Comprehensive molecular profiling of single-cell proteome via gel electrophoresis and 3D single-molecule imaging

22 Recent advances in shotgun proteomics and immunoassays have yielded powerful single-cell 23 proteomics technologies. However, current methods lack the sensitivity required to comprehensively 24 quantify protein abundances in individual cells. Here, we present single-cell PAGE-PISA, an ultra-25 sensitive proteome profiling strategy that combines gel electrophoresis with 3D single-molecule 26 fluorescence imaging. Our approach labels all proteins in single cells with fluorescent dyes, separates 27 them by electrophoresis, and counts with single-molecule resolution. This technique quantified over 28 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 2 107 protein copies from a single mammalian cell with the sensitivity to detect low-abundance proteins 29 down to 104 copies per species. Single-cell PAGE-PISA successfully classified cells into distinct cell 30 types based on their proteomic profiles. Furthermore, our single-cell proteome data strongly 31 correlated with predicted developmental states during cardiomyocyte differentiation, providing 32 complementary information to single-cell transcriptome data. Together, single-cell PAGE-PISA 33 enables highly sensitive and quantitative proteome profiling at the single-cell level, capturing subtle 34 proteomic differences that distinguish diverse cellular states. 35

36 Heterogeneity occurs at multiple levels of molecular biology and is inherent in many biological 37 processes. For many years, the central dogma of molecular biology has been explored primarily 38 through conventional bulk analyses. However, such approaches provide an average measurement 39 (e.g., protein abundance, gene expression) across a population of cells. While bulk analyses have been 40 useful for distinguishing between diseased and healthy tissues1, they often reflect dominant biological 41 traits. To address this issue, single-cell technologies have been developed to capture the unique 42 molecular profiles of individual cells and provide insights into cellular heterogeneity. One widely 43 adopted approach is single-cell transcriptome profiling, which captures RNA expression levels across 44 a broad range of transcripts at the single-cell resolution, enabling clustering of diverse cell types and 45 states within complex populations2,3,4. Despite its potential, RNA expression levels cannot accurately 46 predict protein abundance5,6,7,8,9. This is because protein abundance is highly dynamic and greatly 47 influenced by various processes, such as protein degradation and post-transcriptional and translational 48 modifications10,11. Therefore, it is imperative to perform direct proteome profiling at the single-cell 49 level to comprehensively capture the protein abundance and modifications that are not reflected by 50 RNA expression. 51 Currently, there are two major approaches to quantifying protein expression levels in single cells, 52 which are mass spectrometry (MS) and antibody-based analysis12. So far, MS-based analysis serves as 53 the gold standard for proteomic studies, attributable to its ability to identify, characterize, and quantify 54 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 3 proteins with high multiplexity13. However, MS does not provide a comprehensive proteomic analysis 55 for single cells due to its limitations in sensitivity. This makes accurate quantification of proteins 56 using MS even more difficult for single cells because of its low protein abundance, between 50 and 57 300 pg in single mammalian cell14, and the fact that proteins cannot be amplified like nucleic acids. 58 This limitation was significantly addressed by the development of SCoPE-MS15, which enables the 59 quantification of over 1,000 protein groups per single cell. Since then, many MS-based single-cell 60 proteomics technologies have been developed to improve the detection sensitivity16,17,18,19,20,21,22,23,24. 61 Yet, these methods predominantly quantified highly abundant proteins (>104 copies per cell), while 62 low abundant proteins (101‒103 copies per cell) remain challenging to quantify6. 63 Apart from MS-based analysis, other single-cell protein analysis techniques have also been developed 64 based on antibody labelling25,26,27. In 2014, Herr et al. developed single-cell Western, which involves 65 isolating single cells into microwells, lysing them in situ, separating proteins by electrophoresis, and 66 immobilizing them for protein abundance analysis of target proteins using antibodies. This approach 67 offers high-throughput analysis by enabling simultaneous assay of 1,000‒2,000 single cells in less 68 than 4 hours. This represents an improvement in both sample throughput and measurement time 69 compared to most MS-based single-cell proteomics approaches, albeit sensitivity remains relatively 70 the same. Recent modifications of single-cell Western using nitrocellulose blotting and enzyme-71 antibody conjugates have lowered the detection limit to approximately 103 molecules per protein 72 species28. Despite this progress, its sensitivity is still insufficient for analyzing low-abundance 73 proteins, which remains a key impediment of single-cell Western. 74 To overcome these limitations, single-molecule fluorescence microscopy offers precise detection and 75 quantification of target molecules. Recently, our group developed a custom-built light-sheet 76 microscope called planar illumination microscope for single-molecule imaging for all purpose 77 (PISA)29, which enables 3D single-molecule imaging of all target molecules within a sub-millimeter 78 sample depth. By placing the sample plane above the optical systems for light-sheet imaging and the 79 two objective lenses for illumination and detection below the coverslip at a tilted angle, PISA 80 facilitates counting the number of molecules in the entire volume of a mm-sized biological specimen. 81 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 4 In particular, we have shown that PISA can image sub-mm-thick gel at the single-molecule level, 82 achieving attomolar sensitivity. This remarkable sensitivity opens up the possibility for the analysis of 83 single-cell lysates using polyacrylamide gel electrophoresis (PAGE), as even trace amounts of protein 84 from individual cells can be detected. 85 Here, we present single-cell PAGE-PISA, a highly sensitive strategy for single-cell proteome profiling 86 that integrates PAGE with 3D single-molecule fluorescence imaging using PISA. In this strategy, 87 individual cells are isolated, and proteins are fluorescently labelled, separated by PAGE, and 88 quantified at the single-molecule level. For labelling, we employed N-hydroxysuccinimide (NHS)-89 ester dyes that react with the primary amines of proteins. As the labelling does not rely on antibody, it 90 enables unbiased profiling analysis of the overall cellular proteome without predefined targets, 91 although it does not provide direct identification of the proteins represented in each band. This 92 generates highly sensitive and quantitative electrophoretic band patterns that reflect global protein 93 abundances in individual cells, enabling downstream analyses such as clustering, trajectory inference, 94 and identification of functional cell states solely based on proteome-level variation. To demonstrate 95 the potential of our development, we applied single-cell PAGE-PISA on standard cell lines, PC-3 and 96 U2OS, and explored the temporal changes of cellular proteomes during cardiomyocyte differentiation 97 from human induced pluripotent stem cells (hiPSCs) by tracking their progression through different 98 developmental stages. The resulting proteome profile revealed gradual proteomic changes across 99 single cells that aligned with the progression from pluripotency to differentiated states. Unsupervised 100 clustering and pseudotime analysis further uncovered intermediate subpopulation along the 101 differentiation trajectory, which are often difficult to resolve using transcriptomic analysis or marker-102 based approaches. 103

