Proteoform profiling of endogenous single cells from rat hippocampus at scale

Abstract

We perform intact proteoform profiling of 10,809 endogenous single cells from the rat hippocampus using single-cell proteoform imaging mass spectrometry (scPiMS). scPiMS directly extracts whole proteins and demonstrates high throughput for MS-based single-cell proteomics compared with existing approaches. We develop an informatics workflow dedicated to this datatype and use it to assign neurons, astrocytes or microglia cell types according to their proteoform signatures.

Single-cell proteomics (SCP) captures the heterogeneity in cell type, state and function obscured by population-averaged protein measurements^1,2. It allows for the characterization of the phenotypes and states of diverse cell types comprising complex tissues^3,4. In contrast to antibody-based molecular probes for single-cell protein detection, mass spectrometry (MS)-based SCP is label free, high content and non-targeted, which is complementary for proteome mapping and the discovery of protein markers^2,5,6.

In the absence of molecular amplification strategies for proteins, MS-based SCP has major challenges in cell processing and protein characterization⁷. Notable advances in SCP have been achieved using ‘bottom-up’ proteomics platforms that identify and quantify thousands of proteins in single cells^8,9. These include SCoPE-MS¹⁰, deep visual proteomics¹¹ and nanodroplet processing in one pot for trace samples¹². However, current SCP technologies typically have a throughput of a few hundred cells per day, limiting their ability to measure thousands to millions of single cells in analogy to single-cell RNA sequencing (scRNA-seq)¹³.

We recently introduced proteoform imaging MS (PiMS) for proteoform profiling from thin sectioned tissue^14,15, taking advantage of the high sensitivity of individual ion MS (I²MS) in conjunction with a liquid sampling probe to detect intact proteoform molecules from tissue or surface-immobilized single cells^14–16. Orbitrap-based I²MS enables proteoform detection with 10× higher resolving power and >500× greater sensitivity over traditional Fourier-transform Orbitrap MS^17–19. This combination has enabled the detection of small bursts of highly localized proteoforms up to ~70 kDa from freshly frozen tissue sections at ~20-μm spatial resolution^14,15. Application of PiMS to single cells (scPiMS) could bypass the throughput bottleneck in SCP by avoiding rate-limiting steps such as single-cell compartmentalization, proteolytic digestion and chromatographic separation²⁰. Moreover, scPiMS directly works with intact proteoforms, the exact molecular forms of proteins in cells. With their strong phenotypic linkage to tissues, proteoforms are promising for the classification of complex mixtures of endogenous cell types from tissues^21–23 but this has not been achieved to date in SCP.

Here, we developed proteoform profiling and assignment in endogenous single cells dissociated from rat brain hippocampus²⁴, reaching 10,809 total single cells at a throughput of ~1,000 cells per day in two independent tranches (dataset I, 5,272 cells; dataset II, 5,537 cells). Using a developed informatics workflow, we assigned >93% of the detected single cells to 1 of 3 main cell types, and found proteoform signatures that distinguish the astrocytic population (15%) from microglia (6%) and neurons (72%) using their proteoform profiles (7% unassigned), which were further correlated and verified with high-resolution optical microscopy²⁵.

Single cells from rat hippocampus were obtained using enzymatic and mechanical disaggregation and drop-cast onto glass slides with a target surface density of three cells per mm² (ref. 26; Fig. 1a, left). Proteoform detection in single-brain cells was first verified by targeted scPiMS experiments (Fig. 1b, Supplementary Fig. 1 and Supplementary Video 1), wherein an ~200-μm dynamic liquid bridge was brought into contact with an optically identified single cell for continuous extraction and ionization of intact proteoforms¹⁶ (Fig. 1a, top right, II). We observed a sharp increase in the ion count chronogram immediately after the liquid bridge reached the single cell (Fig. 1a, bottom right). The number of protein ions per spectrum peaked within 2 s, followed by a sharp drop, resulting in a ~7-s peak of proteoform signals above the baseline¹⁷ (Fig. 1a, bottom right). No additional signals were generated while the liquid bridge was parked on the cell for an additional 20 s. The same protein extraction behavior was verified by dozens of these targeted experiments on single-brain cells (Supplementary Fig. 2).

Fig. 1 |. The scPiMS workflow for proteoform detection and identification from single cells.

a, Left: scanning approach. Right: snapshots showing the relative location of a cell before (I), on (II) and after (III) analyte extraction by a liquid bridge. Scale bars, 50 μm. The extraction profile of the single cell (a cellogram) is also shown. b, The aggregated cellogram containing 529 single-cell features in 1 run. c, The corresponding aggregated mass spectrum for b. d, Mass spectrum of the ions collected from a single cell. e, Workflow for the main steps involved with the scPiMS process. f–h, Collector’s curves for dataset I for the number of charge-assigned ions (f), proteoforms detected algorithmically as well-formed isotopic distributions (g) and proteoforms assigned to the rat hippocampal ion library using PAScore at 10% FDR (h). PfRM, proteoform. Panels a and e created with BioRender.com.

