Draft:Deep Visual Proteomics

The study of proteins has been a key focus in biology for over a century, but the field of proteomics emerged in the late 20th century..{{Cite journal |last=Bradshaw |first=Ralph A. |date=2023-05-04 |editor-last=Wade |editor-first=John |title=On the development of proteomics: a brief history |url=https://www.publish.csiro.au/CH/CH23012 |journal=Australian Journal of Chemistry |language=en |volume=76 |issue=8 |pages=418–428 |doi=10.1071/CH23012 |issn=0004-9425}} Early protein research relied on biochemical techniques such as chromatography and electrophoresis to separate and analyze proteins. In the 1970s, scientists used two-dimensional gel electrophoresis which allowed them to separate thousands of proteins in a sample based on their size and charge

The development of mass spectrometry (MS) for protein analysis in the 1980s and 1990s revolutionized proteomics{{Cite journal |last1=Domon |first1=Bruno |last2=Aebersold |first2=Ruedi |date=2006-04-14 |title=Mass Spectrometry and Protein Analysis |url=https://www.science.org/doi/10.1126/science.1124619 |journal=Science |language=en |volume=312 |issue=5771 |pages=212–217 |doi=10.1126/science.1124619 |pmid=16614208 |bibcode=2006Sci...312..212D |issn=0036-8075}}.Techniques like matrix-assisted laser desorption/ionization (MALDI) and electrospray ionization (ESI) enabled researchers to identify and quantify proteins with high precision.{{Cite book |url=https://onlinelibrary.wiley.com/doi/book/10.1002/9783527335961 |title=MALDI MS: A Practical Guide to Instrumentation, Methods, and Applications |date=2013-11-13 |publisher=Wiley |isbn=978-3-527-33331-8 |editor-last=Hillenkamp |editor-first=Franz |edition=1 |language=en |doi=10.1002/9783527335961 |editor-last2=Peter-Katalinic |editor-first2=Jasna}}{{Cite journal |last=Salata |first=Oleg |date=2005-01-01 |title=Tools of Nanotechnology: Electrospray |url=https://www.eurekaselect.com/article/6582 |journal=Current Nanoscience |volume=1 |issue=1 |pages=25–33 |doi=10.2174/1573413052953192 |bibcode=2005CNan....1...25S |issn=1573-4137}} The sequencing of the human genome in the early 2000s further accelerated proteomics by providing a reference for protein identification.{{Cite journal |last1=Venter |first1=J. Craig |last2=Adams |first2=Mark D. |last3=Myers |first3=Eugene W. |last4=Li |first4=Peter W. |last5=Mural |first5=Richard J. |last6=Sutton |first6=Granger G. |last7=Smith |first7=Hamilton O. |last8=Yandell |first8=Mark |last9=Evans |first9=Cheryl A. |last10=Holt |first10=Robert A. |last11=Gocayne |first11=Jeannine D. |last12=Amanatides |first12=Peter |last13=Ballew |first13=Richard M. |last14=Huson |first14=Daniel H. |last15=Wortman |first15=Jennifer Russo |date=2001-02-16 |title=The Sequence of the Human Genome |url=https://www.science.org/doi/10.1126/science.1058040 |journal=Science |language=en |volume=291 |issue=5507 |pages=1304–1351 |doi=10.1126/science.1058040 |pmid=11181995 |bibcode=2001Sci...291.1304V |issn=0036-8075}}

In the 21st century, advances in liquid chromatography, mass spectrometry, and bioinformatics made proteomics more comprehensive and high-throughput.{{Cite journal |last1=Pelletier |first1=Simon J. |last2=Leclercq |first2=Mickaël |last3=Roux-Dalvai |first3=Florence |last4=de Geus |first4=Matthijs B. |last5=Leslie |first5=Shannon |last6=Wang |first6=Weiwei |last7=Lam |first7=TuKiet T. |last8=Nairn |first8=Angus C. |last9=Arnold |first9=Steven E. |last10=Carlyle |first10=Becky C. |last11=Precioso |first11=Frédéric |last12=Droit |first12=Arnaud |date=2024-05-06 |title=BERNN: Enhancing classification of Liquid Chromatography Mass Spectrometry data with batch effect removal neural networks |journal=Nature Communications |language=en |volume=15 |issue=1 |page=3777 |doi=10.1038/s41467-024-48177-5 |issn=2041-1723 |pmc=11074280 |pmid=38710683|bibcode=2024NatCo..15.3777P }} Scientist could research proteins in higher resolution and analyze more complex protein mixtures. These advancements, along with high-resolution microscopy, artificial intelligence, and laser microdissection, have made Deep Visual Proteomics (DVP) possible, enabling the detailed mapping of proteins within individual cells while preserving their spatial organization.{{cn}}

