DNAzymesGEN

Computational pipeline for DNAzyme discovery using generative models, structural priors, and machine-learning screening.

Despite their broad therapeutic and biotechnological potential, deoxyribozymes remain challenging to design rationally due to few consensus sequences, context-dependent activity, target inaccessibility, and a lack of large curated datasets for machine learning. This repository provides a multi-stage pipeline that combines:

• Synthetic pre-training on MFE-bounded or EDS-matched sequence distributions
• Fine-tuning on experimentally validated DNAzymes (Sequence Craft)
• Large-scale sequence generation from trained WGAN-GP checkpoints
• Hierarchical screening with an activity classifier and structural filters

Repository: GenerativeMolMachines/DNAzymesGEN

Publication. This work was accepted to an ICML 2026 workshop in Seoul:

In silico evaluation of pre-training strategies based on synthetic data for functional DNA generation
Ivan Golovkin, Stepan Druzhininskii, Daniil Litunovskiy, Amir Taldaev, Brilenkov Evgeny, Nikita Serov

---

Repository layout

Directory	Description
Data/	Training datasets: EDS, MFE, Sequence Craft, and negative controls
scripts/	Pipeline entry points: training, generation, evaluation, MFE analysis
scripts/env/	GPU environment setup (gpu_env.sh)
scripts/training/	Training orchestrator and fine-tuning launcher
scripts/generation/	Batch sequence generation shell scripts
scripts/evaluation/	Checkpoint JSD evaluation
scripts/analysis/	Post-hoc MFE statistics and bootstrap tests
improved_wgan_training/	DNA sequence WGAN-GP core modules (train, generate, evaluate)
FT/	Fine-tuning helpers and MFE utilities used during generation
checkpoints/	Pre-trained and fine-tuned model checkpoints
generated/	Generated sequence pools, screening outputs, and analysis artifacts
screening_pipeline/	Classifier-based screening and secondary-structure filtering
statistics/	Dataset statistics scripts
visualization/	Embedding and UMAP visualization utilities

See generated/WORKFLOW.md for checkpoint selection rationale and the full training/generation timeline.

---

Prerequisites

• Linux with NVIDIA GPU (CUDA 11.8 recommended)
• Git LFS — required for checkpoints and large datasets
• Conda (Miniconda or Mambaforge)
• NUPACK 4 — not included in this repo; install separately for MFE calculations and screening

Clone with LFS

git lfs install
git clone https://github.com/GenerativeMolMachines/DNAzymesGEN.git
cd DNAzymesGEN
git lfs pull

---

Environment setup

GPU environment (TensorFlow / WGAN)

conda create -n dnazymes-gpu python=3.10 -y
conda activate dnazymes-gpu
conda install -c conda-forge cudatoolkit=11.8 cudnn=8.9 -y
pip install -r requirements-gpu.txt

Update scripts/env/gpu_env.sh if your conda environment path differs from the default.

NUPACK (optional, for MFE filtering and screening)

Download and install NUPACK 4 according to the vendor instructions, then ensure the Python bindings are importable:

python -c "import nupack; print(nupack.__version__)"

NUPACK is required for:

• MFE-filtered generation (scripts/generation/run_generation.sh)
• Post-hoc MFE analysis (scripts/analysis/compute_mfe_*.py, scripts/analysis/mfe_bootstrap_test.py)
• Screening pipeline secondary-structure steps

Unfiltered generation (scripts/generation/run_generation_unfiltered.sh) does not require NUPACK.

Screening classifier dependencies

pip install scikit-learn lightgbm pandas numpy tqdm

---

Reproducing the pipeline

Shell scripts resolve paths relative to the repository root and work from any clone location.

Step 1 — Pre-training

Train WGAN-GP generators on synthetic EDS or MFE datasets:

# EDS pretrain (→ checkpoints/eds/)
bash improved_wgan_training/run_pretrain.sh eds 0

# MFE pretrain (→ checkpoints/mfe/)
bash improved_wgan_training/run_pretrain.sh mfe 1

Or use the orchestrator:

python scripts/training/run_training.py --scenario eds      # EDS pretrain only
python scripts/training/run_training.py --scenario mfe      # MFE pretrain only
python scripts/training/run_training.py --scenario eds_ft   # EDS pretrain + Sequence Craft finetune
python scripts/training/run_training.py --scenario mfe_ft   # MFE pretrain + Sequence Craft finetune

Published checkpoints (already in checkpoints/):

