P. aeruginosa PAO1 (ATCC BAA-47) was maintained on Luria-Bertani agar containing 10 g/L NaCl (LB10) (644520, Difco). S. aureus 25923 was maintained on TSB agar (236920, Difco). Cultures were routinely cultured in 10 mL LB10 media (244620, Difco) at 37°C with 200 rpm shaking overnight.

The flowers, fruits, leaves, stems, and roots of M. dodecandrum were sampled from the NTU Community Herb Garden. Leaves used solely for bioactivity screening were transported in dry ice and lyophilized using a freeze dryer (Labconco). Plant organs used for both bioactivity and transcriptomic studies were transported in liquid nitrogen.

Plant material was ground in liquid nitrogen to a fine powder. One hundred mg of the powder was aliquoted into a screw cap tube. Three 3 mm steel beads and 500 µL of 80% HPLC-MS grade methanol were then added. The sample was subsequently homogenized for two cycles at 2000 rpm (PowerLyzer 24, Qiagen). The steel beads were removed, and the samples were placed on a shaker for 10 mins at 1500 rpm at room temperature in the dark. The samples were centrifuged at 10 min, 12000 × g. The supernatant was transferred into a new 2 mL Eppendorf tube and kept on ice. Five hundred µL of 80% HPLC-MS grade methanol was added to the remaining pellet, and the shaking and centrifuge step was repeated twice until 1.5 mL of supernatant was collected. The supernatant was dried completely using a centrifugal vacuum concentrator (Centrivap, Labconco) and dissolved in 100 – 500 µL of DMSO before use.

Dried methanol extracts derived from 100 mg of plant material or different plant organs were dissolved in 500 µL DMSO to give a stock solution of 200 mg/mL plant material. Overnight bacterial cultures were subcultured 1:10 in 10 mL of LB10 broth until the bacteria reached the exponential phase. Exponential phase bacteria were then diluted in Mueller-Hinton broth (MHB) to a final concentration of 5 × 10⁵ CFU/mL, and 2 µL of the stock DMSO solution was added to 100 µL of MHB to a final concentration of 4 mg/mL starting plant material or 2% v/v DMSO. Evaluation of HPLC fractions was carried out at 2% v/v DMSO, with each fraction corresponding to ~10 mg/mL mass equivalent of the crude methanol extract.

Freeze-dried leaves of M. dodecandrum (11 g) were milled and extracted twice with 400 mL of methanol (MeOH). The combined MeOH extract was dried under reduced pressure to afford 3 g of dried MeOH extract. Crude MeOH extract (2 g) dissolved in 5 mL of MeOH was fractionated using a Sephadex LH20 open column (5 x 35 cm). Compounds were eluted sequentially with 1 L of MeOH, 50% aqueous acetone, and acetone to generate 13 fractions. Biological testing results showed activity in only one fraction (58K, 120 mg) which was eluted in the acetone eluent.

50 mg of active fraction 58K was reconstituted in 15% aqueous MeOH, sonicated, and centrifuged. The insoluble pellet was collected (labeled as PPT, 5 mg) while the supernatant was injected into a Waters XTerra Prep MS C18 Column (10 x 300 mm) on an Agilent 1260 Infinity II Preparative HPLC system. The mobile phase consisted of water (A) and acetonitrile (B), both with 0.1% formic acid. The gradient elution started at 15% B for 5 mins, 15-30% B in 20 mins, 30-60% B in 25 mins, held at 60% for 10 mins, then increased from 60-100% B in 2 mins and washed at 100% B for 10 mins. The flow rate was 20 mL/min and the detection UV wavelength was at 254 nm. 8 fractions were generated from the preparative HPLC. All HPLC fractions, PPT and parent fraction 58K were submitted for biological evaluation.

Purified fractions were analyzed on an Agilent UPLC1290 coupled with a Quadrupole Time-of-Flight (Q-TOF) system. Separation was carried out with a reversed-phase C18 column (2.1 x 50 mm) at 0.5 mL/min, using a 10 mins linear gradient with 0.1% formic acid in both solvent A (water) and solvent B (acetonitrile). The typical QTOF operating parameters were as follows: negative ionization mode; sheath gas nitrogen flow, 12 L/min at 275°C; drying gas nitrogen flow, 8 L/min at 250°C; nebulizer pressure, 30 psi; nozzle voltage, 1.5 kV; capillary voltage, 1.5 kV. Lock masses in negative ion mode: TFA anion at m/z 112.9856 and HP-0921 TFA adduct at m/z 1033.9881. 1H NMR spectra were acquired on a Bruker DRX-400 NMR spectrometer with a 5-mm BBI Cryoprobe.