We next scaled up the scPiMS experiment by performing parallel line scans over a 0.5-cm² region at a scan rate of 15 μm s⁻¹, in which >90% of the cells were sampled as singletons with minimal signal crossover from their neighbors¹⁶ (Supplementary Fig. 3), During ~7 h of sampling time, we obtained 529 single-cell extraction events with the resultant ‘cellogram’ (Fig. 1b, Methods and Extended Data Fig. 1). In the aggregated mass spectrum (Fig. 1c), we used an intact mass tag (IMT) approach to annotate a few abundant proteoforms by checking against a list of rat hippocampal proteoforms within a strict mass tolerance of ±3 ppm at 3σ (ref. 27; Methods): thymosin-β4, adenosine triphosphate (ATP) synthase subunit ε, high-mobility-group nucleosome-binding domain 2, myelin basic protein and a few histone H1 and H2A proteoforms^27,28 (Supplementary Fig. 4). A typical single-cell mass spectrum containing 8,833 single-proteoform ions is shown in Fig. 1d.

We next proceeded to collect dataset I with a ~2-fold higher throughput across several slides at a twofold faster scan rate of 30 μm s⁻¹ (that is, a nominal processing rate of ~50 cells per h)¹⁶. This experiment took ~120 h of unattended acquisition time, from which 5,272 single-cell events were identified using a single-cell versus multiple-cell differentiator algorithm (Fig. 1e, Supplementary Data 1 and Supplementary Fig. 5). Using video microscopy (900× magnification), 200 randomly selected cellogram peaks were spot-checked and validated as true single-cell sampling events.

Sensitive proteoform detection is critical to capture protein-level expression differences in a cell population from tissue dissociates. An average of 4,798 proteoform ions per cell were obtained for dataset I (Supplementary Fig. 6). This was enabled by an optimized sample preparation workflow in which we removed the residual buffer content in cell-free regions without disrupting morphology and localization of the cells (Methods). Relative to dataset II, this workflow achieved a threefold reduction in random noise collected from blank slide areas and better distinguished single-cell peak profiles in the total cellogram (Supplementary Figs. 7 and 8 and Methods). In dataset I, we reached an 86% success rate of converting against 6,116 optical features into 5,272 registered single-cell detection events (Supplementary Data 1). This is a sharp improvement over dataset II (5,537 single cells; Supplementary Fig. 9) obtained with previous sample preparation protocols, where only ~50% of the optically registered single-cell features were detected in the scPiMS data (Supplementary Figs. 7 and 8 and Supplementary Data 2).

Dataset I contained a total of ~29 million verified single-proteoform ions. To assign integer charge states to these ions, the dataset was processed with a reference proteoform library obtained from the hippocampal region of a brain section from the same animal¹⁴ (Extended Data Fig. 2). This step totaled 22.3 million charge-assigned proteoform ions assigned to proteoform isotopic distributions ranging from 5 to 70 kDa (Fig. 1f). The aggregated scPiMS spectrum yielded a total of 566 detected proteoform features (Supplementary Data 3) and a collector’s curve for them is shown in Fig. 1g.

To match the scale of the scPiMS data acquired at a throughput of ~1,000 cells per day, we developed an informatics pipeline to facilitate data analysis and visualization of this unique datatype. The key component of this pipeline is ‘scapp’ (Fig. 1e, middle), a C# application that allows for one-step processing and visualization of scPiMS datasets containing thousands of single-cell proteoform (scProteoform) profiles. Following mass spectrum generation, the scapp assigns single ions to a list of known proteoforms using the proteoform assignment score (PAScore) algorithm¹⁶, generating a cell × proteoform matrix with cell-specific mass spectra (for example, Fig. 1d).

The PAScore uses a statistical framework to assign scProteoforms to a given cell, using a probability-based scoring metric. Individual ions for a given cell are compared against a list of known proteoforms, matched and then scored against theoretical isotopic peaks, with the combinatorial detection of multiple ions providing evidence of a proteoform assignment to a single cell (Methods).

We used a two-phased approach to assign proteoforms for matching ions in dataset I. First, we found 165 proteoform identifications using an IMT search from a rat hippocampal database with ±3 ppm tolerance at 3σ (ref. 29; Supplementary Data 5, Extended Data Fig. 2 and Methods). Second, we used an empirical false discovery rate (FDR) model that established a threshold for PAScores for each cell (Methods), leading to an average of 86 assigned scProteoforms per cell (PAScore matrices in Supplementary Data 3 and overview of the PAScores in Extended Data Fig. 3). The rate of accrual for the 165 identified proteoforms is shown in Fig. 1h. We noted that the rate of rise for this collector’s curve for proteoform assignment increases far more sharply versus that for de novo collection (compare Fig. 1g,h). Among the 4,798 ions charge-assigned per cell in this dataset, an average of 1,289 ions were assignable to one of the 165 proteoforms. In addition, we confirmed 23 proteoform identifications by subjecting them to top-down fragmentation of tissue derived from the same animal¹⁴ (Supplementary Data 4, Supplementary Fig. 10 and Extended Data Fig. 2).

