High-throughput label-free single-cell proteomics enabled by multicolumn NanoLC

Abstract

Mass spectrometry (MS)-based single-cell proteomics (SCP) enables proteome-wide analysis at single-cell resolution, offering insights into cellular heterogeneity, biological processes, and disease mechanisms. However, conventional nanoLC-MS workflows are constrained by long gradients and low throughput, limiting their application to small cohorts. Here, we present a multicolumn nanoLC–MS platform that achieves 5-minute separation windows with 100% duty cycles at ~100 nL/min, enabling the analysis of up to 288 single cells per day. The system provides stable peptide separation, negligible carryover, and robust retention-time reproducibility across nearly 1200 consecutive injections. We successfully analyze more than 4000 samples at nearly 288 samples per day (SPD) throughput. The platform identifies ~5000 proteins per 250 pg digest injection and an average of ~4000 proteins per single HeLa cell with maxima exceeding 5100, which is comparable to state-of-the-art longer-gradient workflows. Quantitative benchmarking with mixed-species standards confirm the quantitative accuracy is largely maintained during extended acquisition. Applying the workflow to 783 LPS-treated RAW264.7 macrophages reveals substantial cellular heterogeneity under a uniform inflammatory stimulus. Together, these results establish a robust and scalable platform for high-throughput SCP, enabling thousands of single-cell analyses within a single study while maintaining deep proteome coverage and biological interpretability.

Single-cell proteomics reveals cellular diversity, but the throughput has been limited. Here, the authors develop a high-throughput multicolumn nanoLC system that analyzes up to 288 cells per day, identifies ~ 4,000 proteins per cell and profiles inflammatory responses in macrophages.

Introduction

Quantifying cellular and tissue heterogeneity at the single-cell level across large populations offers critical insights into the protein-level diversity that are inaccessible through bulk measurements, ultimately enabling improved diagnostics and more effective therapeutic strategies^1–8. Mass spectrometry-based approaches now enable profiling of protein expression at the single-cell level, and recent improvements in ion optics, detector speed, and acquisition strategies have substantially deepened coverage^9–23. These developments have made it possible to reveal new aspects of cellular heterogeneity^24–28.

Despite recent progress, throughput remains a critical bottleneck for SCP, as increasing analysis speed often comes at the expense of proteome coverage. To capture the full diversity of cellular states, thousands of single cells may need to be analyzed, and current LC-MS workflows (typically less than 100 samples per day²⁹ (SPD)) struggle with such large-scale studies. Multiplexed strategies using isobaric^30–32 or nonisobaric tags^33,34 have increased throughput by pooling labeled single cells, but ratio compression and isotopic interference often compromise quantitative accuracy^35,36, and proteome depth typically remains limited. The introduction of the TMTpro 35-plex set, for example, has enabled the analysis of up to 1018 single cells per day, even with only ~30 LC runs in that timeframe³⁷. However, the number of reported proteins quantified per cell was lower than that of current label-free approaches, averaging 712 proteins per cell. Label-free SCP can reach greater proteome coverage, but at the cost of substantially lower throughput. Conventional NanoLC-MS typically restricts analyses to only a few dozen single cells per day²⁹.

One throughput limitation for label-free SCP lies in the inefficiency of conventional LC workflows, where significant portions of the analysis cycle are spent on overhead tasks, such as sample pickup, column equilibration, gradient delay, and column cleanup that leave the mass spectrometer idle. Shortening gradients can reduce separation time, but overhead steps become proportionally more dominant, further lowering overall duty cycle. Increasing flow rates through microLC or capillary-based LC setups shortens run times but sacrifices the sensitivity required for single-cell-level inputs^38–40. Thus, a major challenge for the field is to achieve high throughput without compromising sensitivity.

Here, we present a multicolumn NanoLC-MS platform that achieves 5-min separation cycles at ~100 nL/min flow rate, supporting throughput of up to 288 single cells per day. By multiplexing two analytical columns with dedicated storage loops^41,42, the system maintains a continuous 100% MS duty cycle while requiring only minimal hardware modification.

We show that this design delivers negligible carryover, stable retention times across thousands of injections, and proteome coverage comparable to that of lower throughput analyses. Benchmarking against technical standards, mixed-species controls, and single-cell samples demonstrates that the high throughput does not compromise sensitivity or reproducibility. Finally, we illustrate the biological utility of the platform by capturing cell-to-cell heterogeneity in LPS-stimulated macrophages at single-cell resolution. Together, these results establish a practical and scalable solution to one of the most pressing challenges in SCP: enabling robust, high-throughput proteome profiling at single-cell resolution. By bridging the gap between sensitivity and speed, the 5-min multicolumn LC-MS system transforms SCP from a specialized method into a platform capable of addressing large-scale biological questions.

Results

Multicolumn LC-MS platform for high-throughput SCP

SCP has advanced rapidly in recent years, yet the throughput of LC-MS, particularly for label-free analyses, remains a limiting factor. In a conventional LC-MS setup, each sample undergoes sequential steps including sample pickup, column equilibration, loading onto a trap column, gradient elution from the trap to the analytical column, and column cleanup. Among these, only the gradient elution step generates productive mass spectra, and that may even include less informative spectra due to gradient delays. To address this inefficiency, we multiplexed these steps by using two analytical columns, and we avoided gradient delays through timing optimization and accelerating gradient generation via storage loops with minimal system complexity, adding only a single isocratic pump and one additional valve compared to standard systems⁴³.

Each sample is processed through the following parallelized steps: sample pickup, loading to the trap column, gradient elution from the trap column to a storage loop, and pushing from the storage loop to the analytical column for MS analysis (Supplementary Movies 1 and 2). Completing these processes within a 5-min cycle requires precise coordination of multiple operations (Supplementary Movie 3).

The separate steps are illustrated in the timetable (Fig. 1A), highlighting the individual stages of an LC-MS run with multiplexing implemented in two key areas. First, two analytical columns are operated alternately, each paired with a dedicated storage loop (as shown in the dual 6-port valve setup in Fig. 1B). While one column elutes sample to the MS using the pre-formed gradient from storage loop #1, the next gradient is generated and loaded into storage loop #2, and vice versa. As illustrated in Fig. 1C, the electrospray potential is applied to only one emitter at a time while the other is grounded, holding a droplet at its tip. This voltage control, managed by a simple adapter, ensures that regeneration or cleanup eluents from one column do not interfere with electrospray ionization on the other. Additional parallelization is applied to sample pickup, autosampler cleanup, trap column loading, and gradient elution to the storage loop (as illustrated in the 10-port valve in Fig. 1B). Since sample pickup and gradient elution share the same valve position, they can occur simultaneously. Likewise, autosampler cleanup and sample loading occur in parallel. In this case, during the first 4 min of each cycle, while the eluent of the previous sample transfers from the trap column to the storage loop, the autosampler picks up the next sample. After valve switching, the autosampler is cleaned while the newly drawn sample is loaded onto the trap column.

