This repository accompanies our manuscript:
Hölter, N.; Schlosser, L.; Hümpel, C.; Spierling, L.; Glorius, F.
A Hybrid Machine Learning and Marcus Theory Approach for Rationalizing and Predicting Triplet Energy Transfer.
Manuscript in preparation; publication link to follow.
It builds on the original EnTdecker platform:
Schlosser, L.; Rana, D.; Pflüger, P.; Katzenburg, F.; Glorius, F.
EnTdecker—A Machine Learning-Based Platform for Guiding Substrate Discovery in Energy Transfer Catalysis.
J. Am. Chem. Soc. 2024, 146 (19), 13266–13275.
DOI: 10.1021/jacs.4c01352.
The ML framework in this repository predicts, from a SMILES string, the adiabatic triplet energy ET, the reorganization energies λS0-T1 and λT1-S0, and the triplet spin populations of a molecule. Each prediction comes with an uncertainty. Combined with Marcus theory, these predictions give the driving force and the barrier of triplet energy transfer from a photocatalyst.
Requires Python ≥ 3.12.
git clone https://github.com/GloriusGroup/MarcusEnTdecker.git
cd MarcusEnTdeckerWith uv (recommended):
uv sync --extra ml # ML models, Python API and the ent-ml command line
uv sync --extra webapp # the same, plus the web applicationWith pip:
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[ml]" # or ".[webapp]"The trained model (ml/data/model_final.pt) is part of the repository.
Command line:
uv run ent-ml predict "C=Cc1ccccc1" # triplet and reorganization energies, uncertainties
uv run ent-ml predict --file smiles.txt --format csv --out predictions.csv
uv run ent-ml rate --quencher "C=Cc1ccccc1" \
--photocatalysts data/computational_data_processed/photocatalyst_properties_calc.xlsx \
--photocatalyst 1 # driving force, barrier, rate constant
uv run ent-ml --helpPython:
from ml.prediction import predict
from ml.predict import RatePredictor, photocatalyst_properties
predict(["C=Cc1ccccc1"]) # DataFrame: triplet and reorganization energies, uncertainties, spin populations
pc = photocatalyst_properties(
"data/computational_data_processed/photocatalyst_properties_calc.xlsx", 1) # PC1
result = RatePredictor().predict(pc, quencher="C=Cc1ccccc1")
result.dG_kcal, result.dG_barrier_kcal, result.k_entThe photocatalyst properties (ET, λ) come from DFT. ml/README.md describes the model, its
training and the full API.
EnTdecker serves the model in the browser and computes the EnT barrier for any molecule against the 35 photocatalysts of our library. It is hosted at https://entdecker.uni-muenster.de. To run it locally:
uv sync --extra webapp
uv run uvicorn webapp.app:app --port 5000 # then open http://127.0.0.1:5000The data-processing pipeline and benchmark analyses are in data/ and benchmark/;
data/README.md and benchmark/README.md describe them.
The Zenodo archive includes the code, models, processed data and both raw-data folders:
data/experimental_data_raw/ and data/computational_data_raw/. The GitHub checkout
excludes these raw folders; download them from the Zenodo deposit to run the full pipeline.
uv run reproduce.py # all stages: data, benchmark, further_analysis, webapp
uv run reproduce.py benchmark # one stageAfter extracting the Zenodo archive, run the commands above from its repository folder.
The raw inputs are already in data/. In a GitHub checkout without the raw inputs,
reproduce.py skips extraction and rebuilds the remaining stages from the supplied tables.
Tests: uv run --extra webapp --group dev pytest.
- Current manuscript: publication link to follow
- Data and archived code: Zenodo
- DFT calculations: ioChem-BD
- Web application: EnTdecker
If you use this code or data, please cite our manuscript and the original EnTdecker publication listed above.
The model code is pyrates by Leon Schlosser, carried
here as the ml package; ml/README.md records its origin.