstereo.git
cd NPstereo

Create and activate a conda environment:

conda create -n npstereo python=3.10
conda activate npstereo

Install dependencies:

pip install -r requirements.txt

Data

Training and evaluation data are derived from the COCONUT database, filtered for entries with at least one DOI reference.
We include:

• Scripts to extract and process the dataset (data/)
• Data augmentation methods:
  • SMILES randomization
  • Partial stereochemistry removal
• Download links to processed datasets on Zenodo. This data is too heavy to include on GitHub, however following the code you should be able to reproduce the dataset from scratch.

Model Training

Models are trained using OpenNMT-py:

Example training command:

onmt_build_vocab -config configs/npstereo.yaml -n_sample -1
onmt_train -config configs/npstereo.yaml

Key architectures:

• C1 – baseline (canonical SMILES only)
• A2–A50 – canonical + randomized SMILES
• NPstereo – partially assigned → canonical
• M65 – combined augmentation (randomization + partial assignment)

Running Predictions

To predict stereochemistry for new molecules:

python predict.py   --model seed-0/npstereo/npstereo_step_100000.pt   --src input_smiles.txt   --output predictions.txt

Input: SMILES without full stereochemistry (absolute or partially assigned).
Output: Predicted isomeric SMILES with stereochemistry assigned.

For best performance:

• Use canonical SMILES as input for NPstereo
• For non-canonical SMILES, use models trained with randomization (A50 or M65)

Evaluation

We provide evaluation scripts in evaluation/ to reproduce the metrics reported in the manuscript:

• SMILES validity
• Full-assignment accuracy
• Per-stereocenter accuracy
• Class-wise breakdown by NP structural category

License

MIT

Contact