Protein Imputation

scpFormer had its largest qualitative imputation advantage for low-abundance or nonlinear markers. In the Levine masked-protein benchmark, scpFormer achieved mean Pearson correlation 0.751, outperforming random forest, k-nearest neighbors, and linear regression. For FLT3 imputation, scpFormer reached correlation 0.452 while conventional methods were near 0.20. scpFormer's zero-shot imputation on MIS-C produced the best biological conservation and clustering agreement among baseline imputation strategies. Because naturally missing proteins lacked ground truth, zero-shot clinical imputation was evaluated through topology and biological manifold recovery with scIB. For imputation, scpFormer adapts the same self-decoder objective used during pretraining to reconstruct missing protein expression values. The MIS-C zero-shot experiment used GSE166489 without dataset-specific fine-tuning.