vPro-MS enables identification of human-pathogenic viruses from patient samples by untargeted proteomics

Categories: Disease & Virus

August 1, 2025

vPro-MS workflow for rapid and untargeted virus identification

The vPro-MS workflow for untargeted virus detection by proteomics consists of three main parts: sample preparation, LC-MS measurements and data analysis (Fig. 1). The workflow takes about 2 h to perform from sample preparation to result (vPro-MS report) and has a throughput of 60 samples per day (SPD). Samples analyzed in this study were prepared using a slightly modified S-Trap protocol and the resulting peptides were loaded manually onto EvoTips. The 1 h digestion of proteins into peptides is the longest step in the sample preparation procedure and the whole workflow. In general, alternative sample preparation strategies should work similarly well, but care must be taken to select lysis buffers and conditions that effectively inactivate viruses²⁶. Samples were analyzed using the 60 SPD method on an Evosep LC system (24 min per sample) and a diaPASEF data acquisition scheme on a timsTOF HT mass spectrometer. This LC-MS system offers the robustness necessary to measure thousands of samples in a routine environment. Peptide sequences were identified from the LC-MS data using DIA-NN and the vPro peptide library of the human virome. Human-pathogenic viruses were detected from the DIA-NN output using the vPro-MS R script and metadata of the vPro peptide library. The reliability of virus detection is monitored by calculating a confidence score (vProID), and the results are summarized in a tabular report. Samples can be analyzed in batch mode, during which the peptide identification with DIA-NN is the speed-limiting step. This step can be computationally parallelized. In our computational setting (Intel® Xeon® Platinum 8160, 24 cores, 192 GB RAM) the data analysis was performed regularly in less than 10 min per sample.

**Fig. 1: Overview of the vPro-MS workflow for virus identification by untargeted proteomics.**

Construction of the human virome peptide spectral library

The strategy of the vPro-MS data analysis workflow is to construct a peptide spectral library covering the complete human virome (vPro peptide library, Fig. 2). This library is then used to identify peptide sequences in DIA-MS data. The peptide spectral library was constructed from all protein sequences of human-pathogenic viruses available in the UniProt Knowledgebase (1,463,727 proteins, release: 2023_1). Initially, proteins without an associated proteome ID were removed. Structural proteins are best suited for the detection of viruses, as they are the most abundant of all virus proteins^27,28,29. Therefore, proteins were filtered according to their GOCC (Gene Ontology Cellular Component) terms and only structural proteins, such as core, envelope and virus membrane proteins, were retained (49,657 proteins). An in silico peptide library was predicted from those proteins using DIA-NN to filter the resulting tryptic peptides according to their detectability in terms of m/z values, retention time and ion mobility (126,788 peptides). The lowest common ancestors (lca) of those peptides were analyzed in R using taxonomic information from UniProt and only species-unique peptides were kept. The International Committee on Taxonomy of Viruses (ICTV) classifies viruses into different hierarchical levels of order, family, subfamily, genus and species. Species may also contain different stable genetic variants called strains or clades. Taxonomic information below the species rank was aggregated into a subspecies rank. This resulted in a species-level annotation for all viruses, which did not necessarily include additional subspecies information. Conversely, all subspecies information is associated with the species information. This is why there are more different subspecies (193) than species (187) in the vPro peptide library. e.g., the species Dengue virus may be connected with the subspecies information Dengue virus type 1, 2, 3, 4 or NA (meaning no subspecies information). Peptides matching to either the human proteome or common contaminants, BSA and trypsin, were removed, and the remaining sequences were annotated with protein-level (proteomeID, UniprotID, gene name, protein name), precursor-level (m/z, charge state, iRT, IM, M(ox), CAM) and taxonomic (species, subspecies) information. The peptides were exported as a table (vPro.Peptide.Library.txt), which served as the database for virus detection in DIA-NN precursor output tables. The library consisted of 121,977 peptide sequences from 331 human-pathogenic viruses (Supplementary Data 1) and was constructed from 20,386 virus proteomes. This represents 94% of all human-pathogenic viruses in the UniProtKB. The peptides were also exported in fasta format (vPro.Virus.fasta) along with the peptide.fasta files of the human proteome and common contaminants. These files were used to predict a peptide spectral library of the human virome in DIA-NN (vPro-lib.predicted.speclib). A detailed taxonomic summary of the library is provided in Supplementary Data 2.

**Fig. 2: Data flow chart of vPro-MS for virus identification by proteomics.**

Assessing the confidence of virus identification

The vPro peptide library covering the complete human virome was used to identify peptide sequences in DIA-MS data. This strategy posed a major challenge for post-processing of the identification results in order to achieve a high specificity when trying to identify very few viral peptides in a sample within a library of more than 100,000 viral sequences. Therefore, confidence of virus identification was supervised with the development of the vProID score.