Current proteomics methods mostly rely on bulk mass spectrometry (MS) to measure protein levels across large groups of cells. However, this approach averages the data from all the cells in a sample, making it difficult to detect differences between individual cells. As a result, important variations in protein expression that may play a role in biological processes can be missed. There are existing spatial proteomics techniques, including immunohistochemistry (IHC) and imaging mass spectrometry (IMS).{{Cite journal |last1=Zhu |first1=Shaobo |last2=Schuerch |first2=Conrad |last3=Hunt |first3=Jennifer |date=2015-01-01 |title=Review and Updates of Immunohistochemistry in Selected Salivary Gland and Head and Neck Tumors |url=http://meridian.allenpress.com/aplm/article/139/1/55/100478/Review-and-Updates-of-Immunohistochemistry-in |journal=Archives of Pathology & Laboratory Medicine |volume=139 |issue=1 |pages=55–66 |doi=10.5858/arpa.2014-0167-RA |issn=1543-2165 |pmid=25549144}}{{Cite journal |last1=Buchberger |first1=Amanda Rae |last2=DeLaney |first2=Kellen |last3=Johnson |first3=Jillian |last4=Li |first4=Lingjun |date=2018-01-02 |title=Mass Spectrometry Imaging: A Review of Emerging Advancements and Future Insights |journal=Analytical Chemistry |volume=90 |issue=1 |pages=240–265 |doi=10.1021/acs.analchem.7b04733 |issn=0003-2700 |pmc=5959842 |pmid=29155564}} However, IHC is limited to detecting a predefined set of proteins, and IMS lacks the sensitivity and depth of MS-based approaches.{{cn}}

Deep Visual Proteomics (DVP) takes a different approach by combining imaging techniques, artificial intelligence, laser microdissection, and mass spectrometry. Instead of analyzing all proteins together, DVP allows researchers to examine proteins in specific cells while preserving their spatial organization. This helps in studying how protein expression varies between different cell types within the same tissue and how proteins are distributed in conditions such as cancer.{{cn}}

In Deep Visual Proteomics (DVP), the first step is preparing biological samples to keep their structure intact for detailed imaging and protein analysis. This includes choosing the tissue, fixing it, staining it with special markers, and placing it on slides for later analysis.{{cn}}

DVP works well with formalin-fixed paraffin-embedded (FFPE) tissues, which are commonly stored in biobanks, as well as fresh or frozen tissues.{{Cite journal |last1=Gerdes |first1=Michael J. |last2=Sevinsky |first2=Christopher J. |last3=Sood |first3=Anup |last4=Adak |first4=Sudeshna |last5=Bello |first5=Musodiq O. |last6=Bordwell |first6=Alexander |last7=Can |first7=Ali |last8=Corwin |first8=Alex |last9=Dinn |first9=Sean |last10=Filkins |first10=Robert J. |last11=Hollman |first11=Denise |last12=Kamath |first12=Vidya |last13=Kaanumalle |first13=Sireesha |last14=Kenny |first14=Kevin |last15=Larsen |first15=Melinda |date=2013-07-16 |title=Highly multiplexed single-cell analysis of formalin-fixed, paraffin-embedded cancer tissue |journal=Proceedings of the National Academy of Sciences |language=en |volume=110 |issue=29 |pages=11982–11987 |doi=10.1073/pnas.1300136110 |doi-access=free |issn=0027-8424 |pmc=3718135 |pmid=23818604|bibcode=2013PNAS..11011982G }} The tissue is stained with specific markers that highlight certain proteins, allowing researchers to identify different cell types. Multiplexed immunofluorescence staining can be used to label multiple proteins at once, giving researchers a better understanding of the tissue's complexity.{{Citation |last1=Francisco-Cruz |first1=Alejandro |title=Multiplex Immunofluorescence Assays |date=2020 |work=Biomarkers for Immunotherapy of Cancer |volume=2055 |pages=467–495 |editor-last=Thurin |editor-first=Magdalena |url=https://link.springer.com/10.1007/978-1-4939-9773-2_22 |access-date=2025-03-06 |place=New York, NY |publisher=Springer New York |language=en |doi=10.1007/978-1-4939-9773-2_22 |isbn=978-1-4939-9772-5 |last2=Parra |first2=Edwin Roger |last3=Tetzlaff |first3=Michael T. |last4=Wistuba |first4=Ignacio I. |pmid=31502166 |editor2-last=Cesano |editor2-first=Alessandra |editor3-last=Marincola |editor3-first=Francesco M.}}