Model	Checkpoint	Notes
EDS pretrain	checkpoints/eds/model-1400	Best available EDS snapshot
MFE pretrain	checkpoints/mfe/model-3000	Lowest train JSD on MFE data
EDS finetune	checkpoints/eds_ft/model-999	Finetuned on Sequence Craft (549 seq)
MFE finetune	checkpoints/mfe_ft/model-999	Finetuned on Sequence Craft
SC-only pretrain	checkpoints/sequence_craft/model-700	Trained from scratch on Sequence Craft

Step 2 — Fine-tuning

Fine-tune a pre-trained checkpoint on Sequence Craft:

bash scripts/training/run_finetune.sh 0 \
  checkpoints/eds_ft \
  checkpoints/eds/model-1400 \
  checkpoints/eds_ft/train.log

Parameters: ITERS=1000, SAVE_INTERVAL=100, learning rate 5e-5 (set inside gan_language.py finetune mode).

Step 3 — Checkpoint evaluation

Compare checkpoints against Sequence Craft n-gram statistics (JSD):

bash scripts/evaluation/run_evaluation.sh
# → checkpoints/evaluation_results.json

Step 4 — Sequence generation

Generate 250k sequences per model (length filter only, no MFE):

bash scripts/generation/run_generation_unfiltered.sh

Outputs land in generated/*_nofilter/generated_sequences.csv.

For MFE-filtered generation (requires NUPACK):

bash scripts/generation/run_generation.sh

Single-model generation:

cd improved_wgan_training
python generate_sequences.py \
  --label my_run \
  --checkpoint ../checkpoints/eds_ft/model-999 \
  --num-sequences 250000 \
  --no-mfe-filter \
  --skip-mfe-calc

Step 5 — Post-hoc MFE analysis

python scripts/analysis/compute_mfe_nofilter_comparison.py
python scripts/analysis/mfe_bootstrap_test.py

Step 6 — Screening

Run the hierarchical screening pipeline on generated pools:

cd screening_pipeline
bash run_all_screening.sh

Or score sequences with the LightGBM classifier directly:

cd screening_pipeline/classifier_new_version
python run_model.py --input /path/to/sequences.csv

---

Key scripts

Script	Purpose
improved_wgan_training/gan_language.py	WGAN-GP pretrain / finetune for DNA sequences
improved_wgan_training/generate_sequences.py	Batch sequence generation from checkpoint
improved_wgan_training/evaluate_checkpoints.py	JSD-based checkpoint comparison
FT/customed_tunning.py	Alternative fine-tuning entry point
FT/generate.py	Alternative generation entry point
scripts/training/run_training.py	End-to-end training orchestrator
scripts/training/run_finetune.sh	Fine-tune launcher
scripts/generation/run_generation_unfiltered.sh	Batch unfiltered generation
scripts/evaluation/run_evaluation.sh	JSD checkpoint evaluation
scripts/analysis/compute_mfe_nofilter_comparison.py	Post-hoc MFE comparison
screening_pipeline/run_screening.py	Screening pipeline driver

---

Data sources

File	Description
Data/EDS/distrib_result.csv	EDS-matched synthetic sequences
Data/MFE/seq2 - seq2.csv	MFE-bounded synthetic sequences
Data/Sequence_Craft/SequenceCraft_dataset.csv	549 experimentally validated DNAzymes
Data/Negatives/	Negative control sequences

---

Citation

If you use this code or data, please cite:

@inproceedings{golovkin2026insilico,
  title   = {In silico evaluation of pre-training strategies based on synthetic data for functional {DNA} generation},
  author  = {Golovkin, Ivan and Druzhininskii, Stepan and Litunovskiy, Daniil and Taldaev, Amir and Evgeny, Brilenkov and Serov, Nikita},
  booktitle = {ICML 2026 Workshop},
  address = {Seoul, South Korea},
  year    = {2026},
  note    = {Accepted}
}

Paper. In silico evaluation of pre-training strategies based on synthetic data for functional DNA generation — Ivan Golovkin, Stepan Druzhininskii, Daniil Litunovskiy, Amir Taldaev, Brilenkov Evgeny, Nikita Serov. Accepted at an ICML 2026 workshop, Seoul.

---

License

Apache-2.0 — see LICENSE.

The DNA sequence WGAN code in improved_wgan_training/ is adapted from Improved Training of Wasserstein GANs (image demo scripts are excluded from this repository).