license: mit
tags:
- chemistry
- graphormer
- molecular-property-prediction
- redox-potential
- regression
- pytorch
Graphormer Redox Predictors
Last updated: 10-NOV-2026
Model summary
| Field | Details |
|---|---|
| Developer | Microsoft Corporation |
| Description | Two non-generative molecular property regression models that estimate one-electron reduction or oxidation potential, in volts versus the standard hydrogen electrode, from a molecule represented as SMILES. |
| Model architecture | Twelve-layer Graphormer encoder (graphormer_base) with 768-dimensional embedding, 32 attention heads, and a scalar regression head. |
| Parameters | 1-500M |
| Inputs | SMILES molecular strings; up to 128 non-hydrogen atoms per molecule. |
| Outputs | A scalar reduction or oxidation potential per valid molecule. |
| Context length | N/A |
| Training dates | Sep 2025 |
| Release date / Release date in the EU (if different) | Nov 2026 |
| License | MIT |
| Model dependencies | Microsoft Graphormer; orfb-design |
| Acceptable use policy | N/A |
1. Model overview
Graphormer Redox Predictors are two non-generative molecular-property models that estimate one-electron reduction or oxidation potentials from molecular structures represented as SMILES. They are based on Microsoft Graphormer, a Transformer architecture adapted to graph-structured data through graph-aware attention encodings. The underlying model was pretrained for molecular modeling on PCQM4Mv2 and subsequently fine-tuned with separate scalar regression objectives for reduction and oxidation potential.
The fine-tuning enables rapid redox-potential screening of small organic molecules, but it also introduces limitations associated with the training distribution and quantum-chemical labels. Predictions are screening estimates rather than experimentally validated measurements and do not explicitly account for solvent, pH, supporting electrolyte, temperature, concentration, electrode conditions, kinetic reversibility, or chemical stability.
Key model capabilities
- Predicts one-electron reduction potential from SMILES in volts versus SHE.
- Predicts one-electron oxidation potential from SMILES in volts versus SHE.
- Processes batches of up to 32 molecules.
- Supports molecules containing up to 128 non-hydrogen atoms.
- Returns per-molecule predictions and validation status.
2. Usage
2.1 Primary use cases
Graphormer Redox Predictors are intended for research and engineering workflows that rank or triage small organic molecules according to predicted one-electron redox potential. The models are intended for human-in-the-loop scientific decision support, with independent computational or experimental validation before selecting candidates for synthesis or deployment.
2.2 Out-of-scope use cases
The models should not be used as substitutes for quantum-chemical calculations, electrochemical measurements, chemical-safety testing, or regulatory determinations. They should not autonomously control decisions involving synthesis, scale-up, procurement, manufacturing, or deployment of chemical materials.
The models were not designed or evaluated for inorganic materials, polymers, mixtures, reactions, proteins, molecules containing more than 128 non-hydrogen atoms, or arbitrary non-SMILES inputs. Predictions should not be assumed to apply across all solvents, pH values, supporting electrolytes, temperatures, concentrations, reference electrodes, protonation states, tautomers, stereoisomers, or conformations. The models are also not intended for medical, clinical, biological-activity, environmental-safety, or consumer-facing decisions.
2.3 Distribution channels
Some of Graphormer Redox Predictors' distribution channels include:
- Hugging Face
2.4 Technical requirements and integration guidance
The supported implementation runs in a Linux-based containerized Python environment. Required dependencies include PyTorch, Fairseq, Graphormer, Open Graph Benchmark, PyTorch Geometric, and the associated molecular-graph preprocessing stack. A C/C++ compiler is required by the PyTorch CPU kernel fuser.
Integration is appropriate for batch virtual screening and human-in-the-loop scientific decision support. It is not appropriate as an unaudited autonomous gate for synthesis, scale-up, procurement, safety, or regulatory decisions.
2.5 Supported languages
Not applicable
2.6 Responsible AI considerations
Developers should document the model version, checkpoint, molecular representation, prediction conditions, known uncertainty, and downstream selection criteria. High-impact candidates should be reviewed by qualified scientists and independently verified using appropriate quantum-chemical calculations and experimental electrochemistry. Model outputs must not be represented as experimentally validated measurements.
