Neural Model-Based Similarity Prediction for Compounds with Unknown Structures

EUGENIO BORZONE; LEANDRO EZEQUIEL DI PERSIA; MATIAS GERARD

Arequipa

Conferencia; 5th International Conference on Applied Informatics (ICAI); 2022

ITI Research Group

Compounds similarity analysis is widely used in many areas related to cheminformatics. Its calculation is straightforward when compounds structures are known. However, there are no methods to get similarity when this information is not available. Here we propose a novel approach to solve this problem. It generates compound representations from metabolic networks, and are use a neural network to predict similarity. The results show that generated embeddings preserve the neighborhood of the original metabolic graph, i.e. compounds participating into the same reactions are close together in the embedding space. Results for compounds with known structures show that the proposal allows to estimate the similarity with an error of less than 10%. In addition, a qualitative analysis of similarity shows that the prediction for compounds with unknown structure provides promising results using the generated embeddings