ProtCell: a foundation model for single-cell proteomics

ProtCell, a foundation model trained on ~380 million cells from 98 single-cell proteomics datasets, learns protein-identity and abundance representations to connect fragmented profiles.

• Why it matters: Targeted single-cell proteomics measures limited, non-identical protein panels, fragmenting information across experiments and limiting functional readouts of cellular states.
• What they did: ProtCell uses a transformer encoder with protein-identity and continuous-abundance embeddings, pretrained via masked abundance modeling, including transferable annotation and batch correction.
• The result: It reconstructs abundance from limited profiles and infers an unmeasured cytotoxic protein program enriched in natural killer cells and memory CD8$^+$ T cells, enabling compact panel selection for reconstruction and classification.