Deep learning ~11+ a new perspective on disease-related miRNA research

Introduction

Validated in the database miRNA- The disease association is seriously inadequate , Using traditional biological experimental methods to identify new miRNA- The cost of disease association is high and has certain blindness .

Background introduction

at present , Deep learning has been widely used in the study of various mechanisms of disease , This article is brought to you by Xiaobian today , This paper presents an automatic encoder （DFELMDA） A new computational method for deep forest ensemble learning to predict miRNA Association with disease . Article on 2022 It was published in 《Briefings in Bioinformatics》 On , The influence factor is 11.622, The title of the article is ：Identification of miRNA–disease associations via deep forest ensemble learning based on autoencoder.

Data is introduced

The data sets used in this study are from HMDD v2.0 Database download , The data includes 495 Kind of miRNA、383 It's a disease and 5430 An experimental miRNA- Disease related .

Result analysis

01

Deep forest ensemble learning model based on self encoder ‍

In this study , Came up with a DFELMDA To predict the miRNA Association with disease .DFELMDA There are three main steps ：（i） A new feature representation strategy is proposed , To get the same miRNA- Training models with different representations of disease associations ,（ii） be based on miRNA And disease build two deep self encoders , Used to extract low dimensional feature representation ‍（iii）RF Two types of miRNA- Disease related , And incorporated into the final results ‍.DFELMDA Flow chart of 1 Shown .

chart 1

02

Self encoder training

In this study , Trained two with the same structure （ chart 2） Automatic encoder for , For from miRNA And the low dimensional representation of extracted features from diseases . say concretely , Model training involves two processes ： Encoding and decoding . In the coding phase , Two types of miRNA- The high-dimensional feature representation of disease association is fed to the encoder , To compress features and reduce dimensions . In the decoding phase , The decoder attempts to represent the low dimension H Revert to the same appearance as the input feature representation .

chart 2

03

adopt RF forecast miRNA Association with disease

In order to avoid the influence of feature dimension and feature quality on miRNA- Adverse effects of disease association prediction , This study chooses RF As a classifier . In this study , The experimental data set is output by the automatic encoder 128 The set of dimensional eigenvectors represents . Given training data , Steps are as follows ：1） Sample several samples from the training set in the form of putting back , Conduct K Sub sampling , Training out K A classified regression tree （CART） Decision tree .2） The optimal partition variable calculated by Gini coefficient , Build through node splitting CART Decision tree .3） By repeating the previous steps K Times obtained K individual CART Decision tree .4) according to K CART The result of the decision tree , Predict by most rules miRNA- Disease related .RF The schematic diagram of the is shown in Figure 3 Shown .

chart 3

04

Methods to compare

Cross validation experiments were conducted in this study , Use conventional indicators to study DFELMDA Performance of , And implement case studies to further evaluate DFELMDA forecast miRNA- The ability to relate to disease . To evaluate DFELMDA In discovering potential miRNA- Excellent performance in disease association , take DFELMDA With several advanced methods （TCRWMDA、RLSMDA、 Based on kernel ridge regression miRNA- Disease association prediction （EKRRMDA）、 Improved collaborative filtering miRNA- Disease association prediction （ICFMDA） And for miRNA A graphic automatic coder model for Disease Association prediction （GAEMDA）） Made a comparison .

DFELMDA Realized 5 times CV Of ROC The curve is as shown in the figure 4 Shown . Obviously , Compared with the other five methods ,DFELMDA stay AUC Aspect has the best performance . stay 5 times CV in ,DFELMDA Of AUC achieve 0.9552, And three-layer heterogeneous networks combined with unbalanced random walk MiRNA- Disease association prediction algorithm (TCRWMDA)、RLSMDA、 Integration based on kernel ridge regression MiRNA- The cause of disease AUC Correlation prediction (EKRRMDA)、 Improved collaborative filtering miRNA- Disease association prediction (ICFMDA) And for miRNA- Graph auto encoder model for Disease Association prediction (GAEMDA) Respectively 0.9208、0.8737、0.9307、0.9043 and 0.9353.

chart 4

For further verification DFELMDA The ability of , This study conducted 10 times CV . Pictured 5 Shown ,DFELMDA achieve 0.9560 The average of AUC, namely 10 times CV The average value of is 0.9584、0.9581、t0.9614、0.9628、0.9582、0.9502、0.9582、0.9567、0.9571 and 0.9532.

chart 5

05

Comparison with different classifier models

To further evaluate the performance of this method , This study compares it with four different classification models [ Decision tree 、KNN、 Naive Bayes and deep neural networks (DNN)] Made a comparison . result , Decision tree 、KNN、 Naive Bayes and DNN Got AUC Respectively 0.9150、0.9285、0.9222 and 0.9285. Of different classifier models ROC The curve is as shown in the figure 6 Shown .

chart 6

06

Case study

For further proof DFELMDA Identifying new miRNA- Accuracy in Disease Association , This model is implemented in case studies of complex human diseases , From HMDD A tumor of the colon (CNs)、 Lung tumor (LNs) And breast cancer (BNs) . Known from the database miRNA- Disease association as DFELMDA Training set of , According to the prediction results, the candidate of the studied disease miRNA Prioritize . thereafter , stay HMDD、dbDEMC and microRNA Cancer Society database （miRCancer） Select from the database before 50 Candidates miRNA And check them one by one . surface 1 For in CN、LN and BN Found in 10 position miRNA .

surface 1

This study also selected 14 Plant and more miRNA Related specific diseases . As shown in the table 2 Shown ,DFELMDA Achieved considerable AUC value , Especially Barrett movement and vascular disease ‍,AUC Respectively 0.9579 and 0.9670. in summary , It is not difficult to see from the above results ,DFELMDA Proven ability in cross validation and case studies .

surface 2

Editor's summary

In this study, we developed a method of DFELMDA To infer miRNA- Disease related . First , A new feature representation strategy is applied to obtain the same miRNA- Different types of representations of disease associations （ come from miRNA And diseases ）. then , Two models based on miRNA And disease depth self coder to extract low dimensional feature representation . Last , adopt RF Predict two types of miRNA- Disease related , And combine them into the final result . Experimental results and case studies show that ,DFELMDA Is a powerful computing tool , Can be used for new miRNA- Disease association prediction .