This repository contains models for predicting miRNA-mediated repression described in our paper.
- python 3.6 or higher
- python packages listed in requirements.txt (swap out tensorflow-gpu with tensorflow if not using GPUs). To install, run
pip install -r requirements.txt. - RNAplfold from the ViennaRNA package
This module contains code for building and training the combined CNN and biochemical models in train.py and monitoring the progress using Tensorboard. The parse_data_utils file contains helper functions for parsing tfrecords information and assembling it into the correct input format. The function models.seq2ka_predictor() builds the CNN for predicting KD from miRNA and target sequences.
Here is an example of how to predict relative KD values using the trained model from our paper:
python cnn/generate_12mer_kds.py \
--name all \
--mirdata sample_data/inputs/mirseqs.txt \
--mirlen 10 \
--passenger \
--load_model cnn/trained_model/model-100 \
--outfile sample_data/outputs/kds/MIR_kds.txt
This module folds target sites in many different sequence contexts to calculate the basal accessibility of each site. We recommend using the only_canon flag to only calculate this value for canonical sites and avoid very long compute times.
First partition 12-nt kmers into 10 files:
python rnaplfold/partition_seqs.py \
--mirseqs sample_data/inputs/mirseqs.txt \
--nbins 10 \
--outdir sample_data/outputs/SA_background/sequences \
--only_canon \
--passenger
Then generate 200 random contexts for each sequence in each file and fold them. You may want to use a Makefile or snakemake to automate this step:
for mirname in mir122_pass mir133_pass ; do \
for ix in 0 1 2 3 4 5 6 7 8 9 ; do \
bsub -R "rusage[mem=4096]" \
python rnaplfold/get_SA_bg.py \
--sequence_file sample_data/outputs/SA_background/sequences/canon_"$mirname"_"$ix".txt \
--temp_folder sample_data/outputs/SA_background/bg_vals/canon_"$mirname"_"$ix"_TEMP \
--num_bg 200 \
--num_processes 24 \
--outfile sample_data/outputs/SA_background/bg_vals/canon_"$mirname"_"$ix"_bg_vals.txt; \
done; \
done
Finally, parse RNAplfold outputs for all background sequences. Because the contexts are random, the final results will be very slightly different each time.
python rnaplfold/combine_results.py \
--mirseqs sample_data/inputs/mirseqs.txt \
--nbins 10 \
--num_bg 200 \
--infile_seqs sample_data/outputs/SA_background/sequences/canon_MIR_IX.txt \
--infile_bg sample_data/outputs/SA_background/bg_vals/canon_MIR_IX_bg_vals.txt \
--outfile sample_data/outputs/SA_background/bg_vals_processed/canon_MIR_bg_vals.txt \
--passenger
The sequence and partitioned bg_vals can be deleted at this point.
This module contains code that preprocesses data for both the biochemical model and the CNN. For best results, supply PCT scores.
First, navigate to temp folder and fold ORF + UTR3 sequences using RNAplfold
mkdir sample_data/outputs/rnaplfold/TEMP
cd sample_data/outputs/rnaplfold/TEMP
RNAplfold -L 40 -W 80 -u 15 < ../../../../sample_data/inputs/orf_utr3.fa
Then, navigate back and process results for easier querying later
python rnaplfold/process_mRNA_folding.py \
--transcripts sample_data/inputs/transcripts.txt \
--indir sample_data/outputs/rnaplfold/TEMP \
--outdir sample_data/outputs/rnaplfold/rnaplfold_orf_utr3/
The TEMP files can be deleted at this point. To calculate all features:
for mirname in mir122 mir133 mir122_pass mir133_pass ; do \
python get_features/write_sites.py \
--transcripts sample_data/inputs/transcripts.txt \
--mir "$mirname" \
--mirseqs sample_data/inputs/mirseqs.txt \
--kds sample_data/outputs/kds/"$mirname"_kds.txt \
--sa_bg sample_data/outputs/SA_background/bg_vals_processed/canon_"$mirname"_bg_vals.txt \
--rnaplfold_dir sample_data/outputs/rnaplfold/rnaplfold_orf_utr3/ \
--pct_file sample_data/inputs/pcts.txt \
--overlap_dist 12 \
--upstream_limit 15 \
--outfile sample_data/outputs/features/"$mirname".txt ; \
done
This module contains code for building, training, and using the biochemical model and biochemical+ models.
To use our trained parameters to predict scores:
for mirname in mir122 mir133 ; do \
python biochem_model/predict.py \
--features sample_data/outputs/features/"$mirname".txt \
--features_pass sample_data/outputs/features/"$mirname"_pass.txt \
--model biochem_model/trained_models/biochemplus.json \
--freeAGO -6.5 \
--freeAGO_pass -7.5 \
--outfile sample_data/outputs/predictions/"$mirname".txt ; \
done
The predictions will change slightly for an mRNA depending on the cohort of all other mRNAs because the structural accessibility score for noncanonical sites of a miRNA is determined by the average value of all canonical sites in the given mRNAs. If you chose to calculate background values for all sites of a miRNA instead, this would no longer be true.