Fig. 1. Two-column NanoLC system overview.

A Minute-by-minute timing. B Detailed connections of the multicolumn system. C Photomicrograph of two emitters positioned in front of the inlet of a FAIMS Pro interface.

With these combined strategies, we achieve true 5-min analysis cycles with a 5-min separation window and maintain a near-100% MS duty cycle at a low flow rate of ~100 nL/min. One minor exception is that there is a brief (~21 s) wait period between two injections caused by the downloading process of the MS acquisition software. This may be remedied by collecting data for a number of samples in the same raw file and then splitting the file after collection. As shown in Table 1, we can obtain a chromatogram every 5 min after the 3rd 5 min segment once the operation of the system is fully parallelized. This configuration enables high-throughput analysis of up to 288 SPD for SCP or other low-input analyses. In contrast to alternative high-throughput solutions, such as Evosep, whose Whisper Zoom methods enable 20–120 SPD at higher flow rates, reduced sensitivity and lower duty cycles, the present system preserves ultra-low-flow operation that benefits single-cell proteomics. The ability to simultaneously achieve high throughput and high sensitivity positions this platform in a distinct performance regime within SCP and proteomics more broadly.

Table 1.

Timeline for the 5-min multicolumn NanoLC-MS platform

1st 5 min	2nd 5 min	3rd 5 min	4th 5 min
Sample pickup	Sample pickup	Sample pickup	Sample pickup
Loading to Trap	Loading to Trap	Loading to Trap	Loading to Trap
Gradient elution from Trap #1 to storage loop #1	Gradient elution from Trap #2 to storage loop #2	Gradient elution from Trap #1 to storage loop #1	Gradient elution from Trap #2 to storage loop #2
Pushing from storage #2 to column #2	Pushing from storage #1 to column #1	Pushing from storage #2 to column #2	Pushing from storage #1 to column #1
Pre-blank 1	Pre-blank 2	1st chromatogram	2nd chromatogram

Bold text highlights the processing steps for a single sample.

Separation performance characteristics and long-term robustness

For a 5-min separation with ~100% duty cycle, it is essential that the actual separation window fits entirely within this time frame. The separation window of the multicolumn system was evaluated using 250 pg human protein digest (single-cell level). By adjusting the accelerated gradient and the timing of the 6-port valve switch, we can ensure that the separation window is well aligned with the 5-min MS acquisition time. That is, the separation window is neither too narrow, which would underutilize the 5-min acquisition time, nor too wide to extend beyond it, which could result in sample elution loss. As shown in Fig. 2A, the peptide elution window is well distributed across the 5-min MS acquisition time, enabling the high throughput of ~288 samples per day with an actual 5-min analysis time per sample.

Fig. 2. Separation performance evaluation of the multicolumn LC system.

A Base peak chromatogram of 250 pg HeLa digests with a 5 min cycle time. B Base peak chromatograms of 385 pg HYE sample followed by three blank runs, indicating negligible column carryover. C Retention time reproducibility for 4 peptides over 1158 continuous analyses of 250 pg HYE mixed-species standards. Each point represents the retention time measured for a peptide in a single analysis (n = 1158 technical replicates). For each of the 4 peptides, elution times from one column are shown in orange, and those from the other column are shown in blue. Retention time changes during the analysis are driven primarily by temperature fluctuations in the lab space. The box plots show median retention times (25th–75th percentiles) for Column 1 and Column 2 were 1.305 min (1.198–1.403) and 1.316 min (1.205–1.419) for peptide 1; 2.391 min (2.296–2.507) and 2.398 min (2.283–2.504) for peptide 2; 3.246 min (3.143–3.370) and 3.217 min (3.099–3.341) for peptide 3; and 4.200 min (4.085–4.347) and 4.132 min (4.010–4.288) for peptide 4, respectively. Whiskers extend to the minimum and maximum values within 1.5 × IQR: 0.935–1.659 and 0.994–1.648 min for peptide 1; 2.093–2.737 and 2.083–2.759 min for peptide 2; 2.926–3.626 and 2.893–3.613 min for peptide 3; and 3.817–4.593 and 3.784–4.577 min for peptide 4, respectively for Column 1 and Column 2.

Sample carryover between injections is a critical concern in proteomics, as it can lead to false positives and quantitative inaccuracy. To address this potential concern, we evaluated the carryover of the system using mixed-species human/yeast/E. coli (HYE) samples containing 385 pg total protein. As shown in Fig. 2B, the average intensity of carryover runs was 0.61% compared to the sample. This low carryover is attributed to rigorous washing after the accelerated gradient for trap column cleanup (Supplementary Fig. 2) and the use of 80% ACN in the isocratic “pushing pump” for analytical column cleanup. The low carryover ensures minimal cross-contamination, making it suitable for high-throughput analysis of large sample cohorts.

We next assessed retention time stability over extended operation by analyzing 1158 mixed-species HYE samples without interruption and extracting the retention times for four sample peptides across all runs. As shown in Fig. 2C, for a single analytical column, the standard deviation of retention time ranged from 0.14 min (8.4 s) to 0.19 min (11.4 s). Some of this was due to temperature fluctuations in the room, as the columns were unheated. The RT standard deviation is maintained when the retention times for both channels are combined. The mean retention-time differences for the four representative peptides between Column 1 and Column 2 were −0.81 s, 0.32 s, 1.87 s, and 3.47 s, respectively. This demonstrates that the system offers adequate RT stability and high robustness, making it capable of continuously analyzing thousands of samples.

Column-to-column consistency supports pooled analysis

Since all experiments in this study were analyzed across two analytical columns, spanning multiple plates and acquisition days, it was essential to verify that technical variability did not mask or distort biologically meaningful differences. In addition, because isolated single cells in subsequent experiments were reconstituted using an autosampler prior to being injected into the LC, we also aimed to determine the additional technical variation introduced from the reconstituted sampling procedure relative to the standard sampling. To address both questions, we analyzed large-scale technical replicates generated from a 250 pg HeLa digest standard, a commonly accepted proxy for the approximate protein content of a single mammalian cell^9,44. We assessed whether proteome coverage and intensity-based variation were comparable across analytical columns and sampling approaches, and whether the collected LC/MS data could be fairly combined for downstream analysis.

As a baseline assessment, we analyzed 200 technical replicates of 250 pg HeLa digest standard prepared at a concentration of 250 pg/μL, and 1 μL was injected per run. There were 100 consecutive injections on each analytical column. The pooled dataset yielded a median coverage of 4984 protein groups, with a maximum of 5172 protein groups. Separately, Column 1 and Column 2 produced median coverage of 5017 and 4955 protein groups, respectively, corresponding to a difference of just 1.25%. After data filtration, the number of quantifiable proteins for the 250 pg HeLa digest standard is 4536, 4410, and 4466 for Column 1, Column 2, and the pooled dataset, respectively. The corresponding ratios of quantifiable to identified proteins were 90.4, 89.0, and 89.6%. (Fig. 3A).