104 Establishment of the single-cell PAGE-PISA workflow 105 To provide a streamlined proteomic strategy for single cells, we established a complete workflow of 106 single-cell PAGE-PISA: (1) cell preparation, (2) manual isolation of single target cells, (3) one-pot 107 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 5 sample preparation in PCR tubes, including cell lysis and protein labelling with dye, (4) SDS-PAGE 108 for protein separation, and (5) volumetric single-molecule imaging of dye-labelled proteins in 109 polyacrylamide gel using PISA (Fig. 1). 110 In the workflow, cells are first dissociated from the culture dish, washed several times, and suspended 111 in phosphate-buffered saline (PBS) solution (Fig. 1a). Then, a droplet of cell suspension is deposited 112 onto a slide glass, from which the single cells are manually isolated using an inverted microscope 113 equipped with a TOPick 1-cell handling system (i.e., low-binding coated micro-glass needle with a tip 114 diameter of 30 µm and aspirating volume of 100 pL) (Fig. 1b, Supplementary Fig. 1a). The isolated 115 single cells are dispensed into PCR caps containing a droplet of PBS and spun down to ensure the 116 droplet settles at the bottom of the tube. With manual single-cell isolation, we can precisely control 117 cell occupancy by eliminating the possibility of droplets containing multiple or no cells. The entire 118 process of identifying, picking up, and releasing each target cell takes less than two minutes. 119 Furthermore, we perform single-cell sample preparation on ice without direct contact between pipette 120 tips and protein samples (Fig. 1c). Instead, all reagents, including lysis buffer, protease inhibitor, Cy5-121 NHS ester dye, and quencher solution, are dispensed along the tube wall without touching the samples 122 directly and mixed by spinning down. This ‘all-in-one’ sample preparation method enhances protein 123 recovery for quantitative single-cell proteome profiling by considerably minimizing potential loss due 124 to surface adsorption and sample transfer. 125 Next, to separate proteins by molecular weight, we use a commercial 20-well polyacrylamide gel 126 designed for low sample volume with a capacity of up to 8 µL per loading well (Fig. 1d, 127 Supplementary Fig. 1b). The use of a standard polyacrylamide gel with an 8 cm length, instead of 128 preparing a miniaturized gel like the single-cell Western system26, improves the separation resolution 129 of the band profiles by resolving proteins with close molecular weights. Meanwhile, the narrow 130 loading well of approximately 4 mm width further concentrates the protein samples within each band 131 along the migration path. As protein bands from single cells cannot be observed by the standard gel 132 imager, we optimized the electrophoresis conditions for single cells (e.g., time, voltage, and current) 133 using the bulk cell lysates in a preliminary experiment. The resulting band profiles from bulk cell 134 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 6 lysates, visualized by the standard gel imager allowed us to determine the exact migration pattern and 135 position of each protein band on the polyacrylamide gel (Supplementary Fig. 2a). After single-cell 136 electrophoresis, the gel is carefully cut, transferred onto a film substrate, and placed on a microscopic 137 stage (Supplementary Fig. 1c, d). 138 Finally, we perform 3D single-molecule imaging using PISA to detect and quantify the dye-labelled 139 proteins, which can be observed as clear diffraction-limited spots during imaging (Fig. 1e). We focus 140 on the middle molecular weight region, approximately from 20‒40 to 100‒150 kDa, which 141 corresponds to a 5 cm gel length (Supplementary Fig. 1d). The region of interest is determined by the 142 size of the metal stage and the balance between measurement time and sufficient proteome 143 information. The single-molecule imaging is conducted with a scan speed of 80 µm/s, resulting in 144 approximately 10 minutes of imaging time per single cell. Fluctuations in the number of dye-labelled 145 proteins are observed along the migration path during imaging, which reflects protein bands with 146 differing protein abundances across molecular weights (Supplementary Fig. 2c, Supplementary Movie 147 1). Detecting and quantifying the dye-labelled proteins at the single-cell level allows us to gain 148 insights on protein expression profiles, cell type classification, and pseudotime analysis (Fig. 1f). 149 To evaluate the sensitivity and quantitative accuracy of the PISA measurement, we imaged and 150 quantified the number of dye-labelled proteins (referred to as protein count) from highly diluted bulk 151 HeLa and U2OS cell lysates, equivalent to a single-cell concentration. Note that the bulk cell lysates 152 were diluted to single-cell level based on the estimated initial number of the cell suspension. We 153 found that the reconstructed electropherogram obtained by single-molecule imaging was in good 154 agreement with the SDS-PAGE band profile visualized by the standard gel imager (Supplementary 155 Fig. 2b and Supplementary Fig. 2a, respectively). Quantitative analysis revealed that the protein 156 counts across molecular weights were consistent in three technical replicates, with a median 157 coefficient of variation (CV) of 6.5% and 7.8% for HeLa and U2OS, respectively (Supplementary Fig. 158 2d, e). The analytical reproducibility was assessed by computing the Pearson correlation of the protein 159 counts across molecular weights between replicates, which yielded coefficients larger than 0.96 160 (Supplementary Fig. 2f). To assess the linearity of the method, we performed an experiment using 161 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 7 highly diluted bulk cell lysates equivalent to 1, 2, 4, and 8 cells. Bulk cell lysates were used to 162 minimize the variability from cell-to-cell differences in protein content or cell cycle stages, thus 163 ensuring controlled and consistent protein input. Our result demonstrated excellent linearity (R² = 164 0.998) between cell equivalents and total protein counts (Supplementary Fig. 2g, h). The ability to 165 maintain linearity down to single-cell equivalent shows robust quantitative accuracy and great 166 detection sensitivity of our system, even in highly diluted samples. 167 Lastly, to determine the sensitivity of single-cell PAGE-PISA to detect low-abundance proteins, we 168 conducted a lysate spike-in experiment using transferrin (77 kDa) as a model protein and evaluated its 169 detectability within a complex cell lysate environment (Supplementary Figure 3a). After subtraction 170 of the negative control, we successfully detected as little as 5 fg of transferrin, corresponding to 171 approximately 7 × 10⁴ protein molecules (Supplementary Figure 3b). These results demonstrate that 172 single-cell PAGE-PISA can reliably detect and quantify proteins present at copy numbers ranging 173 from 10⁴–10⁵ within a complex cell lysate background. 174 Single-cell proteome profiling of mammalian cells 175 Next, we performed single-cell proteomic analysis using single-cell PAGE-PISA on two different 176 tumour cell lines, U2OS and PC-3 (Fig. 2). The single cells were manually isolated, prepared in PCR 177 tubes, and the dye-labelled proteins were separated by SDS-PAGE and imaged by PISA. The imaging 178 was performed at the middle molecular weight region, covering approximately 40‒50% of the cellular 179 proteome. The electropherogram was reconstructed from the PISA images to provide a comprehensive 180 view of the protein expression profiles across multiple single cells (10 biological replicates for each 181 cell type) (Fig. 2a). In practice, our single-cell PAGE-PISA could quantify dye-labelled proteins from 182 a 2 mm3 gel volume (0.2 × 0.2 × 50.0 mm3). This translates to 103‒104 dye-labelled proteins per band, 183 corresponding to approximately 2 fg, and a total of 105 dye-labelled proteins per single cell. Ideally, 184 imaging the entire migration path (4.0 × 1.0 × 50.0 mm3) could theoretically increase these numbers 185 to 105‒106 dye-labelled proteins per band, which corresponds to approximately 200 fg, and a total of 186 107 dye-labelled proteins in a single mammalian cell (Fig. 2b, c, Supplementary Fig. 4a). This 187 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 8 detection sensitivity far surpasses that of conventional staining methods for SDS-PAGE like 188 Coomassie Brilliant Blue (CBB) and silver staining, with detection thresholds of 8‒10 ng30 and 0.1‒189 0.5 ng31,32,33 per protein band, respectively. Thus, these methods were insufficient to visualize protein 190 bands from single-cell samples, unlike single-cell PAGE-PISA which can detect protein bands with 191 atto-gram amounts29. Thereby, the detection sensitivity was improved by 3‒5 orders of magnitude 192 over silver staining and CBB, enabling the detection of even 1% of the total protein abundance. 193 Compared to the estimated 6 × 109 protein copies per HeLa cell34, the total protein copies quantified 194 by single-cell PAGE-PISA fall short by two orders. The difference between practical and theoretical 195 values may stem from technical limitations during quantification. For instance, residual dyes that are 196 not completely quenched may weakly interact with the polyacrylamide gel during electrophoresis, 197 causing a slight increase in the overall background signal (Supplementary Fig. 5a, b, Supplementary 198 Movie 2). As a result, some dye-labelled protein molecules that are not distinctive from the 199

