The appeal of deep visual proteomics is easy to state — identify cells by their appearance, isolate exactly those cells, and measure what proteins they contain. High-resolution microscopy could already do the first part and ultra-sensitive mass spectrometry the last. Understanding what actually had to be solved in between is more instructive than either half, and it tells you what implementing this would require.
Key Takeaways
|
What Problem Does Deep Visual Proteomics Solve?
A connection problem rather than a measurement problem — and that distinction is the whole story. The Nature Biotechnology paper introducing the method is unusually direct about this: having combined sub-micron-resolution imaging, artificial-intelligence-based single-cell phenotyping, and an ultra-sensitive proteomics workflow, the authors report that the key challenges turned out to be the accurate definition of single-cell boundaries and cell classes, and the transfer of the automatically defined features into proteomic samples ready for analysis.
Read that second challenge carefully, because it is the one that had no existing solution. An algorithm can identify which pixels belong to which cell. A microdissection microscope can cut along a contour. But the contour exists in the coordinate space of one instrument, and the cutting happens in the coordinate space of another, and nothing connected them. The software introduced to address this, named BIAS for Biology Image Analysis Software and described as coordinating scanning and laser microdissection microscopes, is essentially a bridge between two machines that were not designed to talk to each other.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
That reframing is worth holding onto, because it explains the timing. Deep learning segmentation, laser microdissection, and single-cell-sensitivity mass spectrometry all existed before this method was described. What did not exist was the plumbing. It also tells you where the difficulty will be if you attempt something similar: not in choosing an instrument but in making two instruments agree about where a cell is.
Stage | What Happens | The Technical Problem |
Imaging | High-resolution scanning of cell culture or archival FFPE tissue, with immunofluorescence or immunohistochemical staining | Producing images rich enough to classify phenotype, across formats and dimensionalities |
Segmentation | Deep-learning-based definition of nuclear and cytoplasmic boundaries | Accurate single-cell boundaries. Benchmarked here against three alternative methods using the F1 metric |
Classification | Machine-learning identification of cell types and states from extracted features | Defining cell classes that correspond to something biologically real |
Transfer | AI-defined contours passed into the microdissection microscope’s coordinate space | The problem with no prior solution, and the reason coordinating software was needed |
Excision | Automated laser microdissection of the selected contours into collection plates | Cutting accurately without damaging the material, addressed with a definable offset |
Proteomics | Ultra-sensitive LC-MS on very small isolates, acquired in data-independent mode | Depth from minimal material, as covered in the single-cell and microdissection articles |
Table 1. The workflow with the technical problem at each stage rather than a description of the stage. The transfer row is the innovation; the others are components that already existed.
AI-Guided Cell Selection
Two capabilities operate in sequence here, and they are worth separating because they fail differently.
Segmentation defines where each cell is. The published implementation uses deep learning applied to nuclei and cytoplasm in both normal-appearing and cancer cells and tissue, with image style transfer learning, and its accuracy was benchmarked using the F1 metric against three alternative approaches. Segmentation errors propagate directly into the proteomics: a boundary drawn wrongly means material from the wrong cell enters the sample, and no downstream analysis can detect that.
Classification then decides which cells are of interest, using machine learning on features extracted from the segmented objects. The coordinating software handles multiple two- and three-dimensional microscopy formats, performing pre-processing, segmentation, feature extraction, and phenotype classification in one place. In practice, published work has combined this classification with expression levels from immunofluorescence and immunohistochemical staining for more precise cell classification, so the phenotype is defined by both morphology and marker signal.
Selected by the AI Alone, or After Instruction The paper notes that cellular or subcellular objects of interest are selected by the AI alone or after instruction. That is a genuine fork rather than a technicality, and it determines what a result can claim. In the supervised mode, you tell the system which phenotypes matter and it finds them, which means the analysis inherits your prior about what is interesting. That is entirely legitimate and it is how most published applications operate, but the resulting proteome describes categories you defined. In the unsupervised mode, the system groups cells by appearance without instruction, which can in principle surface phenotypes nobody specified. The second is the more interesting claim and the harder one to make convincingly, since a cluster of similar-looking cells is not automatically a biologically meaningful class. |
This is also the concrete form of the marker-dependency point made in Spatial Proteomics by Mass Spectrometry: LCM, Single-Cell, and Imaging Approaches. The proteome measurement is untargeted, but in supervised mode the cell selection is not, and reporting which mode was used is part of describing the method honestly.
From Contour to Sample
The excision step contains a detail that overview coverage consistently omits and that anyone implementing this needs to know.
Contours generated upstream are aligned in the microdissection microscope’s own software before cutting, and the cut path can be offset from the contour by a definable amount specifically to prevent potential laser-induced damage to the material being collected. That is a small parameter with a real consequence: cut exactly on the boundary and you risk ablating the outer edge of the cell you are trying to measure; cut too far outside it and you collect neighbouring material. The offset is a tunable compromise between contamination and damage, and it is the kind of parameter that separates a working protocol from a disappointing one.
