Cells were maintained in wholly synthetic Sautons media (2% glycerol, 0.4% L-asparagine, 0.2% glucose, 0.2% citric acid, 0.05% mono-potassium phosphate, 0.05% magnesium sulfate, 0.015% Tween 80, 0.005% ferric citrate, 0.00001% zinc sulfate at pH 7.4). Briefly, 190 ml of Sautons medium was inoculated with a 10 ml starter culture (approximately 10⁸ bacteria/ml). The flasks were sealed and incubated at 37°C and 5% CO₂ with gentle agitation until OD₆₀₀ reached 0.9 (approximately 6 weeks).

Proteins were extracted in a Biosafety level 3 facility in line with health and safety guidelines. Briefly, the cell pellet was separated from the culture filtrate by centrifugation at 4000 × g for 15 min in a bench-top centrifuge. Cell lysis was carried out by boiling the cell pellet in 1% SDS buffer (1% SDS, 100 mM Tris-HCl pH 7.6, 0.1 mM dithiothreitol (DTT), 1 mM PMSF) for 30 min. Cell debris was separated from the protein containing supernatant by centrifugation at 10,000 × g for 15 min in a bench top centrifuge and the supernatant containing the protein was transferred into a clean tube. Protein extracts from cell lysates were concentrated and buffer exchanged to 2 M urea buffer using 3 kDa MWCO filters (Millipore). Culture filtrate proteins were concentrated and buffer exchanged into 2 M urea using 15 ml 10 kDa MWCO filters. Protein concentration was determined using the BCA assay kit (Thermo Scientific). A 10 kDa MWCO filter was used for the culture filtrate instead of a 3 kDa because this is the lowest filter size available for large volumes, however, according to manufacturer's product specifications (Millipore), proteins as low as 3 kDa are still retained on 10 kDa MWCO filters.

Proteins were separated according to molecular weight using an SDS PAGE gel system. The separating gels were made from 10% acrylamide: bis-acrylamide, 0.375 M Tris-HCl (pH 8.8), 7.5% SDS, 0.5% ammonium persulphate and 0.1% TEMED. The stacking gels consisted of 4% acrylamide: bis-acrylamide, 0.125 M Tris-HCl (pH 6.8), 0.1% SDS, 0.5% ammonium persulphate and 0.1% TEMED. 40 μg of each sample (culture filtrate and intracellular protein) was mixed with an equal volume of 2x sample buffer and heated at 65°C for 5 min. Electrophoresis was performed from anode to cathode at 100 V using a BioRad mini-Protean II gel system until the bromophenol blue dye reached the bottom of the gel.

Visualization of the proteins on the gel was performed using Coomasie brilliant blue R250 for 1 h (50% methanol, 10% acetic acid and 0.1% Coomasie brilliant blue R250). Destaining of the gels was carried out by incubating on a shaker overnight at room temperature in destaining solution (10% methanol, 10% acetic acid).

Each gel lane for each strain sample was divided into 5 pieces (i.e., 5 culture filtrate fractions and 5 intracellular protein fractions, hence a total of 10 fractions per strain). Each gel piece was cut into smaller cubes and washed twice with water followed by 50% (v/v) acetonitrile for 10 min. The acetonitrile was replaced with 50 mM ammonium bicarbonate and incubated for 10 min. Washes with 50 mM ammonium bicarbonate were repeated twice to remove acetonitrile. All the gel pieces were then incubated in 100% acetonitrile until they turned white, after which the gel pieces were dried in vacuo. Proteins were reduced with 10 mM DTT for 1 h at 57°C. This was followed by brief washing steps of ammonium bicarbonate followed by 50% acetonitrile before proteins were alkylated with 55 mM iodoacetamide for 1 h in the dark. Following alkylation, the gel slices were washed with ammonium bicarbonate for 10 min followed by 50% acetonitrile for 20 min, before being dried in vacuo. The proteins in the gel cubes were digested with trypsin (Promega) at 37°C overnight in a 1:50 trypsin: protein ratio. The resulting peptides were extracted twice with 70% acetonitrile in 0.1% formic acid for 30 min and then dried and stored at −20°C. Dried peptides were dissolved in 5% acetonitrile in 0.1% formic acid and 10 μl injections were made for nano-LC chromatography.

