Predicting psychiatric drug subjective response using graph convolutional neural networks

FEDERICO ZAMBERLAN; ENZO TAGLIAZUCCHI; FEDERICO ISSOGLIO; CARLA PALLAVICINI

Congreso; XXXIV Reunión Anual SAN; 2019

Sociedad Argentina de Investigación en Neurociencias

Precise identification of patient's feelings, state of mind and mood is of paramount importance when diagnosing psychiatric disorders and evaluating the response pharmacological treatments. In general, an accurate prediction of the subjective experience of a person under the influence of psychoactive drugs, which is the result of the drug's complex interactions with a number of neural receptors, implies a technical and methodological challenge. These subjective experiences may be objectively described from unstructured written reports, using natural language processing algorithms. Graph convolutional deep neural networks is a booming and powerful machine learning strategy applying sequential convoluting operations on graph encodings (such as pharmaceutical compounds or natural language reports) extracting relevant features hidden in the input samples. In these networks, each of the convolution layers extracts local patterns or sub-features that have eluded previous strategies. Both molecules and speech structure can be represented as graphs, in which nodes could correspond to atoms or words, and edges correspond to bonds or grammatical relationships, respectively. We were able to extract relevant features of the drug-receptor interactions and predict their response, for instance, in terms of natural language descriptions of patients or users of the compound. The current work represents an effort towards linking computational chemistry, medicine and language processing.