The tissue is then mounted on polyethylene naphthalate (PEN)-coated slides designed for laser microdissection (LMD). These slides provide the necessary support to isolate individual cells with high precision. For FFPE tissues, additional steps like deparaffinization (removing the wax used for preservation) and heat-induced epitope retrieval (uncovering protein markers) are performed. This ensures the tissue is properly prepared for staining and analysis.{{cn}}

This detailed preparation preserves the tissue's natural structure and allows researchers to study proteins in the context of their original environment, which is essential for understanding how cells function in living organisms.{{cn}}

In Deep Visual Proteomics (DVP), high-resolution imaging is used to capture detailed images of tissue samples with techniques like confocal or wide-field fluorescence microscopy. These images show individual cells, nuclei, and subcellular structures, preserving the spatial relationships between cells and their environment. This detail helps identify cellular features and links them to proteomic data.{{cn}}

AI-driven analysis is then applied to process the images. Using BIAS (Biology Image Analysis Software) and deep learning tools like nucleAIzer, the software segments cells and classifies them based on size, shape, and staining patterns. This helps identify different cell types and rare subpopulations. The AI models are trained on both real and synthetic images to improve their accuracy while lowering labour costs to generate images of real cells.{{cn}}

File:Dvp2.jpg

After automated cell selection, a laser capture microdissection system (e.g., Zeiss PALM MicroBeam or Leica LMD7) precisely cuts out targeted cells with sub-micron accuracy. Path-finding algorithms guide the laser along the cell boundaries identified by the software, preserving spatial details down to about 200 nanometers and minimizing damage. This approach prevents contamination from neighboring cells, making it suitable for high-sensitivity proteomics studies.{{cn}}

After proteins are extracted, Deep Visual Proteomics (DVP) utilizes Data-independent acquisition (DIA)—particularly the diaPASEF (parallel accumulation–serial fragmentation combined with data-independent acquisition) method—to simultaneously profile a broad range of peptides, enabling the detection of low-abundance proteins.. Ion Mobility spectrometry adds another separation layer based on peptide ion properties, enhancing resolution and increasing the identification of proteins that might otherwise go undetected[15]. To ensure consistent and accurate results, the workflow incorporates internal standards and technical replicates, maintaining reliable and reproducible protein quantification across single-cell

