RNAincoder: a deep learning-based encoder for RNA and RNA-associated interaction

Abstract

Ribonucleic acids (RNAs) involve in various physiological/pathological processes by interacting with proteins, compounds, and other RNAs. A variety of powerful computational methods have been developed to predict such valuable interactions. However, all these methods rely heavily on the ‘digitalization’ (also known as ‘encoding’) of RNA-associated interacting pairs into a computer-recognizable descriptor. In other words, it is urgently needed to have a powerful tool that can not only represent each interacting partner but also integrate both partners into a computer-recognizable interaction. Herein, RNAincoder (deep learning-based encoder for RNA-associated interactions) was therefore proposed to (a) provide a comprehensive collection of RNA encoding features, (b) realize the representation of any RNA-associated interaction based on a well-established deep learning-based embedding strategy and (c) enable large-scale scanning of all possible feature combinations to identify the one of optimal performance in RNA-associated interaction prediction. The effectiveness of RNAincoder was extensively validated by case studies on benchmark datasets. All in all, RNAincoder is distinguished for its capability in providing a more accurate representation of RNA-associated interactions, which makes it an indispensable complement to other available tools. RNAincoder can be accessed at https://idrblab.org/rnaincoder/

Real-World Molecular Out-Of-Distribution: Specification and Investigation

“…Although our experiments have focused on small molecules, the principles explored herein can be extended to other drug modalities such as RNA, protein, and macrocycles, which encounter the same OOD problems. We might need to consider other representations − and other splits, but the overall idea of comparing deployment to the training distribution can give us more realistic expectations of model generalization in those modalities. The focus on realistic shifts rather than hypothetical ones offers a practical framework to benchmark algorithmic progress.…”

Evaluating Machine Learning Methods of Analyzing Multiclass Metabolomics

“…After obtaining the raw data set from metabolite measurements, there are various data manipulation steps used in multiclass metabolomic methods, consisting of data filtering, imputation of missing values, data normalization, marker identification, sample separation, classification, and so on. , Among these analytical chains, there are several to dozens of machine learning methods that can be applied in each step. , For instance, the methods of data filtering include the “80% rule”. There are many imputation methods of missing values consisting of KNN ( k -nearest neighbor) and so on.…”

“…18,19 Among these analytical chains, there are several to dozens of machine learning methods that can be applied in each step. 20,21 For instance, the methods of data filtering include the "80% rule". There are many imputation methods of missing values consisting of KNN (k-nearest neighbor) and so on.…”