We next sought to identify major cell types from dataset I on the basis of their proteoform profiles (Extended Data Fig. 4). Using the 5,272 cell × 165 proteoform dataset described above, we performed gene set variation analysis (GSVA)³⁰ (Methods) to detect pathway activity changes over the population of cells. Using 147 REACTOME and KEGG biological pathways mapped by the cell × proteoform matrix (Supplementary Data 3), we transformed scProteoform detection into pathway-informed cell type classification (Methods). Using the pathway-adjusted PAScore of the 15 pathways that were within the top 10% highest variation across cells (Supplementary Data 3), we classified 4,927 (93%) of the 5,272 single cells into three major groups using unsupervised clustering (Fig. 2a,b and Supplementary Fig. 11). Figure 2b shows the principal component analysis (PCA) of the classified 4,927 single cells highlighted according to the assigned cell clusters (Supplementary Fig. 12). Cluster 1 was assigned to neurons as the largest cell population in the hippocampus³¹. The neuronal population demonstrates high heterogeneity in proteoform profiles, attributed to complex developmental stages and the current depth of proteome coverage³² (Fig. 2a,b). Notably, glucose metabolic enzymes represent the majority in the pathways driving the separation of cluster 3 from the remainder of the population (Extended Data Fig. 5). We assigned cluster 3 to astrocytes as the major glial cell population in the hippocampus. Cluster 2 was assigned to microglia as a major immune cell population in the central nervous system (Fig. 2a,b). Upon manual inspection using optical images captured during the scPiMS experiment, several classified cells demonstrated morphology typical of their assigned cell types (Fig. 2b). The overall cell typing results here (72% neurons, 15% astrocytes, 6% microglia and 7% others) are consistent with previous scRNA-seq data on mouse hippocampus³¹.

Fig. 2 |. Cell type and proteoform assignments obtained from the 5,272-cell dataset I.

a, A heatmap of 4,927 single cells grouped using their assigned proteoforms aligned with the most variable biological pathways. Inset: pie chart showing the percentage of cell clusters assigned as neurons (cluster 1 of 3,817 cells), microglia (cluster 2 of 317 cells), astrocytes (cluster 3 of 793 cells) and unclassified (345 cells). HIV, human immunodeficiency virus. b, PCA of the 4,927 assigned single cells in a. Each dot represents a single cell positioned by the highest two principal components extracted from the 15 pathways enriched with variable proteoforms shown in a. Cells are colored according to the assigned type in a: neuron, orange; microglia, light blue; astrocyte, green. A few example cells with optical morphology identification are shown on the sides of the PCA plot. Scale bars, 25 μm. c–g, Mass spectra of assigned proteoforms (monoacetylated ARP5L (c), ALDOA (d), monoacetylated GFAP (e), PP5-TPR (f), ARF1 (g)) created from example single cells and aggregated single cells. In each panel, a mass region corresponding to the proteoform labeled at the top is displayed. Top: spectrum showing the single-cell mass spectrum (dark) with the single ions assigned to the scProteoform annotated in wine-red dots with drop lines. The mirrored spectrum shows the theoretical isotopic distribution for the proteoform as reference. Bottom: cell-type-specific mass spectrum (neuron, orange; astrocyte, green; microglia, blue) constructed from 50 single cells with the highest PAScores for that proteoform. The theoretical isotopic distribution is overlaid as a background. These selected proteoforms drive the separation of the cell type shown in the corresponding panels. The dashed lines in c–g indicate the level of single ion in the corresponding spectra.

Despite proteoform heterogeneity observed in the neuronal population, we found some proteoforms with consistent enrichment across the entire neuronal population (for example, monoacetylated actin-related protein 2–3 complex subunit 5-like protein (ARP5L); Fig. 2c and Extended Data Fig. 6). ARP5L detection at the single-neuron level was accomplished with 4–5 ions per cell (Fig. 2c); assignments were supported by the constructive proteoform detection when building mass spectra using a progressively higher number of single-neuronal proteoform profiles (data for 50 neurons in Fig. 2c and data for 300 neurons in Extended Data Fig. 7). ARP5L is part of the ARP 2–3 complex and has an important regulatory role in the formation of the actin cytoskeleton that facilitates neuronal migration anddendrite branching. In the future, deeper scProteoform data may distinguish neuronal subpopulations^33,34.

Astrocytes are known to express heightened glucose metabolism to provide nutrients to their surrounding neurons³⁵. Aside from aldolase A (Fig. 2d and Extended Data Fig. 8), a few other identified glycolytic enzyme proteoforms also demonstrated consistently higher detection frequencies and average PAScores for cells in cluster 3 (Extended Data Fig. 9). This cluster also reported elevated detection of glial fibrillary acidic protein proteoforms (Fig. 2e and Extended Data Fig. 9), a classic immunofluorescence marker for astrocytes in the brain. The microglia-assigned cell population showed elevated detection of tetratricopeptide repeat domain of protein phosphatase 5 (Fig. 2f), adenosine diphosphate ribosylation factor 1 (Fig. 2g) together with its monoacetylated form, excitatory amino acid transporter 2 and 26S proteasome non-ATPase regulatory subunit 11 (Extended Data Fig. 10). From the pathways enriched with proteoforms in astrocytes, primary metabolism was elevated within astrocytes, consistent with scRNA-seq³¹ and the phenotypic assessment of this cell type³⁶.

