siWalk

SiWalk identifies effector siRNA positions within a given precursor sequence using a Random Forest classifier trained on structural and sequence features. It operates on sequence alone — no expression data required.

Citation

Please cite the paper for your work using this tool.

Title: Localization of Phased Effector siRNAs within Precursor Sequences Using Feature-Informed Machine Learning

Cite: Chao-Jung Wu, Abdoulaye Baniré Diallo

BibTex:

@article{... ...
}

---

Quick Start: Predict effector siRNA from a precursor sequence

This is the primary functionality of siWalk. Given a precursor RNA or DNA sequence, siWalk scans it to identify the most likely effector siRNA position.

1. Set up the environment

Download siWalk from GitHub: github/bioinfoUQAM/siWalk.

Option A — Local machine:

Install ViennaRNA (provides RNAfold), then:

pip install numpy pandas scipy statsmodels scikit-learn matplotlib
chmod +x /path/siWalk/lib/miranda

Option B — HPC (Compute Canada):

cd /path/siWalk/workdir/
module load StdEnv/2020 viennarna/2.5.1 gcc/9.3.0
module load blast+/2.14.0 samtools/1.17 bowtie/1.3.0 scipy-stack
chmod +x ../lib/miranda
virtualenv --no-download env
source env/bin/activate
pip install --no-index --upgrade pip
pip install --no-index statsmodels sklearn

2. Run the example

The script siWalk_predict_siRNA_location.py takes two arguments: a precursor sequence and DicerCall (the expected siRNA length, typically 21).

Example

cd /path/siWalk/src/

priseq=CTTGACCTTGTAAGGCCTTTTCTTGACCTTGTAAGACCCCATCTCTTTCTAAACGTTTTATTATTTTCTCGTTTTACAGATTCTATTCTATCTCTTCTCAATATAGAATAGATATCTATCT
DicerCall=21
python siWalk_predict_siRNA_location.py $priseq $DicerCall

Note: Both DNA and RNA sequences are accepted. For best performance, limit precursor sequences to ≤ 200 nucleotides.

3. Expected output

The predicted position is printed to stdout:

yourPrecursor, predicted siRNA start: 75, end: 93, score: -154.733188

The stdout shows the top-ranked candidate. The full list of 6 candidates can be found in yourPrecursor.effector_localization_top6_recommendation.tsv.

Output files are written to /path/siWalk/output/<timestamp>/:

File	Description
yourPrecursor.effector_localization_indication.tsv	Scores for all candidate positions
yourPrecursor.effector_localization_top6_recommendation.tsv	Top 6 ranked siRNA candidates
yourPrecursor.effector_localization_indication.png	Graphical summary of candidate scores

The top-6 recommendation file has the following columns:

Column	Description
Position p	Candidate start position on the precursor
Start S(p)	Start-position score (sum of weighted indications)
End E(p)	End-position score
Sum of Indications	Combined score (higher is better)
Best length	Predicted siRNA length (nt)
End position	Predicted end position
ML predicted	Whether this position was flagged by the ML classifier

---

Dependencies

Languages: python/3.10.2 (scipy, statsmodels, sklearn, numpy)

Packages: spark/3.3.0, bowtie/1.3.0, samtools/1.17, viennarna/2.5.1, perl/5.30.1, blast+/2.14.0

Distributed with siWalk: miRanda, miRCheck

Environments: The localization mode (Quick Start above) runs on any local machine or login node. Modules A and X (precursor prediction and model training) require a Spark-enabled server (developed on Compute Canada Narval).

Supporting Files

File	Description
dbs/demoOnly_background.tsv	Small subset of the training set for quick testing
dbs/background.tsv	Full Arabidopsis training set (185,539 segment samples); download from Zenodo (see below)
dbs/TAIR10_genomic.fa	Arabidopsis genome; download from Zenodo (see below)
dbs/mature.fa	miRNAs from miRBase v22.1
dbs/GSM1087998_C0RC1.txt	Example sRNA-seq library in readcount format (GSE44622)
dbs/feature_definition.txt	Definitions of all variables in the analysis
dbs/complete_feature_list.tsv	Full feature list for reference
model/GBA100.pkl	Gradient Boosting + AdaBoost model, top 100 all-type features
model/GBAs100.pkl	Gradient Boosting + AdaBoost model, top 100 structural features
model/RFAs100.pkl	Random Forest + AdaBoost model, top 100 structural features; download from Zenodo (see below)
model/Arabidopsis_feature_importance_n_correlation.tsv	Feature evaluation file paired with GBA100.pkl
model/Arabidopsis_structure_feature_importance_n_correlation.tsv	Feature evaluation file paired with GBAs100.pkl and RFAs100.pkl

