Predicting single-cell transcriptomic responses to genetic perturbations is central to functional genomics and virtual-cell modeling. Existing approaches, however, typically predict an entire expression profile as a whole, leaving the order in which individual gene responses are generated unmodeled. To address this problem, we introduce \textbf{D2R2} (Discrete Diffusion with Regulation Reinforcement), which reformulates perturbation prediction as regulation-guided gene-wise progressive generation. A Masked Discrete Diffusion Model represents expression as ordinal tokens and reconstructs a fully masked profile step by step, allowing generated gene responses to condition those that remain masked. A Regulatory Policy Module initializes the generation policy from a gene regulatory network inferred from control cells and adapts it to the perturbation and current partially generated state. Then, group-relative policy optimization refines only the ordering policy using final perturbation-effect agreement as reward. Across Norman19 and VCC-H1, D2R2 achieves the best performance on all five metrics on Norman19 and remains competitive on H1. Controlled ablations holding the generator and generation budget fixed show that biological-prior ordering improves over random ordering and is more reliable than uncertainty-based heuristics, whereas reversing the biological-prior ordering degrades every metric. Biological analyses further show that the refined policy prioritizes regulatory genes early while promoting perturbation-specific transcription factors and responsive genes. These results establish gene generation order as an effective, controllable, and biologically interpretable dimension of single-cell perturbation prediction.
Figures & tables
Figure 1: Paradigm comparison. Each row summarizes paradigm properties and traces gene-expression generation, with web nodes and aligned vector entries denoting genes and their expression values. We compare traditional single-pass prediction, DDPM-based global denoising, and two forms of D2R2 : random unmasking (MDDM) and regulation-guided unmasking.
Figure 2: Overview of the D2R2 framework. The left panel shows the iterative generation loop: after the MDDM predicts the selected expression tokens, RPM observes the updated partial state and recomputes which genes should be resolved next. The upper-right panel shows the biological prior used to initialize the policy, while the lower-right panel illustrates GRPO that optimizes the order to each perturbation.
Differential Expression
Absolute Expression
Dataset
Method
Pearson Δ↑
Cos. LogFC ↑
Cos. LogFC Rank ↓
Cos. PCA ↑
Sym. KL ↓
Norman19
SAMS-VAE (NeurIPS 2023)
0.553 ± 0.079
0.560 ± 0.043
0.028 ± 0.010
0.251 ± 0.042
50.888 ± 4.215
Biolord (Nat. Biotechnol. 2024)
0.404 ± 0.069
0.388 ± 0.051
0.046 ± 0.010
0.035 ± 0.016
0.526 ± 0.040
STATE (NeurIPS 2025)
0.467 ± 0.020
0.379 ± 0.020
0.525 ± 0.017
-0.009 ± 0.039
0.273 ± 0.014
Cell Flow (bioRxiv 2025)
0.489 ± 0.077
0.373 ± 0.066
0.329 ± 0.013
0.084 ± 0.066
0.207 ± 0.024
Squidiff (Nat. Methods 2025)
0.291 ± 0.001
0.226 ± 0.002
0.427 ± 0.007
0.110 ± 0.000
2.861 ± 0.019
Table 1: Main results on Norman19 and H1. All methods are evaluated over five independent training runs using seeds {0,1,2,3,42} , and results are reported as mean ± standard deviation. For D2R2 , MDDM checkpoint is fixed across runs, while RPM is independently trained for each seed. Metrics are grouped by differential and absolute expression. Best and second-best results are bolded and underlined, respectively.
Figure 3: Violin plots showing the distribution of evaluation metrics across perturbations in Norman19. Mann-Whitney U test, * P < 0.05, *** P < 0.001.
Order
Pearson Δ↑
Cos. LogFC ↑
Cos. LogFC Rank ↓
Cos. PCA ↑
Sym. KL ↓
Random
0.623 (ref.)
0.720 (ref.)
0.032 (ref.)
0.597 (ref.)
0.183 (ref.)
Model-Uncertainty Orders
Confidence high
0.590 ( −5.3 %)
0.537 ( −25.4 %)
0.030 ( +5.2 %)
0.353 ( −40.9 %)
0.410 ( −124.4 %)
Confidence low
0.672 ( +7.9 %)
0.522 ( −27.5 %)
0.068 ( −114.1 %)
0.640 ( +7.2 %)
0.203 ( −11.0 %)
Entropy high
0.624 ( +0.2 %)
0.452 ( −37.2 %)
0.073 ( −130.9 %)
0.564 ( −5.6 %)
0.206 ( −12.5 %)
Entropy low
0.591 ( −5.1 %)
0.538 ( −25.3 %)
0.030 ( +5.2 %)
0.355 ( −40.5 %)
0.399 ( −118.5 %)
Table 2: Generation-order ablation on Norman19. All methods use the same frozen MDDM. Random averages multiple random orders. Cells show score and relative performance gain over Random, with signs aligned so that positive always indicates improvement; for Cos. LogFC Rank and Sym. KL, this corresponds to relative error reduction. Orange intensity indicates a positive gain. Percentages are computed before rounding.
