N-Alkylation / N-Arylation Regioselectivity Predictor

A machine-learning workflow for predicting the most likely nitrogen reaction centre in N-alkylation and N-arylation reactions of poly-nitrogen nucleophiles.

Developed by Alice Oram, AstraZeneca. With contributions from Christoph Bauer, Isabel Arranz and Per-Ola Norrby

---

Overview

Given a molecule SMILES as input, the predictor:

1. Validates the input (reactive-nitrogen count, molecular weight, salt, metal, charge).
2. Generates 3D coordinates with Corina.
3. Converts coordinates to xyz format with OpenBabel.
4. Deprotonates (neutral molecules) or optimises (anionic molecules) the structure with CREST / xTB.
5. Calculates ionisation potentials for the anionic and neutral forms with Multiwfn.
6. Calculates atomic descriptors (proximity shells, partial charges, polarizability) with kallisto.
7. Combines the descriptors with RDKit fingerprints and feeds them into a trained XGBoost classifier.
8. Outputs the predicted reactive nitrogen and the probability for each nitrogen centre, and saves a labelled molecule image (N_alkylation_probability.png).

---

Requirements

Python packages

Install the Python dependencies with:

pip install -r requirements.txt

External tools

The workflow requires the following external executables to be available on your system (see Configuration for how to point the code at them):

Tool	Version tested	Purpose
Corina	3.60	SMILES to 3D mol2 coordinate generation
OpenBabel	3.1.1	mol2 to xyz format conversion
CREST	latest	Deprotonation / reprotonation
xTB	latest	Geometry optimisation and molden files
Multiwfn	3.8 (no-GUI)	Ionisation potential calculation

---

Installation

git clone https://github.com/<your-org>/n-alkylation-arylation.git
cd n-alkylation-arylation
pip install -r requirements.txt

---

Configuration

All configurable paths are set in config.py in the repository root. Edit this file before running the predictor:

# Root directory where per-run sub-folders will be created
WORKING_DIR = "/path/to/your/working/directory"

# Paths to external tool executables
MULTIWFN_EXEC = "/path/to/Multiwfn_noGUI"
CREST_EXEC    = "/path/to/crest"
XTB_EXEC      = "xtb"          # or full path if not on PATH
CORINA_EXEC   = "/path/to/corina"
OBABEL_EXEC   = "obabel"       # or full path if not on PATH

Alternatively, each executable path can be overridden with an environment variable (e.g. export MULTIWFN_EXEC=/path/to/Multiwfn_noGUI).

The trained XGBoost model (model/XGBoost_fp.pkl) and the Multiwfn input file (util/multiwfn_IP.in) are bundled with the repository and require no additional configuration.

---

Usage

python main.py

The script will prompt for:

• Molecule SMILES - the input nitrogen nucleophile.
• Molecule name - used as the run folder name and file prefix.

Example input:

Molecule SMILES: Cc1ccc(NC(=O)c2ccccc2)cc1N
Name the molecule: example_molecule

Output files are written to <WORKING_DIR>/<molecule_name>/:

File	Description
<name>.smi	Input SMILES file for Corina
<name>.mol2	3D structure (Corina)
<name>.xyz	xyz coordinates (OpenBabel)
deprotonated.xyz / xtbopt.xyz	Optimised anionic structure
protonated.xyz	Neutral (re-protonated) structure
multiwfn_IP_Deprot.out	Multiwfn output for the anionic form
multiwfn_IP_Reprot.out	Multiwfn output for the neutral form
all_features.csv	Full feature matrix used for prediction
N_alkylation_probability.png	Labelled molecule image with per-nitrogen probabilities

---

Repository structure

n-alkylation-arylation/
├── config.py               # User-editable configuration (paths, executables)
├── main.py                 # Main workflow script
├── requirements.txt        # Python dependencies
├── model/
│   ├── XGBoost_fp.pkl      # Trained XGBoost classifier
│   └── README.md           # Model card
├── util/
│   ├── cheminformatics.py  # Input validation and visualisation
│   ├── corina_util.py      # Corina wrapper
│   ├── fp_util.py          # RDKit fingerprint generation
│   ├── kallisto_util.py    # Kallisto descriptor calculations
│   ├── multiwfn_IP.in      # Multiwfn input commands
│   ├── multiwfn_util.py    # Multiwfn wrapper and output parser
│   ├── obabel_util.py      # OpenBabel wrapper
│   └── xtb_util.py         # xTB / CREST wrapper
└── LICENSE

---

Known limitations and TODOs

• Nitrogen indices are taken from the deprotonated structure. If the atom ordering changes during reprotonation the merge of IP_anion and IP_neutral on nitrogen_index may fail. The function check_atom_index_change prints a warning when this occurs; a full fix is planned.
• The workflow currently stops with an AssertionError / ValueError if any validation check fails. Graceful error recovery is planned.
• The model was trained on molecules with MW < 500 g/mol and charge 0 or -1. Predictions outside this domain should be treated with caution.

---

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.