Sentences, Words, Attention: A “Transforming” Aphorism of miRNA Discovery
miRNAs are prime regulatory small RNAs (sRNAs) having ~21 bases length which post-transcriptionally regulate most of the genes and stand critical for most of the processes of eukaryotic cells including their development and specialization. miRNAs can be found in the intronic as well as intergenic regions. Mature miRNAs are derived from longer precursor miRNA molecules (pre-miRNAs) which are double-stranded RNAs (dsRNAs) with terminal hairpin loop. miWords is a novel deep learning method specifically designed for pre-miRNA detection. The model is an attention based genomic language processing transformer and context scoring deep-learning approach to accurately identify pre-miRNAs in plants that can automatically learn suitable features from the raw data, without manual feature engineering.
miWords can be used directly from this web server. This server implements trained models and can process both individual sequences or fasta files. The server generates a table deciphering T-Score for every sequence provided.
The latest version of the package can be downloaded from the GitHub repository.
1. Python3.8.1 or higher
2. Numpy
3. keras
4. tensorflow
5. plotly
6. pandas
7. RNAfold (ViennaRNA package download it from here: ("https://www.tbi.univie.ac.at/RNA/)
8. python module multiprocessing, Bio, bayesian-optimization, RNA, viennarna, XGBoost 1.6.2 version or higher
9. bedops
10. Bedtools
1. M1.sh = Module 1 execution script. (Sequence having length less than 400 base)
2. M2.sh = Module 2 execution script. (Sequence having length more than 400 base)
3. miWords.py = Python script for detecting pre-miRNAs from sequences provided.
4. miWords.h5 = Trained model transformer part.
5. miWords.sav = Trained model XGBoost part.
6. test = dummy fasta sequence for Module 1.
7. genomic_sequence = dummy fasta sequence for Module 2.
8. tokenizer_penta.pickle = build tokenizer.
9. reve.py = Python script for sequence preprocessing.
10. hyper_param.py = Python script build model implementing hyperparameter tuning
11. file_format_GPU = file format of input for script predict_GPU.py. file containing seq_id, sequence (sequence length of >= 200 base), and dot bracket ("(">"M", ".">"O", ")">"N"). All in one line separated by tab for a single instance.
12. score.h5 = scoring profile based CNN trained model
13. bimodal.h5 = scoring and rpm profile based bimodal CNN trained model (Download form (https://scbb.ihbt.res.in/miWords/bimodal.h5)
python3 hyper_param.py file_for_tuning
file_for_tuning: file containing label (0/1), sequence (sequence containing pre-miRNAs and non-pre-miRNAs), and dot bracket ("(">"M", ".">"O", ")">"N"). All in one line separated by tab for a single instance.
To detect the pre-miRNAs, In parent directory execute following command:
Sequence having length less than 400 base module 1:
sh M1.sh <folder path> <fasta file>
eg: sh M1.sh folderpath test
Sequence having length more than 400 base module 2:
sh M2.sh <folder path> <fasta file> <Option A>
eg: sh M2.sh folderpath genomic_sequence A
To generate line plot, execute the following command:
python3 make-plot.py seq1.csv (filename) (generated plots are interactive)
pre-miRNA detection module gives output in following format
1. seq.txt = Chunks wise probability score of the sequence provided.
2. plot = folder containing "csv" files to construct line plot.
3. merge = folder containing "csv" files to construct line plot overlapping sequence.
4. param.txt = Optimized hyparameters for Transformers.
5. merge.txt = output of module 2 (Sequence, Secondary Struture, Position wise T-Score)
6. genomic_sequence.bed = output of module 2 (BED6 format)
7. sequence_feature.tsv = Classification result of the sequence provided.
If provided with read data(sRNA-seq):
User needs to process the fastq file into bed6 format by instructions mentioned below (Note: User can select their own mapping software)
Step 1: Build Genome: hisat2-build [options]* <reference_in> <ht2_index_base>
Step 2: Mapping: hisat2 [options]* -x <ht2-idx> {-1 <m1> -2 <m2> | -U <r>} [-S <sam>]
Step 3: To convert SAM to BED6 format: sam2bed <read.sam |cut -f1-6 >read.bed
if multiple condition are available then merge them into one bed file: cat *.bed >read.bed
For bimodal CNN
sh M2.sh <folder path> <fasta file> <Option B> <read data in BED6 format>
eg: sh M2.sh folderpath genomic_sequence B read.bed
1. read.bed = read alignment data in BED6 format (converted from SAM) (generated by user).
2. genomic_sequence = dummy fasta sequence utilized for Bimodal CNN
3. genomic_sequence.bed = output from Bimodal CNN (BED6 format)
4. genomic_sequence_rpm.txt = Final result (ID, Start, End, Strand, Sequence, Secondary Struture)
To plot:
python3 make-plot_rpm.py seq1.csv (filename) (generated plots are interactive)
Citation: Gupta S, Shankar R* (2023) miWords: Transformers based composite deep-learning for highly accurate discovery of pre-miRNA regions across plant genomes. Briefings in Bioinformatics, 2023. https://doi.org/10.1093/bib/bbad088