Skip to content

Latest commit

 

History

History
29 lines (22 loc) · 2.22 KB

Zaslavskyi et al (2019).md

File metadata and controls

29 lines (22 loc) · 2.22 KB
Raw

  • https://doi.org/10.1016/j.comtox.2019.01.001

  • Aim:

    • Proposing a systematic ensembling method for virtual toxicity screening, a prediction model for toxicity-related bioactivity pre-screening

  • Data:

    • The TOX21 and TOXCAST datasets obtained from the DeepChem package
    • For the individual models they trained, they utilised physical chemicals descriptors, PubChem molecular fingerprints, and SMILES n-grams

  • Approach:

    • XGB, DNNs, and GCNs
    • The multi-task gradient boosting model was trained on an entire dataset simultaneously by stacking all task-specific data together and adding a descriptor representing the identity of the corresponding task
    • GCNs can be seen as a more general version of molecular circular fingerprint, where fingerprints are learned directly from the dataset

  • Model optimization:

    • All model hyperparameters were optimized using the validation set

  • Model performance:

    • Three types of cross-validation experiments were conducted based on random, scaffold and index splits
    • For random splitting, all chemicals were divided randomly into three groups: training (80%), validation (10%) and test sets (10%)
    • When using scaffold splitting, all chemicals were grouped according to their scaffolds, such that chemicals with the same scaffold were always assigned to the same set (training, validation, or test)
    • Index based splitting follows the natural order of rows in the dataset, it uses first 80% rows as the training set and the following 10% and 10% as validation and test sets

  • Feature importance:

    • For NNs, they used a permutation feature importance technique. For each feature, they randomly shuffled the values associated to this feature and the k most correlated features, and computed the average AUC degradation over 5 folds, for k=0, 5 and 10. The lower the AUC, the more important the feature

  • Key learnings:

    • Showed that models trained directly on SMILES features can be competitive with those trained on molecular fingerprint representations for both TOX21 and TOXCAST
    • Many of the most important fingerprint descriptors are related to the size of the molecule, indicating that size characteristics might be a critical indicator for bioactivity prediction