Minimum inhibitory concentrations (MIC) were determined using broth microdilution methods as described in (Wiegand et al., 2008). Briefly, plant extracts or equivalent concentrations of DMSO were diluted in MHB to four times the desired final concentration. 50 µL of the samples were added to 50 µL of MHB media on the first column of a 96-well plate, then serially diluted two-fold down the row. Untreated media-only growth control wells were also included. Overnight cultures of PAO1 were subcultured in LB10 medium and further incubated at 37°C with 200 rpm shaking. Exponential phase cells were then diluted in MHB to 1 × 10⁶ CFU/mL. Fifty µL of the bacterial culture was added to each well to a final concentration of 5 × 10⁵ CFU/mL, with the final concentration of plant extract falling between 0 – 50 mg/mL (0 – 5% DMSO). Sterility control wells, consisting of only MHB media, and blank control wells, containing 0 – 50 mg/mL of plant extract and MHB media, were also included. The 96-well plate was incubated for 24 h at 37°C with OD readings taken at 600 nm every 10 mins using a microplate reader (Tecan M200). Readings from blank control wells were subtracted from wells with bacteria added to account for high background readings at high plant extract concentrations. MIC was similarly determined for 0 – 512 µg/mL of the antibiotics and antimicrobial compounds triclosan, rifampicin, ceftazidime, and gentamicin. MIC was defined as the first concentration at which no growth is observed after 24 h incubation.

Antibacterial activities across experiments were compared using the area under the curve (AUC), which represents bacterial growth over a period of 18 h. The AUC is calculated using Graphpad prism (version 9.3.0). The same parameters (t = 0 - 1080 mins) were used for all data. Eighteen hours were used for the AUC calculations even though OD data were collected over 24 h, as cell growth has already plateaued and increased noise was observed at later time points, possibly due to evaporation of the small volume of medium used. The AUC values were plotted and compared using one-way ANOVA with multiple comparisons in Graphpad prism.

Overnight cultures of PAO1 were subcultured in LB10 medium and further incubated at 37°C with 200 rpm shaking. Exponential phase cells were diluted to 1 × 10⁶ CFU/mL in MHB media and treated with 0 – 2 × MIC of plant extract. Vehicle controls were treated with equivalent volumes of 0 – 5% v/v DMSO. At each timepoint, samples were collected for serial dilution in PBS and CFU enumeration. Five µL of each dilution was spot plated onto LB10 agar plates and incubated at room temperature overnight. The first spots containing between 3 – 30 colonies were counted. The assay was carried out independently twice with two technical replicates each time.

The effect of plant extract on membrane integrity was evaluated using the LIVE/DEAD stain (L7012, Thermo Fischer) and microtiter plate reader (Tecan) as described in the commercial protocol. Briefly, sub-cultured exponential phase bacterial cells were washed in 1 × PBS (PBS). 5 × 10⁷ CFU/mL of bacterial cells diluted in PBS were subsequently treated with a non-inhibitory concentration of 0 – 4 mg/mL of filtered plant extract or equi-volume of DMSO for 10 mins. Samples were then treated for a further 30 mins with LIVE/DEAD stain readings taken with excitation at 485 nm and emission at 535 (red) and 635 nm (green). Dead cells were prepared via incubation of bacteria at 65°C for 30 mins. The percentage of live cells is determined using a standard curve generated through a mixture of live and dead cells and the calculation of the green/red fluorescence ratio as stated in the commercial protocol. When LIVE/DEAD stain is used in microscopy, 2 × 10⁷ cells were seeded into a well of a poly-L lysine pre-treated ibidi 8-well chamber slide (80826, ibidi). The cells were allowed to attach for 10 mins before each well was washed and treated for 10 mins with 0 - 16 mg/mL of plant extract or equivolume of 0 - 1.6% v/v DMSO. Following treatment, each well was washed again and stained with LIVE/DEAD stain as described in the commercial protocol. At least three images per sample were taken using an epifluorescence microscope (Axio Observer Z1, Carl Zeiss) using the FITC filter for imaging SYTO 9 fluorescence and AF 568 filter for imaging propidium iodide (PI) staining. Images of fluorescence staining were analyzed using Image J.

