Preprint: an RNA-based model predicted that old cells would respond differently to the knockout of the same gene
Preprint: an RNA-based model predicted that old cells would respond differently to the knockout of the same gene
On October 13, the authors posted a preprint describing DKO, a computational model that uses a single cell’s RNA profile to predict the effects of knocking out an active gene. Across data from humans, mice, fruit flies and nematodes, predicted responses varied more among old cells in most comparisons. This is a computational finding: the model infers changes in RNA profiles from observational data.
Single-cell analyses usually compare average gene activity between young and old cells. The authors asked a different question: if one active gene is computationally knocked out in each cell’s profile, how similar are the predicted effects across cells of the same type?
The study rests on the hypothesis that random molecular damage accumulated with age makes coordinated gene activity less stable. If so, the loss of the same gene could cause different RNA changes in otherwise similar cells.
DKO was trained on gene activity profiles from cells of the same type. The model then removes one active gene at a time from an individual cell’s set of active genes and generates a predicted RNA profile. The difference between the original and predicted profiles indicates how much the model expects that cell’s profile to change. By repeating this calculation across many cells and genes, the authors identified genes whose knockout produced more variable predicted responses across cells.
An August comparison of eight virtual knockout methods distinguished these models by what they measure: some select candidate genes for experiments, others model changes in cell state, and others predict changes in an individual gene’s activity. DKO belongs to the second group and predicts changes across the entire RNA profile.
The authors first tested DKO on artificial gene networks. They gradually increased the proportion of randomly altered connections in these networks and could calculate the exact effects of each knockout. DKO received only the original profiles, yet predicted the same increase in response variability. The authors then applied it to real single-cell data.
In 41 of 45 comparisons of cell populations, the predicted effects of knocking out the same gene varied more among old cells; the difference was statistically significant in 40 cases. The analysis included human ovary and skeletal muscle, mouse tissues, fruit fly brain and cells from the nematode Caenorhabditis elegans. The authors also trained DKO separately on young mouse bone marrow cells alone and then on old cells alone. In both cases, predicted responses remained more variable among old cells.
The authors also looked for this pattern within individual cell populations. In pancreatic glucagon-producing cells from four of five participants, predicted gene knockout effects in cells with a higher proportion of the age-associated mutational signature S1 differed more from a reference derived from cells with a lower proportion of that signature. They found a similar association in connective tissue cells from the hearts of old mice. In cultures of human umbilical vein endothelial cells and WI-38 fibroblasts, a higher senescence score was associated with greater divergence in predicted responses. Senescence is a state in which a cell stops dividing and changes its functional program.
The authors interpret this pattern as a loss of stability in gene regulation with age. They propose using DKO to select genes and biological pathways for subsequent experimental gene knockouts in cells.