REPRESS: A Deep Learning Model for Cell-Type-Specific Post-Transcriptional Gene Regulation

REPRESS (Regulatory Element PRediction of post-transcriptional Events using Sequence Signals) is a deep learning model that predicts cell-type-specific microRNA (miRNA) binding and mRNA degradation directly from RNA sequence.

Trained on millions of sites from AGO2-CLIP, miR-eCLIP, and Degradome-Seq data across multiple human and mouse cell types, REPRESS learns rich representations of post-transcriptional gene regulation. It captures both canonical and non-canonical miRNA binding, integrates sequence context and site multiplicity, and generalizes across tissues and modalities. The version of REPRESS released in this repository has 8356199 parameters with the list of AGO2-CLIP, miR-eCLIP and Degradome-Seq cell lines shown in the Published Datasets used for model training section below.

Key Features

• Predicts miRNA binding and mRNA degradation at single-nucleotide resolution
• Captures cell-type-specific regulatory signals using endogenous sequence features
• Learns non-canonical targeting rules and contextual repression mechanisms
• Outperforms prior methods on 7 orthogonal benchmarks, including:
  • Variant effect prediction on miRNA binding
  • Out-of-distribution generalization to reporter assay data
  • Discovery of novel repressive elements across transcripts

Applications

• Functional annotation of non-coding variants
• Mechanistic insights into miRNA and RBP interactions
• Rational design of RNA therapeutics

Model Architecture

Prerequisites

• Anaconda or Mamba (Python ≥3.8)
• Optional (for GPU support):
  • NVIDIA GPU with a driver that supports CUDA ≥11.2
  • Note: The CUDA toolkit and cuDNN libraries are automatically installed in the Conda environment; no system-wide CUDA installation is needed.
• Other Tools
  • wget (used to download initial data).

Steps to setup environment

The following commands will create a Conda environment, install required packages, and set up the project. GPU support will be enabled automatically if compatible hardware and drivers are present.

Note: The commands below use mamba for faster dependency resolution. You may substitute mamba with conda if preferred.

# Create the environment from the environment.yml file
mamba env create -f environment.yml

# Activate the environment
mamba activate repress

# Install the project in editable mode
pip install -e .

# Make the build script executable and run it
chmod +x build.sh
./build.sh

Run Analyses

All data required for the analyses with be under the data/ subfolder in each of the analysis and resulting plots will be under the plots/ folder.

# Run REPRESS to generate track plots and ISM for UTRN example
python repress/track_plots_and_ism_analysis/generate_track_plots_and_ism.py

# Run REPRESS to generate degradome predictions for ENO2
python repress/degradome_prediction_analysis/generate_degradome_predictions.py

# Run REPRESS on Slutskin MPRA analysis
python repress/slutskin_mpra_analysis/repress_slutskin_mpra_analysis.py

# Run REPRESS to generate wild type and mutant tracks with variant or oligo treatement
python repress/oligo_track_plots_analysis/generate_oligo_variant_tracks.py

# Run REPRESS on miRNA variant effect prediction analysis
python repress/variant_effect_prediction_analysis/get_variant_predictions.py
python repress/variant_effect_prediction_analysis/plot_variant_analysis.py

Usage

Note: The following code must be executed from the repository root.

import os
from genome_kit import Genome
import matplotlib.pyplot as plt

from repress.plot_tracks import make_track_plot
from repress.model_wrapper import REPRESS

# ------------------------------------------------------------------
# 1. Load a REPRESS model and its associated cell-line metadata
# ------------------------------------------------------------------
#   - "repress/repress_model"       → directory in repo with weights
#   - "repress/cell_line_csv.csv"   → CSV mapping cell-line names to indices
model = REPRESS(
    path="repress/repress_model",
    cell_type_csv="repress/cell_line_csv.csv"
)

# ------------------------------------------------------------------
# 2. Set up reference genome and locate the region of interest
# ------------------------------------------------------------------
# Load GENCODE v29 reference genome annotation
genome = Genome("gencode.v29")

