RaPTOR — Radical Philicity Prediction

Project directory for the RaPTOR paper: a D-MPNN that predicts the Parr electrophilicity index of open-shell radicals from structure alone.

Philicity throughout is ω = (I + A)² / (8(I − A)), in eV, computed from DFT ionization potential and electron affinity. The dataset spans carbon-, nitrogen-, oxygen- and sulfur-centred radicals.

Live model: https://huggingface.co/spaces/mklavin/RaPTOR

---

Headline numbers

Dataset	142,779 radicals (C 90,927 / S 20,511 / N 18,495 / O 12,846)
Split (train/val/test)	128,731 / 12,695 / 1,353
Validation MAE	0.0743 eV
Validation RMSE	0.1269 eV
Inference cost	~0.9 ms per molecule, CPU

Production architecture: depth=10, message_hidden_dim=1250, ffn_hidden_dim=750, atom_ffn_hidden_dim=750, ffn/atom_ffn_num_layers=2, dropout=0.0, batch_size=64, max_lr=3e-4, 50 epochs, patience 20, ensemble size 1.

---

Directory map

01_data_generation/    ORCA/xtb pipeline that produced the DFT dataset
02_data_preparation/   filtering, augmentation, preprocessing
03_dataset/            raptor_dataset.csv — the 142,779-radical dataset
04_modeling/           hyperparameter search, training, the trained models

Each directory has its own README describing its contents.

---

Reproduction chain

01  functions.py + submit_batches.py + submit_philicities.py
        -> per-molecule ORCA jobs -> read_files.py
                |
02  compile_and_filter.py      -> the compiled DFT table
    preprocess_for_modeling.py -> cleaning, artifact rejection
                |
04  chemprop_all_prep.py       -> production_model/training_inputs/
    chemprop_grid_*_train.sh   -> production_model/best.pt   (4-stage grid search)
    chemprop_loeo_*            -> leave-one-element-out zero-shot generalisation
                |
03  build_raptor_dataset.py    -> raptor_dataset.csv

Note the last step is numbered 03 but runs last: the published dataset carries a predicted_philicity column, so it can only be assembled once a model exists.

The intermediate tables between 02 and 04 are not shipped — the scripts read them by absolute path from the original working tree. See Deliberately not included.

---

Two things that silently corrupt predictions

Both are load-bearing and easy to lose in an edit:

1. The training target is log1p(philicity). Model output must go through expm1 to be read as eV.
2. chemprop's build_dataloader sets drop_last=True whenever len(dataset) % batch_size == 1, silently dropping the last molecule. Inference must pass drop_last=False.

The deployed raptor_inference.py in the Space repo handles both correctly and is the reference implementation — see Deployment below.

---

Deployment

The model is served from a Hugging Face Space, which is its own git repository and is not duplicated here:

• Space: https://huggingface.co/spaces/mklavin/RaPTOR
• Remote: git@hf.co:spaces/mklavin/RaPTOR

That repo holds api.py, routes_ui.py, raptor_inference.py, mimis_drawing.py, query_log.py, requirements.txt, Dockerfile, static/, and raptor_model/best.pt — the same checkpoint as 04_modeling/production_model/best.pt (md5 7e259e8d3b7ef583fce8a6fd6545e891).

---

Licence

CC BY 4.0 (LICENSE), covering the code, the dataset and the model weights alike. For academic use, please cite the RaPTOR paper.

---

Environment

Training and inference: chemprop 2.2.3, torch 2.5.1, rdkit, pandas, numpy. The pinned inference stack is requirements.txt in the Space repo.

Scripts carry absolute paths into the original working tree (/insomnia001/depts/tekle_smith/users/MKL/project_5/...). They are preserved verbatim so the provenance of every number is auditable; adjust the path constants at the top of each script to re-run elsewhere.