The number of specific peptides varied greatly between the different virus species within the library. HIV-1 was by far the largest part of the library with 68,103 peptides, which equals 56% of all entries. As peptides are usually identified in proteomics with a controlled error rate (often ~1%), this means that few HIV-1 sequences were identified in almost every sample just by chance. In order to control the reliability of virus identification, we introduced a confidence score, named vProID (Viral Protein Identification). Care was taken to ensure that the sensitivity of virus detection was not impaired. This score is the log10 value of the number of identified peptides for a certain virus proteome divided by the number of expected peptides for this virus proteome just by chance in a specific sample. Afterwards, the top-ranked proteome for each virus species was kept, and a vProID threshold was applied to filter out unreliable virus identifications. Proteomes with a single matching peptide were also removed at this step. An example for filtering the top-ranked virus proteomes per species based on the vProID score is provided in Supplementary Data 3. The discriminatory power of the vProID score to distinguish between random and true virus identifications for the analysis of nasal swab samples is shown in Fig. 3 and Table 1. In total, 58 swab samples positive for SARS-CoV-2 covering a Ct-range of 18–35, and 8 virus-negative swab samples were tested for 331 human-pathogenic viruses, which is equivalent to 66 × 331 = 21,846 individual PCR tests. The DIA-NN report was filtered using 16 different parameter sets including variations of FDR, min. CScore and min. peptides per virus, either with or without applying a vProID score threshold. This resulted in 357 to 2362 viral and between 527,111 and 805,681 human peptide identifications, depending on the filter settings. Please note, that for diagnostic purposes each sample needs to be treated independently and therefore redundant peptide identifications among different samples are summed. The comparison of the DIA-NN report before and after filtering, as well as the vProID score distribution, is visualized in Fig. 3A, B using the example of parameter set 10 (1% FDR, 0.95 min CScore, 2 min. peptides), which was used for the entire remaining manuscript. The specificity on sample- or virus-level and the sensitivity for SARS-CoV-2 detection are reported in Table 1. Specificity is defined as the ability of a test to identify true negatives. As samples were tested for multiple targets (331 viruses), the specificity of vPro-Ms was calculated by referring to the samples or viruses. Specificity was defined as the number of true negative (TN) virus tests or samples divided by the sum of true negative and false positive (FP) virus tests or samples (specificity = TN/(TN + FP)). The results show that regardless of the parameters, applying the vProID score significantly improved specificity and reduced false positive virus identifications by up to 100% (Fig. 3C). In contrast, when only applying a 1% FDR filter, 959 out of 21,846 individual virus tests were incorrect, which equals about 14 false virus identifications per sample. Without vProID score filtering, perfect specificities of 100% on test and virus-level were only achieved using two parameter sets where a minimum number of 5 peptides per species was required (Table 1). However, the use of such a strict filtering severely reduced the sensitivity by 24%. When the same filter parameters were used in conjunction with vProID scores, perfect specificities on sample and virus-level were achieved requiring only 2 peptides per species. As a result, vPro improved the sensitivity with which viruses can be identified with maximum specificity up to 28% (Fig. 3D). These results underline, that the vProID score is able to separate random from true virus identifications with high sensitivity even when using a rather moderate FDR of 1% and a highly imbalanced peptide library covering the human virome.

**Fig. 3: Impact of the vProID score on the sensitivity and specificity of virus identification.**

Table 1 Impact of the vProID score on the sensitivity and specificity of virus identification for various filtering strategies of peptide identification results

