DeepRESIS

DeepRESIS is a command-line tool for predicting ncRNA-drug resistance relationships. You provide an ncRNA FASTA file, a drug SMILES file, and an ncRNA-drug pair file. The tool produces resistance prediction results and top-k gene ranking outputs.

Quick Start

On a fresh machine:

git clone https://github.com/idrblab/DeepRESIS.git
cd DeepRESIS
bash scripts/bootstrap_all.sh
conda activate deepresis
deepresis --help

This is the main installation path. It installs the environment, downloads the required large files from Hugging Face, generates deepresis.toml, and validates that the runtime is ready.

What the Bootstrap Script Does

bash scripts/bootstrap_all.sh will automatically:

1. create a conda environment
2. install Python dependencies
3. install the R runtime and required R packages
4. install the R package LncFinder
5. install DeepRESIS in editable mode
6. download 5 model checkpoint files from Hugging Face
7. download drug_gene.csv and ncrna_gene.csv from Hugging Face
8. generate deepresis.toml
9. validate the runtime and downloaded assets

Large model and gene matrix files are not stored in the GitHub repository. They are fetched automatically during bootstrap.

Files You Need

DeepRESIS needs three input files.

1. ncRNA FASTA file

Example:

>ncRNA_1
AUGCUAGCUAGCUA
>ncRNA_2
GGCUAAGCUU

The FASTA record ID becomes the ncrna_id.

2. Drug SMILES file

Example:

drug1    CCO
drug2    CCN(CC)CC

The first column is drug_id, and the second column is the SMILES string.

3. ncRNA-drug pair file

Example:

ncRNA_1    drug1
ncRNA_2    drug2

The first column is ncrna_id, and the second column is drug_id.

Important: every ncrna_id in the pair file must exist in the FASTA file, and every drug_id in the pair file must exist in the SMILES file.

Run Your First Prediction

After bootstrap finishes, run prediction with the generated config file:

deepresis predict \
  --config ./deepresis.toml \
  --fasta /path/to/test.fasta \
  --smiles /path/to/test_cid.txt \
  --pairs /path/to/test_pairs.txt \
  --output-dir ./outputs \
  --topk 300

You do not need to pass --model-dir or --gene-matrix-dir in the normal workflow.

Output Files

DeepRESIS writes three output files in the output directory.

1. resistance_predictions.tsv

This file contains resistance prediction results for each requested ncRNA-drug pair.

Columns:

• ncrna_id
• drug_id
• score
• label
• label_id

2. topk_genes.tsv

This file contains the top-k ranked genes for each pair when gene ranking is available.

Columns:

• ncrna_id
• drug_id
• rank
• gene_id
• score
• drug_gene_score
• ncrna_gene_score

3. gene_ranking_warnings.tsv

This file records why gene ranking could not be generated for specific pairs.

Columns:

• ncrna_id
• drug_id
• message

Demo Example

You can run the repository demo after installation:

bash examples/run_demo.sh

Or run the same example explicitly:

deepresis predict \
  --config ./deepresis.toml \
  --fasta ./examples/test.fasta \
  --smiles ./examples/test_cid.txt \
  --pairs ./examples/test_pairs.txt \
  --output-dir ./demo_outputs

Expected behavior for the current sample:

• resistance_predictions.tsv contains exactly 3 requested pairs
• topk_genes.tsv may contain only the header
• gene_ranking_warnings.tsv may contain 3 warnings because the sample circRNA IDs are not present in ncrna_gene.csv

Troubleshooting

conda: command not found

Conda is not installed or not on your PATH.

Next step: install Miniconda or Anaconda first, then rerun bash scripts/bootstrap_all.sh.

Hugging Face download failed

The machine cannot access Hugging Face, or the connection was interrupted.

Next step: check internet access and rerun the bootstrap command.

Missing model checkpoint files

One or more downloaded checkpoint files are missing under artifacts/models.

Next step: delete the incomplete artifacts directory and rerun bash scripts/bootstrap_all.sh.

Missing gene matrix files

drug_gene.csv or ncrna_gene.csv is missing under artifacts/gene_matrix.

Next step: delete the incomplete artifacts directory and rerun bash scripts/bootstrap_all.sh.

RNAfold not found

The ViennaRNA binary is not available in the installed environment.

Next step: rerun the bootstrap command and confirm that environment creation completed successfully.

LncFinder check failed

The R package LncFinder was not installed correctly.

Next step: rerun the bootstrap command. If it still fails, inspect the R installation step printed by the script.

Advanced Options

Custom environment name

bash scripts/bootstrap_all.sh deepresis_prod

Custom environment name, asset directory, and config path

bash scripts/bootstrap_all.sh deepresis_prod /data/deepresis_assets /data/deepresis.toml

Override Hugging Face source URLs

Normally you do not need this. If needed, you can override the Hugging Face directory URLs:

export DEEPRESIS_MODELS_URL="https://huggingface.co/swallow-design/DeepRESIS/tree/main/model_pharameter"
export DEEPRESIS_GENE_MATRIX_URL="https://huggingface.co/swallow-design/DeepRESIS/tree/main/gene_matrix"
bash scripts/bootstrap_all.sh

The bootstrap script will automatically normalize tree/main or blob/main URLs to real resolve/main download URLs.

Python API

from deepresis.pipeline import run_prediction

run_prediction(
    fasta_path="test.fasta",
    smiles_path="test_cid.txt",
    pairs_path="test_pairs.txt",
    output_dir="outputs",
    config_path="deepresis.toml",
    topk=300,
)