No drug measurements in the target cell: GRAFT predicts drug transcriptional responses from genetic perturbations, reaching mean Top20 correlation 0.734 across five LINCS lines
Synopsis
When a target cell line has no drug measurements, GRAFT links that line's genetic perturbation responses to drug responses measured in source cells and predicts the transcriptional changes a drug would induce. In leave-one-line-out evaluation across five LINCS cancer lines, GRAFT-D0 reached a mean Top20 Pearson correlation of 0.734, above the source mean of 0.493 and the strongest comparator PRnet at 0.689; across five external cellular contexts the equally weighted mean Top20 correlation was 0.606 against a source mean of 0.416. The method rests on a conditioning state built from how the target's genetic responses differ from source references, and predictions still improve when target genetic measurements are few, though gains vary with data quality and source coverage.
Interpretation
A drug's transcriptional effect differs across cellular contexts, and GRAFT lets a target cell line's existing genetic perturbation responses stand in for missing drug measurements, predicting that line's transcriptional response to many drugs. Earlier routes either used only unperturbed baseline expression, trained genetic-to-chemical transfer within individual cell lines, or adapted genetic and chemical perturbations separately; none showed how genetic responses measured in the target cell could guide drug prediction without target drug measurements. Across the five LINCS cancer lines A375, A549, HT29, MCF7 and PC3, holding out one line as target and using the other four as sources, GRAFT-D0 reached a mean Top20 Pearson correlation of 0.734 over 644 drugs and 976 genes, against 0.493 for the source same-drug mean and 0.689 for PRnet, the strongest D0 comparator, with the advantage holding across all five targets and three seeds.
Prediction first compresses the difference between the target cell's genetic responses and the source reference into a conditioning state, then uses that state to adjust the source cells' average response to each drug. The conditioning state is summarized by cross moments among genetic response residuals, perturbation identities and source-reference features, converted by a hypernetwork into scaling and shifting parameters applied to drug features through feature-wise linear modulation, and reused across drug queries. Disabling the conditioning state in the trained model dropped Top20 correlation from 0.734 to the source mean of 0.493, and replacing target genetic responses with source-cell responses failed to recover the same gains, indicating the gains come from the target-relative-to-source difference.
Perspective
The demonstrated setting uses five LINCS cancer lines as sources and five public-dataset cellular contexts as targets, predicting mean transcriptional responses over a 976-gene panel; GRAFT-Atlas spans 13 cellular contexts and 7,856 drug catalog entries, with ten contexts using GRAFT-Adapt and RPE1, neurons and LUHMES using D0 predictions.
HAP1 showed lower prediction correlations and no significant drug-anchor gains in paired tests, with its genetic inputs combining stimulation conditions while drug signatures came from another platform; K562 and HAP1 also have small query panels and candidate anchor pools drawn from different datasets, raising uncertainty in their gains. In the iPSC analysis, across-donor response-magnitude rankings largely reflected a shared magnitude tendency, and the functional comparisons reused genetic data supplied to the model, so independent drug-treatment experiments are still needed to test predicted drug-specific differences.