All experiments were performed on a Thermo Scientific EASY-nLC II coupled to an LTQ Orbitrap Velos mass spectrometer (Thermo Scientific, Bremen, Germany) equipped with a nano-electrospray source. For liquid chromatography, separation was performed on an EASY-Column (2 cm, ID 100 μm, 5 μm, C18) pre-column followed by a, EASY-column (10 cm, ID 75 μm, 3 μm, C18) column with a flow rate of 300 nl/min. The gradient used was from 5–15% B in 5 min, 15–35% B in 90 min, 35–60% B in 10 min, 60–80% B in 5 min, and kept at 80% B for 10 min. Solvent A was 100% water in 0.1% formic acid; solvent B was 100% acetonitrile in 0.1% formic acid. MS/MS data was acquired from the Orbitrap Velos in Top 20 CID mode.

Raw data was captured from the mass spectrometer and converted to MS2 files using MakeMS2 software (Thermo Scientific). The data was then analyzed using Crux (McIlwain et al., 2014) and Mascot (Cottrell and London, 1999), and the output of MS2PIP (Degroeve and Martens, 2013) was used additional features for the Percolator algorithm. Spectra were obtained from each fraction of the gel (a total of 5 fractions per strain) and were viewed using Peaks v5.3. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (Vizcaíno et al., 2016) partner repository with the dataset identifier PXD004165.

Proteins were extracted in the BSL3 facility in line with health and safety guidelines as described in Section Protein extraction. Protein concentration was determined using a BCA assay according to manufacturer's protocol (#23227, Thermo Fisher Scientific). Protein preparation was performed using a filter aided sample preparation (FASP) method. Briefly, 200 μg of each protein sample was placed into a 10 KDa molecular weight cut off filter (MWCO) (Millipore). Protein cysteine residues were alkylated in the dark for 30 min in 10 mM iodoacetamide. Iodoacteamide was then removed by centrifugation at 14,000 × g for 15 min. Buffer exchange was performed twice with 8 M urea in 0.1 mM Tris-HCl pH 8.5 by centrifugation at 14,000 × g for 15 min in a refrigerated bench-top centrifuge at 18°C. The urea buffer was then exchanged for 0.05 M ammonium bicarbonate by centrifugation. Sequencing-grade modified trypsin (#608-274-4330, Promega) was added at a ratio of 1:100 enzyme: substrate and incubated overnight at 37°C in a wet chamber. The peptides were finally collected through the filter into a clean collection tube by centrifugation at 14 000 × g for 10 min.

To stop the tryptic digest the pH was lowered to 2 using 50% trifluoro acetic acid (TFA) followed by an incubation for 15 min at 37°C with shaking at 500 rpm. The peptide solution was desalted with C18 reversed-phase columns (Pierce #89870-25) according to the manufacturer's instructions. Briefly, the C18 columns were activated with 50% methanol, followed by equilibration with 5% ACN: 0.5% TFA. After loading the sample, the columns were washed 3 times with 5% ACN: 0.5% TFA. Finally, peptides were eluted with 70% ACN, dried under vacuum, and re-solubilized in 0.1% FA to a final concentration of 8 μg/μl.

All SRM experiments were performed on a TSQ Vantage triple quadrupole mass spectrometer (Thermo Fisher Scientific) equipped with a heated electrospray II ion source. For liquid chromatography and separation of peptides, a Synergi 4 μ Hydro RP 150 × 4.60 mm 80 Å pore size C18 column (serial # 630710-14) was used with a column flow rate of 300 μl/min. The gradient used was from 5–15% B in 5 min, 15–35% B in 90 min, 35–60% B in 10 min, 60–80% B in 5 min and kept at 80% B for 10 min. Solvent A was 100% water in 0.1% formic acid, and solvent B was 100% acetonitrile in 0.1% formic acid.

The mass spectrometer was operated in positive mode using electrospray ionization with a voltage of 3500 V. The capillary temperature was set to 350°C and the collision gas pressure to 1.2 mTorr. Up to 336 transitions per run were acquired with a cycle time of 3 s and a dwell time of at least 20 ms. Collision energies were calculated per individual peptide transition ion using Skyline software and further optimized by a series of energy ramping experimental steps (10 steps of 5 V) to obtain the optimum energy of each transition. MS/MS data was acquired from the CID mode. Raw data was captured from the mass spectrometer and analyzed using Skyline software.

The proteins chosen for SRM analysis were based on relevance in pathogenicity and/or virulence as stated in literature as shown in Supplementary Table 2. To design SRM assays, peptides were chosen for analysis based on their prior observation in our discovery experiments. Although, each protein was typically observed by two or more peptides in the discovery experiment, the two best performing peptides were chosen to confirm each protein in the SRM assay, with the exception of 2 proteins (Rv1818c and Rv0833) (Supplementary Table 3). Due to the small size and non-redundant nature of the M. tuberculosis proteome, 3 ion transitions were set as the minimum number required to identify a peptide in an SRM assay.