Swimming and swarming assays with P. aeruginosa were carried out similarly to (Ha et al., 2014a; Ha et al., 2014b). Briefly, 0.3% w/v or 0.5% w/v of 1X M8 agar (1X M8 solution: 42.3 mM Na₂HPO_4, 22 mM KH₂PO_4, 8.56 mM NaCl) were prepared for use in swimming and swarming assay respectively. Plant extracts or equivalent volumes of DMSO were added to the well with the highest concentration of the extracts, then serially diluted 4-fold down the wells before the agar solidified to a final concentration of 0 – 4 mg/mL plant extract. The agar was dried under laminar flow for 45 mins and used immediately after drying. Overnight cultures of PAO1 were used for inoculation. The agar was inoculated either by dipping an Eppendorf pipette tip into the culture and then stabbing it midway of the depth of the agar (swimming assay) or by spotting 1 µL of the culture on the surface of the agar at the centre of the well (swarming assay). Following inoculation, the plates were incubated at 30°C overnight and the images were taken 16 h after inoculation. Swimming and swarming areas were assessed using Image J by setting an appropriate scale and selecting the region of interest. The resulting values are plotted in Graphpad (9.3.0) and analyzed using one-way ANOVA with multiple comparisons against the control group.

Following M. dodecandrum sample collection and grinding into a fine powder (as described in ‘Metabolite extraction’), RNA was extracted from 100mg of a sample using Spectrum™ Total Plant RNA Kit (Sigma) Protocol A following manufacturer’s instructions. Three biological replicates were collected for each adult organ, while single replicates were collected for the young flowers. To extract RNA from bacteria (PAO1), overnight cultures of PAO1 were first subcultured and grown to mid-log phase. Cells were adjusted to OD=0.4 and then treated with equivolume of 2X antibiotic dissolved in MHB medium to a final concentration of OD=0.2, MIC=1X (or 25 mg/mL plant extract and 128 µg/mL triclosan), and a final volume of 500 µL. The samples were incubated at 37°C statically for 30 mins. Following treatment, 1 mL of RNA protect reagent (76506, Qiagen) was added to 500 µL of the sample. The sample was treated as described in the commercial kit and the dried pellet was stored at -80°C. RNA was extracted using the RNeasy plus mini kit (Qiagen) as per the kit’s instructions. For both plant and bacteria, quality control of all extracted RNA was carried out by Novogene (Singapore). Each sample was evaluated for its quantity, integrity, and purity using agarose gel electrophoresis, Nanodrop, and Agilent 2100 Bioanalyzer. Library construction was performed by Novogene where mRNA of plants was enriched from total RNA with oligo(-dT) magnetic beads. No ribosomal RNA depletion was performed for bacteria library construction due to high sequencing depth. The library was quantified with Qubit and real-time PCR, and then sequenced using Illumina NovaSeq 6000, with paired-end sequencing of 150 base pairs (bp) per read and a sequencing depth of approximately 60-70 million reads.

Low-quality RNA-seq reads were removed, and the remaining reads were trimmed with fastp (v0.23.2)(Chen et al., 2018). Reads were then aligned to the PAO1 genome (NC_002516.2) and read counts were obtained using EDGE-pro (Magoc et al., 2013). Due to poor overall read alignment, one DMSO-treated sample (replicate 1) was discarded (data not shown). One-third of the reads from the remaining two DMSO-treated samples were removed and collated to form a new third replicate. To evaluate sample-level variability, read counts were transformed using variance-stabilizing transformation, followed by hierarchical clustering of samples using Euclidean distance as the distance metric (Love et al., 2014). Differentially expressed genes (DEGs) between each antibiotic-treated sample compared to the DMSO control were identified with read counts using DESeq2. Genes with | l o g 2 fold change| ≥ 1 and Benjamini-Hochberg (BH) adjusted p-value< 0.01 were considered significant (Benjamini and Hochberg, 1995) for multiple-testing correction. Gene ontology (GO) enrichment analysis was performed via the GO webtool on each list of significant DEGs against the PAO1 GO biological process data set by applying Fisher’s Exact test with BH correction (adjusted p-value< 0.01)(Mi et al., 2019; The Gene Ontology Consortium, 2021).

