Interpretation of the paper: a convolutional neural network for identifying N6 methyladenine sites in rice genome using dinucleotide one hot encoder

Article address ：https://www.sciencedirect.com/science/article/abs/pii/S0925231220315137
DOI：https://doi.org/10.1016/j.neucom.2020.09.056
Periodical ：Neurocomputing（ Two district ）
Influencing factors ：5.719
Release time ：2020 year 9 month 2 Japan
Web The server ：http://iRicem6A-CNN.aibiochem.net
data ：1760 sample Download link ;154000 sample Download link

1. An overview of the article

N6- Methyladenine （N6-Methyladenine,m6A） Is one of the important epigenetic modifications , With all kinds of DNA Process control . Through traditional methods, the whole genome m6A Analysis is the foundation , But it will take a long time . The author puts forward a new scheme ：iRicem6A-CNN, It is used to identify m6A site , This scheme uses dinucleotide （2-mer）One-hot Coding technology , The input tensor is generated by convolution neural network for prediction , Five times the prediction accuracy of cross validation and independent testing （ACC） Respectively reached 93.82% and 96.19% , Performance is better than other available predictors . Experimental results show that , Only based on dinucleotide One-hot Of iRicem6A-CNN It can show high performance , And compared with single nucleotide （1-mer）One-hot The model is more stable 、 More robust performance .

2. background

N6- Methyladenine （N6-Methyladenine,m6A） yes DNA An important chemical modification product of , It widely exists in various organisms from eukaryotes to prokaryotes , And with DNA Copy 、DNA Repair is related to transcriptional regulation .DNA Methylated genome analysis has become the next generation sequencing technology , In particular, single molecule real-time sequencing technology is more and more widely used .m6A The genome-wide distribution of loci has been better characterized , This has led to a better understanding of its biological function . for example , Genome wide m6A Site studies revealed m6A Different regulatory functions in different eukaryotes , And show that in prokaryotes m6A As a sign , Used to distinguish intruders DNA And hosts DNA.
2018 year ,zhou Used by others smrt It is proved that 0.2% Adenine of is m6A Methylated , Since this discovery , Various machine learning based computational rice genomes m6A Methods have sprung up .2019 year ,chen And others have developed im6A-Pred, One is based on support vector machine （SVM） Methods , This method is based on including 1760 The benchmark data set of samples is trained , Accuracy rate （ACC） achieve 83.13% . And then , In traditional machine learning algorithms ( Such as svm、 Random forests （RF） And Markov chain model （markov chain model） Other methods have been developed to identify the rice genome m6A site , These methods include ：im6A-DNCP、MM-m6Apred、SDM6A、iN6-methylat and iDNAm6A-rice, among iDNAm6A-rice Of ACC The highest , by 91.7% .
In this study , The author developed a new method ：iRicem6A-CNN, To improve the prediction of rice genome m6A Accuracy of loci . The author used a dinucleotide One-hot code , take DNA Sequence is transformed into tensor , Then enter a well-designed CNN Model optimization . stay 5 Times cross validation ACC by 93.82% , In independent testing ACC by 96.19%. Experimental results show that , Use binary One-hot Coded iRicem6A-CNN Than single nucleotide One-hot Coded iRicem6A-CNN It has higher robustness and accuracy . The result of index comparison shows ,iRicem6A-CNN It has good performance , This is not only because it can stably identify positive samples , Also because it can identify negative samples more accurately .

3. data

By chen Et al lv Et al. Established two widely used rice genomes m6A Benchmark data set , They are marked as Chen-rice-m6A and Lv-m6A-rice.Chen-rice-m6A Data set from 1760 Sample composition , Half of them are positive samples , The other half is negative , It has been widely used in the reporting model based on non deep learning algorithm .Lv-m6A-rice Data set from 154000 Positive samples and 154000 Negative samples , And be lv Others used it iDNAm6A-rice On the way , By Yu Others used it SNNRicem6A On the way . The sequence length in the two data sets is 41 Base （bp）, There is an adenine in the center （A）. The author considers CNN The model needs a lot of data , So using Lv-m6A-rice Data set for model training , use Chen-rice-m6A Data sets for independent testing , To facilitate comparison .

4. Method

4.1 Feature code

4.2 Model framework

5. result

5.1 Model comparison based on different encoders

5.2 Compare with the most advanced models

6. Conclusion

ad locum , For rice genome m6A Locus developed a new method based on deep learning ：iRicem6A-CNN. This method inputs DNA The sequence is first converted to a dinucleotide One-hot Coding tensor . The author proved by experiment , Dinucleotide One-hot The performance of the coding model is better than that of the one split one hot coding model , And it shows stronger robustness under different prediction probability thresholds . The model is applied to rice genome m6A Site detection , It turns out that , This model has high 5 Times the accuracy of cross validation (93.82%) And independent test accuracy (96.19%) , It's the rice genome m6A One of the best predictors of loci . The author's analysis and comparison show that ,iRicem6A-CNN It can not only accurately predict m6A Positive samples , And it can reduce the error rate of negative sample recognition . Besides , Also for iRicem6A-CNN It provides a user-friendly web server .