Fig. 2

Pareto front in \(({\overline{U}}_n, {\overline{V}}_n)\) space estimated via Pareto MTL with MINE and associated t-SNE plots. Panel A shows all 12 Pareto candidates (left) and the culled non-dominated points (right) in the \(({\overline{U}}_n, {\overline{V}}_n)\) space on TM-MARROW dataset. Panel B, C and D show the t-SNE plots of the latent \(\varvec{z}\) for the third, sixth and tenth candidate on test set, respectively. Each point in the t-SNE plots indicates a cell and the points are colored by batches (left) and pre-annotated cell types (right). As expected, from Panel B to Panel D, we see increasingly better clustering performance at the cost of more batch effect