Applying Deep Learning and Multi-Task Learning to predict biological target profiles for drug candidates. A short project I pursued within SABS CDT, supervised by Garrett Morris (Dept. of Statistics) and Thierry Hanser (Lhasa Ltd). Here is a quick documentation for my scripts/data/results and how to use them.
Data : The files Interactions_Trainset.tab, Interactions_Validset.tab and Interactions_Testset.tab contain the interactions used for training, validating and testing respectively. The test-set was formed as a roughly balanced selection between active and inactive compounds. More about the data can be found in my report. All these files contain triplets as target-ID, compound-ID, pIC50, one line per interaction, and have occured after processing interactions_kinases_all.tab which contains all the interactions fetched for 797 kinases.
Most of the computational experiments conducted for this project are implemented in the following Python notebooks:
-
Project_SingleTL_CrossVal: This jupyter notebook compares four single-task methods -- RF, Lasso regression and two NN approaches -- for regressing bioactivity data, with one model per target. RF, LR and the first NN are deployed from Scikit-learn and the second NN (lebelled asmyNN) is implemented with Keras. -
Project_MTL_CrossVal: This is similar to the previous notebook but from a multi-task perspective, where we train one model for all the 110 targets. Parameter and architecture selection is conducted with cross-validation. We also explore the use of dropout in a MTL NN. Dropout is useful as it can make the model probabilistic and thus offer the option to produce confidence levels for each prediction. -
Project_Benchmarking: A notebook that applies and compares all six methods on the validation set and measures time complexity. There's no training here and pre-trained models need to be loaded. -
Project_Self-training: The application, and comparison, of two methods that are able to self-train for regression. -
Project_Final_Test: This is used for the final evaluation of our project. It compares RF with MTL on their ability to impute new values. Important notice: The file containing fingerprints for every compound is needed here but not included in the git-repo due to size limitations.
Demos : is a folder with html versions of some of the .ipynb for presenting the pipeline+results quickly.
For all the above, the only input we used as features are the ECFP4, calculated by RDkit. We also ran a few experiments using feature-based fingerprints (FCFP) and the corresponding materal is under the FCFP folder. The rest of the folders are less important and description is skipped at this point.
For more details as well as references please have a look at the report (which is still in-progress!).