LncDC

LncDC: a machine learning based tool for long non-coding RNA detection from RNA-Seq data.

Pre-requisite

python 3 >= 3.9; pandas == 1.5; numpy == 1.23; scikit-learn == 1.1.3; xgboost == 1.7.1; imbalanced-learn == 0.9.1; biopython >= 1.79; tqdm >= 4.64

ViennaRNA (Optional, only required when using secondary structure features for prediction) Can be installed by:

1. CONDA: conda install -c bioconda viennarna
2. Or install from the official ViennaRNA website: https://www.tbi.univie.ac.at/RNA/

Conda environment

1. Download and install Anaconda. https://docs.anaconda.com/anaconda/install/

2. Create an environment

conda create -n lncdc python=3.9     

Here, 'lncdc' is the name of the conda environment, which can be replaced by any names.

3. Activate the environment (before we run LncDC)

conda activate lncdc

4. Install ViennaRNA for secondary structure features extraction (Optional)

You can install the package by CONDA

conda install -c bioconda viennarna

Or you can install it from the official ViennaRNA website: https://www.tbi.univie.ac.at/RNA/

To confirm that ViennaRNA is properly installed, you can test it by:

python
>>> import RNA

ViennaRNA is successfully installed if there are no error messages poped up.

Installation

1. Download the source code from https://github.com/lim74/LncDC

2. Unzip the package

unzip LncDC-master.zip

3. Go to LncDC-master directory

cd LncDC-master

4. Install LncDC

python setup.py install

If you want to install the required side packages in a certain conda environment, make sure the environment is activated.

Requisite check

python test_requirements.py

Usage and Examples

How to conduct a prediction

python lncDC.py -i input.fa -o output -x hexamer_table.csv -m model.pkl -p imputer.pkl -s scaler.pkl -r -k ss_table -t number_of_threads

-i The inputfile with RNA sequences in fasta format. The fasta file could be regular text file or gzip compressed file (*.gz).
-o The output file that will contain the prediction results in csv format. Long noncoding RNAs are labeled as 'lncrna', and message RNAs are labeled as 'mrna'. Default: lncdc.output.csv
-x (Optional) Prebuilt hexamer table in csv format. Run lncDC-train.py to obtain the hexamer table of your own training data. Default: train_hexamer_table.csv
-m (Optional) Prebuilt training model. Run lncDC-train.py to obtain the model trained from your own training data. Default: XGB_model_SIF_PF.pkl
-p (Optional) Prebuilt imputer from training data. Run lncDC-train.py to obtain the imputer from your own training data. Default: imputer_SIF_PF.pkl
-s (Optional) Prebuilt scaler from training data. Run lncDC-train.py to obtain the imputer from your own training data. Default: scaler_SIF_PF.pkl
-r (Optional) Turn on to predict with secondary structure features. Default: turned off.
-k (Optional) Prefix of the sequence and secondary structure kmer tables. Need to specify -r first. For example, the prefix of secondary structure kmer table file 'mouse_ss_table_k1.csv' is 'mouse_ss_table'. Run lncDC-train.py to obtain the tables from your own training data. Default: train_ss_table
-t (Optional) The number of threads assigned to use. Set -1 to use all cpus. Default value: -1.

Examples for prediction

1. Predict with the default model (human).

cd LncDC-master/
python bin/lncDC.py -i test/human_test_toy.fasta -o lncdc_human.csv -t 8

Here, the human_test_toy.fasta file includes human RNA sequences in fasta format. The prediction results will be stored in the file lncdc_human.csv. 8 threads are used for this prediction.

2. Predict by the default human model with secondary structure features. This may take a longer time because secondary structure calculation by RNAfold (The main program of ViennaRNA) is time-consuming.

cd LncDC-master/
python bin/lncDC.py -i test/mouse_test_toy.fasta -o lncdc_mouse.csv -r -t 8

Here, the mouse_test_toy.fasta file includes mouse RNA sequences in fasta format. The prediction results will be stored in the file lncdc_mouse.csv. -r is specified so the program will perform prediction with secondary structure features added. 8 threads are used for this prediction.

3. Predict with the self-trained model (without secondary structure features).

cd LncDC-master/
python bin/lncDC.py -i test/mouse_test_toy.fasta -o lncdc_mouse.csv -x data/train_hexamer_table.csv -m data/XGB_model_SIF_PF.pkl -p data/imputer_SIF_PF.pkl -s data/scaler_SIF_PF.pkl -t 8

The input file in this prediction is mouse_test_toy.fasta, which contains mouse RNA sequences in fasta format. The output file is lncdc_mouse.csv. The parameter '-x' is applied, and the hexamer table train_hexamer_table.csv is provided. '-m' is used, follewed by the self-trained model XGB_model_SIF_PF.pkl. The self-trained model only used sequence intrinsic and protein features in this case. The imputer file imputer_SIF_PF.pkl and the sclar file scaler_SIF_PF.pkl are provided with parameter '-p' and '-s', respectively. 8 threads are used for the prediction.