Quality control and read trimming of raw RNA-seq data were performed with fastp (v0.23.2). To obtain normalized transcript abundances (transcripts per million; TPM) of the RNA-seq samples, the reads were pseudoaligned against the coding sequences of the M.dodecandrum genome (PRJCA005299 from the National Genomics Data Center)(Hao et al., 2022) and quantified using Kallisto (v0.46.1)(Bray et al., 2016). To evaluate sample-level variability, samples were clustered using Spearman’s correlation as the distance metric - d x , y =   1   −   | ρ x , y | , where x and y are vectors comprising TPM expression profiles for all transcripts of a pair of RNA-seq samples respectively and ρ is the Spearman correlation coefficient. The final gene expression matrix was derived after filtering out genes with an average TPM of 0 across all organs. Genes were translated into protein sequences using the transeq module (EMBOSS v6.6.0.0)(Rice et al., 2000).

The following analyses were performed in M.dodecandrum

Enzyme annotation and pathway prediction were performed following the pipeline described by (Hawkins et al., 2021). Briefly, protein sequences were annotated with E2P2 (v4), pathway membership was predicted using PathoLogic in PathwayTools (v26) with default settings, and semi-automated validation of the final pathways was performed by SAVI (v3.1). All pathways assigned to the manual-to-validate list by SAVI were retained except for PWY0-501 and PWY-5723, following recommended validation steps by SAVI. From SAVI’s list of pathways to add, PWY-5173 was discarded as it is a deprecated ID in the current version of MetaCyc used by PathwayTools.

To determine if genes within an SM pathway are more co-expressed than by chance, Pearson’s correlation coefficient (PCC) for each gene-pair within a pathway was first calculated. To simulate random pathways, a list of pathway-labeled gene-pairs was generated. The pathway labels were then randomly shuffled amongst the gene-pairs. The PCC for each gene-pair was calculated, and the median PCC of each random pathway was extracted. This was repeated 100 times.

To identify subclusters of a pathway, genes within a pathway were hierarchically clustered based on their pattern of expression across samples using the dendrogram function from scipy with default parameters (where the distance threshold is 70% of the maximum dendrogram distance). A representative subcluster was only identified for a pathway if the subcluster consisted of genes that represented at least 75% of the annotated reactions in the pathway in the plant (degree of completeness). As some pathways have multiple genes catalyzing the same reaction, subclusters of differing sizes can have the same degree of completeness. In these cases, the largest subclusters will be chosen as representative subclusters. For pathways with no representative sub-cluster, the original pathway was the representative ‘sub-cluster’ in downstream analyses.

To investigate patterns of inter-pathway coexpression within and between SM classes, we generated a pathway coexpression network. First, gene-pairs of all SM genes were calculated using the ‘pearsonr’ function from the scipy package. Then, p-values were corrected with the BH method (Benjamini and Hochberg, 1995). Gene-pairs with PCC ≥ 0.6 and BH-adjusted p-value (two-tailed)< 0.05 were considered significant. Subsequently, the gene nodes within the gene co-expression network were labeled with their respective SM pathways (as determined by pathway prediction). Multiple pathway labels will be assigned to genes that are predicted to be involved in multiple SM pathways. Labels of “Superpathways”, defined as pathways that combine individual pathways, will only be assigned to a gene if the gene is not already labeled with their respective subpathways.

Next, to determine if the intra-pathway coexpression was better than random, the network was then subjected to the following permutation analysis: for each pathway, the number of intra-pathway edges was counted. Pathway labels were then randomly shuffled and the number of intra-pathway edges was counted. The following empirical p-value was calculated for each pathway:

where N is the number of permutations, I is an indicator function that takes a value of 0 or 1 if the statement is false or true, respectively. This was repeated 10 000 times. Pathways with BH-adjusted p-values of< 0.05 were considered to be coexpressed (gene members of the particular pathway are coexpressed with each other). Inter-pathway coexpression between pathway-pairs were also determined similarly by counting inter-pathway edges that connect genes across any given two pathways.

To generate a pathway coexpression network, edges were used to connect SM pathways (represented as nodes) that were coexpressed with each other. Nodes were named according to the SM classes each pathway belongs to. Each node has a suffix of “_X” where X represents a single pathway from an SM class. e.g., if a PWY-A belongs to Y number of SM classes, then there will be Y number of PWY-A nodes in the network (“PWY-A_1, “PWY-A_2”, …, “PWY-A_Y”) while each edge joins 2 significantly coexpressed pathways.

