The bacterium X. hortorum pv. pelargonii (Xhp) strain 305 was isolated from geranium plants in Israel and was a gift from Dr. Shulamit Manulis-Sasson from the Agricultural Research Organization (ARO), Volcani Center Israel (Barel et al., 2015). Genomic DNA of X. fragariae (Xfrg) Fap21 (BioSamble SAMN05505397) (Henry and Leveau, 2016) was kindly provided by Dr. Joël Pothier (Zurich University of Applied Sciences). For the translocation assays, we used X. euvesicatoria (Xeu) hrpG* ΔavrBs2 (Roden et al., 2004). For cloning, we used NEB 5-alpha Escherichia coli that were obtained from New England Bio-Labs inc.

Strains of E. coli and Xanthomonas were grown in Luria–Bertani (LB), broth or agar, at 37°C and 28°C, respectively. The antibiotics used were spectinomycin (Sp; 100 μg/ml), kanamycin (Kan; 50 μg/ml) and gentamicin (Gm; 10 μg/ml). All antibiotics were from Sigma-Aldrich.

Pepper plants (Capsicum annuum) ECW20R (Kearney and Staskawicz, 1990) were grown in the greenhouse at 25°C and kept in long‐day conditions (16 h light, 8 h dark).

Genomic DNA of Xhp305 was isolated from 3 ml of overnight culture using Wizard^® Genomic DNA Purification Kit – Promega. Microbial De novo sequencing was performed at Novogene Co., Ltd. using both PacBio (PacBio Sequel II) and Illumina (NovaSeq 6000) platforms. The shotgun genomic library for short-read sequencing and the library for long-read sequencing were prepared by the service provider, who also performed quality control. The Illumina sequencing yielded 17,914,078 paired-end reads of length 150 bp. The PacBio sequencing yielded, after trimming, 245,660 subreads, with mean length of 11,006 bp, N50 of 13,343 bp, for a total of 2,704 Mbp.

Effectidor v1.04 (Wagner et al., 2022b) was used for T3Es predictions in each of the two genomes. The pipeline within Effectidor is divided into the following steps: (1) Defining the positive T3Es in the input genome either based on the input supplied by the user or based on homology to previously validated T3Es from various strains (this dataset can be viewed and downloaded from https://effectidor.tau.ac.il/data.html. For the analysis done in this work, version 1.04 was used). The homology criteria are E-values smaller than 10^-10 and at least 70% identical matches. If less than five effectors are identified based on this cutoff, the last criterion is reduced by 10 (i.e., 60% identical matches are required instead of 70%) until a minimum of 40% identical matches; (2) Defining the negative set (i.e., non T3Es encoded in the input genome) based on homology to proteins of E. coli K12 MG1655 (accession GCF_000005845.2); (3) Feature extraction. The features used in Effectidor vary based on the provided input. While the only mandatory input in Effectidor is a FASTA file containing all the ORFs records in the genome, additional inputs allow extraction of features outside the gene sequence alone. In our analysis we provided the following additional inputs, available in the advanced options of Effectidor: (3.1) GFF file, which holds information about the location of all the genes in the genome and allows Effectidor to extract genome organization features; (3.2) FASTA file of the full genome, which allows, together with the GFF file, to search for the PIP-box regulatory element in the promoters of the genes; (3.3) ZIP archive with FASTA files holding protein records of Luteimonas sp. MC1825 (accession GCF_014764385), Lysobacter capsica 55 [accession GCF_001442785 (de Bruijn et al., 2015)], Pseudoxanthomonas suwonensis J1 [accession GCF_000972865 (Hou et al., 2015)], Stenotrophomonas maltophilia K279a [accession GCF_000072485 (Crossman et al., 2008)], and Xylella fastidiosa 9a5c [accession GCF_000006725 (Simpson et al., 2000; Marques et al., 2001)] that were used as input for the proteomes of closely related bacteria without T3SS. This input allows to run homology searches against these proteomes, and the results of these searches often serve as informative features for the machine-learning classifier, as T3Es are not expected to be found in these genomes, whereas many of the non-T3Es – are. In addition to these inputs, we conducted all searches including the optional feature that predicts the presence of the type 3 secretion signal in the protein sequence (Wagner et al., 2022a); (4) Training a machine learning classifier. Following the feature extraction step, several classifiers (i.e., Linear Discriminant Analysis, Naïve Bayes, Support Vector Machine, Logistic Regression, K Nearest Neighbors, and Random Forest) are trained on the labeled data (i.e., T3Es and non-T3Es defined in the first step). The labeled data are split into train and test sets, the classifiers are first trained in cross-validation on the training set (including feature selection) and are finally evaluated on the test set. The evaluation method used in Effectidor is the Area Under the Precision-Recall Curve (AUPRC). The temporary best classifier is defined as the one with the highest AUPRC on the test set. All classifiers are then evaluated according to the following criteria: (4.1) AUPRC measured on the test set is smaller than that achieved by the temporary best classifier by no more than 30%; (4.2) The range of the prediction scores of the genes in the training set is at least 0.75, to ensure that not all samples are classified as negative/positive; (4.3) The AUPRC measured on the train set and on the test set are compared, and the difference must be smaller than 0.25, to reduce chances of overfitting. The classifiers that meet all these criteria, are then merged to form a final voting classifier; (5) Applying the final classifier to identify potential novel T3Es in the genome. The final voting classifier is then applied to produce a score between 0 and 1 for each ORF in the genome. This score reflects the likelihood for this ORF to encode a T3E. Of note, in case all classifiers were dropped for not meeting some of the criteria mentioned in (4.3), the final classifier used to produce these predictions is a vote over all classifiers. In this case a message is sent to the user of Effectidor. Genomes with a small number of known effectors are more susceptible to it.