Data analysis was carried out using Skyline software (MacCoss Lab software). Intra-assay ambiguity (CV) for each peptide was based on the calculated average protein concentration for a set of technical duplicate injections of each sample. Inter-assay CV was calculated for each peptide from across three biological replicates of the 7 strains. Quantitation of each peptide was carried out using the area under the curve for the peptide transitions assayed in Skyline. Retention times for the peptide standards were obtained by pre-assessment on the MS. For peptides without standards, the retention times were obtained from predictions made by Skyline software and gated at 5 s from the predicted retention time.

The protein databases used to generate theoretical spectra were strain specific individual non-redundant fasta files obtained from Ensembl (www.ensembl.org), with the exception of LAM for which there is no Ensembl annotation, and W-Beijing whose annotated file is not sufficient for downstream cross strain comparison. For LAM, the UniProt fasta file was used (www.uniprot.org) and for W-Beijing, the H37Rv Ensembl fasta file was used. The parameters were standard across both searches and included carbamidomethylation of cysteine residues as a fixed modification and oxidation of methionine residues as a variable modification. Two missed cleavages were allowed and peptide mass tolerance was set at 10 ppm whilst fragment mass tolerance was set to 0.5 Da. Decoy databases were used for FDR analysis and a cut-off was set at 5% for protein identifications.

The PSMs obtained from the search were used to predict the expressed protein repertoire of each sample. MS/MS spectra from the 10 fractions per strain were searched with each individual search engine. Combining results from multiple search engines yields higher protein identifications (Shteynberg et al., 2013), and therefore all the proteins identified from each individual search engine were combined and redundancy was removed to give one complete non-redundant dataset per study organism.

To carry out an effective cross strain comparison, it was crucial to ensure that as much of the theoretical proteome as possible was observed by discovery MS. After obtaining a non-redundant dataset for each strain using strain specific databases, it became necessary to convert all the protein IDs into a standard protein ID by orthology mapping to allow effective cross strain comparison. To achieve this, the total non-redundant IDs from each strain obtained by searching against its individual Ensembl fasta file were then mapped back to the Ensembl H37Rv protein IDs. These were all in turn mapped to UniProt accession numbers and Tuberculist “Rv” loci numbers to facilitate downstream analysis with tools such as GO analysis and pathway mapping. It was observed that there is a slight deficiency in ortholog mapping data between databases (Ensembl and UniProt), as well as shortcomings in ortholog mapping between strains. These discrepancies lead to a minor loss of information, as represented in Figure 2. The discrepancy in ortholog mapping was much more pronounced when mapping protein IDs to H37Rv from the more distantly related M. avium, which lies outside the MTBC group; this resulted in loss of approximately 50% of the biological information in downstream comparisons to M. avium. Other strains showed relatively small losses in the number of experimentally observed proteins mapped to H37Rv orthologs, for instance the LAM strain had approximately 600 proteins with no orthologs in H37Rv which were therefore not included in the cross-species comparison.

With congruent IDs, strains were cross-compared to obtain a comprehensive qualitative comparison as summarized in Figure 3 using the Venn diagram tool Venny [267]. Protein IDs from the 4 Mycobacterium tuberculosis strains were compared (Figure 3A) and, as expected, the majority of observed proteins were found to be shared amongst all M. tuberculosis strains, with less than 5% being strain specific. A total of 1938 proteins comprise the shared proteins among the 4 M. tuberculosis strains, perhaps representing a M. tuberculosis core proteome. A second diagram was generated comparing the collective, non-redundant proteomes of the four M. tuberculosis strains to those of M. bovis, BCG, M. avium (Figure 3B). Surprisingly, M. avium had many proteins in common with the MTBC strains, with only 12 unique proteins apparently unique to M. avium; this may simply reflect though the deficiencies in ortholog mapping between more distantly related organisms. The four M. tuberculosis strains also share 989 common proteins with M. bovis and BCG which they do not share with M. avium. A group of 168 was observed uniquely in the M. tuberculosis strains and not found in M. bovis, BCG or M. avium; these proteins were therefore earmarked as candidate virulence factors to be further explored.