# Gene we care about
gene_name = "ENO2"

# Get the Gene object for ENO2
# (Genome.genes is an iterable of Gene objects)
gene = [g for g in genome.genes if g.name == gene_name][0]

# Choose a specific transcript (second isoform in the list)
transcript = gene.transcripts[1]

# Grab the first 3′-UTR interval of that transcript
utr3 = transcript.utr3s[0]

# ------------------------------------------------------------------
# 3. Choose the cell line context for the prediction
# ------------------------------------------------------------------
cell_line = "deg_A549"          # must match an entry in cell_line_csv

# ------------------------------------------------------------------
# 4. Run the model on the interval
# ------------------------------------------------------------------
# predict_interval returns a list (one score per input interval);
# we take the first element because we passed a single interval
prediction_score = model.predict_interval(
    interval=utr3,  # Interval / Intervals of interest.
    genome=genome,  # Genome of the species of interest.
    transcript=transcript,  # transcript(s) that the intervals belongs to
    cell_lines=cell_line  # A list of length k corresponding to the cell lines to be included in the output
)[0]

# ------------------------------------------------------------------
# 5. Plot the degradome prediction tracks
# ------------------------------------------------------------------
fig, ax = make_track_plot(utr3, prediction_score, gene, genome, figsize=(8, 2),
                        scale_marker=0.4, custom_transcript=transcript,
                        vertical_markers=[6923337], vertical_markers_colors=["red"],
                        ylabels=["A549 degradome"], labelpad=40,)

os.makedirs("temporary_plots", exist_ok=True)
plt.savefig("temporary_plots/ENO2_degradome_track_plot.png", dpi=600, bbox_inches="tight")
plt.close()

Published Datasets used for model training

Datasets used for miRNA binding components of REPRESS.

Name	GSE	Study ID	GSM ID	SRX ID	Protocol	Aligner	Type	Species	Notes	Source
BC1	GSE32109	PRJNA154855	GSM796037, GSM796038	SRX097115, SRX097116	PAR-CLIP	bowtie1	bcell	human		Previously Published
BC3	GSE32109	PRJNA154855	GSM796039, GSM796040	SRX097117, SRX097118	PAR-CLIP	bowtie1	bcell	human		Previously Published
BCBL	GSE43909	PRJNA188176	GSM1074233, GSM1074234	SRX220837, SRX220838	PAR-CLIP	bowtie1	bcell	human		Previously Published
DG75	GSE43909	PRJNA188176	GSM1074231, GSM1074232	SRX220835, SRX220836	PAR-CLIP	bowtie1	bcell	human		Previously Published
HEK293	GSE28859	PRJNA153959	GSM714646, GSM714647	SRX058621, SRX058622	PAR-CLIP	bowtie1	kidney	human		Previously Published
HeLa	GSE29943	PRJNA140779	GSM741173, GSM741174, GSM741175	SRX083305, SRX083306, SRX083307	PAR-CLIP	bowtie1	cervix	human		Previously Published
A2780	GSE129076	PRJNA529911	GSM3693008, GSM3693009	SRX5604726, SRX5604726	PAR-CLIP	bowtie1	ovary	human		Previously Published
MCF7		PRJNA230871		SRX388831	PAR-CLIP	bowtie1	breast	human		Previously Published
hESC		PRJNA80179		SRX103431	PAR-CLIP	bowtie1	stem	human		Previously Published
A673	GSE80494	PRJNA319049		SRX1716186, SRX1716187	PAR-CLIP	bowtie1	muscle	human		Previously Published
22RV1	GSE137072	PRJNA564505	GSM4066540, GSM4066541, GSM4066542	SRX6817511, SRX6817512, SRX6817513	HITS-CLIP	bowtie2	prostate	human		Previously Published
HCT116	GSE146688	PRJNA611621	GSM4404081, GSM4404082	SRX7883128, SRX7883129	eCLIP	bowtie2	colon	human		Previously Published
Huh7	GSE73057	PRJNA295996			CLEAR-CLIP	star	liver	human		Previously Published
A549					miR-eCLIP	star	lung	human		This Study
K562					miR-eCLIP	star	bone	human		This Study
HEK293T					miR-eCLIP	star	kidney	human		Previously Published
Mouse ESC	GSE108795	PRJNA428611	GSM2913321, GSM2913322	SRX3533994, SRX3533995	PAR-CLIP	bowtie1	stem	mouse	mouse embryonic stem cells	Previously Published
Mouse myotubes	GSE108795	PRJNA428611	GSM2913323, GSM2913324	SRX3533996, SRX3533997	PAR-CLIP	bowtie1	muscle	mouse	cells differentiated from Myoblast → Myotubes	Previously Published
Mouse BMDM	GSE63199	PRJNA266979	GSM1543768	SRX758007	PAR-CLIP	bowtie1	bone marrow	mouse	C57BL/6 mouse	Previously Published
Mouse retina	GSE165832	PRJNA697981	GSM5050738, GSM5050741	SRX9980473, SRX9980474	PAR-CLIP	bowtie1	retina	mouse	CD-1 mice, P0	Previously Published
Mouse brain	GSE129885	PRJNA533090	GSM3724178	SRX5694011	eCLIP	star	brain	mouse	C57BL/6 mouse, P0	Previously Published
Mouse iNeuron	GSE140838	PRJNA591126	GSM4188573	SRX7202046	seCLIP	star	CNS	mouse	ESC, day 10 differentiation	Previously Published
Mouse liver					miR-eCLIP	star	liver	mouse	C57BL/6J mouse, 8-week-old	Previously Published

