scGPT and Geneformer distinguished cell ages in flies, worms, mice, and humans

scGPT and Geneformer distinguished cell ages in flies, worms, mice, and humans

On July 27, the authors released a preprint describing how they fine-tuned scGPT and Geneformer on 1,3 million individual cells from four species. Across the full dataset, the models assigned cells to young, middle-aged, or old groups with accuracies of 78,6% and 80,5%.

A fly head, a whole worm, mouse tissues, and human blood cells provide very different snapshots of gene activity. The authors mapped genes from these species to 2 337 unambiguous human counterparts and gave both models the same task: determine the age group of each cell.

scGPT uses gene activity levels, whereas Geneformer uses genes ranked by activity. Both models detect an age-related signal, but they identify different sets of genes as its basis. When the data were split by donor, accuracy was 51–73%. When a model was trained on one species and tested on another, accuracy was 29–41%.

The most consistent result came from scGPT. In all four species, the model ranked RPL12, a ribosomal protein gene, as the top feature. Ribosomes assemble proteins according to RNA instructions, and ribosomal genes were particularly informative for this classification. When the authors masked 51 ribosomal genes, scGPT accuracy fell by 22–28 percentage points. Masking a random set of 51 genes reduced it by only 0,3 points.

Geneformer also frequently ranked ribosomal genes among the top features in mice and humans. In flies and worms, it highlighted genes involved in signaling, DNA packaging, and ubiquitin ligases, which are proteins that mark other proteins for degradation. The authors attribute this difference to the models' inputs: scGPT uses the magnitude of gene activity, whereas Geneformer uses its rank order.

RPS13 was the only gene among the top 50 features from both models in every one of the four species. Features identified by both models provide specific candidates for experiments that can be repeated across different organisms.