In this work, the interference between molecules of toxicological interest and metabolic pathways of two types was analyzed: Nuclear Receptor (NR) signaling, Stress Response (SR) pathways. In the first case, it deals with proteins, which are macromolecules consisting of chains of amino acids (20 specific molecules),located in the cell, capable of recognizing the presence of specific hormones and consequently improve cellular activity and development. In the second case, it deals with a wide variety of adaptive responses of the cell to maintain the own physiological state despite external stresses (mechanical, physical, chemical, environmental, et alia). In both, the interaction between molecules, when known, it can be traced back to the quality, quantity and organization of the chemical elements in the molecular structures of the cell and of the reagents; yet not this interaction is not directly and universally predictable.
Tox21 is an initiative of the US government, aimed at developing the drug cologia in the 21st century, thanks to both experimental and new technologies analytical. Human beings are exposed to enormous amounts and varieties of chemical agents of little known toxicity, simultaneously about 30% of promising experimental drugs in animal tests prove toxic to the human being. This highlights the need to analyze quickly and effectively ciently large amounts of data different in origin and format (exactly the 3 V of Big Data, Velocity, Variety, Volume). In 2014 it was therefore launched created a Tox21 Data Challenge, a set of sub-problems whose objective generic was to predict exactly the outcome of 12 tests for about 10 000 molecules, 6 NR type assays and 6 SR type assays, the outcome of which may be would end if each molecule interfered with the 12 metabolic pathways or not. The 12 original datasets, divided into train, validation and score set, differs as the same molecule may or may not be independently active on different ways, and because not all assays for molecule-metabolic pair give unique or clear results. Such datasets are highly unbalanced, where the rare class is represented by the case molecules having an interaction. Data are made publicly available in SMILES format here.Following the approach and algorithm presented in (Maya et al. 2018), the SMILES strings were transformed into 42-column Feature Matrices. 21 Columns were needed for one-hot-encoding of SMILES native symbols, in addiction to 21 columns to collect features of the atoms, such as the type of element (categories H, C, O, N, or Other), the charge and the electronic valence, the number of bonds with Hydrogen atoms, the chirality (right, left, or other), the saturation and hybridization of orbitals. After identifiying the longest molecule of the dataset having 400 SMILES characters, each matrix has been zero-padded to equalize the dimensions. The Feature Matrices so constructed has been used as inputs of the 1-Dimensional Convolutional Neural Networks (1-D CNNs), while for Long Short Term Memory units (LSTMs) it was preferred to avoid approaching the problem as multivaried (42 possibly unrelated features), reducing the matrices to one-hot-SMILES string encodings. In particular, columns with the additional atomic characteristics have been removed, in addition to the element type. Each row has been replaced with a "flag" of its index with a non-null element in order to learn effective embedding at the same time to LSTM training.
Two different working lines were followed for the two methabolic pathways, SR and NR, in oder to choose the best architectures and models. In the end, the founded models were applied to all the 4 datasets. Several CNN and LSTMS models were used, in order to find a benchmark model for each of the approaches, finally applying them on the test data. Here, unbalanced data are managed by testing the effect on results of both oversampling and weighted classes techniques are applied. Here, only CNN models are applied, after undergoing a Bayesian Optimizations of the hyperparameters. First, a reduction of the search-space was was carried out, by trying different architectures evaluated on a 3-fold cross validation. Following, the resulting best architecture was optimized using bayesian Optimization. Here, unbalanced data were managed assessing the effectiveness of oversampling, weighted classes and SMOTE techqniques.