The following large files can be downloaded from Zenodo and placed as indicated:

curl -L -O "https://zenodo.org/records/17258908/files/TAIR10_genomic.fa?download=1"  # → dbs/TAIR10/
curl -L -O "https://zenodo.org/records/17258908/files/background.tsv?download=1"     # → dbs/
curl -L -O "https://zenodo.org/records/17258908/files/RFAs100.pkl?download=1"        # → model/

---

Auxiliary Modules

The following modules require a Spark-enabled HPC environment.

Generate the RFAs100 model

Alternatively to downloading, generate it locally (~30 minutes):

cd /path/siWalk/src/
annotated_training_file=../dbs/background.tsv
python siWalk_pickle_localization.py $annotated_training_file

Rename the resulting file to RFAs100.pkl and move it to /path/siWalk/model/.

Module A. Precursor prediction from sRNA-seq libraries

A1. Create a SAM file

Align an sRNA-seq library to the genome using bowtie:

genome_file=/path/TAIR10_genomic.fa
index_prefix=/path/bowtie_index/TAIR10_genomic
bowtie-build -f --quiet $genome_file $index_prefix
samtools faidx $genome_file

path_siWalk=/path/siWalk
filename=GSM1087998_C0RC1
infile=$path_siWalk/dbs/$filename.txt   # in readcount format
raw=$path_siWalk/input/$filename.raw
python $path_siWalk/src/rc2raw.py $infile > $raw

samfile=$path_siWalk/input/$filename.sam
bamfile=$path_siWalk/input/sorted_$filename.bam
bowtie -r $raw -x $index_prefix --threads 64 -v 0 --mm -a --sam --no-unal > $samfile
samtools sort $samfile > $bamfile
samtools index $bamfile

A2. Generate features

Edit submit_siWalk_precursor_features.sh to set:

INPUT_SAM_PATH=/path/siWalk/input
genome_file=/path/TAIR10_genomic.fa

Submit the job from the workdir folder:

cd /path/siWalk/workdir/
sbatch submit_siWalk_precursor_features.sh

Output files will be available in /path/siWalk/output/$jobID/$filename/. The main output is the annotation file: $filename.contig_features.tsv.

A3. Predict precursor segments

cd /path/siWalk/src/
annotation_file=/path/siWalk/output/$jobID/$filename/$filename.contig_features.tsv
python siWalk_classify_precursors.py $annotation_file

The output prefix.prediction.tsv includes these key columns:

Column	Description
CONTIG	5250-nt genomic region (e.g., 3__5860000_5865250)
eff_strand	Strand of the most abundant effector siRNA
eff_seq	Sequence of the most abundant effector siRNA
segment	Predicted start and end of the siRNA-generating locus
Predicted_Class	True if predicted as siRNA precursor, False otherwise

Module X. Train a machine learning model

X1. Prepare the training dataset

• Compute features for each precursor candidate using Module A.
• Add a consistent column: True for confirmed precursors, False for non-precursors.
• The resulting file is the annotated_training_file for the next step.

X2. Generate the model

Top 100 all-type features (produces the GBA100-type model, for precursor prediction only):

cd /path/siWalk/src/
annotated_training_file=../dbs/background.tsv
python siWalk_pickle_precursor.py $annotated_training_file

Top 100 structural features (produces the RFAs100-type model, suitable for both precursor and effector prediction):

cd /path/siWalk/src/
annotated_training_file=../dbs/background.tsv
python siWalk_pickle_localization.py $annotated_training_file

Each run produces:

• output/prefix_*_model.pkl — trained model with timestamp
• output/prefix_*_feature_importance_n_correlation.tsv — ranked feature importances and correlations