Unsupervised Learning based on Deep Learning Applied to the Identification of Applicability Domain of QSAR Models

MARTINEZ, MARÍA JIMENA; CRAVERO, FIORELLA; DIAZ, MÓNICA F.; PONZONI, IGNACIO

Buenos Aires

Conferencia; IV International Society for Computational Biology Latin America Bioinformatics Conference (ISCB-LA 2016); 2016

International Society for Computational Biology (ISCB) y Asociación Argentina de Bioinformática y Biología Computacional (A2B2C)

Background For a QSAR/QSPR (Quantitative Structure-Activity Relationship) model is a key point to define their Applicability Domain (AD). Namely, know for which compounds the predictions will be reliable and which are not. Only compounds that fall within this domain could be evaluated reliably. Furthermore, if there is a new compound to predict, we could know whether this one will have a precision similar (or not) to that reported in the method validation. A strategy to define the applicability domain of a model is to organize the process in two stages. In the first stage, the compounds of the training set that are structurally similar to the new compound are identified, for this the data should be grouped by similarity. To identify these groups of data there are many methods, but in which we focus here is deep learning, which is a machine learning technique in which novelty is in how to combine different unsupervised methods. Deep learning can be implemented using for example, neural networks or autoencoders. Finally, in the second stage of the process, it proceeds to evaluate the ability of the predictor on the new compound, applying a set of statistical tests designed for this purpose.Conclusions The main contribution of this work was to focus on the first stage and by using techniques of deep learning identify the similarity of the compounds. The potential of these techniques was evaluated using generic properties for the process of design and development of new drugs.