In conclusion, we report a workflow for SCP of endogenous cells isolated from primary mammalian tissue, reaching the scale of 10⁴ cells. Using proteoforms in pathways with the highest variation across a cell × proteoform matrix, we demonstrate the initial foray into capturing the heterogeneity of brain hippocampus cell types without prior cell labeling or sorting, opening the door to the discovery of cell-type-specific proteoform signatures. While conceptually similar to direct MS studies of single cells³⁷, this work translates the potential of scProteoform technology into a scalable approach to MS-based SCP for label-free profiling with proteoform specificity up to ~70 kDa. scPiMS also converts the ‘no digestion’ philosophy of proteomics into a key advantage that opens the throughput bottleneck of SCP¹. Higher throughput will enable proteoform signatures from rare cells to be captured. As shown here, the scPiMS platform uses a reference set of proteoforms (Human Proteoform Atlas³⁸) to both accelerate the proteoform assignment step and the depth of the analysis obtained in each cell. By providing proteoform-level information, scPiMS will advance single-cell biology by mapping cell types underlying disease phenotypes in future work.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41587–025-02669-x.

Methods

Tissue preparation

Three 2-month-old male Sprague Dawley outbred rats (Rattus norvegicus; Inotiv) were used in experiments. Animals were fed ad libitum and housed in a 12-h light cycle. Animals were asphyxiated using CO₂. All procedures were performed in accordance with animal use protocol approved by the University of Illinois Institutional Animal Care and Use Committee and in compliance with both federal and ARRIVE guidelines for the humane treatment of animals.

Immediately after killing, animals were perfused transcardially with ice-cold modified Gey’s balanced salt solution (mGBSS) containing (in mM): 1.5 CaCl₂, 5 KCl, 0.2 KH₂PO₄, 11 MgCl₂, 0.3 MgSO₄, 138 NaCl, 28 NaHCO₃, 0.8 Na₂HPO₄ and 25 HEPES (pH 7.2). One third of the hippocampus from one hemisphere was surgically dissected for single-cell preparation. Intact hemispheres of the same brains as used for cell population collection were frozen and sectioned for intact tissue PiMS imaging. Rat hippocampus tissue punches and thin sections were prepared according to published protocols³⁹. Rat brain hemisphere slices 16 μm thick were cut at −16 °C using a cryostat (Leica CM3050 S, Leica Biosystems). Brain slices containing hippocampal areas of all three animals were deposited on the same slide. Sections were rinsed twice with 200-proof ethanol (Lab Alley). Tissue samples were stored at −80 °C before imaging analysis by scPiMS using the published approach¹⁴.

Preparation of individual cell populations

We used the procedures for single-cell isolation developed for single-cell metabolomics, adapted here for scPiMS^24,40. One third of the hippocampus from one hemisphere was surgically dissected and treated with the papain dissociation system (Worthington Biochemical) for 120 min at 34 °C with superficial oxygenation. After the treatment, tissues were gently rinsed twice with ice-cold mGBSS. Mechanical tissue dissociation was performed in ice-cold mGBSS supplemented with Hoechst 33342 (1 μg ml⁻¹). The individual cell suspension was deposited onto indium tin oxide (ITO) glass slides (Delta Technologies) to achieve a low-density cell population with individual cells spaced by an average of >200 μm. Cell populations from three different animals were deposited onto separate marked areas on ITO glass slides. Cells were sedimented and adhered to slide surface for 10–20 min. The surrounding cell medium was quickly replaced with mGBSS containing 33% glycerol. After a brief incubation period for ~10–20 s, most of the solution present on ITO glass slides was removed. The cells for dataset II were prepared as described above. For the cells for dataset I, slides were additionally rinsed twice with 200-proof ethanol (Lab Alley).

Fluorescence microscopy experiments

Images of populations of single cells were taken in mixed bright-field and fluorescence mode using a Zeiss Axio M2 microscope (Zeiss) equipped with an AxioCam ICc5 camera, X-cite Series 120 Q mercury lamp (Lumen Dynamics) and an HAL 100 halogen illuminator (Zeiss). The DAPI (excitation, 335–383 nm; emission, 420–470 nm) dichroic filter was used for fluorescence excitation. A ×2.5 objective was used for fast acquisition with a 13% overlap between individual images allowing proper image stitching and formation of a single view for the entire glass slide. Images were processed and exported as large .tiff files using ZEN software version 2, blue edition (Zeiss). Numbers of cells on the slide in regions of interest were determined using QuPath⁴¹ and ImageJ⁴².

scPiMS probe fabrication and ion source conditions

A custom-designed nanospray desorption electrospray ionization (nano-DESI) source was used for all data acquisition. The experimental details of nano-DESI MS imaging were described elsewhere^43,44. Briefly, the nano-DESI probe comprises a flame-pulled fused silica primary (outer diameter (OD), 40 μm; inner diameter (ID), 20 μm; Molex) and a nanospray capillary (OD, 150 μm; ID, 40 μm) with the spray side of the nanospray capillary positioned toward the MS inlet. The nano-DESI probe uses a dynamic liquid bridge formed between the capillary junction and the glass surface to extract analytes when brought into contact with the glass surface. The liquid bridge is dynamically maintained by solvent propulsion from the primary capillary and instantaneous vacuum aspiration through the nanospray capillary. All samples were electrosprayed in positive ion mode under denaturing conditions in a 60–39.4% acetonitrile–water and 0.6% acetic acid solution compatible with both protein extraction and ionization. The solvent flow rate was kept in the range of 400–500 nl min⁻¹. The ion source conditions on the MS instrument were set as follows: ESI voltage, +2.8 kV; in-source collision-induced dissociation, 15 eV; S-lens radiofrequency level, 70%; capillary temperature, 360 °C.