Datasets used for mRNA degradation components of REPRESS.

Cell Line	Num Reps	Protocol	Aligner	Type	Species	Source
A549	4	Degradome-Seq	star	lung	human	This Study
HepG2	3	Degradome-Seq	star	liver	human	This Study
iNeuron glutamatergic	3	Degradome-Seq	star	CNS	human	This Study
K562	3	Degradome-Seq	star	bone	human	This Study
PHH	6	Degradome-Seq	star	liver	human	This Study
Yecuris liver	3	Degradome-Seq	star	liver	human	This Study
Mouse cortex	3	Degradome-Seq	star	brain	mouse	This Study
Mouse liver	3	Degradome-Seq	star	liver	mouse	This Study
Mouse cns e18 day4	3	Degradome-Seq	star	CNS	mouse	This Study
Mouse cns e18 day11	3	Degradome-Seq	star	CNS	mouse	This Study

License

This material is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. If you use this software or data in your non-commercial work, please cite our paper [1].

For inquiries regarding commercial use please contact: [email protected]. Deep Genomics has filed one or more patent applications related to technical aspects of this work including PCT/IB2025/052017.

Citing this work

If you use REPRESS in your research, please cite:

@article{Kanuparthi2025.05.15.654105,
	author = {Kanuparthi, Bhargav and Pour, Sara E. and Findlay, Scott D. and Wagih, Omar and Gutierrez, Jahir M. and Gao, Rory and Wintersinger, Jeff and Lin, Junru and Gabra, Martino and Bohn, Emma and Lau, Tammy and Cole, Christopher B and Jung, Andrew and Celaj, Albi and Soares, Fraser and Gray, Rachel and Vaz, Brandon and Delfosse, Kate and Lodaya, Varun and Bhargava, Sakshi and Ly, Diane and Yusuf, Farhan and Kron, Ken and Hoffman, Greg and Gandhi, Shreshth and Frey, Brendan J.},
	title = {Sequence based prediction of cell type specific microRNA binding and mRNA degradation for therapeutic discovery},
	elocation-id = {2025.05.15.654105},
	year = {2025},
	doi = {10.1101/2025.05.15.654105},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/05/16/2025.05.15.654105},
	eprint = {https://www.biorxiv.org/content/early/2025/05/16/2025.05.15.654105.full.pdf},
	journal = {bioRxiv}
}