Cellular perturbation prediction drops explicit variable values: RELAY reads shared-variable relations and reaches 0.495 PDCorr on PBMC-3D

Predicting cellular perturbation responses at unmeasured conditions can rest on measured responses that share some experimental variables with the target, without first encoding perturbation identity, cell type or donor as vectors. Earlier methods conditioned on explicit representations of variable values; learned identifier embeddings cannot represent values absent from training, descriptor-based approaches depend on external priors whose coverage varies by variable type, and variables such as donor identity lack standardized descriptors. Across four chemical, genetic and cytokine benchmarks in the unseen-combination setting, RELAY has the highest PDCorr and lowest MSE on all four and beats all supervised baselines on PDS; on PBMC-3D it reaches PDCorr 0.495 against 0.225 for the strongest supervised baseline, STATE.

The model compresses each measured control–perturbed pair into a small set of tokens, adds a relation embedding for the subset of variables that pair shares with the target, and lets target control cells read these demonstrations through cross-attention. The shared-variable set determines the relation embedding, for example the sum of the perturbation and donor embeddings when both are shared; variable values themselves never enter the model input. On PBMC-3D in the unseen-combination setting, keeping both state and relation embeddings gives PDCorr, PDS and REACH of 0.495, 0.833 and 0.420; removing the relation embedding lowers them to 0.170, 0.599 and 0.040, and removing the state embedding gives 0.303, 0.582 and 0.162.

Training compares predicted and observed cell-population distributions with an energy-kernel maximum mean discrepancy, so no cell-to-cell correspondence between control and perturbed populations is required. Target control cells supply the baseline state while the context memory supplies how those states are expected to change under the target condition, and the prediction is emitted as a residual from the target control state. All four benchmarks use identical splits and preprocessing, including library-size normalization to 10⁴, log transformation and a highly variable gene panel selected from training-visible cells only; results are reported as means and standard deviations over three seeds.

Evaluation covers four chemical, genetic and cytokine perturbation benchmarks, Tahoe-100M, Replogle–Nadig, PBMC-2D and PBMC-3D, with variables limited to perturbation identity, dose, cell type and donor; the authors note that high-quality perturbation datasets with enough cells per condition and dense coverage of multidimensional spaces including dose and time are scarce, which constrains broader evaluation.

In the unseen-variable-value setting, performance is generally lower than in the unseen-combination setting; Mean perturbation has the highest PDS on the three 2D benchmarks, and Untreated has lower MSE and MAE than RELAY on Replogle–Nadig and PBMC-3D despite near-zero PDCorr. The attention analysis describes how attention is distributed rather than establishing which demonstrations cause better predictions, and control and perturbed tokens exchange information through pair self-attention, so they are not independent representations of the two states.