Research map / Probabilistic and scientific modeling / Molecules and proteins

STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein fitness and variant effect prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes STRAND, a generative model that predicts single-cell transcriptional responses to CRISPR perturbations by conditioning a latent diffusion bridge transport on DNA sequence embeddings at the target locus, enabling zero-shot prediction at unseen genomic loci.

Problem

Gene-level perturbation models map all interventions within a gene to the same representation, so they cannot distinguish perturbations at different loci (e.g., alternative TSSs or regulatory elements) and cover only ~1.5% of the genome.

Approach

Represent a perturbation by encoding the DNA sequence at its genomic locus (Flashzoi/Borzoi embeddings with a perturbation-anchored cross-attention module) and use it to condition an I2SB-style latent bridge transport from control to perturbed RNA latents (SCimilarity space), with OT pairing, residual hurdle decoding and CLIP-style DNA-RNA alignment.

Key results

• In the low-sample setting STRAND achieves the best average rank (1.86 across cell lines per Table summary) and discrimination score 0.72 vs 0.54 for STATE (33% improvement on K562).
• On the zero-shot unseen-gene benchmark STRAND attains average rank 1.71, the best overall.
• On cross-cell-line transfer STRAND improves Delta Pearson by about 0.14 over STATE on K562 transfer; generative transport reduces E-distance by 38.9% and raises R^2 by 27.6% versus a deterministic concatenation baseline on Combined.

Surprising findings

• Simple baselines (Linear, PerturbMean) achieve low MSE but near-chance discrimination scores (~0.5), showing a global-accuracy versus heterogeneity trade-off.
• STATE performs only marginally better than an identity baseline on unseen cell-line transfer.

Limitations

• Perturbation site is approximated as a +/-200 bp window around the canonical TSS for CRISPRi genes; sgRNA sequence and exact target coordinates are not used.
• Evaluation limited to CRISPR-i datasets in three cell lines (HepG2 excluded for low statistical power).
• In silico case studies are hypothesis generation without experimental validation.

Experimental setup

Datasets: PerturbQA CRISPRi single-cell perturbation data (K562, Jurkat, RPE1, HepG2 excluded) Combined cross-cell-line dataset

Benchmarks: PerturbQA zero-shot splits PerturbQA low-sample splits Combined cross-cell-line leave-one-out

Metrics: Discrimination score (Discr. Cos) Energy distance Dir. Match Prec.@N Ov.@N Delta Pearson MSE Pseudo-bulk R^2 Average rank

Baselines: GEARS BioLord STATE Linear PerturbMean

Builds on: I2SB HiRef CLIP Borzoi SCimilarity

Closest papers by meaning

• CytoWave: Perturbation-Centric Pretraining for Single-Cell Response Prediction (NeurIPS 2026 · similarity 0.89)
• PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects (NeurIPS 2026 · similarity 0.87)
• PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.86)
• Designing Cell-Type-Specific Promoter Sequences Using Conservative Model-Based Optimization (NeurIPS 2024 · similarity 0.85)
• C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation (NeurIPS 2026 · similarity 0.85)
• DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA (NeurIPS 2024 · similarity 0.84)