STRAND: Sequence‑conditioned transport for single‑cell perturbations

Predicted responses differ by perturbed locus and by cell line

Each panel below reports in silico perturbations, in which STRAND masks a window of the reference sequence and predicts the resulting single-cell state.

STRAND attains the best average rank on few-shot and unseen-gene benchmarks

The benchmarks compare STRAND with gene-level perturbation models on CRISPRi Perturb-seq screens in K562, Jurkat and RPE1 cells. The three tabs correspond to few-shot prediction with five cells of each test gene in training, prediction for genes held out from training, and transfer to a cell line held out from training.

How does STRAND compare with gene-level baselines?

STRAND predicts perturbed expression by transporting control cells, conditioned on the DNA sequence at the target

STRAND combines two frozen pretrained encoders with three trainable modules. The Borzoi DNA encoder², in its enhanced Flashzoi version³, embeds the 524 kb sequence window in 128 bp bins, and the perturbation module combines these embeddings with the mask into a perturbation condition. The SCimilarity RNA encoder⁴ places each control cell in a latent space of cell states, where an I²SB bridge⁵, a learned stochastic path from control to perturbed states, transports the cell according to the perturbation condition. A decoder then predicts the expression of each gene from the transported state.

How does STRAND work?

What does each component contribute?

Each variant is retrained with one module replaced and all other components and losses held fixed. In the first variant, the perturbation module is replaced by DNA embeddings that are averaged over the window and kept fixed; in the second, the latent transport module is replaced by a deterministic MLP predictor that takes the control and DNA embeddings as input. Values are means over test perturbations, and average rank is the mean rank of each variant across the seven metrics (Table S10, p.27 of the preprint).

Future Direction: STRAND-guided Prioritization of Candidate Perturbation Sites

We envision combining STRAND's predicted perturbation effects with published experimental findings and genomic annotations to prioritize candidate regions for follow-up experiments.

1. Define the queryThe user specifies the target gene, the cell line and the genomic region to scan.
2. Run a STRAND scanSTRAND predicts the perturbation effect of each fixed-width window as the window moves across the selected region.
3. Shortlist candidate regionsThe scan groups neighboring high-scoring windows into regions and keeps those with the largest predicted perturbation effects.
4. Retrieve evidenceFor each candidate region, experimental studies, regulatory annotations and results from earlier perturbation screens are collected.
5. Prioritize candidatesAn expert evaluates STRAND predictions alongside the available biological evidence and ranks candidate regions for further investigation.

Paper and code

@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}
}