Welcome to the user information page for RIGR.
60% fewer features, resonance-invariant, and delivering the same high performance!
RIGR is introduced and discussed in our work: RIGR: Resonance-Invariant Graph Representation for Molecular Property Prediction(paper, preprint). It is designed to impose resonance invariance for molecular property prediction tasks and is accesible as a featurizer in Chemprop v2 (available in versions 2.1.2 and above).
-
For CLI users: RIGR is available as a choice for the multi-hot atom featurization scheme. To use RIGR, add the following argument to your training or inference script.
--multi-hot-atom-featurizer-mode RIGR
This sets the atom and bond feature set to be resonance invariant and applies to both non-reaction and reaction featurization. See an example bash script here. If required, the overall molecular charge descriptor can be added to the learned embedding by using
--molecule-featurizer charge.NOTE: Ensure consistent featurization scheme for train, predict, and hpopt.
-
For Jupyter Notebook users: Refer to our example notebook to use RIGR in a notebook environment.
RIGR ensures a unified graph representation of different resonance structures of the same molecule, including non-equivalent resonance forms. By using RIGR, users can avoid manually selecting a specific resonance form for molecules or radicals exhibiting resonance. If you find RIGR helpful in your research, please cite our paper.
RIGR uses only the subset of atom and bond features from Chemprop that remain invariant across different resonance forms. The tables below indicate which atom and bond features are present and absent in RIGR.
| Feature | Description | Present in RIGR? |
|---|---|---|
| Atomic number | The choice for atom type denoted by atomic number | ☑️ |
| Degree | Number of direct neighbors of the atom | ☑️ |
| Formal charge | Integer charge assigned to the atom | ☐ |
| Chiral tag | The choices for an atom's chiral tag (See rdkit.Chem.rdchem.ChiralType) |
☐ |
| Number of H | Number of bonded hydrogen atoms | ☑️ |
| Hybridization | Atom's hybridization type (See rdkit.Chem.rdchem.HybridizationType) |
☐ |
| Aromaticity | Indicates whether the atom is aromatic or not | ☐ |
| Atomic mass | The atomic mass of the atom | ☑️ |
| Feature | Description | Present in RIGR? |
|---|---|---|
| Bond type | The known bond types: single, double, or triple bond | ☐ |
| Conjugation | Indicates whether the bond is conjugated or not | ☐ |
| Ring | Indicates whether the bond is a part of a ring | ☑️ |
| Stereochemistry | Stores the known bond stereochemistries (See BondStereo) | ☐ |
The table below provides details on which branch corresponds to specific analyses in our paper:
| Branch Name | Purpose |
|---|---|
main |
Main active branch of Chemprop with RIGR implemented as an optional featurization scheme |
rigr |
The branch used for training all the rigr models in the paper |
native |
The branch used for training all the native and native+aug models in the paper |
rigr_charge |
Same as rigr but with additional molecule level charge featurizer — used for most of the property prediction benchmarks |
rigr_charge_stereo_chiral |
Same as rigr_charge but with bond stereochemistry and atom chirality features — used for RGD1 benchmark |
Relevant citations if this work is useful to you:
@article{doi:10.1021/acs.jcim.5c00495,
author = {Zalte, Akshat Shirish and Pang, Hao-Wei and Doner, Anna C. and Green, William H.},
title = {RIGR: Resonance-Invariant Graph Representation for Molecular Property Prediction},
journal = {Journal of Chemical Information and Modeling},
volume = {65},
number = {20},
pages = {10832-10843},
year = {2025},
doi = {10.1021/acs.jcim.5c00495},
URL = {https://doi.org/10.1021/acs.jcim.5c00495},
eprint = {https://doi.org/10.1021/acs.jcim.5c00495}
}