The genome assembly of Xhp305 was carried out using both long (PacBio Sequel II) and short (Illumina NovaSeq 6000) reads. Using the long reads of PacBio we could close the circular genome, and the short Illumina reads were used to polish the assembly. The assembly resulted with three circular molecules: a chromosome of 5,216,813 bp, and two plasmids of 188,317 bp, and 51,091 bp, with average coverage of 36X, 35X, and 54X, respectively. The assembly polishing using the Illumina reads resulted with a few corrections in the chromosome and in the smaller plasmid. The average coverage of the assembly measured by mapping the Illumina reads was 454X, 467X, and 1,674X for the chromosome and plasmids, respectively. The coverage of the plasmids relative to the chromosome, measured both using the PacBio and the Illumina reads, suggests that the larger plasmid has a copy number of one, while the copy number of the smaller plasmid is either two or three, depending on the coverage of the PacBio versus Illumina, respectively. The average GC content of the chromosome is 0.64, while the average GC contents of the larger and smaller plasmids are 0.59 and 0.62, respectively. The chromosome and plasmids hold 4,357, 191, and 62 coding sequences, respectively. Genome assembly features are available in Table 1 . The sequencing data and assembled genome were deposited to NCBI and can be found in BioProject PRJNA926924. The annotation available in NCBI was done using the internal PGAP annotation pipeline of NCBI. It differs from the annotation we obtained using Prokka, mainly in the prediction of translation start sites. The genomic features estimated here are based on the Prokka annotation. Our downstream analysis was done using Prokka annotation, and this annotation is available in the supplementary data.

