🧬 biology

CAPRA predicts single-cell transcriptional responses to genetic perturbations with response anchoring and residual correction

The paper introduces CAPRA, a novel response-anchored model that predicts single-cell transcriptional responses to genetic perturbations by combining control states, gene embeddings, and learned corrections, thereby achieving superior accuracy and interpretability in both single- and double-gene tasks compared to existing methods.

Original authors: Dong-Qing Wei, Zhongcheng Fang, Qi Gao, Yajing Yuan, Han Wang, Hong Tan, Jiayi Li, Zhennan Peng, Lan Yang, Mengyuan Jin, Heqi Sun, Yufang Zhang

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dong-Qing Wei, Zhongcheng Fang, Qi Gao, Yajing Yuan, Han Wang, Hong Tan, Jiayi Li, Zhennan Peng, Lan Yang, Mengyuan Jin, Heqi Sun, Yufang Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery inside a bustling city: the city is a living cell, and the streets are filled with thousands of tiny workers called genes. Sometimes, a specific worker gets sick or is removed (a "genetic perturbation"), and the whole city changes its behavior. Scientists have developed high-tech cameras called "single-cell screens" that can take a snapshot of the city before and after a worker is removed, showing exactly how the neighborhood reacts. This is incredibly useful for understanding diseases and finding cures.

However, there is a massive problem: the city is too big, and the number of possible combinations of sick workers is astronomical. It would take forever and cost a fortune to test every single combination in the lab. So, scientists have tried to build "crystal balls"—computer programs that guess what would happen if we removed a worker they haven't tested yet. Some of these crystal balls use complex math to learn the general "vibe" of the city, while others try to map the relationships between workers. But here's the catch: when these programs make a guess, they often hide why they made that guess. They give you an answer, but you can't see the evidence they used to get there. It's like a detective saying, "I know the butler did it," but refusing to show you the fingerprint or the alibi. This makes it hard to trust the prediction, especially when the computer is just guessing based on vague similarities.

Enter CAPRA, a new tool introduced by researchers at Shanghai Jiao Tong University that changes the game by refusing to hide its work. Think of CAPRA not as a magic crystal ball, but as a very honest, transparent detective who builds their case step-by-step in front of you.

Instead of guessing the whole outcome at once, CAPRA starts with a "Control Anchor." Imagine you have a perfect, untouched photo of the city (the control state). When you want to predict what happens if you remove a specific worker, CAPRA first looks for a photo of what happened when a similar worker was removed in the past. If it finds an exact match in its database, it uses that as a starting point. If it doesn't find an exact match, it finds the closest neighbor in its library (using a smart system called GenePT) and uses that as a "best guess" anchor. This is the Response Anchoring part: it says, "Here is the evidence I'm starting with."

But CAPRA knows that a similar worker isn't the exact same worker. Maybe the new worker has a different personality or works in a different part of the city. So, CAPRA adds a second step called Residual Correction. It asks, "Okay, I have this starting photo, but how does the specific context of this new situation change things?" It calculates the small differences—the "residuals"—needed to tweak the starting photo to fit the new scenario perfectly.

The paper tested this new detective against other famous crystal balls (like Scouter, GEARS, and scGPT) across 13 different datasets, which included both single-worker removals and double-worker removals. The results showed that CAPRA was the most accurate at predicting the direction of the change. In simple terms, if the real city got louder, CAPRA predicted it would get louder, whereas others sometimes guessed it would get quieter or stayed silent. This was true even when the tool had to guess about combinations of workers it had never seen before.

One of the most exciting findings was how CAPRA handled the "noise" of the city. When scientists try to find which genes are affected by a change, they often use a statistical test that can get confused if the computer's predictions are too perfect or too flat (a problem called "variance collapse"). CAPRA is special because it generates predicted cells that still have a realistic amount of "jitter" or natural variation, just like real cells. This allowed researchers to run standard tests on CAPRA's predictions and successfully identify the affected genes, something other methods struggled to do.

The researchers also broke down CAPRA to see which part did the heavy lifting. They found that the "Anchor" (the starting photo) did most of the work, providing a strong hypothesis. However, the "Residual Correction" was crucial when the starting photo wasn't a perfect match, acting like a fine-tuning knob that fixed the prediction. This transparency is key: CAPRA tells you, "I started with this evidence, and I made these specific adjustments."

In summary, CAPRA doesn't just give you a number; it gives you a story. It shows you the evidence it used (the anchor) and the adjustments it made (the correction). By keeping the process visible and accounting for the natural messiness of biological data, CAPRA offers a more reliable way to predict how cells will react to genetic changes, helping scientists prioritize which experiments to run next without wasting time on wild guesses.

Technical Summary: CAPRA – Control-Anchored Perturbation Residual Architecture

Problem Statement
Single-cell perturbation screens (e.g., Perturb-seq, CROP-seq) map genetic interventions to cellular states, but experimental coverage remains sparse due to cost and throughput limitations. This is particularly acute for combinatorial (multigene) perturbations, where the search space grows exponentially. While computational methods exist to extrapolate responses, they face two primary limitations:

Opacity of Evidence: Existing predictors (using latent representations, gene graphs, or foundation models) often absorb the specific response evidence supporting a prediction into latent model representations. This makes it difficult to distinguish predictions grounded in measured single-gene effects from those driven primarily by extrapolation or semantic similarity.
Variance Collapse in Downstream Analysis: When predicted cells are passed to standard single-cell workflows (e.g., Scanpy) for Differential Gene Expression (DEG) calling, predicted-cell variability often collapses. This leads to inflated test statistics for small mean changes, causing affected-gene recovery to fail even when mean-response rankings retain signal.

