Pop-Corn improves cell-composition prediction after gene knockout
Yale and Whitehead Institute researchers posted a preprint showing Pop-Corn predicts post-knockout cell-type composition (correlation 0.802 vs 0.709 for the best competitor), validated on human T cells and mouse liver tissue, with top-5 gene overlap at 22.6%.
On nearly 58,000 human T cells, four leading models (GEARS, scGPT, CPA, and CellFlow) predicted mean post-knockout gene expression reasonably well. When those predictions were translated into cell-type composition, rare cell states were lost. The models predict an average expression profile; when that profile is matched to known cell types, rare states do not resemble the average and are dropped.
On September 26, researchers at Yale University and the Whitehead Institute for Biomedical Research posted a preprint describing Pop-Corn, a method that skips the intermediate step. Pop-Corn treats all cells sharing one knocked-out gene as a single sample and predicts its composition directly, without building an average profile first. The gene target is encoded by the amino acid sequence of its protein product, using the ESM-2 model (Meta AI). Because human and mouse proteins are largely similar, predictions transfer across species.
Correlation of predicted composition with experimental data on T cells: 0.802 for Pop-Corn versus 0.709 for the best of the four competing models (GEARS). That number is a correlation coefficient, not a success rate.
The method was also validated on sections of living mouse liver tissue, where 202 genes were knocked out. In intact tissue, a cell's response depends on the knocked-out gene and on neighboring cells; in dissociated cells, neighbor effects are invisible. With spatial position encoded, Pop-Corn predicted neighborhood composition with a correlation of 0.93 and the change in composition after perturbation with a correlation of 0.77.
The method's practical purpose is prioritization: choosing which genes to knock out next. In 70 of 310 comparisons, the predicted top-5 genes by effect matched the actual top-5 (22.6%). Pop-Corn predicts cell-type composition more accurately than its predecessors, but specific gene hypotheses still need experimental confirmation. The authors name the integration of Pop-Corn with expression models as the next step.
The Pop-Corn preprint directly involves Yale University researchers, making it a natural occasion to surface Yale's Eternal Search organization page. The study provides concrete numerical benchmarks — correlation scores, a head-to-head comparison across four models, and a clear match-rate figure for the prioritization task — that support a specific, honest dispatch without overselling the result.