Compressed sensing expands the multiplexity of imaging mass cytometry Authors: Tsuyoshi Hosogane, Leonor Schubert Santana, Nils Eling, Holger Moch, Bernd Bodenmiller Published: 2025 ABSTRACT AbstractThe multiplexity of current antibody-based imaging is limited by the number of reporters that can be detected simultaneously. Compressed sensing can be used to reconstruct high-dimensional information from low-dimensional measurements. Previously, compressed sensing using composite in situ imaging (CISI) of transcriptomic data leveraged gene co-regulation structure to recover spatial expression of 37 RNA species from images of 11 composite channels. Here, we extend the CISI framework to protein expression data measured by imaging mass cytometry (IMC). CISI-IMC accurately recovers spatial expression of 16 immune and stromal marker proteins from images of 8 composite channels with an average Pearson’s correlation of 0.8 across protein. Training the CISI-IMC framework using data collected on multiple human tissues enables universal decompression of composite data from a wide range of tumor and healthy tissue types. The expression dictionary and barcoding matrix described here are immediately implementable for general immune and stromal cell type classification, but CISI-IMC can in principle be applied to other markers or other antibody-based imaging methods. Our work lays the foundation for much higher plex protein imaging.Subject terms: Machine learning, Proteomics, Molecular imaging, Tumour immunology Introduction Antibody-based protein imaging allows profiling of the spatial distribution of cell phenotypes across tissues, enabling the study of tissue function in health and disease. Detection of biomolecules in standard immunofluorescence imaging is limited by the number of available fluorescent channels, typically five, due to spectral overlap. This limitation has been overcome by performing iterative staining and de-staining using a few fluorescently labeled antibodies in each cycle, with each channel in each round of staining typically corresponding to a specific marker1–6. Iterative methods such as CODEX, 4i, and CyCIF can achieve up to 60-plex imaging2,3,7. Higher multiplexity has been achieved using combinatorial barcoding, where the sequence of channels over imaging rounds creates a unique barcode for each target molecule. For example, CosMx, which employs combinatorial barcoding, has been used to image 108 proteins in a tissue sample8. Limitations of these methods are autofluorescence and tissue damage caused by repeated staining cycles and increasing measurement time with each cycle. Combinatorial barcoding is also confounded by molecular crowding as this technique requires single-molecule detection of each target molecule for accurate decoding. Even higher multiplexity can be achieved by sequencing-based approaches, in which antibodies are labeled with a DNA tag that carries a barcode that can be sequenced to identify the location on the slide9. For example, 300-plex imaging was demonstrated with spatial CITE-seq10. Sequencing-based methods do not suffer from autofluorescence and do not require repeated staining cycles, but currently the resolution is too low (25 μm) to allow single-cell analyses. In contrast to the fluorescence- and sequencing-based methods, mass spectrometry-based imaging methods such as multiplex ion beam imaging (MIBI) and imaging mass cytometry (IMC) enable over 40-plex imaging without sequential staining cycles at a spatial resolution of 1 μm or higher11,12. For both methods, antibodies are tagged with metal isotopes that are simultaneously detected by either secondary ion mass spectrometry (MIBI) or laser ablation and subsequent mass spectrometry (IMC). The multiplexity of MIBI and IMC is limited by the number of available channels to detect individual metal isotopes. MIBI and IMC already enable large scale sample analyses and since the number of measured antibodies does not affect measurement time, approaches to increase multiplexity are highly attractive to enable deeper tissue analysis. Combinatorial barcoding could be implemented by labeling each antibody with a unique combination of metal isotopes, but given the high abundance of proteins, single-molecule detection that is required for accurate decoding is not feasible for these techniques. Expanding MIBI and IMC beyond 40-plex would be possible with a method that can decode combinatorial barcodes without the need for single-molecule detection. In the RNA imaging field, composite in situ imaging (CISI) has been implemented to decode combinatorial barcodes without single-molecule detection by using prior knowledge of transcriptome profiles13,14. Leveraging the biological principle of gene co-regulation, CISI constructs a dictionary of gene expression modules from training data, which enables the decoding into the spatial expression patterns of individual genes. By training on single-cell RNA sequencing data, CISI accurately recovered the spatial abundance of 37 RNA species from the combinatorially barcoded imaging data of only 11 channels14. CISI decoding can be performed on the single-cell or pixel-level and therefore only requires resolution high enough for single-cell analysis. Here, we implement CISI for mass-tag based imaging approaches such as IMC and MIBI to expand multiplexity over the number of available isotope channels. We reason that a protein expression dictionary for IMC data could be constructed since proteins, like mRNAs, show inter-dependencies due to co-regulation. We perform 16-plex standard IMC imaging of multiple tissues using antibodies that mainly identify immune and stromal cells for training data and constructed a dictionary of protein expression patterns. We then implement combinatorial barcoding by mixing 16 antibodies each labeled with unique combination of 8 metal isotopes. We demonstrate that the obtained 8-channel CISI-IMC images for 71 different tissues are accurately decoded back to 16-plex spatial single-cell protein expression data (average Pearson’s correlation of 0.8) using our optimized CISI algorithm. Results CISI-IMC principles Our goal was to compress high-dimensional imaging data into low-dimensional composite images and then recover the original images using computational methods and prior knowledge of the patterns in imaging data (Fig. 1). To obtain low-dimensional, single-cell composite data using IMC, antibodies targeting p proteins are labeled with unique combinations of m isotopes, where m