In silico labeling in single cell multiplexed imaging

Spatially resolved single cell multiplexed imaging are emerging techniques that enable measuring the expression of several tens of proteins at a sub-cellular resolution and has the potential to bridge the gap between molecular profiling of a single cell and tissue organization. Prior to extensive data collection, pilot experiments are performed to optimize a panel of proteins according to the experimental assumptions and the specific research question. The determination of which proteins to include in the panel is not-entirely objective and essentially inclusion of every protein comes at the expense of the exclusion of another protein. We present in silico proteomics, estimating the expression of some proteins according to other correlated proteins. As a proof-of-concept we trained a deep neural network to map 35-to-1 single cell protein expression using a multiplexed ion beam imaging by time-of-flight(MIBI-TOF) Triple-Negative Breast Cancer (TNBC) dataset. In this dataset, we found three proteins, namely, CD45, H3K27me3, and H3K9ac, that can be accurately estimated in terms of protein expression correlation and cell type classification, indicating that new probes can replace these ones without much loss of information. An additional application of in silico proteomics is the identification of single cell anomalies in the proteins-protein mapping. We show that this anomaly-detection can be used to identify and correct erroneous cell type classification. Altogether, in silico proteomics can contribute as a valuable tool to the experimental design and analysis of single cell multiplexed imaging.