scPiMS operation modes

scPiMS experiments were performed in two distinct modes: targeted mode and high-throughput scanning mode. In targeted mode, a single-cell feature was first identified by the ×900 high-magnification microscope (Supplementary Figs. 1 and 2). In particular, a nano-DESI probe was brought into contact with the glass surface at a location ~50 μm away from the single cell to form the liquid bridge and was subsequently moved toward the cell with manually controlled stage motions after MS data acquisition was triggered. The probe was typically parked on the cell for one minute to obtain a targeted cellogram (chronogram of total ion counts). The cellogram typically starts from a drastic rise in protein signal followed by an exponential decay over time, which is characteristic of the extraction kinetics of cellular protein analyzed by scPiMS.

High-throughput scanning mode was performed by moving the glass slide under the nano-DESI probe in parallel lines at a constant linear velocity. The spacing between adjacent lines was set similar to the size of the liquid bridge formed by the probe and the surface to ensure that the total surface area containing single cells was covered. The size of the liquid bridge in nano-DESI can be adjusted in the range from 50 to 400 μm depending on the size of the capillaries used to fabricate the probe. For scPiMS, 200 μm was selected considering the throughput of the experiment to cover the entire area containing single cells. Larger probes will result in multiple adjacent cells on the surface being analyzed at the same time, which obscures the protein profiles from these individual single cells. Statistical modeling was conducted to simulate the percentage of single-cell extraction events using the single-cell coordinates obtained from fluorescence microscopy. In particular, the surface area containing single cells was divided into 200-μm square grids and the numbers of grids that contained one cell or more than one cell were calculated. At the density of cells loaded on the glass slides, ~90% of the cells can be sampled as single-cell features using the 200-μm probe. In specific cases when cells are more densely populated on the slide, a 70% success rate of single-cell sampling is guaranteed without sacrificing the throughput of the experiments (Supplementary Data 1). Both targeted mode and high-throughput scanning mode use an XYZ motorized stage system (Zaber Technologies) controlled by a custom LabVIEW program (adapted from https://github.com/LabLaskin/nano-DESI-stage-dynamic-sampling).

Cellogram optimization

A cellogram obtained in targeted mode reveals the extraction kinetics of the scPiMS probe on single cells and provides guidance to optimizing the parameters in high-throughput scanning mode. Targeted mode cellograms were obtained by holding the probe above a single cell and monitoring the total ion response. The time at which the targeted cellogram fell back to the baseline indicated that all readily extractable proteins were captured (Supplementary Fig. 3). This gave the approximate time that the cell should be exposed to the probe and determined the overall throughput for scPiMS. In high-throughput scanning experiments, extraction of protein content from a single cell before overlapping with any neighboring cells guarantees that the signal carryover between adjacent cells is minimized. The upper limit was used to calculate the ideal exposure time of a single cell to the liquid bridge in the rastering mode (Supplementary Fig. 3). The average exposure time for single cells at a 500 nl min⁻¹ solvent flow rate was found to be 5–7 s from a few targeted cellograms. Therefore, the rastering scan rate was calculated, using the size of the liquid bridge, size of the single cell and exposure time considering the small size of the rat brain single cells (~10 μm) compared to the liquid bridge (200 μm) in the current scPiMS configuration (Eq. (1)):

$$\text{Rastering scan rate}=\frac{\text{Probe size}+\text{single cell size}}{\text{Exposure time}}\cong \frac{\text{Probe size}}{\text{Exposure time}}$$ (1)

A scan rate of 30 μm s⁻¹ (200 μm divided by 7 s) was used for all high-throughput scPiMS runs to minimize sampling carryover from adjacent cells and ensured near-complete protein extraction from the cell.

scPiMS data acquisition

scPiMS data acquisition was performed in I²MS mode on a previously described Orbitrap Q-Exactive Plus MS instrument (Thermo Fisher Scientific)¹⁷. The spectral acquisition rate was set at one scan per second. During scPiMS data acquisition, proteoforms from single cells were sampled and ionized by a nano-DESI probe to generate multiply charged ions distributed across multiple charge states. The ion injection time was optimized such that ions in one detection period were collected in the individual ion regime, which gives a singular ion signal at a defined m/z (or frequency) value. We used a 700-ms injection time and a higher-energy collisional dissociation (HCD) pressure level of 0.5 (UHV pressure < 5 × 10⁻¹¹ Torr) across all scPiMS experiments for the rat hippocampal cells. These settings were found to efficiently capture the most proteoform signals emerging from single cells, while the majority of ion detection periods were dominated by individual ion signals even at extended MS injection times over 800 ms (ref. 16).