To assess the antimicrobial activity of M. dodecandrum, we sampled roots, petioles, leaves, vegetative branches, stems, fruits, and flowers from the Nanyang Technological University herb garden in Singapore ( Figure 1A ), and tested the antibacterial activity of methanol extracts derived from these organs. Bacterial growth inhibition studies indicated that the plant has antibacterial activity against gram-negative P. aeruginosa ( Figure 1B ; Table S2 ) and gram-positive S. aureus ( Figure 1C ; Table S2 ). When tested at 2% v/v (~4 mg/mL extract) against P. aeruginosa, the stems showed the highest activity with a 33.13% (ANOVA, p< 0.0001) inhibition in growth over 18 h as compared to the DMSO controls, followed by flowers (32.56%), fruits (27.9%), leaves (15.54%), and roots (11.26%) ( Figure 1B ). Similarly, the stems showed the highest activity against S. aureus with 50.21% growth inhibition (p< 0.05)( Figure 1C ). The vegetative branch, fruit, roots, and leaves showed moderate activity in a descending order ranging from 30.02 – 24.27% growth inhibition. In comparison, the flowers and petiole showed the lowest activity against S. aureus at 10.67 and 14.15% growth inhibition, respectively ( Figure 1C ).

Next, we evaluated the minimum inhibitory concentration (MIC), defined as the concentration at which no growth is observed after overnight incubation. While leaves did not have the strongest activity amongst all organs ( Figure 1B ), they were the most abundant and easiest to collect. Hence, leaf methanol extract was used to evaluate the MIC of the extract on P. aeruginosa via a broth microdilution method. Complete growth inhibition was achieved when the extract was used at 50 mg/mL ( Figures 1E, F , Table S4 ). The inhibitory effect was concentration dependent, with > 50% growth inhibition by 1.563 mg/mL of the extract compared to treatment with equivalent volumes of DMSO ( Figures 1E, F , Tables S4 , S5 ). While 5% v/v DMSO treatment resulted in longer lag times, growth resumed by 720 min ( Figure 1D ).

The leaf methanol extract was bactericidal at 50 mg/mL, with complete cell death within four hours after treatment ( Figure 1G , Table S6 ). In contrast, with the addition of equivalent 5% v/v DMSO, P. aeruginosa CFU recovered by 24 h (7.00 × 10⁹ CFU/mL) to levels comparable to that of untreated samples (5.65 × 10⁹ CFU/mL) ( Figure 1G ). A lower cell count was measured from 0 – 8 h with treatment with 25 mg/mL extract ( Figure S1 ), but the bacterial density reached untreated control levels by 24 h. Finally, P. aeruginosa treated with 4 mg/mL of the plant extract or equivalent concentrations of DMSO showed similar growth across all time points ( Figure S1 ), suggesting that a higher concentration of the plant extract is needed to achieve growth inhibition of bacteria.

To elucidate how the M. dodecandrum extracts affect the bacterial viability, we used LIVE/DEAD staining to assess potential membrane damage. The LIVE/DEAD stain consists of two nucleic acid staining dyes, propidium iodide (PI) and SYTO9. The green fluorescing SYTO9 can enter all cells, while the red fluorescing PI can only enter cells with compromised or damaged cytoplasmic membranes. PI, however, has a stronger affinity to DNA and hence can displace SYTO9 if present (Stiefel et al., 2015). While the kit is more accurately said to distinguish between cells with intact or compromised cell membranes, it is more often used to infer populations of living or dead cells under the assumption that all membrane-compromised cells are likely dead.

DMSO, when used at equivalent volumes of 0 – 0.8% v/v and a treatment time of 10 minutes, did not result in increased uptake of PI ( Figure 2A , nearly all cells are green), and the proportion of live cells was calculated to be at ~100% across all concentrations of DMSO ( Figure 2B ). In contrast, 10-minute treatment with 0 – 8 mg/mL of the plant extract resulted in a concentration-dependent increase in PI uptake and a corresponding reduction in the proportion of live cells from ~100% in untreated samples to 13.6% in 8 mg/mL treated samples ( Figures 2A, B , nearly all cells are red, Table S7 ).

