Perturbation Expression Prediction

This task evaluates a model's ability to predict expression for masked genes in a CRISPR interference (CRISPRi) perturbation context, given the remaining (unmasked) genes in a cell as input. Genes selected for masking are drawn from those identified as differentially expressed under CRISPRi knockdowns, ensuring that the task focuses on biologically meaningful perturbation effects. Control cells are matched to each perturbed cell based on GEM group and UMI count. Predictions are compared against ground truth by calculating predicted and experimental log fold change (LFC) values for masked genes, providing a direct measure of how well models capture perturbation-induced expression shifts.