ML Workflow for TS-Geometry Prediction

This repository contains the Python scripts used to reproduce the machine-learning analyses for Reaction B and Reaction C reported in the manuscript:

Hammett-Guided Machine Learning for Transition State Geometry Prediction in Nitrosobenzene Additions

Jingxing Cheng,^a Pauline Bianchi,^a and Jean-Christophe M. Monbaliu^a,b,*

Affiliations:
^a Center for Integrated Technology and Organic Synthesis (CiTOS), MolSys Research Unit, University of Liège B6a, Room 3/19, Allée du Six Août 13, 4000 Liège (Sart Tilman), Belgium Homepage: www.citos.uliege.be E-mail: jc.monbaliu@uliege.be
^b WEL Research Institute, Avenue Pasteur 6, B-1300 Wavre, Belgium

Each regression model uses the corresponding Reaction A transition-state geometry together with a single numerical descriptor to predict one geometric parameter—bond length, angle, or dihedral—for Reaction B or Reaction C.

The workflow separates descriptor screening from model development. Descriptor-screening scripts compare a broad candidate set of Hammett-related and DFT-derived substituent and rank them by LOOCV Q². The machine-learning scripts then build the final Hammett-based models (σp+, σp, σp− by default), export fitted equations, and support downstream external validation.

The machine-learning workflow and implementation were developed by Jingxing Cheng under the supervision of Jean-Christophe M. Monbaliu.

E-mail: jc.monbaliu@uliege.be

Repository contents

TS_geometry_ML_prediction/
├── README.md
└──TS_geometry_ML_prediction/
       ├── datasets/
       │   ├── reaction_B_dataset.csv
       │   ├── reaction_B_external-set.csv
       │   ├── reaction_C_dataset.csv
       │   └── reaction_C_external-set.csv
       ├── reaction_B_descriptor_screening.py
       ├── reaction_B_machine_learning_model.py
       ├── reaction_B_method_comparison.py
       ├── reaction_B_external_test_prediction.py
       ├── reaction_C_descriptor_screening.py
       ├── reaction_C_machine_learning_model.py
       ├── reaction_C_method_comparison.py
       ├── reaction_C_external_test_prediction.py
       ├── results/
       ├── reaction_B_descriptor_screening
       ├── reaction_B_ml_results
       ├── reaction_B_method_comparison
       ├── reaction_B_external_test_predictions
       ├── reaction_C_descriptor_screening
       ├── reaction_C_ml_results
       ├── reaction_C_method_comparison
       └── reaction_C_external_test_predictions

How to cite this material

DOI: 10.5281/zenodo.21376455

Scripts

• reaction_B_descriptor_screening.py: compares candidate Reaction B substituent descriptors, including Hammett-related parameters (for example σp+, σp, σp−) and DFT-derived descriptors (for example EHOMO, ELUMO, η, ω, N, f-N, NN, ωN, B1, B5, L, and v%), and ranks them by bond-length LOOCV Q².
• reaction_B_machine_learning_model.py: builds the final Hammett-based Reaction B models (σp+, σp, σp− by default). Unlike the descriptor-screening script, it focuses on Hammett constants, selects the best descriptor per target, and exports LOOCV metrics, model equations, predictions, parity plots, and software versions for downstream use.
• reaction_B_method_comparison.py: compares Ridge, Lasso, Elastic Net, and Bayesian Ridge regression for Reaction B.
• reaction_B_external_test_prediction.py: applies fixed Reaction B equations to an external test set without refitting.
• reaction_C_descriptor_screening.py: compares candidate Reaction C nucleophile descriptors, including Hammett-related parameters (for example σp+, σp, σp−) and DFT-derived descriptors (for example EHOMO, ELUMO, η, ω, N, B1, B5, L, v%, f-c, Nc, and ωc), and ranks them by dihedral LOOCV Q².
• reaction_C_machine_learning_model.py: builds the final Hammett-based Reaction C models (σp+, σp, σp− by default). Unlike the descriptor-screening script, it focuses on Hammett constants, selects the best descriptor per target, and exports LOOCV metrics, model equations, predictions, parity plots, and software versions for downstream use.
• reaction_C_method_comparison.py: compares Ridge, Lasso, Elastic Net, and Bayesian Ridge regression for Reaction C.
• reaction_C_external_test_prediction.py: applies fixed Reaction C equations to an external test set without refitting.

Environment

Python 3.9 or later is recommended.

