Skip to main content
Back to timeline
arXivSource publication:

ReliFlow folds single-gene reliability into combination prediction: Norman MAE falls from 0.198 to 0.178 when constituent singles are withheld

Synopsis

In predicting transcriptional responses to combined perturbations, errors in the additive reference and genuine non-additive interactions are mixed together, especially when single-perturbation evidence is sparse. ReliFlow shrinks the additive anchor using reliability estimates from training-only cross-fitting and splits the velocity field into a known anchor correction and a learned residual transport field. On Norman genetic perturbations, MAE falls from 0.142 to 0.131 when constituent singles are observed in training and from 0.198 to 0.178 when they are withheld; on ComboSciPlex drug combinations MAE falls from 0.191 to 0.176.

Source-provided article image: ReliFlow: Reliability-Calibrated Residual Transport for Combinatorial Single-Cell Perturbation Prediction
arXiv

Interpretation

Combination-perturbation prediction can now separate two things: how well the additive single effects themselves are estimated, and how much interaction a combination adds beyond the additive reference. Earlier approaches treated the additive reference as a fixed starting point, so when single-perturbation evidence was sparse or fully withheld its error was handed to the nonlinear network to repair. On Norman, against the strongest listed baseline scDFM, MAE falls from 0.142 to 0.131 when constituent singles are seen in training and from 0.198 to 0.178 when they are withheld, relative reductions of 7.7% and 10.1%.

The model first assigns each gene's single effect a reliability scale, then uses it to decide how far the additive starting point shrinks and how wide the initial distribution spreads. Reliability comes from cross-fitting on training perturbations only: training perturbation identities are split into folds, a model without that fold's response profiles predicts the fold's effects, and a nonnegative reliability head is trained with a robust loss on log scales. On the A549 ComboSciPlex drug-combination split, MAE is 0.176 versus 0.191 for scDFM, MMD falls from 0.074 to 0.063, and DE-Spearman rises from 0.605 to 0.641.

The conditional velocity is written as two terms: an analytically computable anchor correction and a network-learned residual transport. The network therefore does not have to rediscover the known difference between the two anchors from scratch and can spend capacity on interaction-dependent transport. On the Norman seen-single split, removing reliability weights raises MAE from 0.131 to 0.148 and lowers DE-Spearman from 0.752 to 0.691, the largest gap among the four ablations.

Perspective

Evaluation covers the seen-single and excluded-single condition-disjoint splits of Norman genetic perturbations and five held-out drug pairs in A549 ComboSciPlex; all splits are by perturbation condition rather than individual cell, and the gene set is 5,000 highly variable genes selected from training cells. The intended reader is a researcher modeling single-cell perturbation responses and combination screens.

The reliability scale correlates positively with error within each support group, but that is not a demonstration of probabilistic calibration or nominal uncertainty coverage. The excluded-intervention stratum has the highest error, indicating that extrapolating to components without measured training responses remains difficult. Three random seeds describe run-to-run variation and do not support a strong significance claim.