Collected material lands in multiwell plates, with published work using 384-well formats, after which it enters a low-input proteomics workflow. The handling constraints from that point are identical to those described in Laser Capture Microdissection + LC-MS Proteomics, and the automation now available for that stage is substantial: a reported automated preparation protocol processes 192 samples in three hours, collecting microdissected material directly into low-volume chips and performing lysis, de-crosslinking, and digestion in place.
LC-MS on Phenotyped Cells
The mass spectrometry itself is conventional ultra-sensitive proteomics rather than anything specific to this method, which is worth saying because it means the depth achievable tracks the wider single-cell field rather than lagging it.
Measurements are acquired in data-independent mode, which suits very small samples because it does not depend on selecting precursors for fragmentation in real time from a sparse population. Achievable depth follows the trajectory described in Single-Cell Proteomics by Mass Spectrometry, with the important qualification that a microdissected cell or cell slice contains less material than a whole cell in suspension.
Two design levers govern depth here, and the second is the one most often left implicit.
- Instrument and acquisition sensitivity. The same improvements driving suspension single-cell proteomics apply directly, since the sample entering the instrument is comparable.
- How many cells you pool. Pooling morphologically similar cells buys depth, a strategy the field terms biological fractionation. A 2025 review of single-cell proteomics solutions sets out why pooling is attractive: since protein amplification is not possible, more material is the only route to more depth.
The consequence is that DVP results span a range from genuinely single-cell to morphologically homogeneous pools, and the two are not interchangeable when reading a depth figure. Single-cell Deep Visual Proteomics reported 1,700 proteins from an individual hepatocyte cell slice with position retained, and found half the measured proteome to be spatially regulated with levels changing markedly near the central vein. A pooled experiment on the same tissue would reach greater depth and lose the single-cell resolution that made that finding visible.
Does It Actually Reproduce?
Reasonable evidence exists, and it is worth knowing because capability claims for new methods often arrive without it. Three published results bear on whether the approach does what it says.
- Quantitative reproducibility. Replicate proteome measurements of cells and nuclei demonstrated a robust workflow with high quantitative reproducibility, reported at a Pearson correlation of 0.96. For a workflow with this many stages, that is a meaningful figure.
- Orthogonal validation of subcellular selection. Proteomes of whole cells were very different from those of nuclei alone, in the same manner as subcellular proteomics based on biochemical separation, and this was reflected in enrichment analysis with terms such as plasma membrane, mitochondrion, and nucleus. That is the right kind of evidence: an independent method predicts the difference the new method observes.
- Segmentation benchmarked rather than asserted. Accuracy was quantified using the F1 metric and compared against three alternative segmentation methods, with true positive, false positive, and false negative areas visualised.
The second of those is the strongest, and it is the pattern to look for when assessing any spatial method. If a technique claims to isolate a subcellular compartment, the proteome it returns should resemble what biochemical fractionation of that compartment returns. It did. That is more persuasive than any workflow diagram, and it is the reason the subcellular claims can be taken seriously.
Where Does This Fit in Pathology and Discovery?
Its distinguishing feature for applied work is that it runs on material that already exists. Published applications used archived biobank tissue, including formalin-fixed and paraffin-embedded material, which means the method addresses existing cohorts rather than requiring prospective collection.
Use Case | Why DVP Suits It | What to Watch |
Phenotype-defined populations in archival tissue | Runs on FFPE biobank material, so retrospective cohorts are accessible | Fixation and staining both affect protein recovery |
Disease progression within one specimen | Cells at different stages can be selected separately from the same section | Selection criteria define the stages, so they must be stated |
Subcellular proteome comparison | Nuclei can be excised separately from whole cells, validated against biochemical fractionation | Segmentation accuracy sets the ceiling on compartment purity |
Rare cell populations | Cells are found by appearance rather than needing to be abundant | Depth per cell is limited, so pooling may be necessary |
Discovery of unspecified phenotypes | Unsupervised selection can group cells without prior instruction | A visual cluster is not automatically a biological class |
Table 2. Where imaging-guided proteomics is the appropriate choice, and the caveat attached to each. The last row is the most ambitious use and the least established.
One practical observation for anyone assessing whether to pursue this. The method requires a scanning microscope, a laser microdissection microscope, software connecting them, low-input sample handling, and a sensitive mass spectrometer, together with expertise spanning image analysis and proteomics. That is a substantial stack, and for most laboratories the realistic route is collaboration rather than local implementation. The question worth asking is whether your biological question genuinely requires cells selected by appearance, or whether a region-level approach would answer it, since region-level microdissection is considerably more accessible and reaches greater depth.
For that comparison, and for how imaging-guided selection differs from antibody-based detection of chosen proteins, see MS-Based vs. Antibody-Based Spatial Proteomics: Choosing an Approach. For where this sits among the spatial modalities generally, see Spatial Analysis in Analytical Science: Mass Spectrometry Imaging and Spatial Omics.
This article was produced under Separation Science's AI Editorial Guidelines.
