Fusion of molecular representations and prediction of biological activity using convolutional neural network and transfer learning

Basic structural features and physicochemical properties of chemical molecules determine their behaviour during chemical, physical, biological and environmental processes and hence need to be investigated for determining and modelling the actions of the molecule. Computational approaches such as machine learning methods are alternatives to predict physiochemical properties of molecules based on their structures. However, limited accuracy and error rates of these predictions restrict their use. This study developed three classes of new methods based on deep learning convolutional neural network for bioactivity prediction of chemical compounds. The molecules are represented as a convolutional neural network (CNN) with new matrix format to represent the molecular structures. The first class of methods involved the introduction of three new molecular descriptors, namely Mol2toxicophore based on molecular interaction with toxicophores features, Mol2Fgs based on distributed representation for constructing abstract features maps of a selected set of small molecules, and Mol2mat, which is a molecular matrix representation adapted from the well-known 2D-fingerprint descriptors. The second class of methods was based on merging multi-CNN models that combined all the molecular representations. The third class of methods was based on automatic learning of features using values within the neurons of the last layer in the proposed CNN architecture. To evaluate the performance of the methods, a series of experiments were conducted using two standard datasets, namely MDL Drug Data Report (MDDR) and Sutherland datasets. The MDDR datasets comprised 10 homogeneous and 10 heterogeneous activity classes, whilst Sutherland datasets comprised four homogeneous activity classes. Based on the experiments, the Mol2toxicophore showed satisfactory prediction rates of 92% and 80% for homogeneous and heterogeneous activity classes, respectively. The Mol2Fgs was better than Mol2toxicophore with prediction accuracy result of 95% for homogeneous and 90% for heterogeneous activity classes. The Mol2mat molecular representation had the highest prediction accuracy with 97% and 94% for homogeneous and heterogeneous datasets, respectively. The combined multi-CNN model leveraging on the knowledge acquired from the three molecular presentations produced better accuracy rate of 99% for the homogeneous and 98% for heterogeneous datasets. In terms of molecular similarity measure, use of the values in the neurons of the last hidden layer as the automatically learned feature in the multi-CNN model as a novel molecular learning representation was found to perform well with 88.6% in terms of average recall value in 5% structures most similar to the target search. The results have demonstrated that the newly developed methods can be effectively used for bioactivity prediction and molecular similarity searching