Before running a machine-learning model, Effectidor searches for ORFs with significant sequence similarity to a database of previously validated T3Es from a large set of organisms (see Methods section 2.3). Another option is to provide a list of positives (i.e., known T3Es) as input, in addition or instead of the internal homology search. This list of positives, if supplied, should be in a FASTA format. For the analysis of Xfrg we chose to supply a list of positives instead of the internal search, as the built-in homology search resulted with only half of the known T3Es in this strain, due to high percentage of identity cutoff (70%) that led to missing some of the more distant homologs. The list of positives to consider was supplied by Dr. Doron Teper, accounting for sequence similarity to previously validated effectors from Xanthomonas, Pseudomonas syringae, Ralstonia solanacearum, Acidovorax, and Pantoea sp., sharing identity lower than 70%, as T3Es. While Effectidor can still build a classifier based on a partial effector list, providing the full list of T3Es is preferable for better representation of the T3Es genomic organization, providing larger training set, and thus for more accurate predictions. For Xfrg this list of positives included 47 T3Es ( Table 2 ). For Xhp305, Effectidor was executed without providing a list of positives, and in its internal homology search, it yielded 36 T3Es, which we consider as positive samples for training the machine-learning algorithm ( Table 2 ).

In the next step within Effectidor, a machine-learning classifier was trained on the known T3Es found in the first step. Based on the trained classifiers, other than homology to effectors, among the ten most important features we used in these analyses were amino acid similarity to effectors vs. to non-effectors ( Figures 5A2 , 5B2 ), and score of the secretion signal in the N-terminal region (Wagner et al., 2022a) with 24 and 41 T3Es out of 36 and 47 in Xhp305 and Xfrg, respectively, with a signal score higher than 0.5 ( Figures 5A1 , 5B1 ). Homology to proteins of closely related bacteria without the T3SS was also important with only one and two T3Es in Xfrg and Xhp, respectively, which show high similarity to some of their proteins ( Figures 5A3 , 5B3 ). Specifically, the protein encoded by BER92_03985 of Xfrg showed sequence similarity to OtsA of Stenotrophomonas and Pseudoxanthomonas, the protein encoded by ELAGFFLI_03189 (PML25_15775 + 141bp upstream) of Xhp showed sequence similarity to WQ53_RS02800, glycoside hydrolase family 30 protein of Pseudoxanthomonas, and the protein encoded by ELAGFFLI_01251 (PML25_06220) of Xhp showed sequence similarity to OtsA of Stenotrophomonas from Luteimonas and Pseudoxanthomonas. The genomic organization of T3Es and specifically the distance to the closest known T3E on the genome was the next contributing feature, with 28 and 23 T3Es in Xfrg and Xhp, respectively, which were less than 15 ORFs away from a known effector on the genome ( Figures 5A4 , 5B4 ). In Xfrg the PIP-box was also important for the prediction, with 14 of the T3Es with a complete PIP-box ( Figure 5B5 ), while in Xhp the GC-content was more important than the PIP-box for the predictions ( Figure 5A5 ). Figure 5 shows the distribution of these features’ values among T3Es and non-T3Es in Xhp305 and Xfrg.

Following the training step, the trained classifier was applied to the remaining genes in the given genome, to yield prediction scores, reflecting the likelihood of each of the genes to encode a T3E. This step yielded several high-scoring predictions in each genome ( Tables 3 , 4 ). The features of these highly ranked genes reveal that most of them have a high secretion signal score predicted for the N-terminal region of the protein sequence. Many of them have a perfect or nearly perfect PIP-box in their promoter. In addition, some of them show sequence similarity to known T3Es, or reside in proximity to other T3Es on the genome ( Tables 3 , 4 ). Of note, the sequence similarity to known effector, when present, was not high enough for these putative T3Es to be considered positive in the previous step, i.e., the identity percentage was less than 70%.

Out of the above predictions, several candidates were chosen for experimental validation. Candidates were selected based on predictions rank, lack of significant sequence similarity to known effectors in Xanthomonas, and minimal length (peptides smaller than 75 amino acids were ignored). In addition, two proteins that were identified as T3Es based on homology to previously validated T3Es were used as positive controls. Of note, additional putative T3Es that we did not validate exist (see discussion).