Methodology: CAPRA Architecture
The authors introduce CAPRA (Control-Anchored Perturbation Residual Architecture), a response-anchored model designed to make prediction support explicit and preserve variability for downstream analysis. The architecture operates in three stages:

Response Anchor Construction:
- For a given perturbation, CAPRA samples a control state ( $\mu_{control}$ ) to define the baseline.
- It constructs an expression-space response anchor ( $\Delta_{anchor}$ ) representing the expected perturbation effect.
- Exact Support: If the perturbation gene has an observed single-gene effect in the training bank, that exact effect is used.
- Retrieved Support: If no exact effect exists, CAPRA retrieves a nearest-neighbor single-gene response from the GenePT embedding space (using K-nearest neighbors based on GenePT vectors).
- For double-gene perturbations, slot-level anchors are summed ( $\Delta_{GA} + \Delta_{GB}$ ).
- Crucially, the model records the provenance of the anchor (exact vs. retrieved) as metadata.
Anchor-Conditioned Residual Correction:
- The model learns a residual correction to refine the anchor based on context and composition.
- A shared low-rank response operator learns departures from the anchor, conditioned on the control state ( $Z_0$ ), anchor support features, and the response anchor itself.
- Additive Residual: Captures context-specific corrections.
- Pair Residual (Double-gene only): A symmetric path captures the specific interaction effects of gene pairs that cannot be explained by simple summation.
Prediction Assembly and Calibration:
- The final profile is assembled in expression space: $X_{pred} = \mu_{control} + \Delta_{anchor} + \lambda_{add}r_{add} + \lambda_{pair}r_{pair}$ .
- Variance Calibration: To prevent variance collapse, CAPRA returns deterministic, control-scale-calibrated predicted cells. It adds reproducible within-condition variability derived from control-scale statistics ( $\sigma_{cal}$ ) rather than learning variance from the test set. This ensures compatibility with standard t-test-based DEG calling workflows.

Key Contributions

Explicit Response Evidence: Unlike latent-space models, CAPRA exposes the "starting hypothesis" (the anchor) and its source (measured vs. retrieved), linking predictive performance to explicit response evidence.
Hybrid Anchoring Strategy: It effectively combines exact experimental data with semantic retrieval (via GenePT) for genes lacking direct training support, using residual correction to adapt these anchors to specific contexts.
Calibrated Inference Interface: The model provides predicted cells with preserved within-condition variability, enabling standard DEG recovery workflows that typically fail with other generative models.

Results
The authors evaluated CAPRA on a panel of 13 genetic perturbation datasets (9 single-gene, 4 double-gene) against state-of-the-art methods (Scouter, GEARS, GenePert, scGPT).

Response Direction Prediction: CAPRA achieved the highest mean PCC-delta (Pearson correlation on control-centered changes for top 100 DEGs) in both single-gene (0.592) and double-gene (0.728) tasks, outperforming Scouter and other baselines.
Generalization under Sparse Support: In double-gene tasks, CAPRA maintained superior performance across "seen2" (both genes seen), "seen1" (one gene seen), and "seen0" (neither gene seen) categories. The advantage persisted even in "seen0" scenarios where predictions rely on GenePT-neighbour anchors and residual adaptation.
Specificity: Even when using a "perturbed-reference" scoring strategy (centering on the average perturbation state rather than the control to remove shared post-perturbation structure), CAPRA retained the highest response-direction signal, indicating it captures perturbation-specific information beyond shared patterns.
Affected-Gene Recovery: In standard Scanpy t-test workflows, non-CAPRA methods showed near-zero Jaccard recovery of top DEGs due to variance collapse. CAPRA's calibrated cells achieved significantly higher Jaccard scores (0.093 for single-gene, 0.174 for double-gene), successfully recovering affected-gene signatures.
Component Analysis: Ablation studies confirmed that the response anchor provides the primary hypothesis (especially for single-gene tasks), while residual correction is most critical when single-gene support is incomplete or for modeling gene-pair interactions.

Significance
The paper claims that CAPRA addresses the "design gap" in current perturbation prediction by making the evidence supporting a prediction inspectable. By separating the measured/retrieved response anchor from the learned correction, CAPRA allows researchers to prioritize unmeasured genetic interventions based on the quality of their underlying evidence. Furthermore, by solving the variance-collapse issue, it enables the direct use of predicted cells in standard differential expression pipelines, bridging the gap between computational prediction and experimental validation workflows. The authors note that while the current study is retrospective, the framework is designed to support future prospective selection of perturbations and transfer to other modalities (e.g., chemical perturbations, multimodal readouts).

Technical Summary: CAPRA – Control-Anchored Perturbation Residual Architecture

More like this