4. Predict with the self-trained model (with secondary structure features).

cd LncDC-master/
python bin/lncDC.py -i test/mouse_test_toy.fasta -o lncdc_ss_mouse.csv -x data/train_hexamer_table.csv -m data/XGB_model_SIF_PF_SSF.pkl -p data/imputer_SIF_PF_SSF.pkl -s data/scaler_SIF_PF_SSF.pkl -r -k data/train_ss_table -t 8

The input file is mouse_test_toy.fasta and the output file is lncdc_ss_mouse.csv. In addition to the parameters requried for the prediction without using secondary structure based features, '-r' and '-k' are requried for prediction with secondary structure features. The '-r' parameter will turn on the prediction with secondary structure features, and the '-k' parameter will provide the prefix of the requried train_ss_table(s). For example, the prefix of secondary structure kmer table file 'mouse_ss_table_k1.csv' is 'mouse_ss_table'. The '-t' parameter indicates that the program will use 8 threads for prediction.

How to train a model with my own data

python lncDC-train.py -m mrna.fa -c cds.fa -l lncrna.fa -o output -t number_of_threads -r

-m The file with mRNA sequences in fasta format. The fasta file could be regular text file or gzip compressed file (*.gz). 
-c The CDS sequences of the mRNAs in fasta format. The fasta file could be regular text file or gzip compressed file (*.gz). The order and number of the CDS sequences should be the same as the mRNA sequences.
-l The file with lncRNA sequences in fasta format. The fasta file could be regular text file or gzip compressed file (*.gz).
-o The prefix of the output files, including a hexamer table, a prediction model, an imputer and a scaler. If the '-r' parameter turned on, the output files will also include secondary structure kmer tables.
-t (Optional) The number of threads assigned to use. Set -1 to use all cpus. Default value: -1.
-r (Optional) Turn on to train a model with secondary structure features. This will generate secondary structure kmer tables. Default: turned off.

FASTA format example:

>id or name for transcript 1
AGGGCCAACGAACGCAACACAGGGACATGGGGGACAGAGAGGAATGTCTCTCTACCCCCCAACC
CCCCATGTCTGTGGTGAAGTCGATCGAATTAGTGCTGCCCGAGGATAGAATCTACCTGGGGACC
CCATACTGGCTCCAGCATAAAGGGCAGGTGATCTTAACCCTGAACA
>id or name for transcript 2
AACAGCACCCTGGTGGACCCCATAAGGGCCAACGAACCGGGAATTCCCCCCAACCCCCCATGTC
CGAATTAGTGCTGCCCGAGGATAGAATCTACCTGGCTGGCTCCAGCATAAAGGGCAGGTGATCT
AAGAATTGCAACAAC

NOTE: Suppose you use LncDC to predict lncRNAs from non-model organisms and don't have enough well-annotated lncRNAs for model training. In that case, you could train a model with the data from evolutionary closed model organisms, such as Zebrafish, Yeast, Soybean, Drosophila melanogaster, Caenorhabditis elegans, Arabidopsis thaliana, Oryza sativa, and etc.

Examples for training models

1. Train a model (No secondary structure features) using mouse data

cd LncDC-master/
python bin/lncDC-train.py -m mrna_mouse.fasta -c cds_mouse.fasta -l lncrna_mouse.fasta -o self_mouse -t 8

To train a model with mouse data, we need to provide the mRNA and lncRNA sequences, respectively. The mRNA sequences are stored in the mrna_mouse.fasta file, and their corresponding CDS sequences are stored in the cds_mouse.fasta file. The lncrna_mouse.fasta file includes the long noncoding RNA sequences. 'self_mouse' is the prefix of the output files, which include four files: self_mouse_hexamer_table.csv, self_mouse_xgb_model_SIF_PF.pkl, self_mouse_imputer_SIF_PF.pkl and self_mouse_scaler_SIF_PF.pkl. We set '-t' to 8 so that there are 8 threads will be used for the model training.

2. Train a model (with secondary structure features) using mouse data. This may take a longer time because secondary structure calculation by RNAfold (The main program of ViennaRNA) is time-consuming.

cd LncDC-master/
python bin/lncDC-train.py -m mrna_mouse.fasta -c cds_mouse.fasta -l lncrna_mouse.fasta -o SS_mouse -r -t 8

To train a model with secondary structure features, we only need to add the '-r' parameter. In addition to the outputs with a prefix 'SS_mouse' that are similar to the first example, five ss tables will also be generated, including SS_mouse_ss_table_k1.csv, SS_mouse_ss_table_k2.csv, SS_mouse_ss_table_k3.csv, SS_mouse_ss_table_k4.csv and SS_mouse_ss_table_k5.csv.

Authors

Minghua Li

Contact

lim74@miamioh.edu

Cite this article

Li, M., Liang, C. LncDC: a machine learning-based tool for long non-coding RNA detection from RNA-Seq data. Sci Rep 12, 19083 (2022). https://doi.org/10.1038/s41598-022-22082-7

License

This project is licensed under the MIT License. Copyright (c) 2020 lim74