Fig. 3. Protein coverage and quantitative variability across analytical columns and sampling approaches.

A 250 pg HeLa digest standard was analyzed using either a standard wet sampling approach (blue; n = 200 consecutive runs with 100 runs per column) or an online reconstitution approach (orange; n = 199 consecutive runs). A Protein groups identified per run are shown as a box plot for Column 1, Column 2, and the pooled dataset from both analytical columns. The median protein groups identified are listed above the box plot, and diamonds indicate the number of quantifiable proteins after filtering. The box plots show the median (25th–75th percentiles) of protein identifications: 250 pg HeLa Column 1, 5017 (4962–5078); 250 pg HeLa Column 2, 4955 (4870–4994); 250 pg HeLa Pool, 4984 (4919–5050); 250 pg HeLa Reconstituted Column 1, 4916 (4866–4975); 250 pg HeLa Reconstituted Column 2, 4884 (4815–4937); and 250 pg HeLa Reconstituted Pool, 4900 (4843–4956). Whiskers extend to the minimum and maximum values within 1.5 × IQR: 4789–5172, 4684–5130, 4723–5172, 4701–5118, 4633–5111, and 4673–5118, respectively. B Distribution of protein intensity CV (%) for each condition after data filtering. The y-axis displays the full observed range of CV values. The box plots inside the violin plots show the median CV% (25th–75th percentiles): 250 pg HeLa Column 1, 16.07% (9.50–25.86%); 250 pg HeLa Column 2, 17.20% (10.91–26.29%); 250 pg HeLa Pool, 17.05% (10.58–26.37%); 250 pg HeLa Reconstituted Column 1, 18.48% (12.99–26.80%); 250 pg HeLa Reconstituted Column 2, 18.92% (13.31–26.97%); and 250 pg HeLa Reconstituted Pool, 19.11% (13.42–27.21%). Whiskers extend to the minimum and maximum values within 1.5 × IQR: 3.81–50.40%, 5.13–49.36%, 4.60–50.06%, 5.80–47.35%, 6.58–47.29%, and 7.26–47.89%, respectively.

The median coefficient of variation (CV) of protein intensity was 17.0% across all runs (Fig. 3B), which is in line with measurements in other low-input studies^9,45. Together, the results indicate negligible column-specific bias and justify pooling MS runs from both analytical columns for subsequent analyses without introducing column-specific bias.

To evaluate the sampling approach used for isolated single cells, we next analyzed a reconstituted 250 pg HeLa digest standard, which was prepared by dispensing 0.20 µL of a 1.25 ng/μL HeLa digest standard into a 384-well PCR plate, evaporating to dryness, storing at −20 °C, and then reconstituting to a larger volume in the autosampler before aspirating into the sample loop. As shown in Fig. 3, proteome coverage is nearly identical for wet vs. reconstituted sampling, and CVs are only slightly greater when samples are reconstituted. This supported the use of reconstitution-based autosampling for the remainder of this study. Similar to the regular sampling dataset, the differences between analytical columns remained minimal for both proteome coverage and quantitative reproducibility.

Longitudinal system stability and proteome coverage of single HeLa cells

As mentioned above, the primary goal of this multicolumn high-throughput LC system is to enable the analysis of large-cohort of single cells. It was therefore important to determine whether prolonged operation and repeated analysis of prepared single-cell samples, which contain more contaminants than commercially available protein digest standard, affect system performance over time. We thus analyzed 1230 isolated single HeLa cells together with 138 injections of 250 pg HeLa digest standard that was periodically inserted throughout the acquisition sequence as a system quality control. These measurements were distributed across four sample plates and acquired over 6 days.

We applied a 1.5 × IQR filter, removing 169 single HeLa cells and 3 samples of 250 pg HeLa digest standard, representing 14% and 2% of the population, respectively. The complete run-order plot, including the filtered runs, is shown in Supplementary Fig. 4B. To visualize instrument performance over time, Fig. 4A plots the number of protein groups identified in each retained run as a function of acquisition order. As shown in the plot, the 250 pg HeLa Digest standard (green) is distributed within a narrow band with minimal fluctuation throughout the 6-day acquisition period. In contrast, the single HeLa (blue) shows a substantially broader distribution in proteome coverage due to cell size differences, etc., but the overall distribution remains consistent across the entire acquisition sequence without evidence of temporal drift. The boxplots shown in Fig. 4B further quantify these distributions. The consistent results and low failure rate for both single cells and standard across 6 days of continuous acquisition demonstrate stable system performance.

Fig. 4. Proteome coverage throughout a large-cohort single-cell proteomics analysis.

A Number of protein groups identified per run as a function of acquisition order for isolated single HeLa cells (blue, n = 1061 biological observations from 1 biological replicate) and periodically injected 250 pg HeLa digest standard as quality control (QC) samples (green, n = 138 technical replicates) acquired over 6 days. B Distribution of protein groups identified for single HeLa cells and 250 pg HeLa digest QC datasets. The box plots show the median protein identifications (25th–75th percentiles) were 5079 (5019–5138.75) for 250 pg HeLa digest QC and 4038 (3686.5–4309.5) for single HeLa cells. Whiskers extend to the minimum and maximum values within 1.5 × IQR: 4853–5264 and 2757–5167, respectively.

To place these results in context, we compared them with two studies published in 2025 that also highlighted the high sensitivity of single-cell proteomics using the Orbitrap Astral mass spectrometer. Those studies reported an average of 5102 proteins from 250 pg of HeLa digest standard with an 80 SPD method, 4879 and 3166 proteins from A549 and H460 single cells with a 50 SPD method⁹, and 4567 proteins from HeLa single cells with a 120 SPD method²². The search parameters used for these comparison datasets are provided in Supplementary Data 1. Despite operating at 288 SPD, our platform achieved comparable proteome coverage for both standard and single-cell samples while providing approximately 2–5-fold greater sample throughput.

Quantitative performance with mixed-species standards

Having demonstrated the sensitivity, reproducibility, and robustness of the 5-min multicolumn high-throughput LC-MS system for large-cohort SCP experiments, we next evaluated its quantitative accuracy. We used standards composed of HYE peptides mixed at defined ratios^46,47, which is widely used for benchmarking quantitative accuracy in proteomics because it provides a ground-truth comparison of relative abundances. We analyzed HYE-A (H:Y:E = 65:30:5) and HYE-B (H:Y:E = 65:15:20), with the human component maintained at 250 pg in both mixtures (Supplementary Fig. 4A), and compared the quantitative accuracy between an extended-acquisition dataset (Fig. 5A) comprising 578 interspersed HYE-A/HYE-B pairs acquired over 4 days (1156 LC-MS runs). We also evaluated a short-term reference dataset, comprising 4 consecutively acquired HYE-A/HYE-B pairs (8 LC-MS runs) (Fig. 5B).

Fig. 5. Quantitative performance evaluated using HYE mixed-species standards.