Evaluation of the specificity of vPro-MS for virus identification

The specificity of the vPro-MS workflow was determined by the analysis of 221 samples from 4 different sources covering 17 human-pathogenic viruses (Fig. 4). Two sample panels were analyzed for this study, and the raw data from two further studies of patient samples were downloaded from the PRIDE repository^30,31. The sample types included purified viruses, cell-culture supernatants, respiratory swabs and plasma (Fig. 4A). MS data of all samples were analyzed using the vPro-MS data workflow covering 331 human-pathogenic viruses. This corresponded to the analysis of 221 samples x 331 viurses = 73,151 individual virus tests. Specificity was calculated either on virus or sample-level as the number of true negative (TN) results divided by the sum of true negative and false positive (FP) results (specificity = TN/(TN + FP). The specificities ranged from 99.97–100% on virus-level and from 95.74–100% on sample-level (Fig. 4B). The analysis of the specificity panel revealed 3 additional virus species identifications in one monkeypox and one vaccinia virus sample (Fig. 4C). These two viruses belong to the same virus family of poxviruses, genus orthopoxviruses. In both cases, the top hit based on the vProID confidence scores was the correct orthopoxvirus species. However, in both samples, either one or two closely related orthopoxvirus species were identified as well. Most probably, this resulted from the incompleteness of the sequence database underlying the vPro peptide library. It might occur that certain tryptic peptides were considered to be species-unique, but in fact, the respective sequence also occurs in closely related species, for which the respective isolates are missing in the UniProt database. Therefore, care should be taken if several species of the same genus are identified in one sample. In the nasal swab study of SARS-CoV-2 patients obtained from PRIDE, SARS-CoV-2 was detected in two samples, which were labeled negative. In both samples, the same two SARS-CoV-2 peptides were identified, which resulted in vProID scores of 2.4 and 2.5, respectively. However, as this study was about analyzing SARS-CoV-2 positive samples, it is unlikely, that these additional identifications were random events. If this were true, it must have been expected that any other species would have been identified. Therefore, it is more likely that those additional identifications were either due to peptide carry-over between samples or resulted from an error in the initial characterization of those samples. Overall, the data underline that the vPro-MS data analysis workflow is highly specific for the detection of human-pathogenic viruses in different sample types and data from various laboratories.

Evaluation of the sensitivity of vPro-MS for the identification of SARS-CoV-2

The sensitivity of the vPro-MS workflow for the identification of SARS-CoV-2 was evaluated in 66 upper respiratory swab samples covering a Ct range of 18–35 (58 samples) and including 8 virus-negative samples. Additionally, 6 further peptide libraries were created to analyze the influence on the sensitivity (Table 2). In order to analyze the influence of the library size, the initial vPro peptide library (121,977 viral peptides) was either restricted to SARS-CoV-2 (177 viral peptides) or reference proteomes of HIV (53,600 viral peptides). In another library, in conjunction with the SARS-CoV-2 restriction, human entries were restricted to peptides identified in the swab samples, which reduced the number of human entries from 591,159 to 18,655. This library size matches approximately the size of a library obtained from the measurement of synthetic SARS-CoV-2 peptides, which was used to analyze the difference between in silico and measured spectra. Therefore, all theoretical peptides of the SARS-CoV-2 nucleoprotein were synthesized and spiked into negative swab samples to create a SARS-CoV-2 peptide spectral library, including human respiratory swab peptides. In order to further investigate the influence of the library prediction model, we created another library from the same fasta files used for the initial vPro library prediction using the timsTOF model in AlphaPeptDeep instead of the algorithm implemented in DIA-NN³². The results demonstrate that the library size as well as the prediction software have only minor effects on the sensitivity for virus identification. However, the library created from measurements of negative samples spiked with synthetic virus peptides clearly outperforms all in silico predicted libraries.

Table 2 Impact of vPro library variations on the sensitivity of SARS-CoV-2 identification

The sensitivity panel contained three different SARS-CoV-2 variants of concern (VOCs), namely alpha, delta and omicron. vPro-MS is currently not able to identify VOCs, because the information needed to include VOCs in the vPro library is not available in UniProt. Nevertheless, in order to analyze the potential of untargeted proteomics for VOC identification, a vPro peptide library with additional VOC entries was created. The vPro reports for all libraries are available in the source data of Table 2. The report of the vPro (SARS-CoV-2 VOCs) peptide library contains additional columns for the taxonomic layer VOC. The VOC was correctly identified only in one delta sample by two unique peptides. If the peptide threshold is reduced to one, already 2 out of all 4 VOC identifications are wrong. The fact that only in 4 out of 58 samples a VOC-unique peptide was identified demonstrates that this approach is not very promising. Maybe more sophisticated approaches might be suited for that purpose, but therefore a well curated protein sequence database of all VOCs is mandatory.

The limit of detection (LOD) with 95% confidence of the initial vPro peptide library corresponded to a Ct value of 27 as determined by qPCR (Fig. 5). It should be noted that approximately twice the amount of starting material (27 µL) was used per qPCR reaction compared to an average volume of 12 µL starting material, which was injected per proteomics measurement. However, twice the sample amount corresponds to a very small Ct difference of just 1 Ct. Virus quantification by proteomics and qPCR correlated with an R² value of 0.62. This value is well within the range of 0.54 to 0.82, which was reported in recent studies using targeted proteomics to detect SARS-CoV-2^{17,18,33,34,35}. The library constructed from measurements of synthetic SARS-CoV-2 peptides improved the LOD with 95% confidence to a Ct value of 29.4, which is equivalent to ~5 times less virus being detected compared to the in silico predicted vPro peptide library. This underlines the potential of library construction from synthetic peptides to improve the sensitivity in the future. In principle, such specific libraries could also be constructed from measurements of any virus positive sample.