Deep learning for de novo drug design
Robin Winter, Floriane Montanari, DA Clevert · Ultraschall in der Medizin - European Journal of Ultrasound · 2019
An essential part of computer-aided drug development is the prediction of molecular properties for a specific chemical structure (e.g. the interaction of a compound with a protein) or the design of molecules with the desired properties. Molecular descriptors play a crucial role for building predictive models, since they allow representing chemical information of molecules in a computer-interpretable vector. In this talk, we present (i) a neural machine translation approach for learning a robust and continuous feature representation of chemical structures, (ii) an exhausted benchmarked respect with various human-engineered molecular fingerprints and state-of-the-art graph-convolution models and (iii) exploit the continuity of our descriptor space to generate/optimize new chemical structures from the latent representation.