FateDriver: Weakly supervised dual-view graph learning for transition-specific prioritization of candidate regulators in cellular reprogramming

Cellular reprogramming holds promise for regenerative medicine, but the specific endogenous regulators for certain source-to-target transitions are still poorly understood. Current computational approaches rely on network topology, trajectory dynamics, or cell-state labels to rank candidate regulators, ignoring the experimental factors that drive the transition.

To address this gap, researchers introduce FateDriver, a weakly supervised dual-view graph neural network. This innovative method integrates single-cell transcriptomes of cells and genes, treating experimental factors as sparse gene-level labels. FateDriver then ranks the remaining genes for a specified transition using a positive-unlabeled formulation.

In evaluations using two OSKM datasets and one GMT dataset, FateDriver emerged as a top performer, ranking first or second among six methods in recovering externally curated reference genes. Notably, FateDriver outperformed a supervision-matched non-graph model despite exhibiting weaker co-expression with the provided factors. Additionally, FateDriver successfully identified distinct candidate programs when holding either the source or the target constant, such as a DDX58--interferon axis associated with cells lacking endogenous pluripotency markers.

Furthermore, the framework was applied to two in vivo partially reprogrammed intestinal populations, yielding distinct candidate origins for these populations. In essence, FateDriver transforms scarce experimental insights into transition-specific hypotheses for regulator prioritization derived from single-cell trajectories.