Figure 4: Biological analysis. (a) Proportions of regulatory and regulated genes across generation steps under the RPM-refined generation order. (b) Boxplot comparing ΔOrder between perturbation-specific TFs and non-specific TFs. Mann-Whitney U test, P=1.6×10−15 . (c) Boxplot showing the fraction of DE genes whose generation order was promoted by RPM or remained unpromoted. Each dot represents one perturbation, and paired fractions from the same perturbation are connected by lines. Paired Wilcoxon test, P=1.7×10−9 .
Batch 2000; hidden size 2048 with three layers; learning rate 10−5
Squidiff
Three layers; batch 64; learning rate 10−4 ; EMA 0.9999
scDFM
Hidden size 512; batch 32; learning rate 10−5
Appendix
Table 3: Model-specific configurations used for the baseline methods.
Parameter
Value
Batch size
128
Optimizer
AdamW
Weight decay
10−2
Learning rate (Norman19 / H1)
10−4 / 5×10−4
Warmup fraction
0.01
Gradient clipping
1.0
Appendix
Table 4: MDDM training configuration. Dataset-specific values are reported as Norman19 / H1.
Parameter
Value
Batch size
96
Group size
4
Optimizer
AdamW
Learning rate
3×10−6
Weight decay
10−4
Warmup fraction
0.01
Appendix
Table 5: GRPO training configuration for RPM optimization.
Figure 5: GSEA analysis using the ΔOrder values of all genes as the ranking metric. The GSEA analysis shows that RPM-promoted genes are enriched in cell cycle pathway (a) and G2-M transition pathway (b).
Predicting transcriptional responses to genetic perturbations is fundamental to functional genomics and therapeutic discovery. Recent deep learning models have shown promise in single-cell perturbation response prediction, but they typically generate each response in isolation, without explicitly leveraging related perturbations. We introduce PT-RAG (Perturbation-aware Two-stage Retrieval-Augmented Generation), a plug-in retrieval-and-conditioning module for generative cellular perturbation response. PT-RAG augments an existing perturbation-response backbone with learned access to related perturbation contexts. The key challenge is that relevance is not fixed in this setting: functionally related genes may elicit different effects across cell types. PT-RAG addresses this with a two-stage retrieval mechanism: GenePT-based semantic retrieval first identifies K candidate perturbations, after which a differentiable Gumbel-Softmax selector adaptively selects retrieved contexts conditioned on the control cell state, the query perturbation, and each candidate perturbation. We evaluate PT-RAG on two backbones, a STATE-style generator used as a frozen random reservoir and a fully trained scGPT, across cross-cell-type and cross-perturbation generalization tasks. PT-RAG consistently improves distributional similarity and often overall predictive quality; for example, on scGPT cross-cell-type results, the 2-Wasserstein distance drops by 5.9% relative to scGPT alone. The code to reproduce our experiments is available at https://github.com/difra100/PT-RAG_NIPS.
Andrea Giuseppe Di Francesco, Andrea Rubbi, Rishabh Jain +1
Single-cell perturbation models can reduce costly wet-lab screening by predicting how cells respond transcriptionally to interventions. While recent generative models improve population-level prediction, individual generated cells are not explicitly checked for biological consistency. We introduce PerturbCellRL, a reinforcement learning (RL) framework that post-trains a pretrained single-cell transcriptomic generator using a suite of cell-level verifiers as rewards. These verifiers define four rewards: Pearson top-k similarity, RMSE top-k proximity, DE Spearman, and Pathway activity. The Pathway activity verifier rewards cells whose pathway responses match known perturbation biology. We evaluate PerturbCellRL on multiple genetic and chemical perturbation benchmarks. Across these benchmarks, PerturbCellRL improves over the pretrained flow-matching generator on reward-aligned evaluation metrics and a held-out evaluation metric. Moreover, PerturbCellRL remains competitive with state-of-the-art methods on population-level metrics. Together, these results frame trustworthy single-cell prediction as verifier-guided generative alignment, moving beyond matching expression distributions toward predictions whose single-cell perturbation effects are explicitly checked for biological consistency.
Dongxia Wu, Mingyu Li, Yuhui Zhang +4
Peking University Beijing, China · Stanford University Stanford, CA
Predicting cellular transcriptional responses to genetic perturbations is a central problem in single-cell biology, especially in the zero-shot setting where the perturbed gene or gene combination is unseen during training. A major difficulty is that perturbation effects are not determined by expression state alone: they depend on how the perturbed gene product influences other genes and proteins, how those downstream factors act on cis-regulatory elements, and which regulatory programs are active in the current cell state. To better capture this biological complexity, we propose CisTransCell, a cell-conditioned multi-modal framework for single-cell perturbation prediction that augments each gene with two complementary priors: a regulatory-sequence prior that captures how the gene is controlled, and a coding-sequence prior that captures what the gene product does. By integrating these priors with cellular expression state, CisTransCell models perturbation response as a cascade from gene function to regulatory control to downstream transcriptional change. Experiments on benchmark single-cell perturbation datasets show that CisTransCell achieves strong performance in zero-shot perturbation prediction.
Wei Zhang, Xun Jiang, Yuesi Xi +1
1L3S Research Center, Leibniz Universität Hannover, Germany · School of Clinical Medicine & Laboratory Medicine, Jiangsu University, China · Institute for Information Processing (tnt), Leibniz Universität Hannover, Germany.