GED and mordred molecular descriptor distances prediction using KNN from chemical graphs

Identifier:  TFG:9489

Handle: https://hdl.handle.net/20.500.11797/TFG9489

Authors:  Vega Cuadros, Javier

Abstract:
This final grade project has analyzed the prediction of the distance between chemical molecules using methodologies: the Graph Edit Distance (GED) and the Mordred molecular descriptors, combined with the K-Nearest Neighbors (KNN) algorithm. They have two data bases (ESOL and FreeSolv) and two approaches to calculate the GED: a CPU mitjançant (with the NetworkX library) and another GPU mitjançant (with the Fast Bipartite algorithm). The obtained results show that the calculation with the GPU is much more efficient in time (less than one week vs. three weeks with the CPU) as the calculated values ​​have varied slightly with respect to those of the CPU. KNN models trained with both GED values ​​calculated by GPU will generally obtain better results than those from CPU, with both lower quadratic errors (MSE) and lower absolute errors (MAE), as well as higher coefficients of determination (R²). In all ways, the best models have been trained with the GED values ​​that mostly coincide between the CPU and GPU algorithms, in two subdatasets, one for each data base. It is also going to identify a correlation between Mordred and GED distances, which will help predict the similarity of molecular graphs from more easily computable descriptors. In conclusion, the work demonstrates that it is feasible to predict molecular distances efficiently using KNN and molecular methods, and proposes a useful technique for chemoinformatics based on the developed code.