may fall below the detection threshold and failed to be quantified (Supplementary Fig. 200 5c). Furthermore, if two dye-labelled protein molecules are within the optical diffraction limit (i.e., a 201 few hundred nanometers), especially in highly dense protein bands, they cannot be resolved and will 202 be detected as a single molecule due to the diffraction limit. Nevertheless, the strong mean Pearson 203 correlation of protein counts across molecular weights between single cells from the same cell types (r 204 > 0.89) demonstrates great reproducibility and a prominent level of quantitative accuracy of single-205 cell PAGE-PISA (Fig. 2d, Supplementary Fig. 4b). 206 Next, we conducted a detailed peak fitting analysis to determine the peak capacity and mass 207 resolution of single-cell PAGE-PISA. We identified 21 distinct Gaussian peaks between 35 and 128 208 kDa in a representative single-cell proteome profile (Supplementary Fig. 6a). For proteins in the 35–209 60 kDa range (peaks #1–8), the average mass resolution was 1.34 kDa (FWHM) (Supplementary Fig. 210 6b). For proteins in the 60–128 kDa range (peaks #9–21), the resolution declined to an average of 211 2.69 kDa, likely due to reduced gel migration efficiency for large proteins. 212 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 9 We further assessed the accuracy and reliability of our single-cell measurement by comparing the 213 single-cell protein expression profile to those of highly diluted bulk cell lysates equivalent to a single-214 cell concentration from the same cell types. We averaged the protein counts across multiple single 215 cells (n = 10) and technical replicates of highly diluted bulk cell lysates (n = 3), then compared their 216 mean protein count across molecular weights (Supplementary Fig. 4c). We observed a strong Pearson 217 correlation of mean protein counts between single cells and highly diluted bulk cell lysates for both 218 U2OS and PC-3 cells (r > 0.90) (Fig. 2e). The observed correlation showed that the single-cell protein 219 quantification by single-cell PAGE-PISA was in good agreement with that from bulk cell lysates 220 diluted to the single-cell level, indicating the robustness and reliability of our single-cell 221 measurement. 222 Lastly, we visualized the 2D projections of single U2OS and PC-3 cells, specifically in UMAP 1 vs 223 UMAP 2 and UMAP 1 vs UMAP 3 plots (Fig. 2f). Our results showed clear separation between the 224 two cell lines, indicating that the clustering captured meaningful biological differences and was not 225 confined to higher-order dimensions. This demonstrates that single-cell PAGE-PISA can effectively 226 identify and classify different cell types solely based on their proteome. 227 Single-cell proteome profiling of cardiomyocyte differentiation from human induced pluripotent 228 stem cells 229 The ability of hiPSCs to self-renew and differentiate into various cell types and tissues has garnered 230 significant interest, especially in the pursuit of heart regenerative therapies35,36. Numerous studies 231 have shown that protein expressions of hiPSC-derived cardiomyocytes (hiPSC-CMs) exhibit dynamic 232 fluctuations and temporal changes during differentiation and maturation37,38,39. 233 To highlight the potential applicability and versatility of single-cell PAGE-PISA in detecting global 234 proteome changes during dynamic processes, we performed single-cell proteome profiling analysis on 235 hiPSCs as they progressed from pluripotency through stage-specific transitions during cardiomyocyte 236 differentiation (Fig. 3a). We observed cardiac contraction in hiPSC-CMs as early as day 7, indicating 237 the successful differentiation process. To improve cardiomyocyte purity in culture, the hiPSC-CMs 238 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 10 were carefully dissociated and transferred to a freshly coated plate between day 10 and 12. Cells were 239 harvested on different days, including day 0 (hiPSCs), 16, and 30 (hiPSC-CMs), referred to as D0, 240 D16, and D30, respectively. In total, 49 single cells were collected, prepared, and analyzed with 241 single-cell PAGE-PISA (n = 12, 14, and 23 for D0, D16, and D30, respectively) (Fig. 3b). 242 Quantitative analysis of single hiPSC-CMs from D16 and D30 revealed 1.5‒2.0 times higher total 243 protein abundance than D0, likely due to the increased cell size and complexity in specialized cells 244 (Fig. 3c). In addition, we observed a strong mean Pearson correlation of protein counts across 245 molecular weights among D0 single cells (r = 0.88), implying a consistent single-cell protein 246 expression levels in the undifferentiated state (Fig. 3d). Meanwhile, the mean correlation of protein 247 counts decreased by approximately 13.6% in both D16 and D30 single cells, suggesting stochastic 248 occurrences in individual cells during cardiomyocyte differentiation. Using all the standardized 249 protein counts from 35‒112 kDa, principal component analysis (PCA) separated the cells in a time-250 wise manner along PC1, which accounts for 44.1% of the total variance (Supplementary Fig. 7a). 251 Notably, the data points from different batches did not form distinct clusters in the PCA plot, 252 indicating that batch effects were negligible and did not confound the temporal separation. 253 We next sought to investigate the cell progression from pluripotency towards differentiated states by 254 pseudotemporal ordering of the dynamic cells across a developmental lineage40. Briefly, cells were 255 presented on a diffusion map based on the cell-to-cell transition probabilities, positioning those with 256 higher transition probabilities closer to each other (Fig. 3e, left). The trajectory was inferred by 257 computing diffusion pseudotime (DPT) for each cell relative to the root cell, often considered the 258 starting point of the differentiation process, and cells were ordered along the pseudotemporal 259 trajectory to track their developmental progress during cardiomyocyte differentiation (Fig. 3e, right). 260 As anticipated, the D0 single cells were in close proximity to one another on the diffusion map, 261 indicating homogeneous protein expression profiles among the single cells. In contrast, cells from 262 D16 and D30 appeared more spread out across the developmental trajectory, with D16 and D30 263 primarily dominating the early and terminal end of the trajectory, respectively. This reflects the 264 temporal dynamics of proteome changes as cells differentiate from early to later stages. We performed 265 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 11 UMAP analysis on DPT-assigned single cells and identified three clusters (Fig. 3f, Supplementary 266 Fig. 7b). Precursor stage (n = 13) consisted predominantly of D0 cells (n = 12, 92.3%), while early (n 267 = 23) and late cardiomyocytes (n = 13) were mainly composed of D16 and D30 cells (n = 10 and n = 268 10, 43.5% and 76.9%, respectively). 269 We further visualized the single-cell proteome profiles across different developmental stages to 270 identify differentially regulated protein bands (Fig. 3g, Supplementary Fig. 8a). By comparing the 271 protein expression differences between early and late cardiomyocytes, we identified two protein bands 272 that were differentially regulated in late cardiomyocytes (Fig. 3h, Welch's two-sided, two-sample t-273 test, p 1.2). Notably, the protein bands at 55.0‒57.7 kDa and 47.1‒48.0 kDa 274 were significantly expressed in late cardiomyocytes compared to the precursor stage and early 275 cardiomyocytes (Fig. 3i). However, due to the lack of reported cardiac-specific proteins within these 276 molecular weight ranges, we hypothesize that the observed expression may be associated with certain 277 biological processes which are crucial for cardiomyocyte development. To support our notion, we 278 compared our single-cell proteome data to the single-cell transcriptome data obtained from a public 279 database41. We averaged the transcript counts across single cells for each sampling day: D0 (n = 280 9,146), D15 (n = 2,897), and D30 (n = 3,294), visualized their mean transcriptome profiles, and 281 identified upregulated genes in D30 (Supplementary Fig. 8b, c, Welch's two-sided, two-sample t-test, 282 p 1.4). Among the differentially upregulated genes with the highest transcript 283 count within 55.0‒57.7 kDa and 47.1‒48.0 kDa were ENO1 (47.2 kDa), ATP5B (56.5 kDa), and MT-284 CO1 (57.0 kDa) (Supplementary Fig. 8d). In addition, a previous study has shown that both MT-CO1 285 and ATP5B proteins were significantly