Pretreatment transcriptional state carries condition-specific information about future clonal detection in a lineage-traced melanoma line

Abstract

Background. Whether a cell's molecular state before a perturbation predicts what happens to it afterwards is usually asked retrospectively, after outcome and state have been measured in the same cells. We ask it prospectively at clone level, in one BRAF-V600E melanoma cell line (WM989, GSE279162), in which a barcoded population was split across six observed experimental conditions: Acid, Cisplatin, CoCl2, Dabrafenib, Doxorubicin and Trametinib. Of the clones that experiment recovered, 1,401 carry a pretreatment profile and are therefore analysable prospectively; those are the clones used here. The ranking test uses the 892 of them detected under at least one condition and undetected under at least one other. Results. Within this system, pretreatment gene expression contains condition-specific information about future clonal detection beyond condition identity and captured pretreatment clone abundance, under clone-held-out evaluation whose folds, features and exclusions were fixed before any model was fitted. Under a test preregistered in full — the metric, population, weighting, comparator, null and verdict rule all fixed by digest before any ranking statistic was computed, though after earlier predictive analyses of the same data — a frozen state-by-condition interaction model improves clone-specific ordering of the six conditions over a non-interactive additive model: +0.051605 in equal-clone-weighted within-clone AUROC, 95% CI [+0.037197, +0.065571], with 0 of 1000 full-refit permutation draws reaching the observed value (p < 0.001). Conclusions. A state contribution shared additively across all conditions cannot change their ordering within a clone. Allowing state effects to vary by condition improved that ordering in WM989: part of what a clone's pretreatment state carries bears on which of the six conditions it is still detected after, not only on how detectable it is overall. The outcome is a detection proxy and is not death, sensitivity, resistance or clinical response. The design and the artifacts behind it are released, so that the result can be reproduced from the public data and tested in other lineage-traced systems, where it has not yet been tested. Preprint of the CellFate-Rx Generation 1 manuscript: a preregistered reanalysis of GEO GSE279162, generated and deposited by Schaff et al. (doi:10.1016/j.xgen.2026.101191). No new data were generated, and this manuscript has not been peer reviewed. The complete analysis, the frozen protocols, the stage-by-stage records including negative and failed results, and tools that refuse on any modified artifact are archived at https://doi.org/10.5281/zenodo.22769563, with the code at https://github.com/hagai-tyty/Cellular-Outcomes-of-Perturbations. Manuscript text and figures are CC BY 4.0; the software and the frozen model are under the PolyForm Noncommercial License 1.0.0.