Goal: Predict different chemical properties.
- This is my first AI
- Right now the model does not give reliable predictions (See Problem section)
- If you have suggestions how to improve my code or do things more efficiently I am happy to hear from you
Right now only the file 'all_chem_with_ghs.csv' is available. It contains all chemicals from https://pubchem.ncbi.nlm.nih.gov/ with a CID < 1.000.000 and a GHS symbol.
Formats raw_data files into a usable format for the predicter classes. Has to be executed before the first call of any predicter to ensure the needed data is avaiable
- creates a dict of all appearing units in the used smiles strings.
'C=O' -> {'C':12, '=': 13, 'O':14} - formats smiles strings into a list of integer representation:
'CC' -> [12,12] - formats ghs list into list of boleans:
[2,7] -> [0,1,0,0,0,0,1,0,0] - saves formated data in the main folder as csv or pickle
class ChemAI()
Parent Class of all following Predicters to reduce redundant code.
Model build is strongly influenced by the type of data that is to be predicted.
For non-chemists: There are 9 GHS symbols. They represent dangerous properties of chemical substances (See https://pubchem.ncbi.nlm.nih.gov/ghs/)
class GHSPredicter(ChemAI)
When executed, trains a model that predicts GHS symbols.
def predict_ghs(self, smile)
When given any smiles string, prints a list of GHS symbols with descending probability for the given compound.
clone or download repository
make sure you have the most commen ML packages installed (tensorflow, numpy, pandas)
execute modify_data.py
python modify_data.py
make instance of GHSPredicter:
chem_model = GHSPredicter()
your_smile = 'CC(=O)C'
chem_model.predict_ghs(your_smile)
The current data set has GHS7 for about 80% of the ~27.000 compounds. This leads to the model predicting GHS7 for every molecule first and neglect the others. A larger more balanced data set would be needed to eradicate this unflattering behavior.