expressed in the cavities of cardiac tissue42, consistent with 286 their role in energy production during cardiac development. Considering these genes are 287 quantitatively abundant and significantly expressed at both transcriptome and proteome levels, we 288 believe that their expression levels are reflected in single-cell PAGE-PISA. 289 Comparative analysis between single-cell proteome and transcriptome during cardiomyocyte 290 differentiation 291 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 12 As proteome and transcriptome measurements are not always correlated5,6,7,8,9, we next compared the 292 extent to which our single-cell proteome data obtained by single-cell PAGE-PISA are in agreement 293 with reported single-cell RNA sequencing (scRNA-seq)41 during cardiomyocyte differentiation. The 294 scRNA-seq data was chosen as the sampling days when the single cells were collected closely 295 matched our single-cell PAGE-PISA. We analyzed both datasets with an equal sample size (n = 49) 296 for accurate comparison. Three random samples (RS) were generated from the scRNA-seq data from 297 37‒126 kDa, with each RS consisting of 49 randomly selected single cells from three different 298 sampling days (n = 12, 14, and 23 for D0, D15, and D30, respectively). UMAP analysis revealed a 299 consistent spatial distribution of the cells at both proteome and transcriptome levels (Fig. 4a). The 300 D0protein and D0RNA populations formed a distinct cluster, with cells in close proximity to each other, 301 implying that both populations exhibit stable developmental states with consistent expression profiles. 302 Meanwhile, D16protein, D30protein, D15RNA, and D30RNA populations displayed two separate clusters that 303 were distinctive from D0 but did not otherwise cluster with respect to their sampling days. These 304

indicate that although D16protein, D30protein, D15RNA, and D30RNA populations share similar 305 expression profiles on UMAP space, they reflect transitional states or different developmental stages 306 during cardiomyocyte differentiation. 307 We expect the spatial distribution of these states can be better resolved on a pseudotemporal scale. 308 Hence, we ordered the dynamic cells along a developmental trajectory from pluripotency to 309 differentiated states at both proteome and transcriptome levels (Fig. 4b, Supplementary Fig. 9b, c). As 310 expected, the majority of D30protein and D30RNA single cells were located at the terminal end of the 311 trajectory, distal from the root cell, signifying their progression into a fully differentiated state. While 312 D0protein population clustered closer to the root cell on the diffusion map, we observed greater 313 variations of D0RNA population in all three RS, with cells sparsely distributed across the 314 developmental trajectory. The D0RNA population, however, can be clearly distinguished from D15RNA 315 and D30RNA populations when increasing the sample size from 49 to 3,000 single cells 316 (Supplementary Fig. 9a). This emphasizes the need for a larger sample size of scRNA-seq data43 to 317 capture the full spectrum of biological variability. 318 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 13 Lastly, we sought to investigate how well the inferred DPT correlates with the sampling day (referred 319 to as real-time) during cardiomyocyte differentiation at both proteome and transcriptome levels. To 320 this end, we visualized the DPT-assigned single cells on UMAP and assessed the correlation between 321 the DPT and real-time (Fig. 4c, d). We observed a strong concordance between DPTprotein and real-322 timeprotein (Spearman’s rho = 0.72), while relatively lower concordance between DPTRNA and real-323 timeRNA in all three RS (Spearman’s rho < 0.49) (Fig. 4d). These results demonstrate that the 324 directional changes of the developmental states at proteome level are generally consistent whether 325 determined by DPT or real-time and that both recapitulate the expected temporal changes during 326 cardiomyocyte differentiation. While several D15RNA and D30RNA cells displayed a wide dynamic 327 range of DPT values due to the different progression rate during the differentiation process, it is more 328 pronounced in D0RNA populations, suggesting the presence of subpopulations at the pluripotent state, 329 consistent with previous reports44,45. The observed disparities could be attributed to the difference in 330 stabilities and half-lives of RNA and protein. By comparison, RNA has a relatively shorter half-life 331 and is more susceptible to degradation than proteins46. In turn, it directly impacts the RNA abundance 332 and leads to temporal variations during sampling. In contrast, long protein half-lives and greater 333 stability have led to more consistent expression levels, which can accurately reflect the cellular states 334 and functions. 335 Overall, these findings demonstrate the ability of single-cell PAGE-PISA to provide a direct and 336 robust depiction of developmental status during dynamic processes, even with minimal sample sizes. 337 Its ability to resolve developmental trajectories from as few as 49 cells not only is particularly 338 promising for studies of rare cell populations or samples with limited availability, but it also opens up 339 new avenues in various research fields, including developmental biology, regenerative medicine, and 340 disease modeling. Together, single-cell PAGE-PISA serves as a powerful complementary approach to 341 scRNA-seq by bridging the gaps between transcriptome and proteome and providing a more 342 comprehensive understanding of cellular dynamics at the single-cell level. 343

344 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 14 In this present work, we introduced single-cell PAGE-PISA capable of highly sensitive and 345 quantitative profiling of a wide range of proteins in individual cells. This system was realized through 346 the integration of PAGE for protein separation based on molecular sizes with 3D single-molecule 347 imaging using PISA for protein quantification. The workflow is streamlined and benchtop-compatible, 348 using microliter-scale sample volumes in standard PCR tubes that are easily handled with pipettes, 349 eliminating the need for specialized robotic systems. 350 With single-cell PAGE-PISA, we detected 21 protein bands per single cell and quantified 107 351 molecules covering approximately 1% of the total proteome per cell. This proportion can be further 352 improved to over 10% by using brighter dyes such as SeTau-64747 or advanced spot-recognition 353 algorithms, including deep learning48 (Supplementary Fig. 10). With further improvements in 354 separation resolution and labelling efficiency, we anticipate achieving detection sensitivity down to 355 102–103 molecules per species, enabling the analysis of low-abundance proteins. 356 A key advantage of single-cell PAGE-PISA is its ability to uncover functional subpopulations that 357 may not be distinguishable using conventional markers or transcriptomic profiling. For example, 358 heterogeneous T cell populations often exhibit diverse activation states that are not easily resolved by 359 known surface markers used in cell sorting49,50. By capturing the full proteomic content of individual 360 cells, our method enables the identification of rare or intermediate states based on functional protein 361 features. Another strength is its ability to resolve modification-specific and structural proteoforms that 362 are often inaccessible to MS-based approaches due to sample loss, limited enrichment efficiency, or 363 disruption of protein complexes during sample preparation. This is enabled by integrating Phos-tag51 364 PAGE or native PAGE, which permits direct detection of phosphorylated isoforms and native protein 365 complexes at the single-cell level—information critical for understanding cellular decisions governed 366 by signaling assemblies such as TNF receptor complexes52. Furthermore, by labelling specific 367 proteins of interest in situ, this platform could be extended to spatially resolved subcellular 368 proteomics, enabling the analysis of organelle-associated proteins while preserving intracellular 369 localization. 370 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 15 Although chip-based electrophoresis, as implemented in single-cell Western blotting26, and capillary 371 electrophoresis (CE) can, in principle, be adapted for proteome analysis using PISA, both approaches 372 face significant limitations for single-cell proteome profiling. The format used in single-cell Western 373 offers high-throughput analysis of over 1,000 cells but suffers from limited separation resolution due 374 to its short migration distance (~1 mm), resulting in overlapping bands and reduced quantification 375 accuracy. CE systems, such as the PA 800 Plus (Sciex), can provide higher-resolution separation due 376 to longer capillary lengths. However, their workflows are inherently sequential and typically require 377 1–2 hours per cell, which imposes a major constraint on throughput. Taken together, single-cell 378 PAGE-PISA offers a practical and effective balance of resolution, sensitivity, and throughput for 379 comprehensive proteome profiling at the single-cell level. 