Druggability assessment algorithm based on Composition, Transition and Distribution descriptors and publicly available predictive tools

ALICE, JUAN IGNACIO; RODRIGUEZ, SANTIAGO; ALBERCA, LUCAS NICOLÁS; BELLERA, CAROLINA; TALEVI, ALAN

Virtual

Congreso; XI Argentine Congress of Bioinformatics and Computational Biology (XI CAB2C; 2021

Asociación Argentina de Bioinformática y Biología Computacional

Background:In the framework target-guided drug discovery it is important to be able to assess the druggability ofthe proposed drug target prior to the implementation of a drug discovery project. Druggability is aconcept coined by Hopkins and Groom to refer to the ability of a protein to be modulated by small,drug-like molecules.Results:A dataset of 222 proteins druggable and undruggable was compiled, and it was split into a training setfor model building and an independent test set for model validation. The training set was then used toinfer linear classifiers capable of prospectively discriminating druggable from non-druggable targets.Two algorithms were built and validated for such task. The first one uses CTD (Composition,Transition and Distribution) descriptors, while the second combines CTD descriptors with alreadyreported and validated online druggability assessment tools. 14 druggability predictors were derivedfrom online tools and 147 CTD descriptors were computed using the PyProtein module fromPyBioMed library. Using a combination of feature bagging and forward stepwise feature selection,1000 linear models were built using either a combination of online tools plus CTD or CTD descriptorsalone .The best individual model for CTD descriptors displayed an accuracy of 0.803, a precision of 0.738and recall of 0.939 on the test set, while the best individual model emerging from the combination ofCTD descriptors and online tools showed an accuracy of 0.871, a precision of 0.800 and a recall of0.848.Conclusions:Any target-focused, rational drug discovery initiative starts with the choice and validation of anadequate drug target. Drug target validation implies, among other studies, guaranteeing that thechosen target is druggable. Here, we have reported an algorithm based on CTD descriptors anddruggability descriptors derived from online tools, capable of differentiating, with remarkableaccuracy druggable from non-druggable proteins in a fast and cost-efficient manner.