Data and code for "Accelerating Proton Affinity Prediction with Multi-Fidelity Machine Learning"

Data and code accompanying the paper "Accelerating Proton Affinity Prediction via Multi-Fidelity Machine Learning" by D. Bhattacharya, Y. Liu, V. R. Cooper, and W. F. Reinhart.

This archive contains everything needed to reproduce every figure and every number reported in the manuscript and Supporting Information: the raw quantum-chemistry outputs (PM7 via MOPAC 2016, and B3LYP/def2-TZVP DFT via PySCF 2.12.1 / gpu4pyscf 1.5.2), the derived ML-ready feature tables, the trained-model cross-validation results, the prospective ZINC-screening outputs, and all analysis/plotting scripts.

Contents

data/ — raw and processed datasets

nist1185/ — NIST 1185-molecule B3LYP/def2-TZVP DFT outputs (one JSON per molecule).

kmeans251/ — k-means 251-molecule DFT outputs (one folder per molecule; 251 successful runs).

pm7/ and pm7_source_raw/ — aggregated and raw PM7 proton-affinity CSVs for both datasets.

features/, targets/, processed/ — ML-ready feature matrices, regression targets, and parsed dataset JSON archives.

screening/ — prospective screening of ~821K ZINC molecules (FAISS index, per-iteration candidates / PM7 / DFT outputs).

scripts/ — dataset construction, featurization, model training, SHAP, learning-curve sweeps, and all paper/SI plotting code.

screening/scripts/ — the seven-stage prospective-screening pipeline.

pm7_scripts/, dft_scripts/ — stand-alone, generic PM7 and B3LYP/def2-TZVP proton-affinity pipelines (SLURM templates included) that reviewers or readers can run on their own molecules.

results/, figures/ — CV results, SHAP values, learning-curve JSONs, and every manuscript/SI PDF.

make_figures.py — single entry point that regenerates every figure in the paper from the in-repo results.

README.md — full repository layout, setup, and reproduction instructions.

Quick start

unzip proton-affinity-paper.zip
cd proton-affinity-paper
pip install -r requirements.txt
python make_figures.py # regenerate paper figures

To retrain models or rebuild feature tables, see the usage notes in README.md and scripts/README.md. The stand-alone PM7 and DFT pipelines in pm7_scripts/ and dft_scripts/, respectively, each include its own README.

Methods summary

PM7 semi-empirical calculations were performed with MOPAC 2016 (PM7 PRECISE GNORM=0.001 SCFCRT=1.D-8, CHARGE=1 and UHF for protonated species). B3LYP/def2-TZVP DFT geometry optimizations, analytical Hessians, and IGRRHO thermochemistry (T = 298.15 K, frequency scaling factor 0.9850) were performed with PySCF / gpu4pyscf on NVIDIA A100 GPUs. Full methodological details are in the SI.

Citation

If you use these data or code, please cite the accompanying manuscript.

License

Data are released under CC BY 4.0. Code is released under the MIT License.