380 Although single-cell PAGE-PISA currently requires approximately 7 hours to analyze 20 cells 381 simultaneously, scaling up to thousands of cells will likely be necessary to address broader biological 382 questions. We anticipate that such expansion will be feasible through the use of automated single-cell 383 collection (e.g., fluorescence-activated cell sorting or automated live imaging and picking systems 384 such as ALPS53), high-throughput robotic reagent handling, and rapid, parallel electrophoresis (e.g., 385 10-minute runs at 400 V using bullet PAGE from Nacalai Tesque). Although imaging currently limits 386 the overall throughput, our preliminary tests demonstrate that scan speeds can be increased up to 387 400 µm/s without loss of accuracy29. The integration of high-speed CMOS cameras with larger fields 388 of view could further accelerate data acquisition. Overall, scaling up to thousands of cells is estimated 389 to require approximately 30 hours from cell isolation to PISA imaging. Since single-cell PAGE-PISA 390 uses only commercially available materials, the cost per cell is estimated at $0.60–1.00 and can be 391 reduced to ~$0.11 using homemade gels, making it a highly cost-effective alternative to other single-392 cell proteomics methods. 393 Currently, a major limitation of single-cell PAGE-PISA is the inability to assign molecular identities 394 to individual protein bands, which limits the biological interpretability of the protein expression 395 profiles and constrains the applicability for precise cell phenotyping. To address this limitation, one 396 potential strategy is to incorporate multiplexed protein identification directly within the gel using 397 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 16 antibodies or multicolor tags, similar to In-Gel Western54. Additionally, combining single-cell PAGE-398 PISA with fluorescent antibody-based sorting via FACS and conventional MS can help associate 399 single-cell proteome profiles with specific protein identities. This integrative approach enables 400 molecular annotation of phenotypically meaningful bands and enhances the ability to interpret cell 401 state differences observed in single-cell PAGE-PISA. 402 In conclusion, we have established single-cell PAGE-PISA and demonstrated its potential for ultra-403 sensitive single-cell proteome analysis on diverse cell types, from steady-state to differentiated cells. 404 We strongly believe that our system can help researchers narrow down the number of potential protein 405 species in each band at single-cell level prior to immunofluorescent imaging or MS. By focusing on a 406 smaller subset of proteins, this approach not only will help to reduce the time and resources required 407 for downstream analysis but also minimize the need for extensive antibody libraries, making target 408 protein identification more cost-effective for future biomarker discovery. 409

410 Cell culture 411 U2OS, HeLa (RCB0007), and PC-3 (RCB2145) cells were obtained from the RIKEN cell bank. The 412 U2OS and HeLa cells were maintained in Dulbecco's modified eagle medium (DMEM) (Thermo 413 Fisher Scientific, 10566-016) supplemented with 10% fetal bovine serum (FBS) (Corning, 35-079-414 CV). The PC-3 cells were maintained in RPMI 1640 medium (Thermo Fisher Scientific, 11875-093) 415 supplemented with 10% FBS. Human induced pluripotent stem cells (hiPSCs) were purchased from 416 RIKEN BRC (HPS4290:201B7-Ff) and maintained in mTeSRTM Plus medium (STEMCELL 417 Technologies, 100-0276). All cell lines were maintained in a 5% CO2 incubator at 37°C. 418 Cardiomyocyte differentiation protocol 419 Cardiomyocyte-directed differentiation from hiPSCs was performed and modified based on the 420 previous protocol55. Four days before cardiomyocyte differentiation, the hiPSCs were dissociated 421 using TrypLETM Select CTSTM (Gibco, A12859-01) and cultured in a matrigel-coated 24-well plate at 422 a density of 5 × 104 cells/well in mTeSRTM Plus. On day 0, the hiPSCs were washed with phosphate-423 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 17 buffered saline (PBS) (Nacalai-tesque, 14249-95) and treated with differentiation medium (RPMI 424 1640, GlutaMAX™ (Gibco, 61870036) containing 500 µg/mL human serum albumin (HSA) (Wako, 425 010-27601) and 213 µg/mL L-ascorbic acid 2-phosphate (Nacalai-tesque, 13571-56)) supplemented 426 with 6 μM CHIR99021 (Wako, 252917-06-9). On day 2, the medium was replaced with fresh 427 differentiation medium supplemented with 5 μM IWP-2 (Wako, 686770-61-6). On day 4, the medium 428 was replaced with a fresh differentiation medium without supplemental inhibitors. Medium exchange 429 was performed every two days. Beating was first observed on day 7. 430 Bulk cell lysate sample preparation 431 Cells were harvested, washed with PBS three times, and counted using an automatic cell counter 432 (TC20, BioRad) to obtain a concentration of 1.5 × 103 cells/µL. Approximately 50 µL of lysis buffer 433 (50 mM borate (Nacalai-tesque), 1% Tween 20 (Sigma-Aldrich, 9005-64-5), 1% sodium dodecyl 434 sulfate (SDS) (Wako, 192-13981), adjusted at pH 8.0) and 5 µL of protease inhibitor (Nacalai-tesque, 435 25955) were added to 50 µL of cell suspension. After cell lysis, proteins were labelled with Cy5-NHS 436 ester dye (AA T Bioquest, 151) to a final concentration of 100 µM and incubated at 15ºC and 2000 437 rpm for 45 minutes. The dye-labelled proteins were washed with PBS three times in a centrifugal 438 filter unit (10K) (Millipore, UFC5010BK) and centrifuged at 10,000 × g and 4ºC for 10 minutes. 439 Approximately 50 µL of protein sample was collected after the final wash and stored at -20ºC. 440 Lysate spike-in experiment 441 Approximately 15,000 cell lysates were spiked with 75 ng, 7.5 ng, and 0.75 ng of purified transferrin 442 (Fujifilm Wako, 205-18084), lysed in the presence of protease inhibitor, and labelled with Cy5-NHS 443 ester dye to a final concentration of 100 µM. The samples were washed with 20 mM borate buffer 444 three times in a centrifugal filter unit (10K) and centrifuged at 10,000 × g and 4ºC for 10 minutes. For 445 single-molecule imaging, the lysate spike-in samples were diluted to obtain cell lysates containing 446 500 fg, 50 fg, and 5 fg of transferrin. Negative control was prepared containing only bulk cell lysates 447 without exogenous protein. 448 Single-cell isolation 449 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 18 A droplet of cell suspension was deposited onto a glass slide (Matsunami, S1111) that was placed on 450 the stage of an inverted microscope (CKX41, Olympus) equipped with a TOPick 1-cell handling 451 system, composed of a touch panel, micro liquid pump, and controller (YODAKA Co., Ltd.). The 452 single cells were manually isolated using a 30 µm G-tip low-binding coated micro-glass needle 453 (YODAKA Co., Ltd.) with a minimum handling volume of 100 pL. The isolated single cells were 454 transferred to the PCR cap containing 1 µL of PBS, spun down, and stored at -80ºC. 455 Single-cell sample preparation 456 Manually isolated single cells were lysed on ice with 0.6 µL of lysis buffer and 0.4 µL of protease 457 inhibitor. Proteins were labelled with 0.4 µL of Cy5-NHS ester dye to a final concentration of 1 µM 458 and incubated at 15ºC and 2000 rpm for 45 minutes. To quench the reactivity of excess unreacted dye, 459 0.4 µL of Tide Quencher™ 5WS amine (AA T Bioquest, 2076) was added to the protein sample to a 460 final concentration of 10 µM. The quencher was added not only to stop the reactivity of NHS ester but 461 to effectively quench the fluorescence of unreacted Cy5-NHS ester dye. As such, it omits the need for 462 common dye removal methods such as gel filtration or desalting columns, which often result in 463 significant protein loss. For efficient protein labelling, the molar ratio of dye to quencher solution 464 used in this protocol was 1:10. To avoid protein loss, all reagents were dispensed along the walls of 465 the PCR tubes, spun down, and stored at -80ºC. Negative control was prepared similar to the single-466 cell sample but containing only PBS buffer, lysis buffer, protease inhibitor, dye, and quencher 467 solution. 