This project aims to analyze the quantitative structure-activity relationship to predict acute aquatic toxicity towards the fish Pimephales promelas (fathead minnow) on a set of 908 chemicals using K-Nearest Neighbor (kNN).
The dataset consists of six molecular descriptor as features and their corresponding LC50 values, which are:
| Feature | Description |
|---|---|
| CIC0 | Set of indices of neighbourhood symmetry |
| SM1_Dz(Z) | Set of descriptors calculated from 2D matrices derived from the molecular graph (2D matrix-based descriptors) |
| GATS1i | 2D Geary autocorrelation descriptor |
| NdsCH | count the number of unsaturated sp2 carbon atoms of the type =CH- |
| NdssC | Count the number of unsaturated sp2 carbon atoms of the type =C |
| MLOGP | The octanol-water partitioning coefficient (log P) calculated by means of the Moriguchi model |
Dataset source : here.
During the training process, we used 10-fold cross validation to evaluate the model's performance for a range of values for the number of nearest neighbors (k) from 2 to 9. After training the model with different values of k, we found that the model performed best when k was set to 8.
The results are shown below:
| Metric | Training Score | Testing Score |
|---|---|---|
| R-Squared | 0.683 | 0.600 |
| RMSE | 0.804 | 0.984 |
In order to run the python script and notebook, you will need to have the following packages installed:
- matplotlib
- numpy
- pandas
- scikit-learn
M. Cassotti, D. Ballabio, R. Todeschini, V. Consonni. A similarity-based QSAR model for predicting acute toxicity towards the fathead minnow (Pimephales promelas), SAR and QSAR in Environmental Research (2015), 26, 217-243; doi: 10.1080/1062936X.2015.1018938