Required packages:

numpy
pandas
matplotlib
scikit-learn

Create and activate an isolated environment:

python -m venv .venv

On macOS or Linux:

source .venv/bin/activate

On Windows PowerShell:

.venv\Scripts\Activate.ps1

Install the dependencies:

python -m pip install --upgrade pip
python -m pip install numpy pandas matplotlib scikit-learn

Typical workflow

1. Prepare the input CSV file using the required column names or one of the accepted aliases defined in the corresponding script.
2. Run the descriptor-screening script for Reaction B or Reaction C to compare Hammett-related and DFT-derived descriptors and identify the most predictive candidates.
3. Run the Hammett-focused model-development script for Reaction B or Reaction C to fit the final Ridge models and export equations.
4. Run the method-comparison script to compare the available regression approaches, if required.
5. Run the external-test script to evaluate the fixed equations on an independent external test set.
6. Inspect the generated metrics, predictions, equations, and parity plots.

Running the scripts

Run each script from the TS_geometry_ML_prediction/ directory:

python reaction_B_descriptor_screening.py --input datasets/reaction_B_dataset.csv --output results/reaction_B_descriptor_screening
python reaction_B_machine_learning_model.py --input datasets/reaction_B_dataset.csv --output results/reaction_B_ml_results --alpha 0.1
python reaction_B_method_comparison.py --input datasets/reaction_B_dataset.csv --output results/reaction_B_method_comparison --descriptor "σp+"
python reaction_B_external_test_prediction.py --input datasets/reaction_B_external-set.csv

python reaction_C_descriptor_screening.py --input datasets/reaction_C_dataset.csv --output results/reaction_C_descriptor_screening
python reaction_C_machine_learning_model.py --input datasets/reaction_C_dataset.csv --output results/reaction_C_ml_results --alpha 0.1
python reaction_C_method_comparison.py --input datasets/reaction_C_dataset.csv --output results/reaction_C_method_comparison --descriptor "σp+"
python reaction_C_external_test_prediction.py --input datasets/reaction_C_external-set.csv

The scripts currently expect their input files and configuration to follow the paths and names defined in the source code. Update those paths when necessary before execution.

Model and validation

Bond lengths, angles, and dihedrals are modeled separately.

Internal validation is performed using leave-one-out cross-validation (LOOCV). Training metrics and LOOCV metrics are reported separately to distinguish model fit from cross-validated predictive performance.

Input data

Input files must be CSV tables containing:

• compound identifiers;
• substituent labels;
• Reaction A transition-state geometries;
• target Reaction B or Reaction C geometries;
• selected numerical descriptor columns.

Accepted column aliases are defined near the top of each script. Preserve the original column names or use one of the listed aliases.

Outputs

Descriptor-screening scripts

Reaction B ranks descriptors by bond-length LOOCV Q². Reaction C ranks descriptors by dihedral LOOCV Q² and includes a rank column in descriptor_ranking.csv.

• descriptor_target_metrics.csv: LOOCV Q² and sample size for each descriptor/target combination
• descriptor_ranking.csv: ranked summary table used to build the figure
• figure.png
• figure.pdf

Model-development scripts

• model_metrics.csv
• model_predictions.csv
• model_equations.txt
• best_model_parity.png
• best_model_parity.pdf
• software_versions.json

Method-comparison scripts

• method_target_metrics.csv
• method_summary.csv
• loocv_predictions.csv
• fitted_model_coefficients.csv
• comparison and ranking figures

External-test scripts

• external_test_predictions.csv
• external_test_metrics.csv
• external_test_parity.png
• external_test_parity.pdf

Reproducibility notes

• Model-development scripts use deterministic LOOCV and do not rely on a random train/test split.
• Method-comparison hyperparameters are fixed to reproduce the reported analysis.
• External-test scripts apply fixed model coefficients and do not refit the models.
• For small external test sets, interpret R² together with MAE, RMSE, residuals, and parity plots.
• Software-version differences may lead to minor numerical or graphical variations.

Data and software availability

The scripts are author-generated and may be shared together with the associated machine-readable data. Quantum-chemistry output files, optimized coordinates, and processed datasets are provided in the Supporting Information accompanying the manuscript.

Acknowledgements

This work is supported by the Walloon Region as part of the funding for the FRFS-WEL-T strategic axis. The authors acknowledge the WEL Research Institute (grant WEL-T-CR-2023 A – 05, "Smart Flow Systems"). Computational resources were provided by the “Consortium des Équipements de Calcul Intensif” (CÉCI), funded by the “Fonds de la Recherche Scientifique de Belgique” (F.R.S.-FNRS) under Grant No. 2.5020.11a and by the Walloon Region.

License

The code in this repository is licensed under the MIT License. You are free to use, modify, and distribute the code under the terms of this license.