468 SDS-PAGE 469 Electrophoresis was performed using two different polyacrylamide gels depending on the 470 experimental purposes. For standard gel imaging (Fujifilm, LAS 4000), the dye-labelled protein 471 samples from 7.5 × 103 cell lysates were mixed with 4× SDS sample buffer (240 mM Tris-HCl, 40% 472 glycerol stock (Nacalai-tesque, 17045-65), 8% SDS (Nacalai-tesque, 31606-75), and β-473 mercaptoethanol (Nacalai tesque, 21438-82), adjusted at pH 6.8), heated at 95ºC for two minutes, and 474 separated using a 12-well, 5‒20% precast polyacrylamide gel (Bio-craft, #SDG-571). The loading 475 volume was 12 µL. For single-molecule imaging, the dye-labelled protein samples were prepared, 476 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 19 heated at 95ºC for two minutes, and separated using a 20-well, 5‒20% precast polyacrylamide gel 477 (Bio-craft, #SDG-576). A total volume of 8 µL was loaded for highly diluted bulk cell lysates 478 equivalent to single-cell level and 5 µL for single-cell lysates, both containing 4× SDS sample buffers. 479 To avoid diffusion of free Cy5-NHS ester dye during protein migration, the samples were applied and 480 separated by two loading wells. All gels were irradiated under UV overnight to remove 481 autofluorescence signals that could contribute to the background noise. Electrophoresis was 482 performed at 250 V and 30 mA for 70 minutes. 483 Single-molecule imaging 484 After SDS-PAGE, the polyacrylamide gel was cut vertically into a 5 cm length, covering the majority 485 of the protein bands in the middle molecular weight region, and placed on a UV-irradiated fluorinated 486 ethylene propylene (FEP) film (Daikin Chemical, NF-0025). To avoid gel desiccation and reduce the 487 reflection of excitation illumination on the gel surface during observation, two blank polyacrylamide 488 gels were placed on top of the gel with protein bands, followed by a thin layer of transparent film 489 (Supplementary Fig. 1c, d) This configuration allows for a longer PISA observation of up to four 490 hours. 491 A detailed description of the optical components and designs of PISA can be found in our previous 492 report29. Briefly, PISA was built on a custom microscope body with two water-immersion objective 493 lenses, fluorescence illumination (Special Optics, 54-10-7, NA = 0.66, 28.6×) and detection (Evident, 494 XLUMPLFLN 20XW, NA = 1.0, 20×), that were placed below the coverslip at a tilted angle of 33.8 495 degrees. Bessel beam was generated by passing the laser source via an axicon lens (Mie Optics), 496 which was then reflected by a Galvano mirror (Cambridge Technology, 6215HB), creating a light 497 sheet. The detection port was connected to an EM-CCD camera (Andor, iXon Ultra 897) via an 498 imaging lens and used to image fluorescence single molecules. 499 Imaging was conducted using a 647 nm fiber laser (MPB Communications, 2RU-VFL-P-2000-647) at 500 1,000 mW and detected through a near-infrared band-pass filter (Semrock, FF01-708/75-25). The 501 protein bands embedded in polyacrylamide gel were imaged along a Y’-axis with a 4 μm step size at 502 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 20 50 ms exposure per frame. Images were acquired using a commercial software (Molecular Devices, 503 MetaMorph) and a motorized stage (Prior Scientific, H117), which were controlled by a homemade 504 program written in LabVIEW (National Instruments) and saved in TIFF format for further image 505 analysis. 506 Image analysis 507 All raw images were opened in ImageJ software (v1.51n), pre-processed using the rolling ball 508 algorithm for background subtraction, and imported into Arivis Vision 4D (Zeiss) (Supplementary 509 Fig. 5a, b). Denoising filter was applied to the background-subtracted images to improve the signal-510 to-noise ratio and a blob finder tool was used to detect spots in the 3D image matrix (Supplementary 511 Fig.5c). Segment filter was optionally applied to remove single molecules with low-intensity signals. 512 The images were saved and the acquired dataset containing the information of each spot, 513 corresponding to single dye-labelled proteins, was exported in CSV format for data analysis. 514 Data analysis 515 The exported dataset was analyzed and visualized using OriginPro (OriginLab) and Python (version 516 3.10.11). Frequency count function was used to generate a binned dataset and provide the count of 517 dye-labelled proteins in each bin. The data was organized into a matrix of protein count × cell ID 518 (rows × columns). To compensate for gel-to-gel variability, the proteome profile of each cell was 519 manually aligned and scaled to standardize the peak (protein bands) positions across multiple 520 technical or biological replicates. Molecular weights of unknown proteins were estimated by fitting a 521 standard curve to relate the known molecular weights of proteins in a set of standards to their relative 522 migration. 523 Two different normalization approaches were used to normalize the protein counts. For dimensional 524 reduction such as principal component analysis (PCA), uniform manifold approximation and 525 projection (UMAP), and diffusion map, the protein counts of individual cells were standardized using 526 z-score, with the mean set to zero and standard deviation set to one. UMAP was used to visualize 527 single cells in two (or three) dimensional maps and the clusters were identified using k-means. 528 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 21 Diffusion map was used to visualize the cell trajectory during cardiomyocyte differentiation by 529 employing a random-walk algorithm to estimate the cell transition probabilities based on a weighted 530 nearest neighbors graph40. Trajectory was inferred by computing diffusion pseudotime (DPT) for each 531 cell relative to the root cell, and the pseudotemporal values were assigned to the cells along the 532 inferred trajectory. To identify differentially regulated protein bands, the number of quantified 533 proteins for each molecular weight was normalized by the total protein count in individual cells. 534 Differential protein bands (visualized in volcano plot) were determined by the Welch's two-sided, 535 two-sample t-test (p 1.2). For pairwise comparisons between three or more 536 groups, significant differences were determined by the one-way Welch’s ANOV A with Games-Howell 537 post-hoc test (p < 0.05). 538 The scRNA-seq datasets for D0, D15, and D30 were obtained from the ArrayExpress database at 539 EMBL-EBI under the accession number E-MTAB-626841. Three random samples (RS) were 540 generated from the dataset, each containing 49 randomly selected single cells from D0, D15, and D30 541 (n = 12, 14, and 23, respectively). For a larger sample size, 3,000 single cells were randomly selected 542 from D0, D15, and D30 (n = 1,000 for each sampling day). Differential genes were determined by the 543 Welch's two-sided, two-sample t-test (p 1.4). The analysis included a total of 544 10,484 genes, which represented all genes corresponding to proteins between 37 and 126 kDa. The 545 protein molecular weights were obtained from UniProt. 546 Data availability: Source Data are provided with this paper. 547 Code availability: All custom analysis code generated as part of this work is available from the 548 corresponding author upon reasonable request. 549

550 1. Papalexi, E. & Satija, R. Single-cell RNA sequencing to explore immune cell heterogeneity. Nat. 551 Rev. Immunol. 18, 35–45 (2018). 552 2. Zhang, S., Li, X., Lin, J., Lin, Q. & Wong, K. C. Review of single-cell RNA-seq data clustering 553 for cell-type identification and characterization. Rna 29, 517–530 (2023). 554 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 22 3. Jiang, L., Chen, H., Pinello, L. & Y uan, G. C. GiniClust: detecting rare cell types from single-cell 555 gene expression data with Gini index. Genome Biol. 17, 1–13 (2016). 556 4. Xu, C. & Su, Z. Identification of cell types from single-cell transcriptomes using a novel 557 clustering method. Bioinformatics 31, 1974–1980 (2015). 558 5. Taniguchi, Y . et al. Quantifying E. coli proteome and transcriptome with single-molecule 559 sensitivity in single cells. Science 329, 533–538 (2010). 560 6. Schwanhäusser, B. et al. Global quantification of mammalian gene expression control. Nature 561 473, 337–342 (2011). 562 7. Darmanis, S. et al. Simultaneous multiplexed measurement of RNA and proteins in single cells. 563 Cell Rep. 14, 380–389 (2016). 564 8. Gong, H. et al. Single-cell protein-mRNA correlation analysis enabled by multiplexed dual-565 analyte co-detection. Sci. Rep. 7, 2776 (2017). 566 9. Reimegård, J. et al. A combined approach for single-cell mRNA and intracellular protein 567 expression analysis. Commun. Biol. 4, 624 (2021). 568 10. V ogel, C. & Marcotte, E. M. Insights into the regulation of protein abundance from proteomic and 569 transcriptomic analyses. Nat. Rev. Genet. 13, 227–232 (2012). 570 11. Hanna, J., Guerra-Moreno, A., Ang, J. & Micoogullari, Y . Protein degradation and the pathologic 571 basis of disease. Am. J. Pathol. 189, 94–103 (2019). 572 12. Kim, S., Kamarulzaman, L. & Taniguchi, Y . Recent methodological advances towards single-cell 573 proteomics. Proc. Jpn. Acad. Ser. B 99, 306–327 (2023). 574 13. Mishra, N. C. Methodology for separation and identification of proteins and their interactions. 575