The extended-acquisition dataset and short-term reference dataset were generated in this study using a 288-SPD (samples per day) LC system coupled to an Orbitrap Astral mass spectrometer. The published Astral dataset from Bubis et al.¹¹ was acquired using a 50-SPD LC method coupled to an Orbitrap Astral mass spectrometer. Quantitative performance of the extended-acquisition dataset (A), short-term reference dataset (B), and published dataset from Bubis et al. (C). The top panels show local CV distributions of protein intensity across the three species after data filtration. Lower left panels show protein-level log2 fold changes plotted against protein abundance. The number of quantifiable shared proteins between HYE A and HYE B after filtering is labeled. Lower right panels show KDE distributions of protein-level log2 fold changes. In the lower left panels, solid lines indicate the measured mean log2 fold change and dashed lines indicate the expected log2 fold change. The standard deviation (s.d.) of each KDE distribution is indicated in the corresponding lower right panel.

For the extended acquisition dataset, the number of quantifiable protein groups shared between HYE-A and HYE-B was 3574 for human, 800 for yeast, and 226 for E. coli after filtering (Fig. 5A-left). In terms of quantification reproducibility, the median CV was 17.16%, while the local CV of protein intensity, calculated using rolling averages of 100 proteins across species, ranged from 9.39 to 30.98%, with higher-abundance proteins generally exhibiting lower CV values (Fig. 5A-top). To assess quantitative accuracy, we compared the measured fold changes against the expected log₂ ratios of the HYE standard (Human: Yeast: E. coli = 0:1:−2). The measured mean log₂ ratios were −0.13, 0.87, and −1.97, corresponding to absolute errors of 0.13, 0.13, and 0.03 for human, yeast, and E. coli, respectively (Fig. 5A-left). Kernel Density Estimate (KDE) plots of protein-level fold changes showed distributions centered near the expected ratios for all three species (Fig. 5D-right). These results from the extended-acquisition dataset represent quantitative performance measured across 4 days of continuous operation and demonstrate that quantitative accuracy is largely maintained over this timeframe. The observed errors are within the range typically reported for label-free DIA analyses^48,49, supporting the suitability of the platform for large-cohort quantitative proteomics studies.

The short-term reference dataset represents the quantitative performance of the multicolumn LC-MS platform along. After filtering, the number of quantifiable protein groups shared between the two mixtures was 3525 for human, 848 for yeast, and 560 for E. coli (Fig. 5B-left). The median CV is 11.19%, and the local CV ranged from 4.06% to 24.07% (Fig. 5B-top), while the absolute errors from the expected fold changes were 0.05, 0.03, and 0.05 for human, yeast, and E. coli (Fig. 5B-left), respectively. KDE plots again showed protein-level fold-change distributions centered near the expected values (Fig. 5B-right). Compared with the extended-acquisition dataset, the short-term reference dataset showed lower local CVs and absolute errors, indicating better run-to-run reproducibility and quantitative accuracy. Although the extended-acquisition dataset exhibited apparently narrower protein-level fold-change distributions, this likely reflects averaging across substantially more replicate measurements rather than greater run-to-run precision. These results suggest that prolonged acquisition contributes additional run-to-run variation, which can reduce quantitative reproducibility and accuracy over time. Further reductions in variability may be achievable through additional normalization and batch-correction strategies, although such approaches were not investigated here.

To place these results in context, we compared them with a recently published Astral Orbitrap dataset generated by Bubis et al.⁹. Raw files from that study were downloaded and processed using the same identification and post-search workflow applied to the present study. The study used human and yeast peptide mixtures with a total protein input of 250 pg per sample. Sample A contained H:Y at 60:40, while Sample B contained H:Y at 80:20, corresponding to expected log₂ ratios of −0.42 and 1 for human and yeast, respectively (Supplementary Fig. 4A). After filtering, the number of quantifiable protein groups shared between the two mixtures was 4,983 for human and 1862 for yeast (Fig. 5C-left). The median CV is 8.36%, and local CV values ranged from 5.34% to 17.87% (Fig. 5C-top), and the absolute errors from the expected fold changes were 0.12 for human and 0.07 for yeast (Fig. 5C-left). KDE plots showed fold-change distributions centered near the expected values (Fig. 5C-right). Compared with the Bubis et al. dataset, our extended-acquisition dataset and the short-term reference dataset both showed lower proteome coverage and modestly higher run-to-run variability, as reflected by protein-intensity CVs and the distributions of identified protein groups (Figs. 5A–C and S4B), which is expected given the substantially shorter gradient and higher throughput of the 5-min workflow (288 SPD versus 50 SPD), as well as the added complexity from the inclusion of a third species. Still, despite spanning more than 1100 LC-MS runs over 4 days of continuous operation, measured fold changes remained in close agreement with expected values. Furthermore, comparison with a published Astral dataset processed using the same workflow showed that the substantially increased throughput of our platform is achieved with only modest reductions in proteome coverage and precision while maintaining competitive quantitative accuracy. These findings establish that the platform can support large-cohort single-cell proteomics studies without sacrificing the quantitative performance required for biological interpretation.

Resolving single-cell heterogeneity in LPS-stimulated macrophages

We next assessed the utility of the high-throughput system in resolving biological changes. RAW 264.7 cells are a murine macrophage cell line commonly used to study innate immune responses. Stimulation with lipopolysaccharide (LPS), a component of the outer membrane of Gram-negative bacteria that activates Toll-like receptor 4 (TLR4), induces inflammatory response and provides a well-established model for investigating cellular heterogeneity in response to immune stimulation. To demonstrate the capability of the platform in resolving single-cell heterogeneity, we analyzed more than 1000 individual RAW 264.7 cells treated with either PBS (control) or 100 ng/mL LPS for 24 h prior to analysis. Before cell sorting, the post-wash cell size and viability measured 9.4 μm and 92% for PBS-treated cells, and 12.1 μm and 81% for LPS-treated cells, respectively. After removing 72 failed runs (6.7%, typically due to sample preparation or autosampler failure), a total of 994 single cells were retained for downstream analysis, including 211 PBS-treated cells and 783 LPS-treated cells.

UMAP dimensionality reduction was performed using the global protein expression matrix. Three distinct clusters were observed (Fig. 6A). Approximately 93% of PBS-treated cells localized into one cluster (labeled PBS), while the LPS-treated cells separated into two distinct populations: 66% in the LPS-1 cluster and 33% in the LPS-2 cluster. When the UMAP was colored by sorting gate, analytical column, and batch (Fig. 6B–D), all technical conditions were evenly distributed across the three clusters. These results suggest that the clustering pattern is not primarily driven by technical issues, but instead reflects biologically distinct states induced by LPS treatment.

Fig. 6. Characterization of PBS- and LPS-treated RAW 264.7 macrophages.

