STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations

Authors

• Boyang Fu
• Sameer Gabbita
• George Dasoulas
• Xiang Lin
• Shanghua Gao
• Xiaorui Su
• Soumya Ghosh
• Marinka Zitnik

Preprint | Website

Advances in Neural Information Processing Systems (NeurIPS), 2026

STRAND predicts single-cell transcriptional responses by conditioning on the regulatory DNA sequence at the perturbed locus rather than on a gene identifier, so perturbations that target different genomic sites within the same gene can receive different predicted responses.

Why sequence conditioning

Most single-cell perturbation models represent a perturbation as a discrete gene identifier or a node in a gene graph, which collapses every targeted site inside a gene into one representation and closes the perturbation vocabulary to annotated genes. STRAND instead takes a reference DNA window and a perturbation mask at a user-specified locus, so the perturbation is defined by sequence and position. This supports zero-shot inference at targets unseen during training and locus-resolved inputs across a 524 kb window at 128 bp model resolution.

Figure 1. Perturbations targeting the same gene at different genomic locations (s_p¹, s_p²) can induce distinct transcriptional responses.

How STRAND works

Regulatory DNA and a target mask enter a frozen DNA foundation model; a learnable DNA Perturbation Module turns the masked sequence context into a perturbation condition; a control-anchored latent transport module carries the control cell state to the perturbed state under that condition; and a residual hurdle decoder predicts perturbed expression relative to a reconstructed control baseline.

Figure 2. The frozen encoders are Borzoi (DNA, 128 bp output bins over a 524 kb window) and SCimilarity (RNA). The learnable components are the DNA Perturbation Module g_φ, the Cellular Transport Module e_θ (a control-anchored I²SB bridge), and the gene-level decoder. An auxiliary sequence–expression alignment objective links the DNA condition to the mean transcriptional effect.

Quick start

git clone https://github.com/mims-harvard/STRAND.git && cd STRAND
pip install -r requirements.txt

Requires Python ≥ 3.10 (tested: Python 3.11, PyTorch 2.9.0). Run commands from the repository root; tests: python -m pytest tests.

Point the pipeline at the reference genome, the annotation and your data roots:

export DATA_ROOT="/path/to/data_root"              # curated .h5ad + embeddings
export OUTPUT_ROOT="/path/to/outputs"              # checkpoints + DNA cache
export DNA_REFERENCE_PATH="/path/to/hg38_fasta_dir"  # per-chromosome FASTA (GRCh38)
export GENE_ANNOTATION_PATH="/path/to/gencode.v47.transcript.annotation.gtf"
export HGNC_COMPLETE_SET_PATH="/path/to/hgnc_complete_set.txt"  # gene symbol -> Ensembl ID
export CCRE_BED_PATH="/path/to/ENCFF234XEZ.bed.gz"  # ENCODE cCREs (cCRE bias)
export DNA_CACHE_DIR="$OUTPUT_ROOT/dna_cache"      # DNA embeddings, built in step 02, read by steps 04-05

Train a per-cell-line model with the configuration used in the paper (after data steps 01–03 below):

bash scripts/04_train/train.sh k562   # k562 | jurkat | rpe1 | combined_perturbqa_hvg2k

Predict the test split with a trained checkpoint (writes paired predicted / measured .h5ad files):

python scripts/05_evaluate/evaluate.py \
  --config configs/k562_config.yaml --dname k562 \
  --checkpoint_path /path/to/best_model.pt \
  --result_path outputs/k562_eval

Inspect or build a sequence-defined perturbation directly. strand.locus needs only NumPy, and reproduces the coordinate and mask construction the model uses at inference:

from strand import Locus, masked_tokens

locus = Locus.centered("chrX", 48_786_590, half_width=200, tag="GATA1", dataset="k562")
print(locus.label)                      # chrX:48786390-48786790(+)-GATA1-k562
print(masked_tokens(locus, 156_040_895))  # DNA context tokens selected by the mask

strand.STRAND.from_checkpoint(...) loads a trained checkpoint and exposes predict and scan for single loci and sliding-window scans; see docs below for the model dependencies this requires.

Reproducibility and documentation

The pipeline runs in numbered order:

Step	What it does	Where
01	Curate the PerturbQA screens into per-cell-line .h5ad files	scripts/01_curate/
02	SCimilarity embeddings and OT pairing of control and perturbed cells (the split column chosen here sets zero-shot or few-shot); Borzoi DNA embedding cache	scripts/02_embed_and_match/
03	Train/validation split of the OT training pairs	scripts/03_make_splits/
04	Train STRAND	scripts/04_train/
05	Predict the test split with a trained checkpoint	scripts/05_evaluate/

Also	Where
Datasets, sources and the perturbation-coordinate protocol	docs/data.md
Training configurations (one per cell line, plus Combined)	configs/
STRAND model (paper Fig. 2) and the two Fig. 4 ablations	src/diffusion_ot/models/
DNA Perturbation Module over the Borzoi backbone	PModule/DNA_modeling.py
Baselines (GEARS, BioLord, STATE, STACK)	benchmark/

Perturbation coordinates. Public CRISPRi datasets carry gene-level labels rather than exact guide coordinates, so the default benchmark uses a TSS-centered proxy: gene symbols are resolved to Ensembl IDs, a canonical transcript is chosen by tag priority, and a ±200 bp mask is placed on the TSS inside a 524 kb window. Exact guide coordinates can be supplied when available; the paper reports a sensitivity analysis comparing the two. docs/data.md documents the full procedure.

External dependencies. Not bundled here: the Borzoi/Flashzoi DNA backbone (borzoi-pytorch), the SCimilarity RNA encoder (Genentech/scimilarity), the HiRef optimal-transport solver (raphael-group/HiRef) and the GRCh38 reference with GENCODE annotation.

Scope. Training and evaluation use CRISPRi data in K562, Jurkat and RPE1. HCT116 appears only as an external cross-study transfer evaluation and is not part of the training data. STRAND operates at 128 bp model resolution, set by the DNA encoder's output binning.

Citation

@inproceedings{fu2026strand,
  title     = {{STRAND}: Sequence-Conditioned Transport for Single-Cell Perturbations},
  author    = {Fu, Boyang and Gabbita, Sameer and Dasoulas, George and Lin, Xiang and
               Gao, Shanghua and Su, Xiaorui and Ghosh, Soumya and Zitnik, Marinka},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Questions: please open an issue in this repository or contact marinka@hms.harvard.edu.