A total of 11,250 MS/MS peak lists were searched against the UniprotKB C. sativa database (663 entries) using the Mascot algorithm, a fragment tolerance of ±50 ppm or ±2 Da, and validating the results using a decoy or an error tolerant method (Supplementary Table S1). With a ±50 ppm fragment tolerance, protein N-term acetylation and Met oxidation set as dynamic modifications, and an error tolerant method, 12 proteins were identified [210 (2%) matches] with 11,040 (98%) MS/MS spectra remaining unassigned and a search time of over 24 h. Using the same parameters but changing error tolerance to decoy brought the number of accessions identified to 21 from 213 (2%) matched MS/MS spectra and a very fast search time of 29 s (Table 5). Excessive stringency in Mascot algorithm could justify the low number of database hits [37], as discussed above in the sections pertaining to known protein standards. Relaxing the fragment tolerance to ±2 Da listed 36 proteins based on 355 (3%) assigned MS/MS spectra with a search time of 2.5 min. With a ±50 ppm fragment tolerance, protein N-term acetylation, Met oxidation, and phosphorylations of Ser and Tyr residues set as dynamic modifications and a decoy method, the number of unique proteins identified was 21 (187 matches) over a search time of 2 h. Lifting the fragment tolerance to ±2 Da increased the number of hits to 61 proteins with 590 (5%) MS/MS spectra assigned. Forsaking dynamic modification reduced search times to mere seconds and yielded 20 and 24 identities using ±50 ppm and ±2 Da fragment tolerance, respectively (Supplementary Tables S1 and S4).

To further evaluate the effect of Mascot search parameters, a more extensive but less targeted database was interrogated using the least stringent fragment tolerance (±2 Da) and a decoy method. We chose Swissprot, as it is a curated database currently hosting over 200 million sequences. Without any dynamic modification set, searching the whole taxonomy yielded 94 accessions with 998 (9%) MS/MS matches, and searching only viridiplantae taxonomy (39,800 entries) yielded 80 hits [1181 (10%) matches]. Searching viridiplantae taxonomy and setting protein N-term acetylation and Met oxidation as dynamic modifications listed 141 accessions [1352 (12%) matches]. Finally, still searching viridiplantae taxonomy but adding phosphorylations of Ser and Tyr residues as dynamic modification generated 274 accessions [1863 (17%) matches]. The latter search lasted the longest (53 h) (Supplementary Tables S1 and S4). While the list of proteins extends when a larger database is used in conjunction with more relaxed mass tolerances, we do not yet believe in their relevance, as only one protein (Olivetolic acid cyclase, OAC) actually comes from C. sativa species. Previously, Shen and colleagues demonstrate that more decoys fragments (i.e., false positives) are identified when the mass error increases [44]. Furthermore, lower mass accuracy decreases the specificity of identification [37]. For these reasons, we hereafter only focus on the search result obtained from the uniprotKB database with a stringent fragment tolerance (±50 ppm) (Table 5).

We then searched a larger and more specific database containing all the C. sativa protein sequences we could retrieve from UniprotKB, NCBI, and MPGR, which amounted to a total of 59,525 accessions. Running the Mascot search with a 50 ppm tolerance and no modifications was very quick (1.2 min) and yielded 31 accessions (Table S1). However, imputing variable modifications significantly augmented the search time. For instance, searching for oxidation (M) and N-term acetylation lasted 1 h and produced 36 accessions; search for methylation (K) lasted 2 h and yielded 33 accessions. Anything else turned into day-long searches (Table S1). Setting modifications exponentially increased the search space, thereby exponentially increasing the duration of the search. To overcome this issue, we devised a two steps strategy: (1) In Mascot, we minimized the number of PTMs (none or oxidation M) in order to quickly generate a list of hits. Not including the search outcomes using the second database (Swissprot viridiplantae), a total of 46 unique accessions were identified in medicinal cannabis mature buds bearing numerous PTMs (Table 5); (2) in the Genedata Refiner LC-MS maps of the deconvoluted masses, we explored features in the vicinity of these hits to detect more proteoforms. This allowed the detection of 136 proteoforms (Supplementary Table S5).

NCBI yielded one (2%) accession (AKP55264.1 30S ribosomal protein S16, chloroplastic), MPGR yielded 20 (43%) accessions, and UniprotKB yielded 25 (55%) accessions. This suggests that, even though UniprotKB hosts the smallest number of C. sativa accessions (663) relative to NCBI (1451) and MPGR (57,411), it generated the highest number of identities, a testament to UniprotKB quality and well-deserved status of reference database for proteins. The masses of the 46 identified proteins ranged from 3.8 kD to 17.9 kD. Twenty accessions had a Mascot score above 100, and 36 accessions were identified using more than one MS/MS spectrum (Table 5 and Supplementary Table S5). No missed cleavage was found (M > 0), possibly explaining the low number of identified proteins, as we were expecting natural protein degradation to occur. The protein identified with the highest score was cytochrome b559 subunit alpha (accession A0A0C5ARS8, score of 1641, 29 matches, Supplementary Figure S6). Only nine (20%) of the intact proteins identified by TDP were also identified in our BUP study [1], which highlights the complementarity of both approaches.