A UMAP projection of 974 single RAW 264.7 cells colored by treatment condition. The LPS-treated cells are separated into two clusters, LPS-1 and LPS-2. B–D UMAP projection of the same cells shown in (A), colored by single-cell isolation gate size (B), analytical column used for separation (C) runs order (D), and treatment (E). F Distribution of identifiable protein groups per cell for the PBS (n = 200 biological observations from 1 biological replicate), LPS-1 (n = 526 biological observations from 1 biological replicate), and LPS-2 (n = 268 biological observations from 1 biological replicate) clusters. The box plots show the median protein identifications (25th–75th percentiles) were 3166 (2890–3498) for PBS, 3166 (2929–3415) for LPS-1, and 1765 (1554–2043) for LPS-2. Whiskers extend to the minimum and maximum values within 1.5 × IQR: 2113–4410, 2201–4142, and 1108–2776, respectively. G Distribution of protein intensity coefficient of variation (CV%) after data filtering for the PBS, LPS-1, and LPS-2 clusters. The box plot inside the violin plots shows the median CV% (25th–75th percentiles) was 28.38% (21.80–35.10%) for PBS, 29.80% (23.28–36.89%) for LPS1, and 40.06% (32.02–52.48%) for LPS2. Whiskers extend to the minimum and maximum values within 1.5 × IQR: 10.08–55.06%, 11.74–57.30%, and 17.93–83.18%, respectively. H Venn diagram showing overlap of identified proteins among the PBS, LPS-1, and LPS-2 clusters after filtering.

To further characterize these populations, Leiden clustering was used to formally define cluster membership, and 3 clusters were identified (Fig. 6E). These clusters closely corresponded to the three populations observed in the treatment-colored UMAP. Individual cells were grouped according to their Leiden cluster assignments for downstream analysis. The median number of identifiable proteins per cell in the PBS, LPS-1, and LPS-2 clusters was 3166, 3166, and 1765, respectively (Fig. 6F). After missing-value removal and IQR filtering, the number of quantifiable proteins were 2458, 2402, and 1215, respectively (Fig. 6F). The median protein intensity CVs were 28.4%, 29.8%, and 40.1% for the PBS, LPS-1, and LPS-2 clusters, respectively (Fig. 6G). Compared with PBS and LPS-1, which showed similar proteome coverage and quantitative variation, the LPS-2 cluster exhibited 44% lower proteome coverage and 38% higher intensity variation.

Although the number of identified proteins in LPS-2 was substantially lower than the other two clusters, Venn analysis showed that 95% of the proteins identified in LPS-2 were also detected in at least one of the other clusters (Fig. 6H), suggesting that the clustering reflects differences in protein abundance and cellular regulation across mostly shared proteins.

To further understand the biological differences between clusters, proteins from each comparison pair were ranked by log₂ fold change and then subjected to Gene Set Enrichment Analysis (GSEA) using WebGestalt 2024^50–54. To simplify interpretation, we separated the resulting biological process terms by positive and negative enrichment, manually grouped into broader cellular processes, and visualized by their relative density within each comparison group (Fig. 7A).

Fig. 7. Proteome-level differences between the three macrophage populations.

A Summary of Gene Set Enrichment Analysis (GSEA) results from the proteins shared between each of the comparison groups. The resulting biological process terms were manually grouped into broader cellular processes. Cell color intensity reflects the number of enriched terms within each category. UMAP projections colored by the intensity of representative proteins associated with DNA replication (B, Pcna), immune/inflammatory (C–E, Ifi204, Stat3, Ncf2), proteostasis (F, Ube2n), mitochondria/oxidative (G, Scmc1), and Cytoskeleton (H, Anxa7) processes.

The GSEA analysis showed that LPS-1 was strongly positively enriched in immune and inflammatory processes and negatively enriched in cell cycle and DNA replication pathways compared with PBS. Cytoskeletal, membrane-trafficking, and endocytic processes were also positively enriched in LPS-1, consistent with cellular remodeling accompanying macrophage activation. In contrast, LPS-2 showed no significant positive enrichment of immune and inflammatory pathways, but strong negative enrichment of proteostasis, translational, and cytoskeleton processes together with enrichment of mitochondrial and oxidative pathways. These results suggest that the two LPS-derived clusters represent biologically distinct states rather than differences driven solely by cell size or protein identification depth.

To demonstrate how the pathway-level differences appear across individual cells, representative proteins from the DNA replication (Fig. 7B), immune/inflammation (Fig. 7C–E), proteostasis (Fig. 7F), mitochondria/oxidative (Fig. 7G), and cytoskeleton (Fig. 7H) categories were visualized by coloring the UMAP distribution according to single-cell protein intensity. Immune-associated protein Ifi204 showed strong enrichment in LPS-1, consistent with the inflammatory activation pattern observed in the GSEA analysis^55,56 (Fig. 7B). In contrast, the proteostasis-associated protein Ube2n showed lower intensity in the LPS-2^57,58 (Fig. 7F), while the mitochondria/oxidative-associated protein Scmc1 showed enrichment in LPS-2^59,60 (Fig. 7G). These single-protein intensity distributions further support that the two LPS-derived populations represent distinct biological response states rather than technical separation artifacts.

The strongly regulated proteins in LPS-1 revealed a classical macrophage activation signature driven by Tlr4 signaling. Upregulated proteins included major inflammatory and signaling regulators (Rel⁶¹, Hck⁶², and Stat3⁶¹), phagocytic and antimicrobial effectors (Ncf2⁶³, Irgm1⁶⁴), and antigen presentation machinery (Tap1⁶⁴). In addition, a dense group of interferon-stimulated genes (Isg15⁶⁵, Ifih1^65–74, Slfn5⁶⁸, Ifi204^55,56, and Helz2⁶⁹) suggested activation of a secondary type-I interferon feedback response (Supplementary Data 2). Notably, 77% of these immune-associated biomarkers were not identified in LPS-2, further supporting that LPS-1 represents the primary immune-activated macrophage population.

In contrast, LPS-2 showed lower protein identification and higher cell-to-cell variation compared with PBS and LPS-1. Based on the lower post-treatment viability together with the observed suppression of proteostasis and translational pathways, enrichment of mitochondrial and oxidative pathways, and reduced immune activation signature, we speculate that LPS-2 may represent a stressed or dying cell population induced by LPS treatment. However, because the fraction of LPS-2 cells substantially exceeds the measured dead-cell fraction prior to sorting, the cluster may also include cells undergoing dysfunctional or incomplete inflammatory activation rather than only fully non-viable cells. Follow-up experiments incorporating dead-cell exclusion strategies or viability markers may help further distinguish immune-active and stress-associated populations under different treatment conditions.

The differential protein analysis presented here is intended primarily as an exploratory characterization of cluster-specific biology. Nevertheless, the combined UMAP clustering, GSEA, and representative single-protein intensity patterns consistently support the presence of at least two biologically distinct response states within the LPS-treated macrophage population: an immune-activated state (LPS-1) and a stress-associated state (LPS-2). Importantly, these heterogeneous cellular responses were resolved from nearly one thousand individual cells analyzed using the multicolumn 288-SPD LC-MS workflow. Together, these results demonstrate that the platform maintains sufficient proteomic depth, quantitative consistency, and throughput to reveal biologically meaningful single-cell heterogeneity at scale.