To examine the translocation of predicted effectors, we utilized a reporter system based on the delivery of a truncated form of the Xeu T3E AvrBs2 (amino acids 62–574) into susceptible plant cells. AvrBs2_62–574 lacks a translocation signal, but is sufficient to elicit HR in plants expressing the Bs2 resistance gene (Roden et al., 2004). The deleted translocation signal is supplied (or not) by the cloned candidate effector. The conjugant strains we obtained were tested for elicitation of HR in the pepper line ECW20R, which encodes a functional Bs2-resistance gene.

Our results show that the following candidates induced the HR 36 h post infection ( Figure 6 ): Xhp conjugants: ELAGFFLI_04662 (conjugant of PML25_23150+PML25_23155)/ELAGFFLI_02299 (conjugant of PML25_11405+PML25_11410), identical genes on the chromosome and plasmid, respectively, named hereafter XopBB; ELAGFFLI_03194 (PML25_15800 + 60bp upstream), named hereafter XopBC; ELAGFFLI_00506 (PML25_02555 + 252bp upstream), named hereafter XopBD; and ELAGFFLI_01101 (PML25_05510 - 249bp upstream), named hereafter XopBE. Xfrg conjugants: BER92_21920, named XopBF; and BER92_22150, named XopBG.

Of note, the protein encoded by the gene PML25_02815 of Xhp contains 15 conserved SKW repeats, previously described in the effector XopAD of Xeu (Teper et al., 2016). It tested negative in the translocation assay, see discussion.

No HR was observed in leaf areas inoculated with Xeu strains expressing the other tested constructs ( Tables 3 , 4 ). Xeu expressing ELAGFFLI_00550 (XopAL) and ELAGFFLI_00565 (XopZ2), of the latter we cloned only the first 200 N-terminal amino acids, were tested as positive controls and also induced HR on pepper leaves ( Figure 6 ). The parent strain Xeu hrpG* ΔavrBs2 expressing the AvrBs2_62–574::HA fusion (“empty” vector) was tested on the same pepper leaves, as negative control. As expected, this strain did not cause HR. All in all, four out of six Xhp genes and two out of six Xfrg genes we tested encode proteins that elicited HR response in the pepper line ECW20R, which encodes a functional Bs2 resistance gene ( Figure 6 ). Thus, all six can be defined as novel T3Es.

We next conducted sequence similarity searches to identify the taxonomic distribution of the newly identified T3Es.

The 332 amino acids long XopBB protein from Xhp305 is encoded by two identical genes, on the chromosome and on the smaller plasmid. We searched for homology of XopBB to a list of previously validated T3Es from Xanthomonas, Pseudomonas syringae, Ralstonia solanacearum, Acidovorax, and Pantoea sp, supplied by Dr. Doron Teper (available in the supplementary data), using BLASTp. The best hit was to APS58_0178 of Acidovorax citrulli M6 that we have previously reported as a putative T3E based on sequence similarity to HopF2 (Jiménez-Guerrero et al., 2020). The alignment between XopBB and APS58_0178 shared 50% identical matches on 70% coverage. In a regular BLASTp search, closer putative homologs were found in X. hortorum, X. campestris, and X. hydrangea ( Table 5 ). All these inferred homologs are annotated as hypothetical proteins.

The Xhp305 T3E XopBC shares some sequence similarity with XopAV (Teper et al., 2016) and XopAY (Yang et al., 2015); The validated XopAV from Xeu is a protein of 165 amino acids. In contrast, XopBC is 249 amino acids long. The pairwise alignment between these two proteins is only between the 49 most N-terminal amino acids of XopBC, and a region near the C-terminus of XopAV, where they share only 50% identity. In contrast, XopBC shares 52% identity over 92% coverage with XopAY, which is encoded by a gene adjacent to the gene encoding for XopBC. We therefore hypothesize that XopBC and XopAY are two paralogs. XopBC was annotated as XopAV in the PGAP annotation of NCBI. Nevertheless, since the sequence similarity observed between XopBC and the validated XopAV is between the N-terminal region of XopBC, which is expected to hold the translocation signal, and the C-terminal region of XopAV, which is expected to hold the active part of the effector, and since both of these effectors were found to encode an active translocation signal which enabled them to be translocated into pepper leaves in the translocation assay, we hypothesize that these are two different effectors, and that XopBC is a newly identified effector. A BLAST search using XopBC as query reveals the presence of putative homologs of this protein in X. hortorum, X. campestris, X. arboricola, X. hydrangea, and X. codiaei ( Table 5 ). These proteins are annotated as XopAV, based on a sequence similarity similar to the one we found between XopBC and XopAV. These proteins share a higher sequence similarity with XopBC than with the validated XopAV from Xeu, and with the above results we suggest that they should not be annotated as XopAV, but rather as XopBC.

