matplotlib=3.3.1
numpy=1.19.2=pypi_0
Orange=3.25.1
pandas=1.1.1
scipy=1.5.2=pypi_0
seaborn=0.9.0
sklearn=0.22.1
xgboost=1.3.3=pypi_0
- It is recommended to execute this project in a linux environment
Step 1: Optimization of the structures of reactant and product molecules (substrates and HAT radicals) and calculation of their property features
1. Get stable structure to do the feature calculation: use quantum chemical software, such as Gaussian or Orca, to obtain optimized structures at the level of B3LYP/6-31+G(d,p). The C atoms to be considered for HAT reactivity may be numbered sequentially starting at 1 to facilitate the 3. Feature collection step.
2. Feature calculation: Perform single point energy calculation at the level of M06-2X/def2-TZVPP with natural bond orbital (NBO) calculations to get FMO energies, atomic charge and bond order. As for BDE, solvation energy correction of MeCN need to be evaluated using SMD model with M06-2X/6-31G(d). Buried volume features can be calculated using SpaceDes.ipynb. The script getDescriptors.py and gen3-sp.py may be helpful in transforming the output files (.log) of structure optimization into the input files of feature calculation.
3. Feature collection: Calculate the BDE from reaction enthalpy(Thermal Correction to Enthalpy+Electronic Energy (EE)) and other features. The script CatchDescriptors.ipynb can help read the data used to calculate the FMO energy, atomic charge and bond order features. Besides, an integrated script is on progress.
1. Model load: Load the model HAT-ReactivityPredictor.pkl. Related codes are shown in HAT-ReactivityPredictor.ipynb.
2. Test set preparation: Test set should contain 56 features with specific order which is shown in SampleTest-ExpCH.csv.
3. Reactivity prediction: Load the test set file (.csv) and get ML-predicted HAT barriers. Related codes are shown in HAT-ReactivityPredictor.ipynb.