As previously observed on the protein standards, fragmentation efficiency of cannabis intact proteins depends on the charge state of the parent ion, on the type of MS/MS mode, and on the level of energy applied. We illustrated this using the protein exhibiting the second highest Mascot score (1664), Photosystem I iron-sulfur centre (PS I Fe-S centre, accession A0A0C5AS17), identified with 39 MS/MS spectra. Fragmentation efficiency was assessed using the ProSight Lite program by the percentage of inter-residue cleavages achieved. MS/MS spectra differed in the number of peaks and their distribution along the mass range (Figure 6A,B).

The optimum dissociation of a precursor ion with high charge state (857.31 m/z, z = +11) was achieved with ETD at “mid” energy, whereas a precursor ion of comparable intensity but with lower charge state (1178.55 m/z, z = +8) responded better to CID and HCD at “low” and “high” energy levels, respectively. All MS/MS data considered, fragmenting 857.31 m/z and 1178.55 m/z parent ions yielded 70% and 65% inter-residue cleavages, respectively, and 82% all together (Figure 6C). In order to maximise AA sequence coverage, it was essential to multiply the MS/MS conditions on as many precursor ions as possible. This, of course, limited the total number of different proteins analysed in a top-down approach. Coupling this strategy with an extended separation run should alleviate this drawback.

Most of the identified proteins (17/46, 37%) were involved in photosynthesis (subunits of cytochromes and photosystems I and II, as well as chloroplastic ATPases), then in protein translation (eight ribosomal proteins, 17%). Also identified were six histones. Only one protein belonged to the phytocannabinoid biosynthesis, olivetolic acid cyclase (I6WU39, OAC), previously identified by BUP [1,17]. With 46 identities, this TDP experiment identified fewer individual proteins than our previous shotgun study, which produced 160 accessions using a database from UniproKB [1]; however, the power of TDP lies in the detection of proteoforms of the identified proteins (Table 5 and Supplementary Table S5). In particular, the N-terminus excision of the initial M (NME, Table 5) was determined for 28 (61%) accessions. This valuable and novel information is not found in public databases such as UniprotKB. The PTMs identified in this work were methylation (K), dimethylation (K), acetylation (N-term and K), succinylation (K), phosphorylation (STY), as well as oxidation (M). Seventeen (37%) proteoformes were oxidized (O, Table 5). Our experimental design did not allow for distinguishing between in vivo protein oxidation and artefactual oxidation occurring during sample preparation. Ursem and colleagues demonstrated that centrifugation, freeze-thaw cycles, or long-term storage at −20°C or −80°C did not oxidise proteins [47]. The authors did not test the effect of sample agitation such as vortexing, for instance, whereby aerating the sample can trigger spontaneous oxidation. This might have occurred in our experiment. If needed, a method is available to eliminate this analytical artefact [48]. Limiting the number of steps in sample preparation lowers the chance of artefactual oxidation. A critical review of this approach has been undertaken and documented [49].

Eighteen (39%) proteoforms bore an N-terminus acetylation (NA, Table 5). N-terminal acetylation corresponded to the appending of an acetyl group to the N-terminal amino group in an irreversible manner, which impacted the lifespan, the folding characteristics, and the binding properties of the acetylated protein [50]. This is a widespread protein modification across all taxa, and C. sativa is no exception, as demonstrated in this work. Another fairly frequent PTM in our dataset was the methylation of K residues reported in eight (17%) protein accessions (M, Table 5). The post-translational methylation of K residues catalysed by lysine (K)-specific protein methyltransferases (KMTs) is a very common and important protein modification [51] and is particularly well documented for histones with respect to DNA processing [52]. While we did not identify any KMTs by BUP [1], we suspect they are present and highly active in cannabis buds, given the frequency of methylated proteoforms (Table 5). We identified four accessions with phosphorylations (P, Table 5). Phosphoproteins are ubiquitous to all kingdoms; they have been well documented in human, mice, and yeast but much less in Arabidopsis [53], the reference plant. Plants contain glycoproteins; only a few reports have applied MS to profile intact plant N-glycoproteins, and their analysis involves specific technical steps, such as the deglycosylation of glycoproteins and their recovery by lectin-affinity columns [54], which were not attempted here. However, this warrants future investigation.

Examples of proteoforms detected in this TDP study are given in Figure 7.

Histones were heavily modified [55]—in particular, they were heavily methylated [52]—which we demonstrate in Figure 7 on histones H3.2 and H4. Histones were also acetylated and phosphorylated [56], as shown for histones H2A, H3.2, and H4 (Figure 7). We also found a succinylated proteoform of histone H4. A non-specific lipid transfer protein (nsLTP) also exhibited several methylated states (Figure 7), which has never been shown before. Due to their high allergenicity and cross-reactivity, nsLTPs are major allergens of C. sativa [57]. Perhaps some methylated proteoforms of nsLTP are more allergenic than other. Bet v1-like protein (UniprotKB accession I6XT51) is similar to the major pollen allergen Bet v 1. In our sample, we found this protein to be methylated, acetylated, oxidized, and phosphorylated (Figure 7). We also observed that the same group of proteoforms shifted by 245 Da, which we could not assign to any obvious modification. With a Mascot score of 174, OAC was identified without its initial M residue, which has not been evidenced before. Several methylated and acetylated proteoforms of OAC WEre identified (Supplementary Table S5 and Figure 7), along with a possible allelic variant (Figure 7). However, the latter must be confirmed by top-down sequencing.