In addition, the Orbitrap central electrode voltage was adjusted to −1 kV to improve the ion survival rate in I²MS. Additional relevant data acquisition parameters were as follows: mass range, 500–1,500 m/z; automatic gain control mode, disabled; enhanced Fourier transform, off; averaging, 0; microscans, 1. Time-domain data files were acquired at detected ion frequencies and recorded as STORI files⁴⁵ with the STORI setting enabled.

scPiMS feature selection and data analysis

scPiMS data analysis was performed using MATLAB and C# scripts developed in house. scPiMS raw data were collected as continuous chronograms containing discrete or overlapping cellograms and other peak features from chemical noise on the glass slide collected during the experiments. Cellogram picking was first performed to recognize MS scans corresponding to true single-cell events and eliminate the blank and noise scans from the dataset.

The processed single-ion chronogram was used for peak picking with a peak intensity threshold of 500 ions and a minimum peak distance of eight scans (exposure time × MS scan rate). The peak features picked by the program were further categorized into single-cell-like and multiple-cell-like features using peak characteristics. Specifically, if the half width at half maximum on the right side of a cellogram was more than two scans (3 s), the feature was recognized as a multicell feature. Multicell features were effectively filtered in this step and only the single-cell features were used for further analysis. The number of single-cell features was also validated and matched to the fluorescence single-cell feature count as described above. In all samples prepared using the improved protocol, the number of single-cell-like features from the cellograms matched optically registered single-cell features (Extended Data Fig. 1). The MATLAB code used to perform the analysis is on GitHub (https://github.com/NRTDP/scPiMS). Executable files of the C# code (SingleCellApp.exe) are available on Zenodo (https://doi.org/10.5281/zenodo.14611173)⁴⁶.

The STORI files of individual MS scans from the single-cell-like features were extracted from the dataset and processed into single ions using storiboarD. Scans corresponding to specific cells were grouped together, yielding cell-specific individual ions. Individual ion signals were subjected to charge assignment, allowing for mass calculation. The neutral masses of the protein ions were calculated as follows:

$$\text{Mass}=\left(\frac{m}{z}\times z\right)-\left(z\times M_{\text{proton}}\right)$$ (2)

The charge state (z) is obtained from the slope of induced image current determined by the STORI analysis described in detail elsewhere⁴⁵. Accurate charge assignment of each ion was statistically evaluated by comparing the slopes of its isotopologs across different charge states from the entire dataset. We used a probability metric to filter out ions with a lower probability score and construct a mass-domain isotopic distribution with statistical confidence.

PAScore calculation of scProteoforms and FDR

Individual ions corresponding to each single-cell feature were aggregated into a single-cell spectrum and were used to generate a score for each proteoform in a given cell feature. The proteoform assignment space may be determined through multiple pathways: either isotopic envelopes created using the THRASH deconvolution algorithm⁴⁷ or envelopes created and validated through identified proteoforms from IMT search or MS/MS through the PiMS platform described in detail below. Cell-specific individual ions were searched against the isotopic envelope library and matched to specific isotopic peaks with a tolerance of ±0.3 Da; these ion matches are scored and combined with all ion scores for a given proteoform to yield a single-cell PAScore using the following equation:

$$\text{PAScore}=1-\prod _{k=1}^n\left[1-\left(P_{\text{isotopolog}}\times P_{\text{mass error}}\right)\right]$$ (3)

where $P_{\text{isotopolog}}$ is the expected relative intensity of the matching isotopolog, and $P_{\text{mass error}}$ is a probability score for the mass error between the observed individual ion mass and the theoretical mass of the isotopolog, using the cumulative distribution function (CDF) of a normal distribution with the mean and $\sigma$ determined by the theoretical mass and width of an isotopic peak, respectively.

$$P_{\text{mass error}}=\left\{\begin{array}{c}\mathrm{CDF}\left(m_{\text{ion}}\right)m_{\text{ion}}<m_{\text{iso}}\\ 1-\mathrm{CDF}\left(m_{\text{ion}}\right)m_{\text{ion}}>m_{\text{iso}}\end{array}\right.$$ (4)

where $m_{\text{ion}}$ is the mass of the individual ion and $m_{\text{iso}}$ is the theoretical mass of the matched isotopolog. To control for multiple-hypothesis testing, an empirical FDR procedure was implemented at the single-cell level. Decoy proteoforms (ten per proteoform in the library) were generated using random amino acids and constrained to be the same sequence length as the representative proteoform. These decoys were scored alongside the proteoform hits and rank-ordered and all proteoform hits were given a q value. Proteoforms with a q value less than 0.1 (10% FDR) were kept as assigned in that single cell.

IMT search

The summed mass-domain ion library was converted to .mzML format and processed using a custom version of TD Validator (Proteinaceous) implemented with an MS1 IMT search function. The library spectrum was self-calibrated by +10.25 ppm according to the accurate masses of six identified proteoforms in the mass range of 10–50 kDa. A custom rat proteoform database was constructed from 1,143 proteins in hippocampal-enriched biological pathways extracted from a bottom-up proteomic characterization of mouse brain hippocampus²⁹ (Supplementary Data 4). In particular, the mouse protein Entrez identifiers (IDs) were first converted to rat protein Entrez IDs by gene homology search using g:Profiler (http://biit.cs.ut.ee/gprofiler/). The rat homologous gene list was manually inspected and verified. All SwissProt and TrEMBL proteins from the identified rat genes were considered in the IMT database construction. IMT search was performed in the mass range of 4–70 kDa with a ±3 ppm mass tolerance considering methionine on or off and monoacetylation as possible proteoform modifications in the database (that is, a search space of ~20,000 proteoforms).

