Parallel Generative Topographic Mapping: an Efficient Approach for Big Data Handling

Arkadii, L.; Baskin, I.; Marcou, G.; Marcou, G.; Horvath, D.; Horvath, D.; Beck, B.; Beck, B.; Varnek, A.; Varnek, A.

Авторы: Arkadii Lin, Baskin Igor, Marcou Gilles, Marcou Gilles, Horvath Dragos, Horvath Dragos, Beck Bernd, Beck Bernd, Varnek Alexandre, Varnek Alexandre
Журнал: Molecular informatics
Том: 39
Номер: 12
Год издания: 2020
Издательство: Wiley-VCH Verlag
Местоположение издательства: Weinheim, Germany, Germany
Первая страница: 2000009
DOI: 10.1002/minf.202000009
Аннотация: Generative Topographic Mapping (GTM) can beefficiently used to visualize, analyze and model largechemical data. The GTM manifold needs to span thechemical space deemed relevant for a given problem.Therefore, the Frame set (FS) of compounds used for themanifold construction must well cover a given chemicalspace. Intuitively, the FS size must raise with the size anddiversity of the target library. At the same time, the GTMtraining can be very slow or even becomes technicallyimpossible at FS sizes of the order of 105 compounds –which is a very small number compared to today’scommercially accessible compounds, and, especially, to thetheoretically feasible molecules. In order to solve thisproblem, we propose a Parallel GTM algorithm based onthe merging of “intermediate” manifolds constructed inparallel for different subsets of molecules. An ensemble ofthese subsets forms a FS for the “final” manifold. In order toassess the efficiency of the new algorithm, 80 GTMs built onthe FSs of different sizes ranging from 10 to 1.8 M compounds selected from the ChEMBL database. Each GTMwas challenged to build classification models for 712biological activities (depending on the FS size). With thenovel parallel GTM procedure, we could thus cover theentire spectrum of possible FS sizes, whereas previousstudies were forced to rely on the working hypothesis thatFS sizes of few thousands of compounds are sufficient todescribe the ChEMBL chemical space. In fact, this studyformally proves this to be true: a FS containing only 5000randomly picked compounds is sufficient to represent theentire ChEMBL collection (1.8 M molecules), in the sensethat a further increase of FS compound numbers has nobenefice impact on the predictive propensity of the abovementioned 712 activity classification models. Parallel GTMmay, however, be required to generate maps based on verylarge FS, that might improve chemical space cartography ofbig commercial and virtual libraries, approaching billions ofcompounds
Добавил в систему: Баскин Игорь Иосифович

	ИСТИНА	Войти в систему Регистрация
	Интеллектуальная Система Тематического Исследования НАукометрических данных
	Главная Поиск Статистика О проекте Помощь

ИСТИНА

Интеллектуальная Система Тематического Исследования НАукометрических данных

Parallel Generative Topographic Mapping: an Efficient Approach for Big Data Handlingстатья

Прикрепленные файлы