Moving CRISPRi knockdown responses to a new cell line takes a partial screen: at 2.5% screened, CASTER beats the transferred mean
Synopsis
Across CRISPRi Perturb-seq of essential genes in four human cell lines, the same knockdown's transcriptional response correlated at a median of 0.52 between lines after noise correction, so responses are only moderately shared. CASTER starts from a knockdown's mean response in the other lines and corrects it for the target line using part of that line's screened library; with 30% screened, a linear model and CASTER scored 0.84 and 0.80 on the calibrated scale versus 0.44 for the transferred mean. CASTER led only at 2.5% screened (0.62 versus 0.47), and with control cells alone and no screen it fell below the transferred mean.
Interpretation
The transcriptional response to the same knockdown is only moderately shared across cell lines, with a median cross-cancer-line correlation of 0.52 after noise correction. The uncorrected raw correlation was 0.31, because measurement noise at about 32 cells per knockdown pulled similarity down. After Spearman correction for attenuation using split-half reliabilities, the median rose from 0.31 to 0.52 between cancer lines and from 0.38 to 0.56 between a cancer line and RPE1, with the highest pair being the two TP53 wild-type lines, HepG2 and RPE1.
Screening a small fraction of a new line's library lets a learned model beat the transferred mean response. With no screen of the new line, the transferred mean (PerturbMean) scored 0.43 to 0.44 on the calibrated scale, the baseline any method must clear. With a fixed 70% test set and nested screens from 2.5% to 30%, the linear model's score rose from 0.47 to 0.84 and CASTER's from 0.62 to 0.80, while the transferred mean stayed at 0.43 to 0.44.
CASTER treats the mean response in the other lines as a starting point and then rescales it gene by gene and corrects it cell by cell for the target line. The model reads the target line's control cells as a set and is trained with an energy-distance maximum mean discrepancy loss, with the output layer and scaling initialized at zero so an untrained CASTER equals the transferred mean. On K562, removing the transfer prior dropped Pearson delta from 0.201 to 0.100, while attention and FiLM contributed little.
CASTER's lead comes mainly from starting at the transferred mean, not from cell-to-cell variation among control cells. At 2.5% screened, a ridge anchored on the transferred mean reached 0.56, the transferred mean with a gene-wise scale reached 0.50, and the linear model reached 0.47. Given the line's mean control profile instead of a set of cells, CASTER still reached 0.63, comparable to full CASTER's 0.62, whereas giving it another line's control cells dropped it to 0.36.
Perspective
The conclusions rest on public CRISPRi Perturb-seq data from four human cell lines (K562, HepG2, Jurkat and hTERT RPE-1), with at most 32 cells per knockdown used for training and models trained on CPUs; the authors note that four lines are too few to separate line generalization from lineage or p53 status.
CASTER cannot predict perturbations never observed in any line; the dependency analyses use CRISPR knockout fitness data while the screens use CRISPRi knockdown, and the four screens differ in effector construct, time point and sequencing depth; the work did not compare against a large pretrained perturbation model and did not test any prediction experimentally.
