Сравнение эффективности молекулярных дескрипторов в моделировании отношения «структура-активность»

Abstract

To date, there are two main approaches in drug design based on ligand representation: a) a search for active compounds based on chemical similarity and b) predictions based on of the structure-activity relationship (QSAR) of various datasets. Approaches based on similarity require one known hit compound, while QSAR models can be developed with a sufficiently large data set of biologically active compounds. Previously, the two approaches mentioned above were compared with the kNN-QSAR method. Strictly constructed and tested, the kNN-QSAR models showed the highest predictive power for almost all tested data sets. This conclusion once again confirms the importance of solving the problem of presenting the original structural data set (descriptors), which can improve the quality of prediction. The purpose of this work is to test the use of various sets of descriptors by the kNN-QSAR method, including descriptors generated by Rcdk, Dragon and SPCI software, which allowed to interpret the results of modeling for a set of 90 nitroaromatic compounds; an example of the interpretation of the simulation result is shown.

More by Fatima Adilova

• Graph-based density-variation outlier removal for chronic disease prediction: A leakage-free benchmark and component-wise ablationResults in Engineering · 2026
• Monotonic System Framework for Globally Optimal Structuring of Complex Data2026
• Boosting LLM Adaptation: Superior Stability and Convergence via OLoRA+ Across Diverse NLP BenchmarksInternational Siberian workshops and tutorials on electron devices and materials · 2026
• Генерации графов заданной структуры: от глубоких нейронных сетей к квантовым моделям (на примере создания новых лекарств)Problems of Computational and Applied Mathematics · 2026