MODEL RESULTS
See responses.
See variation.
From unseen perturbations to new cellular contexts.
Explore the evidence for what the model has learned.
Stay close to observed states.
Cover more of their diversity.
01 / UNSEEN PERTURBATIONS
Unseen perturbations.
Faithful responses.
Across 272 unseen perturbations in K562,
we evaluate both mean responses and cell distributions.
Prediction quality needs more than one measure.
View the values +
For scPLAD, three random seeds are averaged within each condition first. FCS uses a shared PCA50 space, up to 100 cells per group, 5NN support, and 10 resampling runs. Scores are not percentages of correct predictions.
02 / POPULATION STRUCTURE
Beyond the mean.
The whole population.
Variation matters alongside shared responses.
In 255 conditions with multiple observed subpopulations,
we test recovery of both their proportions and internal states.
Each point summarizes one method across 255 conditions.
Closer proportions.
Fuller state coverage.
In this comparison, scPLAD combines lower composition error with higher coverage of states within subpopulations.
Figure 3i–j ↗View population-structure values +
These 255 conditions form the multi-subpopulation subset of the 272 unseen perturbations. Subpopulations are clusters of observed expression within each condition, rather than independently validated cell types. GEARS is omitted because verifiable subpopulation metrics are unavailable in this page’s source data. Complete TVD includes unassigned cells.
03 / ACROSS CELLULAR CONTEXTS
From known contexts
to new responses.
Combine perturbation knowledge from source cell lines
with the control-cell background of K562.
Evaluate predictions across 1,086 transfer conditions.
Different metrics, different strengths. scPLAD has higher top-100 DE gene overlap; TxPert has higher response correlation, and PIE has higher AUPRC.
STATE directly predicts 580 perturbations observed in the source contexts and uses mean fallback for the other 506. The charts show aggregate means.
Figure 4a ↗View transfer values and evaluation scope +
Of the 1,086 conditions, 506 are unseen in all three source cell lines and 580 are observed in at least one. These results aggregate all conditions; the full set is not source-unseen.
04 / THE ROLE OF BIOLOGICAL KNOWLEDGE
How does knowledge
shape prediction?
Across the same unseen perturbations,
compare prior configurations and gene orderings
to examine their effects on responses and population structure.
These are three independent comparisons with the labeled baselines, not cumulative steps. All variants use the same fixed mean correction. Values are means across three random seeds.
Figure 3k · Supplementary Figure S3 ↗View the full prior comparison +
Complete priors do not lead on every metric. For example, “+ Network” has higher ΔPCC, while complete priors score higher on expression-range coverage, interval overlap, and covariance-structure metrics in this table.