Doloris

Single-cell RNA sequencing destroys the cell it measures. A perturbation experiment therefore never yields a before-and-after pair for the same cell — only two populations, one control and one perturbed, with no correspondence between them. Most models for predicting perturbation responses paper over this by forcing a pairing, matching perturbed cells to control cells or to a control mean, which biases learning toward the average profile and erases the heterogeneity that makes single-cell data worth collecting.

Doloris, developed at Westlake University with collaborators at Zhejiang University, HKUST (Guangzhou), and several Chinese institutes, takes the unpaired structure as the starting point. It trains two conditional diffusion models — a source model on unperturbed cells and a target model on perturbed cells — that share a Gaussian latent space. Inference runs a control cell forward through the source model into that latent space, then denoises it back out through the target model under the desired perturbation condition, a dual conditional diffusion implicit bridge. No explicit pairing is ever required. The paper was accepted at ICLR 2026.

#Key Features

• Unpaired distributional bridging: Control and perturbed expression distributions are learned separately and aligned only through the shared noise prior, so the model captures the transition between population states rather than fitting cell-to-cell correspondences that do not exist.
• Sparsity masking: A separate mask model predicts which genes are silenced under a given perturbation, and the diffusion loss is computed only over expressed genes. Without it, the model devotes capacity to reproducing the many zeros in the expression matrix and collapses toward low-diversity predictions.
• Heterogeneity-preserving conditioning: During training, the target model is conditioned on the mean and standard deviation of the control population with added noise rather than a bare mean, an ablation the authors show is essential to the model's performance.
• Genetic and chemical perturbations in one framework: Gene perturbation embeddings support multi-gene knockouts and combinatorial effects, while molecular perturbations are conditioned on representations from a pretrained molecular model, so both perturbation types run through the same generative machinery.
• Distribution-level evaluation: The authors argue that RMSE on conditional means is misleading when differentially expressed genes are bimodal, and evaluate with energy distance and earth mover's distance alongside RMSE.

#Technical Details

The source and target models share an architecture and are unified into one implementation differing only in conditioning inputs. Training uses AdamW at a learning rate of 0.001 with batch size 32 and a 500-step diffusion schedule; inference uses 50-step DDIM sampling. Training runs 10,000 steps each on the Adamson and Norman CRISPR datasets and 100,000 on sci-Plex3, all on a single A100 80GB.

On held-out single-gene perturbations in Adamson, Doloris reaches an RMSE of 0.0336 and energy distance of 0.4682, against 0.0544 and 0.8921 for GEARS, 0.0510 and 0.8705 for GRAPE, and 0.0473 for a simple linear baseline. On unseen drug–cell-line–dose combinations in sci-Plex3 it reaches 0.0287 RMSE against 0.0570 for chemCPA and 0.0697 for CPA. Held-out double knockouts in Norman give 0.0385 against 0.0405 for the linear baseline and 0.0533 for GEARS, and out-of-distribution drugs give 0.0547 against 0.0689 for chemCPA. A fine-tuned scGPT performs poorly on these splits. The margins over the linear baseline on double knockouts are narrow, consistent with recent findings that simple baselines remain hard to beat on perturbation prediction; the gaps widen on the distributional metrics and on out-of-distribution drugs.

#Applications

The model targets in-silico screening: prioritising gene knockouts or knockout combinations before running a CRISPR screen, predicting the transcriptional effect of a compound on a cell line and dose not yet tested, and generating expression profiles for perturbations that would be impractical to measure across every cell type. Because generation is distributional rather than mean-seeking, downstream analyses that depend on cell-to-cell variability — subpopulation detection, response heterogeneity — remain meaningful.

#Impact

Doloris reframes perturbation prediction as unpaired distribution translation, borrowing dual diffusion bridges from image translation and adding the sparsity handling that single-cell data demand. The evaluation protocol matters as much as the model: by arguing against mean-based metrics for bimodal genes and reporting energy and earth mover's distances, the paper pushes the subfield toward assessments that reward capturing the shape of a response rather than its centre. The code repository provides training and evaluation scripts for the genetic perturbation datasets but carries no license file and no released checkpoints, so reuse currently means retraining from the published recipe.