Predicting and scoring partial reprogramming: a gene-regulatory-network reanalysis of a pooled Yamanaka-factor screen in aged mesenchymal stromal cells

Partial reprogramming through transient expression of the Yamanaka factors can reverse features ofcellular ageing but driving cells too far toward pluripotency risks loss of identity and tumourigenesis.Identifying which factor combinations rejuvenate safely would benefit from computational prediction,yet the network-based perturbation models available for this purpose were developed for normaldevelopment rather than reprogramming or ageing, and their suitability has not been rigorously tested.Here we assess this approach by applying gene regulatory network inference and in-silico perturbation(CellOracle) to a published pooled Yamanaka-factor screen in young and aged mouse mesenchymalstromal cells, asking both whether the model predicts the measured effects of reprogramming andwhether it can be repurposed to score movement along a data-derived ageing axis. Ageing in thesecells proved to be characterised by myofibroblast-like contractile activation rather than classicalinflammatory senescence, with a published SASP panel failing to track age. The model recovered atranscriptional programme shared across factor combinations, correlating with measured effects atmean r = 0.13 and significantly exceeding a gene-shuffled null (p = 1.8 × 10⁻¹¹), but could notdistinguish combinations from one another, as their measured effects converge (mean pairwise r =0.83); a supervised classifier independently confirmed this, separating cell types at 86% accuracy butcombinations at only 11% (chance 7%). Scored against the ageing axis, all fifteen combinationsmoved cells toward an older, activated state rather than a younger one, most strongly via Myc, andthis pro-ageing signal persisted on an ageing axis constructed to exclude activation genes, indicatingthat ageing and activation are inseparable in these cells. These results delineate both the capabilitiesand the limits of network-based reprogramming prediction and underscore that rigorous null modelsand orthogonal validation are essential to distinguish genuine predictive power from a capturedgeneric response.