Discussion

This study establishes a multicolumn LC-MS system for high-throughput SCP, overcoming a fundamental bottleneck in current SCP workflows. By implementing the LC system with 5-min duty cycle, we demonstrate that high-throughput SCP profiling is achievable for up to 288 single cells per day with minimal compromise in sensitivity or reproducibility. This throughput provides a substantial improvement over prior workflows, overcoming the long analysis times that have historically limited the scale of SCP studies.

Qualitative and quantitative assessments across technical standards, mixed-species benchmarks, and single-cell samples show that the platform achieves deep proteome coverage, up to 5172 proteins per single HeLa cell, while maintaining stringent retention-time stability, minimal carryover, and reproducibility across thousands of samples. Compared with recent Astral-based workflows using 40–120 SPD methods, our system delivers only slightly reduced proteome depth at 2–5-fold greater sample throughput, reshaping expectations for the balance between sensitivity and speed in single-cell proteomics.

The platform also enables resolution of cellular heterogeneity at the single-cell level. Specifically, we profiled 783 individual RAW264.7 macrophages cells exposed to LPS (100 ng/mL, 24 h), enabling characterization of cellular heterogeneity within a uniform inflammatory stimulus and revealing distinct proteomic response states following LPS stimulation, including a major immune-activated population and a stress-associated population. These results demonstrate that the combination of high throughput, proteomic depth, and quantitative reproducibility provided by the multicolumn workflow is sufficient to detect complex cellular responses in a realistic biological system.

Together, these advances directly address two of the most pressing needs in single-cell proteomics: throughput and robustness. The ability to analyze 288 single cells per day with minimal carryover and stable retention times enables large-scale studies of cellular heterogeneity, perturbation responses, and patient cohorts that were previously impractical. Because the system requires minimal hardware modifications (an additional pump and valve assembly), it can be readily adopted on existing low-flow LC setups, lowering the barrier for broad implementation.

In summary, the high-throughput multicolumn LC system is a valuable contribution to facilitate the transition of the SCP field from method development to the biological toolbox. By enabling routine analysis of 288 single cells per day while maintaining high depth, this platform increases the scale at which biological and clinical questions can be addressed, opening the door to proteome-wide investigations in single-cell atlases and translational studies. To date, we have analyzed more than 10,000 samples using this platform for multiple projects, and more than 4000 samples are shown in this manuscript, and with continued advances in mass spectrometer performance, throughput of 500–1000 cells per day at comparable depth appears to be within reach.

Methods

Sample preparations

Cell culture and treatment

HeLa cells were obtained from ATCC (CCL-2) and cultured in 10-cm tissue culture plates using 10 mL of 1× DMEM (Corning, 10-013-CV). Media were supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin. All plates were incubated at 37 °C in a humidified incubator with 5% CO₂. Once HeLa cells reached confluency, they were rinsed with 5 mL of 1× phosphate-buffered saline (PBS), detached using 4 mL of 0.25% trypsin-EDTA (Gibco, 25200-072), and incubated at 37 °C for 4 min. Trypsin activity was quenched by the addition of 10 mL of complete medium.

RAW264.7 cells were obtained from ATCC (TIB-71) and cultured in 10-cm plates using complete Corning DMEM supplemented with 10% FBS, 1% penicillin-streptomycin, and 2 mM Gibco GlutaMAX. When cells reached approximately 60–70% confluency, they were washed once with PBS, and fresh complete medium was added. Lipopolysaccharide (LPS; Escherichia coli O111:B4, Sigma-Aldrich, L5293) was added to achieve a final concentration of 100 ng/mL by adding 50 µL of a 20 μg/mL LPS stock solution. For PBS-treated samples (control), an equivalent volume (50 µL) of PBS was added instead of LPS. All plates were incubated at 37 °C in a humidified incubator with 5% CO₂ for 24 h, then collected as described for HeLa cells above.

Cell wash

All cell types were pelleted separately by centrifugation at 200 × g for 10 min, resuspended in 2 mL of 1× PBS, and passed through a 40 µm filter. The cells were washed further by subsequent repelleting at 200 × g for 5 min, resuspending in 1 mL of 1× PBS and repeating these steps to remove residual media and enzymes. Cell viability and diameter were assessed using Trypan Blue Stain (Gibco, Ref# 15250-061) and the Logos Luna II Automated Cell Counter. All cell types were adjusted to a final concentration of 2 × 10⁵ cells/mL before isolation.

Single-cell isolation

The single-cell sample preparation followed the protocol published previously^70–72. In brief, single cells were dispensed into 0.01% DDM-treated 384-well PCR plates using the Tecan UNO dispenser (Männedorf, Switzerland). During dispensing, size-based gating was applied with a diameter range of 15–17 µm for HeLa cells and 9–11 µm for control RAW264.7 cells. For LPS-treated RAW 264.7 cells, a 9–11 µm and 12–14 µm diameter range was used for isolation. HeLa cells were dispensed onto a separate set of 384-well plates from the RAW 264.7 experiments. For RAW 264.7 cells, either control or LPS-treated cells were isolated onto the same plates, or LPS-treated cells from different diameter ranges were isolated onto separate plates. Plates were centrifuged at 380 × g immediately following dispensing to ensure cells were driven to the bottom of each well. Single cells were randomized across conditions and analyzed to minimize batch effects over an extended period.

Digestion

A one-pot workflow was applied to SCP preparation to avoid sample loss during multiple transfer steps^70–74. Immediately after isolation of all cell types, wells containing single cells were subjected to in-well proteolytic digestion⁷². A 200 nL volume of digest solution was dispensed into each well using the Tecan UNO dispenser. Each aliquot consisted of 10 nL of 0.2 µg/µL Rapid Trypsin/Lys-C (Promega, Cat# V507C) in Resuspension Buffer (Promega, Cat# V181A), 10 nL of 1% DDM in Rapid Digest Buffer (Promega, Cat# VA106A) and 180 nL of Rapid Digest Buffer, resulting in final concentrations of 10 ng/µL Trypsin/Lys-C and 0.05% DDM. Plates were centrifuged, sealed with a silicone mat, wrapped in aluminum foil, and incubated in a humidified oven at 70 °C for 1 h. After digestion, plates were cooled to 4 °C for 15 min, sealed with aluminum sealing foil, and stored at −20 °C until LC-MS/MS analysis.

Protein digest standard

HeLa digest standard was prepared using Pierce™ HeLa Protein Digest Standard (ThermoFisher, Ref# 88329). Dilutions were prepared at 10 ng/µL for experimental spectral library creation, and at 250 pg/µL with 0.1% formic acid for system evaluation. The HeLa Digest standard was also prepared at 1.25 ng/µL and dispensed with the D100 HP dispenser at 200 nL per well to DDM-treated 384-well plates for the evaluation of sample recovery following reconstitution.

