Single-Cell Perturbation Modeling

• Single-cell perturbation modeling is a computational framework that predicts and interprets cell responses to genetic and chemical changes using stochastic dynamics, causal inference, and optimal transport.
• This approach integrates techniques like latent causal diffusion, optimal transport models, and cycle-consistent autoencoders to accurately map regulatory mechanisms and identify drug targets.
• Advanced methods such as Tikhonov regularization and sparse variational autoencoders ensure biological fidelity and scalability in high-dimensional single-cell omics analyses.

Single-cell perturbation modeling encompasses the mathematical and computational frameworks for predicting, interpreting, and inferring the response of individual cells to genetic or chemical perturbations at transcriptome-wide resolution. This modeling remains central to the project of systematically mapping regulatory mechanisms, identifying drug targets, and distinguishing direct versus indirect effects in high-dimensional single-cell omics screens. State-of-the-art approaches unify generative modeling, causal inference, and optimal transport, often deploying specialized architectures to disentangle biological signal from experimental noise, generalize to unseen perturbations, quantify causal effects, and provide mechanistically interpretable outputs. The field is marked by advances in stochastic dynamical systems, diffusion processes, sparse latent-variable models, cycle-consistent autoencoders, and distributional transport methods, each addressing distinct challenges in scalability, interpretability, and biological fidelity.

1. Stochastic Dynamical Systems and the Latent Causal Diffusion Model

The latent causal diffusion (LCD) framework recasts single-cell gene expression as a stationary stochastic differential equation (SDE) observed under measurement noise. Each cell’s noiseless state $x(t)\in\mathbb{R}^d$ (for $d$ genes) evolves under regulatory drift $f(x)$ perturbed by Brownian noise: $$dx_g(t) = [f(x(t))]_g\,dt + \sigma\,dW_g(t),\qquad g=1,\dots,d$$ with $\sigma=\sqrt{2}$ and $f$ parameterized by a neural network, whose hidden layers are shared across genes to encode shared regulatory structure. Perturbations are embedded as vectors $e_q$ added to the hidden state, inducing shifts in the stationary gene-expression distribution. Measurement noise is modeled via a zero-inflated Poisson (ZIP) process with learned dropout and scaling parameters for each gene. Training proceeds in two stages: (i) empirical Bayes sampling per perturbation condition, maximizing marginal likelihood under ZIP decoding; (ii) score-matching of drift parameters on Monte Carlo-recovered gene expression samples. LCD achieves state-of-the-art accuracy on held-out perturbation combinations, especially for non-additive (“synergistic,” “neomorphic”) genetic interaction classes, outperforming additive heuristics and conditional autoencoders in maximum mean discrepancy (MMD), RMSE, and Pearson correlation benchmarks (Lorch et al., 20 Jan 2026).

2. Causal Inference and Identifiability: The CLIPR Linearization Framework

Causal Linearization via Perturbation Responses (CLIPR) renders direct gene-gene causal effects interpretable from learned LCD dynamics. Under a linear drift assumption ($f_q(x)=A x + b + c_q$), the stationary response to perturbation $q$ is fully specified by $A$ and a bias term. CLIPR computes direct effects by assembling initial drift responses ($f^{(0)}_q$) and their equilibrium ($f^{(\infty)}_q$), then solving for the minimum-norm matrix $A$: $A = -F^\infty (F^0)^+$ where $F^0, F^\infty$ are $d\times k$ matrices over $k$ unique perturbations, and $(\cdot)^+$ is the Moore–Penrose pseudoinverse. Tikhonov regularization improves stability. CLIPR is provably identifiable with $k\geq d$ linearly independent perturbations. For nonlinear drift, $f^{(0)}_q$ and $f^{(\infty)}_q$ are estimated by evaluating $f(0;e_q)$ and integrating the ODE until equilibrium, respectively. Applied to genome-scale Perturb-seq, CLIPR yields sparse, modular causal effect matrices that outperform classical differential-expression for identifying direct links, as validated by fold-enrichment for observed DE targets among top predicted edges (Lorch et al., 20 Jan 2026).

3. Distributional and Optimal Transport Approaches

Distributional transport models, including the Conditional Monge Gap (CMonge) and Wasserstein-1 neural optimal transport (W1OT), address the challenge of unpaired single-cell data by learning mappings between control and perturbed cell populations. CMonge conditions the learned transport map on arbitrary covariates (drug, dose, cell type): $$T_\theta: (\mathbb{R}^m_c, \mathbb{R}^k_z) \to \mathbb{R}^k_z$$ with optimization driven by a Sinkhorn divergence plus Monge gap regularizer, yielding state-of-the-art generalization to unseen drugs/doses, robust first- and higher-moment alignment, and competitive performance against effect-based drug representation methods (Driessen et al., 11 Apr 2025).