File:Computational components of DVP.png

In Deep Visual Proteomics (DVP), artificial intelligence (AI) plays a crucial role in analyzing biological samples. The process begins with high-resolution imaging, where detailed pictures of tissue samples are taken. These images are then processed using deep learning models, particularly the BIAS (Biology Image Analysis Software) platform, which automatically detects and outlines individual cells with high accuracy. DVP employs pre-trained deep neural networks trained on synthetic microscopy images, allowing the model to adapt to different tissue types and staining techniques.{{Cite book |last1=Isola |first1=Phillip |title=2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) |last2=Zhu |first2=Jun-Yan |last3=Zhou |first3=Tinghui |last4=Efros |first4=Alexei A. |date=July 2017 |publisher=IEEE |isbn=978-1-5386-0457-1 |pages=5967–5976 |chapter=Image-to-Image Translation with Conditional Adversarial Networks |doi=10.1109/CVPR.2017.632 |chapter-url=https://ieeexplore.ieee.org/document/8100115 |arxiv=1611.07004}} BIAS has been shown in research to carry advantages over other segmentation tools like unet4nuclei, Cellpose, and CellProfiler, distinguishing single-cell features.{{Cite journal |last1=Stringer |first1=Carsen |last2=Wang |first2=Tim |last3=Michaelos |first3=Michalis |last4=Pachitariu |first4=Marius |date=January 2021 |title=Cellpose: a generalist algorithm for cellular segmentation |url=https://www.nature.com/articles/s41592-020-01018-x |journal=Nature Methods |volume=18 |issue=1 |pages=100–106 |doi=10.1038/s41592-020-01018-x |issn=1548-7105 |pmid=33318659}}{{Cite journal |last1=Carpenter |first1=Anne E |last2=Jones |first2=Thouis R |last3=Lamprecht |first3=Michael R |last4=Clarke |first4=Colin |last5=Kang |first5=In Han |last6=Friman |first6=Ola |last7=Guertin |first7=David A |last8=Chang |first8=Joo Han |last9=Lindquist |first9=Robert A |last10=Moffat |first10=Jason |last11=Golland |first11=Polina |last12=Sabatini |first12=David M |date=2006-10-31 |title=CellProfiler: image analysis software for identifying and quantifying cell phenotypes |journal=Genome Biology |volume=7 |issue=10 |pages=R100 |doi=10.1186/gb-2006-7-10-r100 |issn=1474-760X |pmc=1794559 |pmid=17076895 |doi-access=free}}

After segmenting cells, machine learning (ML) algorithms come into play for feature extraction and classification. For example, supervised learning can classify cells using predefined markers like FOXJ1, which identifies ciliated cells, while unsupervised learning groups cells based on similar features without prior labeling.{{Cite journal |last=Lengyel |first=Ernst |date=September 2010 |title=Ovarian Cancer Development and Metastasis |journal=The American Journal of Pathology |volume=177 |issue=3 |pages=1053–1064 |doi=10.2353/ajpath.2010.100105 |pmc=2928939 |pmid=20651229}} ML also quantifies morphological traits, such as cell size and shape, and evaluates how cells interact with their neighbors.{{Cite journal |last1=Berg |first1=Stuart |last2=Kutra |first2=Dominik |last3=Kroeger |first3=Thorben |last4=Straehle |first4=Christoph N. |last5=Kausler |first5=Bernhard X. |last6=Haubold |first6=Carsten |last7=Schiegg |first7=Martin |last8=Ales |first8=Janez |last9=Beier |first9=Thorsten |last10=Rudy |first10=Markus |last11=Eren |first11=Kemal |last12=Cervantes |first12=Jaime I. |last13=Xu |first13=Buote |last14=Beuttenmueller |first14=Fynn |last15=Wolny |first15=Adrian |date=December 2019 |title=ilastik: interactive machine learning for (bio)image analysis |url=https://www.nature.com/articles/s41592-019-0582-9 |journal=Nature Methods |volume=16 |issue=12 |pages=1226–1232 |doi=10.1038/s41592-019-0582-9 |issn=1548-7105 |pmid=31570887}} This approach is particularly useful for identifying rare or novel cell subpopulations, such as those found in heterogeneous tissues like tumors.{{cn}}

Once individual cells are isolated via laser microdissection, their proteins are analyzed using mass spectrometry (MS), generating large proteomic datasets. Bioinformatics tools are then applied to process this data, ensuring protein features are aligned across different samples for comparison. Techniques like Data-Independent Acquisition (DIA) and diaPASEF are used to improve the detection of low-abundance proteins,.{{Cite journal |last1=Meier |first1=Florian |last2=Brunner |first2=Andreas-David |last3=Frank |first3=Max |last4=Ha |first4=Annie |last5=Bludau |first5=Isabell |last6=Voytik |first6=Eugenia |last7=Kaspar-Schoenefeld |first7=Stephanie |last8=Lubeck |first8=Markus |last9=Raether |first9=Oliver |last10=Bache |first10=Nicolai |last11=Aebersold |first11=Ruedi |last12=Collins |first12=Ben C. |last13=Röst |first13=Hannes L. |last14=Mann |first14=Matthias |date=December 2020 |title=diaPASEF: parallel accumulation–serial fragmentation combined with data-independent acquisition |url=https://www.nature.com/articles/s41592-020-00998-0 |journal=Nature Methods |volume=17 |issue=12 |pages=1229–1236 |doi=10.1038/s41592-020-00998-0 |issn=1548-7105 |pmid=33257825}} providing a more comprehensive view of the proteome. After identifying proteins, bioinformatics models map them to biological pathways, revealing insights into cellular functions. For example, in cancer studies, DVP has been used to track changes in protein expression that reflect disease progression, such as mRNA splicing dysregulation in melanoma and altered interferon signaling as tumors progress