to Proteomics 61–102 (John Wiley & Sons, Hoboken, 2010). 576 14. Wiśniewski, J. R., Hein, M. Y ., Cox, J. & Mann, M. A “proteomic ruler” for protein copy number 577 and concentration estimation without spike-in standards. Mol. Cell. Proteomics 13, 3497–3506 578 (2014). 579 15. Budnik, B., Levy, E., Harmange, G. & Slavov, N. SCoPE-MS: mass spectrometry of single 580 mammalian cells quantifies proteome heterogeneity during cell differentiation. Genome Biol. 19, 581 1–12 (2018). 582 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 23 16. Zhu, Y . et al. Nanodroplet processing platform for deep and quantitative proteome profiling of 583 10–100 mammalian cells. Nat. Commun. 9, 882 (2018). 584 17. Li, Z. Y . et al. Nanoliter-scale oil-air-droplet chip-based single-cell proteomic analysis. Anal. 585 Chem. 90, 5430–5438 (2018). 586 18. Woo, J. et al. High-throughput and high-efficiency sample preparation for single-cell proteomics 587 using a nested nanowell chip. Nat. Commun. 12, 6246 (2021). 588 19. Petelski, A. A. et al. Multiplexed single-cell proteomics using SCoPE2. Nat. Protoc. 16, 5398–589 5425 (2021). 590 20. Gebreyesus, S. T. et al. Streamlined single-cell proteomics by an integrated microfluidic chip and 591 data-independent acquisition mass spectrometry. Nat. Commun. 13, 37 (2022). 592 21. Leduc, A., Huffman, R. G., Cantlon, J., Khan, S. & Slavov, N. Exploring functional protein 593 covariation across single cells using nPOP. Genome Biol. 23, 261 (2022). 594 22. Ctortecka, C. et al. An automated nanowell-array workflow for quantitative multiplexed single-595 cell proteomics sample preparation at high sensitivity. Mol. Cell. Proteomics 22, 100665 (2023). 596 23. Wang, Y . et al. Pick-up single-cell proteomic analysis for quantifying up to 3000 proteins in a 597 mammalian cell. Nat. Commun. 15, 1279 (2024). 598 24. Ctortecka, C. et al. Automated single-cell proteomics providing sufficient proteome depth to 599 study complex biology beyond cell type classifications. Nat. Commun. 15, 5707 (2024). 600 25. Bendall, S. C. et al. Single-cell mass cytometry of differential immune and drug responses across 601 a human hematopoietic continuum. Science 332, 687–696 (2011). 602 26. Hughes, A. J. et al. Single-cell western blotting. Nat. Methods 11, 749–755 (2014). 603 27. Grist, S. M., Mourdoukoutas, A. P. & Herr, A. E. 3D projection electrophoresis for single-cell 604 immunoblotting. Nat. Commun. 11, 6237 (2020). 605 28. Alibekova Long, M., Benman, W. K., Petrikas, N., Bugaj, L. J. & Hughes, A. J. Enhancing single-606 cell western blotting sensitivity using diffusive analyte blotting and antibody conjugate 607 amplification. Anal. Chem. 95, 17894–17902 (2023). 608 29. Taniguchi, Y . et al. PISA: versatile microscope for 3D single molecule light sheet imaging. 609 Preprint at https://doi.org/10.1101/2024.12.05.625331 (2024). 610 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 24 30. Patton, W. F. Detection technologies in proteome analysis. J. Chromatogr. B 771, 3–31 (2002). 611 31. Neuhoff, V ., Arold, N., Taube, D. & Ehrhardt, W. Improved staining of proteins in polyacrylamide 612 gels including isoelectric focusing gels with clear background at nanogram sensitivity using 613 Coomassie Brilliant Blue G-250 and R-250. Electrophoresis 9, 255–262 (1988). 614 32. Jin, L. T., Hwang, S. Y ., Yoo, G. S. & Choi, J. K. Sensitive silver staining of protein in sodium 615 dodecyl sulfate‐polyacrylamide gels using an azo dye, calconcarboxylic acid, as a silver‐ion 616 sensitizer. Electrophoresis 25, 2494–2500 (2004). 617 33. Chevallet, M., Luche, S. & Rabilloud, T. Silver staining of proteins in polyacrylamide gels. Nat. 618 Protoc. 1, 1852–1858 (2006). 619 34. Bekker-Jensen, D. B. et al. An optimized shotgun strategy for the rapid generation of 620 comprehensive human proteomes. Cell Syst. 4, 587–599 (2017). 621 35. Rajala, K., Pekkanen-Mattila, M. & Aalto-Setälä, K. Cardiac differentiation of pluripotent stem 622 cells. Stem Cells Int. 2011, 383709 (2011). 623 36. Zakrzewski, W., Dobrzyński, M., Szymonowicz, M. & Rybak, Z. Stem cells: past, present, and 624 future. Stem Cell Res. Ther. 10, 1–22 (2019). 625 37. Ivashchenko, C. Y . et al. Human-induced pluripotent stem cell-derived cardiomyocytes exhibit 626 temporal changes in phenotype. Am. J. Physiol. Heart Circ. Physiol. 305, H913–H922 (2013). 627 38. Hellen, N. et al. Proteomic analysis reveals temporal changes in protein expression in human 628 induced pluripotent stem cell-derived cardiomyocytes in vitro. Stem Cells Dev. 28, 565–578 629 (2019). 630 39. Jabart, E. et al. Single-cell protein expression of hiPSC-derived cardiomyocytes using single-cell 631 westerns. J. Mol. Cell. Cardiol. 149, 115–122 (2020). 632 40. Haghverdi, L., Büttner, M., Wolf, F. A., Buettner, F. & Theis, F. J. Diffusion pseudotime robustly 633 reconstructs lineage branching. Nat. Methods 13, 845–848 (2016). 634 41. Friedman, C. E. et al. Single-cell transcriptomic analysis of cardiac differentiation from human 635 PSCs reveals HOPX-dependent cardiomyocyte maturation. Cell Stem Cell 23, 586–598 (2018). 636 42. Doll, S. et al. Region and cell-type resolved quantitative proteomic map of the human heart. Nat. 637 Commun. 8, 1469 (2017). 638 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 25 43. Liu, Y . et al. Robustness of single-cell RNA-seq for identifying differentially expressed genes. 639 BMC Genomics 24, 371 (2023). 640 44. Narsinh, K. H. et al. Single-cell transcriptional profiling reveals heterogeneity of human induced 641 pluripotent stem cells. J. Clin. Invest. 121, 1217–1221 (2011). 642 45. Nguyen, Q. H. et al. Single-cell RNA-seq of human induced pluripotent stem cells reveals cellular 643 heterogeneity and cell state transitions between subpopulations. Genome Res. 28, 1053–1066 644 (2018). 645 46. Shao, W. et al. Comparative analysis of mRNA and protein degradation in prostate tissues 646 indicates high stability of proteins. Nat. Commun. 10, 2524 (2019). 647 47. Podgorski, K., Terpetschnig, E., Klochko, O. P., Obukhova, O. M. & Haas, K. Ultra-bright and -stable 648 red and near-infrared squaraine fluorophores for in vivo two-photon imaging. PLoS One 7, e51980 649 (2012). 650 48. Liu, X., Jiang, Y ., Cui, Y ., Y uan, J. & Fang, X. Deep learning in single-molecule imaging and 651 analysis: recent advances and prospects. Chem. Sci. 13, 11964–11980 (2022). 652 49. Savas P. et al. Single-cell profiling of breast cancer T cells reveals a tissue-resident memory subset 653 associated with improved prognosis. Nat. Med. 24, 986–993 (2018). 654 50. Slavov N. Single-cell protein analysis by mass spectrometry. Curr . Opin. Chem. Biol. 60, 1–9 (2021). 655 51. Kinoshita, E., Kinoshita-Kikuta, E. & Koike, T. Separation and detection of large phosphoproteins 656 using Phos-tag SDS-PAGE. Nat. Protoc. 4, 1513–1521 (2009). 657 52. Gough, P. & Myles, I. A. Tumor necrosis factor receptors: pleiotropic signaling complexes and their 658 differential effects. Front. Immunol. 11, 585880 (2020). 659 53. Jin, J. et al. Robotic data acquisition with deep learning enables cell image–based prediction of 660 transcriptomic phenotypes. Proc. Natl Acad. Sci. USA 120, e2210283120 (2023). 661 54. LI-COR Biosciences. In-Gel Westerns. LI-COR. 662 https://www.licor.com/bio/applications/in-gel-westerns (accessed December 10, 2024). 663 55. Burridge, P. W. et al. Chemically defined generation of human cardiomyocytes. Nat. Methods 11, 664 855–860 (2014). 665 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 26