Antibacterial compounds can also negatively affect the motility of bacteria, effectively disrupting their pathogenesis (Erhardt, 2016). Microscopy revealed reduced motility of cells treated with M. dodecandrum plant extract compared to the DMSO control groups, even at low concentrations of 1 mg/mL, where most P. aeruginosa cells appeared alive following 10 mins of treatment ( Video 1 and 2 for DMSO control and extract treatment, respectively). To quantify and evaluate the effect of plant extract treatment on P. aeruginosa motility, swarming and swimming motility assays were carried out ( Figures 2C–F ). Swarming motility was significantly impaired at 0.0625 mg/mL (p< 0.05) ( Figures 2C, D ; Table S7 ), while swimming motility was impaired at 0.25 mg/mL (p< 0.0001) ( Figures 2E, F ; Table S7 ). These results indicated that M. dodecandrum leaf extracts exhibit a multi-pronged effect on P. aeruginosa, inhibiting viability, growth, and motility.

Antibiotics can stimulate or depress gene expression in bacteria (Shitikov et al., 2022). Studying gene expression changes in responses to antibiotics can help the understanding of the mode of action of the various drug classes on bacterial adaptive ability, physiology, and metabolism. Since different antibiotics can elicit specific transcriptional responses (Hesketh et al., 2011), we hypothesized that gene expression analysis of P. aeruginosa-treated M. dodecandrum extracts and several antibiotics would allow us to propose the mode of action of the plant extracts.

To see if transcriptome changes caused by our plant extracts are similar to the changes caused by common antibiotics, we performed a comparative RNA sequencing analysis. We treated P. aeruginosa with rifampicin (RNA synthesis inhibitor), triclosan (fatty acid synthesis inhibitor), ceftazidime (cell wall synthesis inhibitor), gentamicin (protein synthesis inhibitor), M. dodecandrum extracts and DMSO control ( Table S8 ). Notably, rRNA depletion was not needed to obtain a high number (on average ~20 million) of reads mapping to protein-coding genes when increasing the sequencing depth to ~40 million reads ( Table S9 ), indicating that this expensive step can be skipped if sequencing depth can be adjusted.

Concentrations of extracts and antimicrobial compounds used for RNA-sequencing were selected based on results from microtiter plate growth curves ( Figure S2 ). While 25 mg/mL of the extract did not fully inhibit P. aeruginosa growth by 24 h ( Figures 1E, H ), it was selected over 50 mg/mL as it inhibited growth to a similar extent within 18 h of growth without strong inhibition of growth associated with higher DMSO concentrations at 50 mg/mL ( Figure 1F ). Rifampicin, ciprofloxacin, and gentamicin were used at 1 × MIC of 32, 0.5, and 2 µg/mL respectively ( Figure S2 ). Lastly, no growth inhibition was observed for triclosan over the range of concentration tested. Higher concentrations of triclosan treatment were not evaluated due to the precipitation of the compound in the media. As such, the highest concentration of 128 µg/mL, where some growth inhibition was observed, was used. DMSO was adjusted to the same concentration of 2.5% v/v for all samples to enable comparison with the control group. The resulting RNA-sequencing data was used to generate the gene expression matrix ( Table S10 ).

Interestingly, the hierarchical clustering analysis of the 18 RNA-seq samples revealed that M. dodecandrum extracts show the highest Euclidean distance to all other samples ( Figure 3A , M. dodecandrum RNA-seq samples form an outgroup). Similarly, rifampicin and triclosan also showed unique transcriptomic responses, while DMSO control, ceftazidime, and gentamicin displayed a similar response ( Figure 3A ). These responses could be explained by the number of significantly differentially up- (red) and down-regulated genes ( Figure 3B , BH-adj. p-value< 0.01). M.dodecandrum elicited the highest number of up- (712) and down-regulated (757) genes, followed by rifampicin (331 up, 283 down) and triclosan (250, 135)( Table S11 ). Conversely, ceftazidime (40, 13) and gentamicin (99, 13) caused the mildest transcriptomic response. Overall, we observed that the M.dodecandrum extracts showed some similarities of the differentially expressed genes with rifampicin and triclosan ( Figure S3 ).

