A hybrid approach for predicting multi-label subcellular localization of mRNA at genome scale

Abstract

In the past, number of methods have been developed for predicting single label subcellular localization of mRNA in a cell. Only limited methods had been built to predict multi-label subcellular localization of mRNA. Most of the existing methods are slow and cannot be implemented at transcriptome scale. In this study, a fast and reliable method had been developed for predicting multi-label subcellular localization of mRNA that can be implemented at genome scale. Firstly, deep learning method based on convolutional neural network method have been developed using one-hot encoding and attained an average AUROC - 0.584 (0.543 - 0.605). Secondly, machine learning based methods have been developed using mRNA sequence composition, our XGBoost classifier achieved an average AUROC - 0.709 (0.668 - 0.732). In addition to alignment free methods, we also developed alignment-based methods using similarity and motif search techniques. Finally, a hybrid technique has been developed that combine XGBoost models and motif-based searching and achieved an average AUROC 0.742 (0.708 - 0.816). Our method - MRSLpred, developed in this study is complementary to the existing method. One of the major advantages of our method over existing methods is its speed, it can scan all mRNA of a transcriptome in few hours. A publicly accessible webserver and a standalone tool has been developed to facilitate researchers (Webserver: https://webs.iiitd.edu.in/raghava/mrslpred/).

A review from biological mapping to computation-based subcellular localization

“… 56 Shubham – XGBoost 2023 Choudhury et al. 57

Figure 3 Published timeline of protein and RNA subcellular localization models (DL RNA models (deep learning RNA models), ML RNA models (machine learning RNA models), ML protein models (machine learning protein models), DL protein models (deep learning protein models)).

…”

“…Choudhury et al. 57 evaluated both models on the basis of machine learning and deep learning, and finally determined on a hybrid technique that combines the XGBoost model and subject search. Short time is one of the main characteristics of this model.…”