We subsequently sought to quantitatively assess a subset of the candidate virulence factors identified by discovery MS and cross-strain comparison using selected reaction monitoring (SRM)—a sensitive, reproducible and quantitative MS technique. Since the design of SRM robust assays can be a lengthy process and the capacity to highly multiplex hundreds of SRM assays remains challenging, we devised a strategy to create a short list of candidate proteins with possible relevance in differential clinical phenotypes observed between the M. tuberculosis isolates for subsequent quantitative proteomic analysis. To do this, we compared our proteomic data on each of the 168 proteins (Supplementary Table 5) that were observed only in the M. tuberculosis strains with 771 gene expression data sets contained in the TBDB, representing varying in vitro models of TB disease. We focussed attention in particular on 7 categories of experiment from the TBDB that aimed to more closely reflect in vivo conditions (e.g., starvation models, macrophage infection models, hypoxic models, etc.), the logic being that consistent over-expression of a protein in one of those categories might plausibly confer a selective advantage to the bacterium in vivo; the categories chosen are listed in Table 2.

For each of the 168 proteins, we assessed whether they were significantly over- or under-expressed in each of the 7 chosen categories of gene expression models and carried out statistical analysis on the gene expression values obtained from the TBDB experiments using packages in R (strategy depicted in Figure 4). Proteins from the 168 protein set whose gene expression showed an average fold change of ≥2 SD from the mean across all datasets in an individual condition were taken as significantly differentially expressed in that condition. Proteins that had significant fold change for fewer than 4 out of the 7 categories were removed from this list, resulting in a final shortlist of 23 proteins, summarized in Table 3. In order to create SRM assays for the shortlisted proteins, the M. tuberculosis proteome library (Schubert et al., 2013) was consulted and validated SRM assays for the proteins of interest were extracted.

A minimum of 2 peptides per protein and 3 transitions per peptide were assessed by SRM for the shortlisted proteins in all seven strains. Intra-assay and inter-assay coefficients of variation were determined for each individual peptide. Intra assay variability was based on 3 technical replicates per strain whilst inter assay variability was assessed based on 2 biological replicates. The signal of each peptide observed was obtained by summing the peak areas of each measured transition for that peptide and then normalizing by the total number of cells per strain at the point of protein extraction. An ANOVA test was used to determine if there was a significant difference in the expression of each of the peptides representing each protein. The single factor ANOVA results (Table 3) shows that of the 23 peptides assessed, 18 had a significant difference in expression between the 7 strains. The proteins were then broadly classified into 4 groups denoting some aspect of the organisms' success in the host (Supplementary Figure 2) in order to aid further interpretation of the data.

The four proteins assayed with roles in drug response are PyrB, PyrC, CarA, and CarB, and all form part of the pyrimidine biosynthetic operon in M. tuberculosis. All with the exception of carB (Rv1384) are more abundant in M. avium compared to the other strains, but within the MTBC the Beijing strain has the highest expression of PyrB, PyrC, and CarA.

Amongst proteins which are known to modulate the host immune response, the uncharacterized hypothetical protein Rv0966 is highly expressed in the LAM strain compared to the other strains. In this functional category, Rv2136c, Rv1002, Rv2703, and Rv2108 are more abundant in the Beijing strain. Rv1818c appears to be more abundant in the CAS strain while in the other strains it is present at comparatively low amounts.

Amongst proteins responsible for the growth of M. tuberculosis in the host, a possible toxin with unknown function, VapC2 (Rv0301), has the highest relative expression in the Beijing strain. PE_PGRS13 (Rv0833), MetC (Rv3340), and conserved hypothetical protein Rv3412 are all more abundant in the LAM strain than in BCG or M. avium, although similarly highly abundant in H37Rv and M. bovis. PPE65 (Rv3621c) appears to be upregulated in both LAM and M. avium strains.

In terms of adaptation to stress, the Acyl-CoA dehydrogenase MbtN (Rv1346) is particularly abundant in the Beijing strain. The protein Rv0901 is apparently less abundant in M. bovis and BCG strains whereas it is relatively abundant in the other strains, particularly so in H37Rv and M. avium. Aspartate kinase (Rv3709c) is more abundant in H37Rv and LAM strains comparatively, and almost entirely absent in M. avium and M. bovis.

The contribution to total signal detected for each protein as measured by SRM in each of the 7 strains is represented in Figure 5. This analysis demonstrates that proteins which were detected in a particular strain in the discovery experiment, e.g., Rv0301 and Rv1002c which were observed only in the Beijing strain, tend to contribute the highest signal when measured by SRM. This appears to be clearly the case for 14 of the 23 proteins assessed (Rv0301, Rv1381, Rv1383, Rv1002c, Rv3709c, Rv1346, Rv3412, Rv2108, Rv3621, Rv0833, Rv0966c, Rv1818c, Rv2136c, and Rv3340).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.