Fig. 7: Ablation studies of scLong demonstrate the benefits of modeling low-expression genes and integrating the Gene Ontology (GO) graph.

a In the ablation study assessing the impact of modeling low-expression genes (LEGs), the full scLong model significantly outperformed two ablation settings: (1) omitting LEGs and the mini Performer encoder (w/o LEG) and (2) assigning random weights to the mini Performer encoder (Random LEG). b In the ablation study evaluating the role of the GO graph, scLong significantly outperformed two ablation settings: (1) removing the GO graph and the graph convolutional network (w/o GO) and (2) replacing the GO graph with a random graph (Random GO). Higher values indicate better performance across all metrics, except for mean squared error. In (a, b), bar heights represent the mean and error bars indicate the standard deviation across n = 5 independent training runs with different random seeds. Results from individual runs are shown as dot points. Source data are provided as a Source Data file. Two-sided t-tests with Benjamini-Hochberg correction were used; see Supplementary Tables 15, 16 for detailed statistics.