The HYE mixture was prepared using the same HeLa digest standard, yeast protein extract (Promega, Cat# V7461), and MassPREP™ E. coli Digest Standard (Waters, Part#186003196). Two mixtures were prepared: HYE A and HYE B. For HYE A, 100 µL each of the three components: HeLa, yeast, and E. coli digests were combined, yielding final concentrations of 20, 9.23, and 1.54 ng/µL, respectively. HYE B was prepared similarly, with final respective concentrations of 20, 4.62, and 6.15 ng/µL. The solution was diluted to achieve a final HeLa component concentration of 10 ng/µL for the creation experimental spectral library and 250 pg/µL for system evaluation.

384-well plate treatment

384-well PCR plates (Eppendorf, Cat# 951020702) were treated with 0.01% n-dodecyl β-D-maltoside (DDM; Sigma-Aldrich, Cat# D4641-1G) in Milli-Q water by dispensing 20 µL into each well using a multichannel micropipette. Plates were centrifuged to ensure DDM coated the well bottoms and were incubated at room temperature for 30 min. The DDM solution was then removed by flicking, followed by inverting the plates and centrifuging at 380 × g to remove residual liquid.

Multicolumn NanoLC setup

The multicolumn LC system was constructed as described previously⁴³ with the following modifications. The trap columns were 100 μm i.d. × 35 mm long and packed with 3.5 µm media having 100 Å pores. Analytical columns with integrated emitter (30 μm i.d. × 50 mm long, 1.7 µm diameter, 100 Å C18 particles with the orifice of ~8 μm) were purchased from MicrOmics Technologies, LLC (Spanish Fork, UT). Mobile phases A and B were LC-MS-grade water with 0.1% formic acid and acetonitrile with 0.1 % formic acid (Honeywell, Charlotte, NC). The minute-by-minute multicolumn NanoLC system setup is shown in Supplementary Fig. 1. Samples were first aspirated into the sample loop (a and b) followed by loading onto the trap column at 5 μL/min for ~1 min (c), and then samples were eluted to the storage loop from the trap column (d). The eluent, including the sample, gradient, and the column equilibration solvent A, were stored inside the storage loop. After switching the 6-port valve, an additional isocratic pushing pump containing 0.1% formic acid in 80% ACN served to push the eluents of the storage loop to the analytical column at the flow rate of ∼100  nL/min (d and e). After that, the pushing pump was directly connected to the analytical column for cleanup (f). Each cycle consists of 5 min of operating time plus a 20 s Xcalibur response time.

Mass spectrometry

All analyses were performed using an Orbitrap Astral mass spectrometer (Thermo Fisher Scientific, San Jose, CA). FAIMS was applied with the carrier gas flow set at 3.5 L/min and CV set at −45 V. The ion transfer tube temperature was set as 200 °C. Full scan was applied with m/z ranging from 400 to 800 at 240 k resolution. Injection time was 100 ms and the AGC target was 500%. RF lens was 45%. Precursor scan range for MS/MS was the same as the full scan range. For single-cell level samples, the MS/MS scan range was set as 150–2000 with DIA isolation window set to 8 Th and generate 50 scan events. For the library of bulk samples, the MS/MS scan range was the same, but the DIA isolation window was set to 4 Th and generate 100 scan events. Normalized HCD collision energy was 25% with AGC set to 800% and maximum injection time set to 12 ms.

Data analysis and bioinformatics

Near-real-time monitoring

Individual MS runs were searched using DIA-NN 2.2.0 Academia through MSConnect⁷⁵ with a Human UniProt database downloaded in December 2024 for both HeLa standard and single-cell dataset; human, yeast, and E. coli UniProt protein databases downloaded in December 2024 were combined into a single search database for the HYE dataset; and the Mouse UniProt database downloaded in March 2026 was used for the RAW264.7 dataset. An experimental spectral library and three Proteomic Data Processors were employed to match the search time to a 5-min LC-MS cycle time for near-real-time monitoring.

The experimental spectral libraries were created using 10 ng peptide standards or bulk-digested RAW 264.7 cells, which were analyzed on the same LC-MS system as the sample runs with the corresponding predicted spectral library. The following parameters were the same for predicted spectral library creation, experimental spectral library creation, and the data search. Trypsin/P was specified as the digestion enzyme, with a peptide length range of 7–54 amino acids and up to three missed cleavages. A maximum of three modifications were allowed per peptide, including N-terminal methionine excision, carbamidomethylation, N-terminal acetylation, and methionine oxidation. The false discovery rate for the precursor was controlled at 1% using DIA-NN. Quantification was performed by DIA-NN’s default at the MS2 level, and cross-run normalization was disabled.

Identification and quantification

While near-real-time monitoring was performed during acquisition using the workflow described above, all reported identifications and quantifications in this study were generated from a final DIA-NN search performed after completion of each experiment. In this final search, all MS runs from the same experiment were analyzed together against the corresponding experimental spectral library using DIA-NN 2.2.0 Academia, with the same search parameters as those used for near-real-time monitoring. For the column-to-column comparison (Fig. 3), single-cell HeLa analysis (Fig. 4), and HYE mixed-species standards (Fig. 5), searches were performed without MBR and without inclusion of the 10 ng standard runs to preserve direct assessment of run-to-run system variation. In contrast, the RAW264.7 dataset (Figs. 6 and 7) was searched with MBR and 10 ng bulk-digested RAW264.7 reference runs to maximize proteome coverage for biological interpretation.

Post-search data processing

Normalization and data filtration:

Downstream data analysis and visualization were performed using in-house Python scripts. Individual MS runs with fewer than 1000 identified protein groups after DIA-NN search were considered failed runs and excluded from downstream analyses. This threshold was used to remove obvious technical failures, including sample preparation and autosampler-related issues, that produced substantially fewer protein identifications than successful runs. The only exception was the large-cohort single-cell HeLa dataset used for system stability assessment in Fig. 4, where run filtering was performed using a 1.5× interquartile range (IQR) filter on the number of identified protein groups. Because the objective of this analysis was to evaluate the distribution of proteome coverage over time rather than perform downstream CV or fold-change calculations, the IQR-based approach was used to remove extreme outliers while preserving the overall population distribution.

The workflows described below were applied to all datasets unless otherwise noted. Protein identifications were determined by counting protein groups with intensity values >0. For CV calculations, proteins with more than 33% missing values within the same experimental group were excluded. Raw DIA-NN protein intensities were then median-normalized within each group prior to CV calculation. The CVs reported throughout this study were calculated after applying a 1.5× IQR filter on protein-level CV values to remove proteins exhibiting extreme variation. Proteins remaining after missing-value filtering and CV-based IQR filtering are referred to as quantifiable proteins throughout this study. Pre-IQR filtering statistics for all datasets are provided in Supplementary Data 3. Across all datasets, >95% of proteins were retained after IQR filtering, indicating minimal impact on protein selection.

