scPILOT transfers a known cellular response to another patient, cell type, or species
scPILOT transfers a known cellular response to another patient, cell type, or species
On August 29, Advanced Science published a paper on scPILOT, a computational model that uses a previously measured cellular response to predict the response in another patient, cell type, or species. During validation, the authors withheld the post-treatment cells from the specific context for which each prediction was made. The model received only the baseline cells, while the withheld measurements provided the final test.
Single-cell sequencing shows which genes are active in individual cells. An experiment measures one group of cells before treatment and another group after treatment. A conventional analysis therefore often reduces the response of the entire group to an average difference, losing variation among individual cells.
The authors use optimal transport, a mathematical method that constructs probabilistic pairings between two groups while minimizing their total discrepancy. It links cells measured before and after a previously studied treatment and estimates the shift in gene activity for each baseline cell.
The model then receives baseline cells from a new patient, another cell type, or another species. It searches previous experiments for groups with similar gene activity, transfers their cell-level shifts to the new group, and gives greater weight to more similar groups. By adding the transferred shift to the baseline state, scPILOT predicts gene activity after the same treatment.
Across seven immune cell types treated with interferon beta, an immune signaling protein, scPILOT achieved a mean R² of 0.945. This metric indicates how closely the predicted mean gene activity levels matched the measured values. The best competing model achieved 0.935. In a test involving eight patients, R² was 0.598, approximately the same as for CellOT, while scPILOT produced the smallest discrepancy between the distributions of individual cells.
Accuracy depended on the similarity of the baseline cell states. The three most similar contexts produced almost the same accuracy as all available contexts combined, while the three most distant contexts made the predictions worse. Across six tumor cell lines, differences between cell lines were greater than the effect of gene knockout. In that setting, scPILOT was comparable to the simple baseline-state prediction when evaluated by mean gene activity levels.
scPILOT uses previous experiments as a source of predictions by matching their cells to a new population. A transferred response becomes more useful as the baseline states become more similar.