Pathway-adjusted PAScore matrix and unsupervised clustering analysis

The PAScore matrix of 165 identified proteoforms after FDR filtering in dataset I was subjected to GSVA using R package GSVA (version 1.50.5)³⁰. GSVA is a non-parametric, unsupervised method for estimating the variation of gene set enrichment through the samples of an expression dataset³⁰. In our study, we considered each single cell as a distinct sample. The analysis was performed using the KEGG and REACTOME biological pathways included in the gene set dataset ‘c2BroadSets’, which contains the c2 collection of canonical pathways from the molecular signature database (MSigDB, https://www.gsea-msigdb.org/gsea/msigdb) of the Broad Institute. Proteoform IDs were mapped to gene Entrez IDs, subjected to rat-to-human gene homology using g:Profiler (http://biit.cs.ut.ee/gprofiler/) and manually validated (Supplementary Data 4). Whenever gene symbols were assigned to multiple proteoforms, we merged the PAScores using the following equation:

$$\text{PAScore}=1-\prod _{k=1}^n[1-(\text{PAScore})]$$

In each cell, the PAScore of each scProteoform is transformed into a pathway-adjusted PAScore by converting PAScore values into cumulative density functions across cells, rank-transforming them and then comparing the rank-transformed values of genes inside each detected pathway to those outside³⁰. The matrix of pathway-adjusted PAScore (147 canonical pathways) was filtered out to keep the top 10% most variable pathways (15 pathways; Supplementary Data 3) and subjected to unsupervised clustering analysis applying the ‘partitioning around medoids’ algorithm⁴⁸ using the ComplexHeatmap R package. A silhouette plot analysis was performed in R and three major clusters were identified to represent different cell types. PCA was performed using the adjusted PAScores from these 15 pathways (factoextra R package). The R code used to perform the analysis is on GitHub (https://github.com/NRTDP/scPiMS). Adobe Illustrator was used to combine and organize figures.

Data analysis for proteoform identification by MS/MS along with identification

Targeted MS/MS experiments were performed in the rat hippocampal region of a brain tissue section from one of the animals using the AutoPiMS workflow¹⁵. Briefly, a survey PiMS line scan along the hippocampal region was obtained and processed using I²MS that gives the m/z, charge state and favorable location for an MS/MS experiment for a list of proteoform targets. MS/MS experiments were performed on an adjacent line region offset by 50 μm according to a multiplexed acquisition method generated from the last step. A 0.8 m/z isolation window was used for most of the targets. The raster scan rate for survey and MS/MS experiments was 3 μm s⁻¹. MS/MS data acquisition was conducted using HCD with fragment ion detection in either traditional ensemble or I²MS mode for <20-kDa and >20-kDa targets, respectively. Ensemble MS/MS experiments were performed on an Orbitrap Exploris 480 MS instrument (Thermo Fisher Scientific) using a mass resolution of 30,000 at an acquisition rate of 14.5 spectra per s and a HCD collision energy of 25–45 eV. MS/MS experiments in I²MS mode were performed on a Q-Exactive Plus as described above using an Orbitrap detection period of 1 s (HCD pressure setting = 0.5)⁴⁹. Typical values for collision energy and injection time used in this study were 12 eV and 700 ms, respectively. The total data acquisition time for each target ranged from 1 to 5 min. For the database search of targets collected in I²MS mode, mass-domain spectra were generated. Top-down MS/MS searches were performed using ProSight Native and TDValidator (Proteinaceous) against a set of proteoforms created from the entire rat SwissProt database. Proteoform identifications were reported using E values and the number of matched fragments for ensemble and I²MS data types, respectively.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Code availability

Compiled scripts used to create and process I²MS files are available on GitHub (https://github.com/NRTDP/scPiMS). Additional software and data that support and corroborate the findings of this study are available on GitHub (https://github.com/NRTDP/scPiMS).

Extended Data

Extended Data Fig. 1 |. Single cell feature selection demonstrated using the data shown in Fig. 1b-d.

From 752 optically-registered cells, 529 single cells were detected based on filtering of distance and scPiMS sampling conditions (annotated in panel (a)). Location of the 529 single cells are registered in the pixel-based graph in panel (a). (b) Cellogram single cell feature selection purely based on peak features (more in Methods). Using a peak threshold of 750 ions and a spacing of 7 scans, 529 single cell-like peak profiles were registered, which was well correlated with the optical registration and statistical modeling.

Extended Data Fig. 2 |. Mass spectrum of (a) the proteoform library obtained from hippocampal-specific PiMS experiment and (b) aggregated 5272-cell dataset I with proteoform annotations.