File:Dvp3.png

Deep Visual Proteomics (DVP) integrates high-content imaging with sensitive proteomic methods to analyze single-cell protein expression while maintaining spatial context, offering detailed insights into tissue organization and disease mechanisms that traditional methods often miss.

DVP helps profile tumor heterogeneity by revealing spatially distinct microenvironments linked to disease progression.. In melanoma, it uncovers proteomic shifts—such as reduced interferon signaling and antigen presentation—that support immune evasion and metastasis. DVP also identifies rare cancer-specific protein signatures, as seen in salivary gland acinic cell carcinoma, which were previously undetectable using standard techniques

By mapping protein expression at single-cell resolution with spatial precision, DVP highlights how cells emerge, differentiate, and interact during tissue formation.. This is especially valuable for studying complex tissues (e.g., epithelial layers and neural networks), where spatial cues are critical for proper development

DVP can re-analyze archived formalin-fixed paraffin-embedded (FFPE) tissue samples despite the usual challenges of Cross-linking and degradation.. This capability enables retrospective studies on diverse clinical specimens, aiding Biomarker discovery and personalized medicine. For instance, FFPE samples from melanoma and salivary gland carcinoma have provided novel insights into disease progression and treatment resistance when examined through DVP

In neuroscience, DVP maps protein expression in brain tissue at Single-cell resolution,hedding light on cellular heterogeneity and the molecular underpinnings of neurological disorders[8]. In immunology, it reveals how immune cells infiltrate tumors or lymphoid tissue, offering a clearer view of immune responses and cell-cell interactions. By preserving spatial details alongside proteomic data, DVP enables a deeper understanding of neurodegenerative diseases, immune evasion in cancer, and Inflammatory processes.{{cn}}

Deep Visual Proteomics (DVP) faces several challenges. One major limitation is in sample preparation, particularly with formalin-fixed paraffin-embedded (FFPE) tissues, where inconsistent protein recovery can lead to variable data quality..{{Cite journal |last1=Kim |first1=Min-Sik |last2=Zhong |first2=Jun |last3=Pandey |first3=Akhilesh |date=March 2016 |title=Common errors in mass spectrometry-based analysis of post-translational modifications |journal=Proteomics |volume=16 |issue=5 |pages=700–714 |doi=10.1002/pmic.201500355 |issn=1615-9853 |pmc=5548100 |pmid=26667783}} While DVP offers high spatial resolution, it is more time-consuming than bulk proteomics, especially when analyzing individual cells, making it difficult to scale for high-throughput studies. Additionally, DVP requires specialized expertise in both imaging and proteomic analysis, which can limit its accessibility and make it challenging for widespread adoption. AI-driven image segmentation and classification may also introduce biases if the models are not well-trained on diverse tissue types, affecting the reliability of results. Although DVP provides in-depth proteome coverage, low-abundance proteins remain challenging to detect due to limitations in mass spectrometry sensitivity

Looking ahead, combining DVP with multi-omics approaches, such as single-cell transcriptomics and spatial metabolomics, could offer a more comprehensive understanding of cellular functions and interactions. As the technology matures, DVP could have a significant impact on clinical diagnostics and precision medicine, driving the discovery of novel biomarkers and therapeutic targets. However, further advancements are needed to overcome current limitations, including improving throughput and making the technology more accessible, in order to fully realize its potential in studying complex biological systems.{{cn}}

{{Reflist}}