The scMLEAge code includes data from test mice in age profiles before validation
The scMLEAge code includes data from test mice in age profiles before validation
On 12 September, npj Aging published a paper on scMLEAge, a program that assigns individual cells to age groups based on the number of RNA molecules counted for their genes. In the repository linked in the paper, the code builds shared profiles before splitting mice into training and test groups.
scMLEAge learns from the ages of donor mice. For each cell type, the program sums RNA molecule counts across cells from the same age group to produce an age profile: the characteristic proportions of reads from different genes. It then compares an individual cell's profile with these age profiles and selects the most likely age group. A Poisson model calculates this likelihood from integer RNA counts.
The authors used Tabula Muris Senis, a dataset of RNA measurements from mouse cells, selecting data from nine tissues in ten males. They reported R² values from 0,17 to 0,95. R², the squared correlation, measures how closely the predicted age group is associated with donor age. In the validation described in the paper, a subset of mice is held out for testing in each of five rounds. The number of genes is selected using the remaining mice, and predictions are then compared with the ages of the held-out donors.
The available repository code follows a different order. It first ranks genes by their association with age using cells from all mice and builds age profiles from those cells. It then splits the donors into five groups. After this split, the code selects the number of genes using cells from the training mice, but relies on the existing ranking and profiles, which already incorporate cells from the mice subsequently assigned to the test group.
The resulting test R² describes a procedure whose shared parameters already incorporate data from every mouse. To assess performance on held-out donors, each of the five runs needs to rank genes, build profiles, and select the number of genes afresh using only the training mice, then apply the fitted procedure to the test mice. In PanSci, another mouse cell dataset, the authors obtained an R² of 0,75 for muscle satellite cells after building a separate clock using data from that atlas.