Mass spectrum of (a) the proteoform library obtained from hippocampal-specific PiMS experiment (a 2.5 mm by 2.5 mm hippocampal region of a 10-micron thin brain section with a total of 38 million ions and 15930 MS scans) and (b) aggregated 5272-cell dataset I with annotations of the 165 identified proteoforms in Supplementary Data 5.5 and the 23 MS/MS-identified proteoforms in Supplementary Data 4. The inset in (a) shows the detailed annotation of MBP (P02688–4) proteoforms at ~14.1 kDa. The proteoforms labeled in gray are abundant proteins found in hippocampal tissues but not part of the hippocampal-enriched biological pathways, therefore they were not included in the IMT database and not used for subsequent cell typing effort. However, these proteoforms were identified using our MS/MS approach and were confirmed to be ubiquitously-expressed brain proteins (for example, ubiquitin, ATP synthase subunits-ATP5E, myelin basic proteins-MBPs, brain acid soluble protein-BASP1) that only have a minor contribution to cell type classification of neurons and glial cells.

Extended Data Fig. 3 |. PAScore summary of the 5272-cell dataset I with 10% FDR and 10 decoy sequences for each proteoform matching event.

(a) pie chart of a total of 869,880 candidate matching events with 45.5% missing values, 49.5% FDR-filtered PAScores, and 5% non-zero PAScores assigned to 165 proteoforms identified from the rat hippocampus. (b) PAScore distribution of FDR-filtered matching events with an average of 1.86 ions matched to a proteoform (blue portion in the pie chart in panel a). (c) Non-zero PAScores assigned to the 165 identified proteoforms plotted according to number of ions (2–8) matched to the proteoforms (red portion comprising 5% in the pie chart in panel a). (d) Distribution of PAScores for the 5% of proteoforms assigned in single cells to one of the 165 identified proteoforms (that is, a total of 43.5k scProteoforms).

Extended Data Fig. 4 |. Principal component analysis of 4927 cells colored according to the cell type assigned.

The pink arrows are the ‘biplot’ vectors representing the coefficients of the biological pathway variables on the principal components shown in Fig. 2b. The Neu, Ast and Mgl labels show the locations of the single cells with optical images in Fig. 2b.

Extended Data Fig. 5 |. Total and proteoform-assigned ion count per cell distribution across cell types in the 5272-cell dataset I.

Total (a) and proteoform-assigned (b) ion count per cell distribution across cell types in the 5272-cell dataset I. Ast, astrocytes; Mgl, microglia; Neu, neurons.

Extended Data Fig. 6 |. Bar and scatter plots of the monoacetylated ARP5L proteoform across cell types that showed consistent elevated PAScores in the neuronal population.

The average unadjusted PAScores are labeled using solid diamond symbols. The upper side of the whiskers indicate PAScore at 99% percentile.

Extended Data Fig. 7 |. Mass spectra of the monoacetylated ARP5L proteoform constructed from 50, 100, 200, and 300 single neurons and individual ion mass spectra from three single neurons showing assignment of the monoacetylated ARP5L proteoform.

(a) Mass spectra of the monoacetylated ARP5L proteoform constructed from 50, 100, 200, and 300 single neurons. The theoretical isotopic distribution is shown in the mirrored panel. (b) Individual ion mass spectra from three single neurons showing assignment of the monoacetylated ARP5L proteoform in its mass range with the PAScores in cells numbered #784, #1247 and #605.

Extended Data Fig. 8 |. Different representation of the mass spectrum of cell #179 shown in Fig. 2d.

Different representation of the mass spectrum of cell #179 shown in Fig. 2d (middle and bottom). The top panel shows a single-ion representation that shows the breakdown of the four-ion composite peak, of which two ions were matched to an ALDOA proteoform and two were not.

Extended Data Fig. 9 |. Bar and scatter plots of detected proteoforms involved in glucose metabolism (REACTOME Stable Identifier: R-HSA-70326) and glial fibril acidic protein (GFAP, a classical immunofluorescent astrocytic marker) across cell types.

Bar and scatter plots of detected proteoforms involved in glucose metabolism (REACTOME Stable Identifier: R-HSA-70326) and glial fibril acidic protein (GFAP, a classical immunofluorescent astrocytic marker) across cell types that showed consistent elevated PAScores in the astrocytic population. Each panel is labeled with the proteoform name and UniProt accession together with a proteoform descriptor. The average unadjusted PAScores are labeled using solid diamond symbols. The upper side of the whiskers indicate PAScore at 99% percentile.

Extended Data Fig. 10 |. Bar and scatter plots of detected proteoforms involved in HIV Infection (REACTOME Stable Identifier: R-HSA-162906) and Host Interactions of HIV Factors (REACTOME Stable Identifier: R-HSA-162909) across cell types.

Bar and scatter plots of detected proteoforms involved in HIV Infection (REACTOME Stable Identifier: R-HSA-162906) and Host Interactions of HIV Factors (REACTOME Stable Identifier: R-HSA-162909) across cell types that showed consistent elevated PAScores in the microglia population. Each panel is labeled with the proteoform name and UniProt accession together with a proteoform descriptor. The average unadjusted PAScores are labeled using solid diamond symbols. The upper side of the whiskers indicate PAScore at 99% percentile.

Data availability

A list of 165 identified proteoforms, .RAW files and processed single-cell data are available on MassIVE repositories under accession codes MSV000096472 and MSV000092360. The biological pathway database for GSVA was retrieved from MSigDB of the Broad Institute (https://www.gsea-msigdb.org/gsea/msigdb). Source data are provided with this paper.