2.7 Known limitations
- Performance may degrade for molecules outside the chemical distributions represented by ZINC-22.
- Results depend on the supplied SMILES representation, including charge, protonation state, tautomer, stereochemistry, and salt handling.
- The models do not explicitly condition predictions on solvent, pH, supporting electrolyte, concentration, temperature, reference electrode, or other experimental conditions.
- The models do not predict electrochemical kinetics, reversibility, chemical stability, aqueous solubility, toxicity, or synthetic feasibility.
3. Quality and performance evaluation
| Model | Split | MAE (V) | RMSE (V) |
|---|---|---|---|
| Reduction potential | Validation (10% split) | 0.093 | 0.22 |
| Reduction potential | Test (10% split) | 0.092 | 0.20 |
| Oxidation potential | Validation (10% split) | 0.092 | 0.18 |
| Oxidation potential | Test (10% split) | 0.092 | 0.18 |
3.1 Benchmarking methodology
Models were evaluated using an 80/10/10 train/validation/test split of the training data.
3.2 Safety evaluation and red-teaming
Models were not evaluated for safety.
4. Data overview
4.1 Training, testing, and validation datasets
Graphormer Redox Predictors were trained on a set of DFT-computed redox potential labels. The full data generation protocol is described in https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15002385/v3.
4.1.1 Size of dataset and characteristics
| Source field | Details |
|---|---|
| 4.1.2.A Text training data size | Not applicable. Text data is not part of the training data. |
| 4.1.2.B Text training data content | Not applicable. |
| 4.1.2.C Image training data size | Not applicable. Images are not part of the training data. |
| 4.1.2.D Image training data content | Not applicable. |
| 4.1.3.A Audio training data size | Not applicable. Audio data is not part of the training data. |
| 4.1.3.B Audio training data content | Not applicable. |
| 4.1.4.A Video training data size | Not applicable. Video data is not part of the training data. |
| 4.1.4.B Video training data content | Not applicable. |
| 4.1.5.A Other training data size | 42,988 |
| 4.1.5.B Other training data content | (SMILES, reduction/oxidation potential) pairs where each SMILES represents a molecule sampled from the ZINC22 database |
| 4.1.6 Latest date of data (acquisition/collection for model training) | 01-AUG-2025 |
| 4.1.7 Is data collection ongoing to update the model with new data collection after deployment? | No |
| 4.1.8 Date the training dataset was first used to train the model | SEP 2025 |
| 4.1.9 Rationale or purpose of data selection | The dataset was constructed to be representative of commercially available small molecules. |
4.2 List of data sources
4.2.1 Publicly available datasets
4.2.2 Have you used publicly available datasets to train the model? Yes
4.2.2 Private non-publicly available datasets obtained from third parties
4.2.2.1A Have you concluded transactional commercial licensing agreement(s) with rightsholder(s) or with their representatives? No
4.2.2.2.A Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1, such as data obtained from providers of private databases, or data intermediaries? No
4.3 Personal data
4.3.1 Was personal data used to train the model? Microsoft follows applicable laws and best practices pertaining to personal data.
4.4 Synthetic data
4.4.1 Was any synthetic AI-generated data used to train the model? No
4.5 Data processing aspects
4.5.1 Respect of reservation of rights from text and data mining exception or limitation
4.5.1A Does this dataset include any data protected by copyright, trademark, or patent? Microsoft follows applicable laws and best practices for processing data protected by copyright, trademark, or patent.
4.5.2 Other information
4.5.2A Does the dataset include information about consumer groups without revealing individual consumer identities? Microsoft follows applicable laws and best practices for protecting consumer identities.
4.5.2B Was the dataset cleaned or modified before model training? Yes
5. Contact
Requests for additional information can be directed to MSFTAIActRequest@microsoft.com.
Authorized representative: Microsoft Ireland Operations Limited 70 Sir John Rogerson's Quay, Dublin 2, D02 R296, Ireland
6. Appendix
Additional details on redox potential label generation and model training can be found in https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15002385/v3.