W1OT simplifies the Kantorovich–Rubinstein dual problem to maximization over a single 1-Lipschitz potential, learning only the direction of transport, and subsequent adversarial training recovers the step-size for an explicit map, accelerating computation 25–45× over standard W2 solvers while matching or exceeding predictive accuracy (Chen et al., 2024).

4. Structured Variational and Cycle-Consistent Models

Sparse mechanism shift variational autoencoders (SAMS-VAE, sVAE+) enforce compositionality and disentanglement by modeling perturbation effect vectors and binary masks in latent space, allowing additive combination and sparse intervention over subspaces. This structure facilitates generalization to novel perturbation combinations and interpretable mapping to biological pathways, as evaluated by average treatment effect (ATE) correlation with empirical differential expression (Bereket et al., 2023, Lopez et al., 2022).

Cycle-consistent autoencoders (cycleCDR) impose reversibility constraints so that drug perturbation vectors induce latent-space translations whose inverse mapping restores pre-perturbation state. This ensures bidirectional realism, prevents mode collapse, and enables generalization to unseen drugs encoded via molecular graphs (Huang et al., 2023).

5. Benchmarking, Mode Collapse, and Evaluation Best Practices

Large-scale benchmarks reveal that naïvely applied fit-based metrics (RMSE, Pearson on $\Delta$) may disguise mode collapse: models predicting population means across conditions can outperform on uncalibrated metrics. The introduction of differentially expressed gene (DEG)-weighted mean-squared error (WMSE) and weighted delta $R^2$ metrics, together with calibration against negative (mean, control) and positive (technical duplicate) baselines, isolates true signal from collapse and reorders trivial predictors at null performance (Mejia et al., 27 Jun 2025). Under WMSE, models are explicitly penalized for failing to capture perturbation-specific DEGs, making this loss preferable in training of new single-cell perturbation predictors.

6. Causal Discovery from Observational and Interventional Data

Causal-differential-network frameworks (CDN) combine neural amortized causal graph inference with axial-attention classifiers to detect intervention targets by contrasting noisy causal graphs recovered from control and perturbed datasets. CDN achieves superior mean-average-precision and recall on large CRISPRi and chemical perturbation screens, reliably identifying on-target genes in high-dimensional settings, and remaining robust to data subsampling (Wu et al., 2024). Extensions include analysis of soft interventions and polynomial data-generating mechanisms.

7. Applications: Interpretable Dynamics, Multi-Modal and Spatial Modeling

Dynamic modeling, exemplified by FLeCS, encodes cell trajectories via ODE systems informed by sparse prior gene networks. Perturbations induce compensatory and indirect effects, with inference performed over deep ensembles of model parameters (Bertin et al., 25 Mar 2025). Celcomen extends causal modeling to spatial transcriptomics, disentangling intra- and inter-cellular gene programs and generating counterfactual spatial maps under arbitrary perturbations, with rigorous identifiability results and biological validation in human tissues and in vivo CRISPR screens (Megas et al., 2024).

Reinforcement learning models using TRPO–PPO multi-stage optimization interpret cell fate responses as control problems on the nonconvex Waddington landscape, employing natural gradients and trust-region updates to escape poor local minima, yielding improved generalization on scRNA-seq and scATAC-seq perturbation tasks (Boabang et al., 14 Oct 2025).

Diffusion-based models (LCD, scPPDM, Unlasting, Departures) integrate score-matched SDE simulation, dual-channel conditioning, and classifier-free guidance for what-if analysis and dose titration, with interpretable latent representations aligning to molecular signature and baseline cell state (Lorch et al., 20 Jan 2026, Liang et al., 8 Oct 2025, Chi et al., 26 Jun 2025, Chi et al., 17 Nov 2025).

8. Future Directions and Persistent Challenges

Key directions involve refining causal inference in the presence of confounders, extending models to handle multi-modal and temporal data, scaling to hundreds of conditions for pan-perturbation generalization, and improving artifact disentanglement (e.g., via counterfactual regularization in CRADLE-VAE (Baek et al., 2024)). Emphasis persists on interpretability, biological plausibility, and robust uncertainty quantification. As single-cell atlases grow and multi-omics integration intensifies, single-cell perturbation modeling is poised to become a mainstay for experimental design, therapeutic screening, and causal inference in molecular biology.