
scBOND and scBOND-Aug demonstrates robustness and superior performance under data perturbations. (A) Robustness evaluation of scBOND and baseline methods under increasing DNA methylation noise levels (1%–5%) on the HumanBrainA data set. (B) Radar plots comparing the translation performance of scBOND, scBOND-Aug, scCross, and MAPLE across four clustering metrics on two paired data sets. (C) Downsampling experiments on the HumanBrainA data set to evaluate robustness under varying training sample sizes, comparing methods of scBOND, scBOND-Aug, scCross, and MAPLE. The x-axis indicates the proportion of training cells sampled.











