TL;DR
ReliFlow is a new generative model that predicts how cells respond to combinations of drugs or genetic edits by explicitly accounting for the "reliability" of individual effects. By separating the errors in baseline estimates from actual biological interactions, it significantly improves the accuracy of single-cell perturbation predictions, especially when extrapolating to entirely new interventions.
What Problem Does This Paper Address?
In single-cell biology, researchers want to predict how a cell's gene expression will change when subjected to multiple "perturbations" (like knocking out two genes at once or applying a drug cocktail). However, the number of possible combinations is astronomical, making it impossible to measure them all.
A common starting point is additivity: if we know the effect of Perturbation A and Perturbation B individually, we assume the combination is roughly A + B. The "residual"—the difference between this sum and the actual result—is where the interesting biological interactions (synergy or antagonism) live.
The challenge is that our estimates for A and B are often noisy or based on sparse data. If a model doesn't know how much to trust its "additive anchor," it might waste its learning capacity trying to "fix" a bad baseline rather than learning the actual biological interactions. This issue is most severe when trying to predict combinations involving a perturbation the model has never seen before.
Methods
ReliFlow uses Conditional Flow Matching (CFM), a generative framework that learns the "velocity field" required to transform a starting distribution (control cells) into a target distribution (perturbed cells).
The authors introduce three key innovations:
- Reliability Estimation: The model uses a "cross-fitting" approach on training data to estimate how much the predicted effect of a single gene varies. It learns a reliability map ($\alpha$) that tells the model which gene-wise shifts are trustworthy and which are uncertain.
- Calibrated Additive Source: Instead of a raw sum, ReliFlow creates a "calibrated anchor." It shrinks uncertain additive shifts toward a learned prior and adjusts the "spread" (variance) of the starting distribution based on this uncertainty.
- Decomposed Transport: This is the architectural "secret sauce." The model decomposes the learning task into two parts:
- A known correction that bridges the gap between the calibrated anchor and the raw additive estimate.
- A learned interaction velocity ($u_\theta$) that focuses solely on the remaining non-additive biological transport.
Datasets
The model was evaluated on two major benchmarks:
- Norman Dataset: Genetic perturbations (CRISPRa) in human K562 cells, involving 105 single-gene and various double-gene perturbations.
- ComboSciPlex: Chemical (drug) combinations in A549 cells, featuring 7 single drugs and 25 drug pairs.
The authors used rigorous "condition-disjoint" splits, ensuring the model was tested on combinations it hadn't seen during training, including an "excluded-single" regime where one of the constituent perturbations was entirely withheld from the training set.
Key Insights and Ideas
- Don't Mix Noise with Biology: By explicitly modeling anchor uncertainty, ReliFlow prevents the neural network from "silently mixing" estimation errors into the nonlinear residual. This makes the learned interactions more biologically meaningful.
- Reliability is the Biggest Lever: In ablation studies, removing the reliability weights caused the largest drop in performance, proving that knowing what you don't know is critical for generalization.
- Analytic Shortcuts Help: Decomposing the velocity field (Figure 2) allows the model to use an exact algebraic correction for the parts of the path we can calculate, leaving the "heavy lifting" of learning complex interactions to the neural network.
Key Results
ReliFlow consistently outperformed state-of-the-art baselines like scDFM and GEARS across multiple metrics:
- Genetic Perturbations (Norman): Reduced Mean Absolute Error (MAE) by 7.7% for seen-single combinations and by 10.1% when constituent singles were withheld.
- Chemical Combinations (ComboSciPlex): Achieved a 7.9% reduction in MAE compared to scDFM.
- Interaction Recovery: The model was significantly better at capturing non-additive effects, improving the Pearson correlation of residuals from 0.574 to 0.638.
Limitations and Future Work
- Computational Cost: The improved accuracy comes with a trade-off; training cost increased by about 19% (from 12.4 to 14.8 GPU-hours) due to the cross-fitting and reliability estimation steps.
- Population vs. Individual: Like most flow-matching models in this space, ReliFlow predicts population-level distributions. It does not identify a unique causal trajectory for a specific individual cell, as single-cell sequencing is a destructive process.
- Extrapolation Limits: While ReliFlow handles "unseen combinations" well, extrapolating to entirely new classes of drugs or genetic pathways remains a difficult frontier for the field.
Summary updated