XopBD of Xhp305 did not have hits to any of the validated T3Es from Xanthomonas, Pseudomonas syringae, Ralstonia, Acidovorax, and Pantoea sp. It has some sequence similarity (60% identity over 69% coverage) to a protein from Xanthomonas campestris pv. raphani 756C, a pathogen of the plant model organism Arabidopsis thaliana, which was previously suggested as a T3E candidate XopAT, based on the presence of a PIP box and a −10 box-like sequence upstream of the coding sequence, low GC content, and eukaryotic motifs (Bogdanove et al., 2011). It has not been validated yet, and its function is unknown. ORFs with higher sequence similarity to XopBD were found in X. hortorum, X. arboricola, X. cucurbitae, X. codiaei, and X. campestris ( Table 5 ). All these putative homologs are annotated as hypothetical proteins.

XopBE of Xhp305 shows distant sequence similarity (47% identity) to XopC1 (Noél et al., 2003). Putative closer homologs were found in X. hortorum and Xfrg ( Table 5 ). The proteins found in X. hortorum are annotated as hypothetical protein whereas in Xfrg they are annotated as hydrolase-like protein or hypothetical protein.

Multiple sequence alignments (MSAs), of the abovementioned T3Es and their homologs from Xanthomonas, produced by Clustal Omega (Sievers and Higgins, 2014; Madeira et al., 2022) using default parameter values, are available in the supplementary data. All the homologs listed in Table 5 and used to produce the MSAs share at least 70% identity and 70% coverage with the respective T3E.

Of the six effector candidates tested in Xfrg, two tested positive in the translocation assay: BER92_21920 (XopBF) and BER92_22150 (XopBG).

XopBF is a hypothetical protein with unknown function. It shares some sequence similarity (identity of ~40%) with the proteins homologous to our newly validated XopBC (“XopAV”, see above) of several strains of Xanthomonas, among which are: X. hortorum (WP_159087131, WP_176339450, WP_180336534, WP_168958006, WP_152025508, WP_268212485), X. campestris (WP_228439322, WP_169705357, WP_273676157), X. arboricola (WP_212583737, WP_080591365, WP_104562523), X. oryzae (WP_019303846, WP_113343065, WP_075244353, WP_044757351, WP_027704160, WP_047339610, WP_041183112, WP_240113023, WP_029217345, WP_113221815, WP_113335989, WP_113000154, WP_069963882), X. codiaei (WP_104539725), and X. prunicola (WP_101363523). Proteins with higher sequence similarity are annotated as hypothetical proteins and are restricted to Xfrg ( Table 5 ).

Finally, the gene product of BER92_22150, XopBG, has various putative homologs, all restricted to sub-strains of Xfrg ( Table 5 ). These are all hypothetical proteins with unknown function. Surprisingly, no homologs were detected in other Xanthomonas strains. The only other putative homology found, to some degree (coverage of 99% and identity of 45%), is a hypothetical protein from Xylophilus ampelinus (WP_146228602), a grapevines pathogen that encodes a T3SS. This pathogen was previously termed Xanthomonas ampelina. It should be noted that this newly identified effector emphasizes the power of Effectidor in discovering novel T3Es without any sequence similarity to known effectors.