HYE dataset (Fig. 5):

Raw DIA-NN intensities were median-normalized within each HYE-A or HYE-B group across all three species. This normalization strategy was used to correct for run-level technical variation while preserving the known relative abundance differences between the HYE-A and HYE-B groups. Protein intensities were then averaged within each mixture, and protein-level CV values were calculated. A 1.5× IQR filter on CV values across all three species was applied to remove proteins exhibiting extreme variation. The remaining proteins were used for all subsequent analyses, including quantifiable protein counts, fold-change calculations, and local CV distributions. The local CV distribution values were calculated from abundance-ranked proteins using a moving window of 100 proteins. Fold changes were subsequently calculated using the filtered protein intensities shared between HYE-A and HYE-B. Because the objective of this analysis was to assess agreement with known ground-truth ratios, fold changes were calculated from the filtered technical replicate data without additional normalization across the two HYE mixtures. KDE plots were generated using protein-level log₂ fold changes.

RAW 264.7 single-cell dataset (Figs. 6 and 7):

UMAP visualization:

Runs have <1000 proteins identified were removed before UMAP dimensionality reduction was performed using proteins retained after missing-value and CV-based IQR filtering. Missing values were imputed using a Missing-Not-At-Random (MNAR) approach with a fixed random seed of 42 after raw protein intensities were log₂-transformed and median-normalized across all cells. The resulting imputed, median-normalized log₂ intensity matrix was transposed into a sample-by-protein matrix and used as input for UMAP dimensionality reduction with two output dimensions. UMAP projections were colored by treatment condition, single-cell isolation gate range, analytical column, sample preparation batch (Fig. 6A–E), or the processed intensity of individual proteins (Fig. 7B–H).

Leiden clustering:

Leiden clustering was performed in Scanpy using PCA scores generated from the protein intensity matrix processed as described above for UMAP analysis. The first 20 principal components were used to construct a k-nearest-neighbor graph with k = 15. Leiden community detection was then applied to the resulting undirected graph with a resolution parameter of 0.30 and a random seed of 42. Clusters were subsequently assigned biological labels (PBS, LPS-1, and LPS-2) based on their treatment composition and relative position in the UMAP projection. These cluster assignments were subsequently used for protein identification, CV analysis, Venn diagram generation, fold-change calculations, and Gene Set Enrichment Analysis (GSEA).

Venn diagrams and fold change:

Venn diagrams (Fig. 6H) were generated using proteins retained after CV-based IQR filtering. For pairwise cluster comparisons (Fig. 7A and Supplementary Data 2), protein intensities were log₂-transformed, median-normalized across the two groups being compared, and imputed using the MNAR method with a fixed random seed of 42. Protein-level fold changes were then calculated from the processed intensity values.

Gene Set Enrichment Analysis (GSEA):

For each set of fold change comparisons (LPS-1 vs. PBS, LPS-2 vs. PBS, and LPS-1 vs. LPS-2), all shared proteins and their log₂ fold change were submitted to Webgestalt 2024^50–54. Mus musculus is selected as the species, and Biological Process noRedundant is used as the functional database. To condense the output into a more interpretable figure, the resulting biological process term entries were categorized into nine overarching biological themes (Fig. 7A). All assignments were individually reviewed and assigned by the authors using the QuickGO browser⁷⁴. Biological term placement was verified by examining term definitions and their hierarchical parent-child relationships, with parent terms used when available to assign a biological theme. Biological terms that could not be confidently assigned through QuickGO were further evaluated using published literature describing the associated biological process in the context of its assigned biological theme. Biological theme counts were then calculated for each of the six groups to generate Fig. 7A. The complete fold-change results, GSEA outputs, assigned biological-theme categories, and the supporting source used for each biological term category assignment are provided in Supplementary Data 2. Representative proteins visualized on UMAP projections (Fig. 7B–H) were selected from biologically relevant categories identified through GSEA enrichment analysis.

Statistical test

Statistical significance testing was not performed for the HYE benchmark analyses because the objective was to evaluate quantitative accuracy relative to known ground-truth ratios. For the RAW 264.7 dataset, differential protein analysis was intended as an exploratory characterization of cluster-specific biology based on fold-change and pathway-level trends rather than formal differential-expression testing.

Reporting summary

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

Author contributions

C.W. and R.T.K. conceived the study. C.W., H.L.L., and R.T.K. designed the experiments. C.W. and S.H. developed and optimized the multicolumn NanoLC-MS platform. T.T. and X.X. provided the NanoLC columns. C.W. performed the chromatographic characterization, system benchmarking, and long-term LC–MS acquisitions. G.D.H., K.S.T., E.E.N., and R.J.S.M. prepared the bulk and single-cell proteomics samples. Y.C. and D.V.H. performed RAW264.7 cell culture and LPS stimulation. C.W. and H.L.L. processed and analyzed the mass spectrometry data. C.W. and H.L.L. prepared the figures and wrote the original manuscript. R.T.K. supervised the study, provided resources, and acquired funding. All authors interpreted the results, reviewed and edited the manuscript, and approved the final version.

Peer review

Peer review information

Nature Communications thanks Karl Mechtler and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

Research reported in this publication was supported by the National Institute of General Medical Sciences of the National Institutes of Health (NIH) under award number R35GM153179, and by the NIH NCI under award numbers R01CA279074 and U01CA271410 (CPTAC). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data availability

The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE⁷⁵ partner repository with the dataset identifier PXD081782.

Code availability

Downstream data filtering, normalization, statistical analysis, clustering, and visualization were performed using in-house Python 3.10 scripts with NumPy v1.26.4, pandas v2.2.3, Matplotlib v3.10.9, seaborn v0.13.2, SciPy v1.10.1, statsmodels v0.14.6, scikit-learn v1.6.1, tqdm v4.67.1, AnnData v0.11.4, Scanpy v1.11.5, and igraph v1.0.0. Custom analysis scripts are available on GitHub at https://github.com/RTKlab-BYU/NC_HighThroughput_LF_SCP. Input files, software requirements, and instructions are also included for reproducing the reported analyses and figures. The code is distributed under the MIT License, with no additional restrictions on access. The specific version of the code associated with this publication is archived in Zenodo and is accessible via https://doi.org/10.5281/zenodo.22031389⁷⁶.

Competing interests

T.T., X.X., and R.T.K. have a financial interest in MicrOmics Technologies, which provided LC columns for this work. The remaining authors declare no competing interests.

Artificial intelligence use statement

Artificial intelligence tools, including OpenAI ChatGPT (GPT-5.5), were used to assist with manuscript editing, code generation, and preliminary organization of Gene Ontology (GO) terms into broader biological themes. All AI-generated text, code, and GO term assignments were independently reviewed, validated, and modified where necessary by the authors. Final analyses, biological interpretations, and conclusions were determined solely by the authors, who assume full responsibility for the accuracy and integrity of the work.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-78059-x.