We sincerely thank all members of our laboratory and Dr. Tomoyuki Ohkawa 666 (Kyoto University) for the useful insights and discussions. L.K. acknowledges the MEXT scholarship 667 program for its support. This work was supported by grants-in-aid for Scientific Research (A) 668 (20H00460), Challenging Pioneering Research (19H05545 and 20K20458), Early-Career Scientists 669 (19K15718 and 22K14800) from Japan Society for the Promotion of Science; and ACT-X 670 (JPMJAX1914), PRESTO (JPMJPR25J4), and CREST (JPMJCR2334) from Japan Science and 671 Technology Agency; and grants from the RIKEN DECODE project, Stage Transition project, RIKEN 672 Incentive Research Projects, Suntory Rising Stars Encouragement Program in Life Sciences 673 (SunRiSE). S.K. acknowledges support from RIKEN’s SPDR fellowship. 674 Author contributions: Y .T., L.K., and S.K. conceived the main idea of the research. Y .T. designed the 675 PISA microscope. S.K. and Y .T. constructed the microscopic system. L.K. and S.K. optimized the 676 condition for microscopic imaging. L.K., S.K., T.H., and M.T. performed mammalian cell culture and 677 cardiomyocyte differentiation from human induced pluripotent stem cells. L.K., Y .T., and S.K. wrote 678 the manuscript, designed the experiments, and prepared the main and supplementary figures. L.K. 679 designed the analysis pipelines and conducted the bioinformatical analysis. L.K. and M.T. performed 680 all supplementary experiments. All authors contributed ideas for data analysis and interpretations and 681 participated in the manuscript revision. Y .T. and S.K. supervised the research and acquired funding. 682 Competing interests: The authors declare no competing interests. 683 Correspondence and requests for materials should be addressed to Yuichi Taniguchi 684 ([email protected]). 685 Additional information 686 Supplementary information consists of 10 Supplementary figures, 2 Supplementary movies, and 1 687 Supplementary table. 688 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 27 Figure 1 | Workflow of single-cell PAGE-PISA. a Firstly, cells are dissociated, washed three times, 689 and suspended in PBS at a concentration appropriate for single-cell collection. b Then, the single cells 690 are manually isolated from a droplet of cell suspension using an inverted microscope equipped with a 691 cell picker (TOPick 1-cell handling system) and transferred into PCR caps containing PBS. The 692 isolated single cells are spun down to ensure they settle at the bottom of the PCR tubes. c After cell 693 lysis, the proteins are labelled with Cy5-NHS ester dye solution at 15°C for 45 minutes. To inhibit the 694 reactivity of excess dye, quencher solution is added to the protein sample, followed by a second 695 incubation at 15°C for 45 minutes. All steps are performed on ice. The molar ratio of the dye to 696 quencher solution used in the single-cell sample preparation is 1:10. d The dye-labelled proteins from 697 each single cell are loaded into the well and separated by molecular size using SDS-PAGE. e After 698 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 28 protein separation, the dye-labelled proteins in the polyacrylamide gel are visualized using PISA with 699 a 647 nm excitation wavelength. Each detected single molecule observed in the PISA image 700 represents a single dye-labelled protein. Fluctuations in the number of single molecules are observed 701 along the migration path, which correlate with protein bands of different abundances. Subsequent 702 image analysis, including background subtraction, denoising filter, and spot detection, enables the 703 reconstruction of a detailed proteome profile for each single cell. Scale bar: 50 µm f Schematic 704 diagrams depicting different single-cell proteome analyses that can be achieved with single-cell 705 PAGE-PISA. In (a, c), the illustration of Eppendorf tube was created with BioRender.com. 706 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 29 Figure 2 | Single-cell PAGE-PISA of single U2OS and PC-3 cells. a Reconstructed single-cell 707 electropherogram of U2OS and PC-3 cells obtained by single-cell PAGE-PISA. The 708 electropherogram was reconstructed based on the standardized protein counts from 35‒128 kDa. b 709 Representative proteome profile of single U2OS and PC-3 cells. The y-axes represent the standardized 710 (left) and raw (right) protein counts from a single cell across molecular weights, respectively. c Total 711 protein count for each cell type. d Mean Pearson correlation of protein counts across molecular 712 weights between single cells, within and across cell types. Error bars represent ± the standard 713 deviation of the mean. e Comparisons of the mean protein count between single cells and technical 714 replicates of highly diluted bulk cell lysates equivalent to single-cell level. The red lines represent the 715 line of best fit, showing a positive correlation between single-cell and bulk measurement. f Two-716 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 30 dimensional UMAP projections based on electrophoretic protein band profiles from individual U2OS 717 (blue) and PC-3 (red) cells, shown as UMAP 1 vs. UMAP 2 and UMAP 1 vs. UMAP 3. Clusters (grey 718 region) were identified using k-means. Number of biological replicates (single cells): n = 10 for U2OS 719 and n = 10 for PC-3, number of technical replicates (bulk cell lysates): n = 3 for U2OS and n = 3 for 720 PC-3 cells. Statistical significance was determined by the one-way Welch’s ANOVA followed by the 721 Games-Howell post-hoc test. *p <= 0.05 722 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 31 Figure 3 | Application of single-cell PAGE-PISA to cardiomyocyte differentiation from human 723 induced pluripotent stem cells. a Schematic protocol of cardiomyocytes transitioning from 724 pluripotent state towards cardiac lineage. hiPSCs: human induced pluripotent stem cells; MD: 725 mesoderm; PC: progenitor cells; CM: cardiomyocytes; AA: ascorbic acid; HSA: human serum 726 albumin. The culture dish represents the day when hiPSC-CMs were transferred to a freshly coated 727 plate. Illustrations were created with BioRender.com. b Single-cell electropherogram of hiPSCs (D0) 728 and hiPSC-CMs (D16 and D30) were reconstructed based on the standardized protein counts from 729 35‒112 kDa. c Total protein count for each sampling day. d Mean Pearson correlation of protein 730 counts across molecular weights between single cells, within and across sampling days. Error bars 731 represent ± the standard deviation of the mean. Number of biological replicates: n = 12 for D0, n = 14 732 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 32 for D16, and n = 23 for D30. e Diffusion maps showing the developmental trajectory from 733 pluripotency to differentiated states. The single cells were coloured according to their sampling day 734 (left) and DPT (right). The trajectory starts from the root cell (blue) that is located at the left terminal 735 and progresses towards the most differentiated cell (yellow). DC and DPT refer to diffusion 736 components and diffusion pseudotime, respectively. f UMAP analysis with k-means clustering of the 737 DPT-assigned single cells. Three clusters were identified, representing different developmental stages, 738 precursor stage (n = 13), early cardiomyocytes (n = 23) and late cardiomyocytes (n = 13). g Average 739 single-cell proteome profile for each developmental stage. h Volcano plot showing upregulated and 740 downregulated protein bands in late cardiomyocytes compared to early cardiomyocytes. Differential 741 protein bands were determined by Welch's two-sided, two-sample t-test (p 1.2). 742 i Comparisons of the quantitative expression levels of two protein bands, 55.0‒57.7 kDa and 47.1‒743 48.0, across three developmental stages. The box plots show median values (central line), interquartile 744 range (box edges), and whiskers extending up to 1.5 times the interquartile range. Statistical 745 significance was determined by the one-way Welch’s ANOVA followed by the Games-Howell post-746 hoc test. Individual data points are overlaid. *p <= 0.05 747 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 33 Figure 4 | Comparative analysis between single-cell proteome and transcriptome during 748 cardiomyocyte differentiation. a UMAP projection of single cells based on the overall similarity of 749 gene or protein expression. b Diffusion maps showing single cells arranged along a pseudotemporal 750 trajectory based on the transition probabilities between two cells, positioning those with higher 751 transition probabilities closer to each other. c UMAP projection of single cells after DPT assignment. 752 d Correlation between sampling day and DPT. For scRNA-seq data, three RS were generated by 753 randomly selecting 49 single cells from the entire dataset. In total 10,484 genes were used for the 754 analysis, corresponding to proteins with molecular weights within 37‒126 kDa. DC refers to diffusion 755 components, DPT refers to diffusion pseudotime, and RS refers to random sample. Number of 756 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint 34 biological replicates for single-cell PAGE-PISA: n = 12, 14, and 23 for D0, D16, and D30, 757 respectively. Number of biological replicates for scRNA-seq: n = 12, 14, and 23 for D0, D15, and 758 D30, respectively. 759 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 15, 2026. ; https://doi.org/10.64898/2026.01.14.699423doi: bioRxiv preprint