To see which biological processes are differentially expressed, we performed a Gene Ontology enrichment analysis (see methods). We observed that the DEGs caused by triclosan resulted in the highest number of expressed GO terms ( Figure S4 ). These terms comprised many metabolic, biosynthetic, respiration, and macromolecule localization processes. All 8 GO terms enriched in the plant extract-treated samples overlap with triclosan and are mostly metabolic and biosynthetic processes (translation, peptide, amide, protein, and nitrogen compounds). Conversely, rifampicin only induced changes in the type VI secretion system ( Figure S4 ). Interestingly, gentamicin caused changes in terms involved in the high-temperature response (folding, response to heat, and temperature stimulus), while ceftazidime responses were external stimulus and DNA damage ( Figure S4 ). Taken together, these results indicate the M.dodecandrum extracts may have an overlapping mode of action with triclosan (fatty acid synthesis inhibitor).

To identify the antibacterial compounds in M.dodecandrum, we performed a bioassay-guided purification on the MeOH extract from the leaves. The extract was further fractionated by preparative reversed-phase HPLC and yielded 8 subfractions (a-h). Antimicrobial activity was observed in the Parent fraction and PPT, however, no activity was found in the a-h fractions ( Figure 4A ), indicating that the activity is caused by more than one synergistically-acting metabolites, or that the active metabolite was diluted during the fractionation.

To study which metabolites might be present in the active fractions, we compared the active material (Parent and PPT) to the three inactive fractions f-h, using ¹H NMR and LC-HRMS. HR-ESI-MS (High-resolution electrospray ionization mass spectrometry) of fractions f-h showed a major m/z of 1402.1149 [z=2]; 784.0776 [z=2] and 935.0793 [z=2], respectively ( Figure 4B ). A dominant fragment ion observed at m/z 301 indicated the existence of a hexahydroxydiphenoyl group (HHDP), which was characteristic of ellagitannins (Era et al., 2020). The MS/MS profiles were also in good agreement with that of hydrolyzable tannin casuarinin (m/z 935.0805 [z=1])( Figure 4C )(Chang et al., 2019). Coincidentally, fraction h presented the same m/z 935.0805, but a doubly charged [M-2H]^2-. Therefore, the mass of h was calculated to be 1872.1738 and the molecular formula was determined as C₈₂H₅₆O₅₂. From the predicted formula, fractions h, f (C₁₂₃H₈₂O₇₈), and g (C₆₈H₅₀O₄₄) were putatively identified as nobotanin analogs. The ¹H NMR spectra ( Figure 4D ) of the major compound in fraction H also resembles spectra containing nobotanins (Chang et al., 2019). However, as none of the eight fractions a-h is active, it is unlikely that the tannins are significantly conferring antibacterial activity.

The overlay of ¹H NMR spectra ( Figure 4D ) showed that the precipitation of the Parent fraction using 15% MeOH/H₂O removed most of the hydrolysable tannins. ¹H NMR of PPT revealed several unidentified signals in the Parent fraction, indicating a complex mixture consisting of phenolic compounds (9 - 11 ppm) and other unidentified compounds that could be responsible for the antibacterial activity.

To investigate the types of metabolic pathways in M.dodecandrum, we analyzed its 32021 annotated protein-coding genes with PathwayTools (Karp et al., 2016). The analysis revealed that 25.09% (8035) of the genes are enzyme-coding and 12.08% (3867) could be assigned to a pathway ( Table S12 ). Most of these genes were associated with general metabolism (GM, 3097 genes, 80.09% of all enzymes), followed by specialized metabolism (SM, 494 genes, 12.77%) and both GM/SM (276 genes, ~7%, Table S12 ). From the 463 predicted pathways, 104 (22.4%) were associated with SM ( Tables S12 , S13 ).

Next, we investigated the types of metabolites produced by the detected SM pathways. The majority of the compounds belonged to phenylpropanoids, of which flavonoids are most abundant ( Figure 5A ). Terpenoids, and more specifically diterpenoids biosynthesis comes in a close second. These observations are concurrent with previous studies which report these compounds in the larger M.dodecandrumcae family (Serna and Martínez, 2015).

To propose which specialized metabolites are most abundant in M.dodecandrum, we analyzed the gene expression of the SM pathways. Since SM pathways are thought to be under strong transcriptional control (Mutwil, 2020), we set out to investigate the gene expression patterns in M.dodecandrum. To this end, we isolated RNA from the same samples used to measure antibacterial activity ( Figure 1A ) and constructed a gene expression atlas for M.dodecandrum ( Figure S5 , Table S14 ). We then ranked the SM pathways within each organ based on its average TPM expression. To identify the 20 highest and lowest-ranked pathways, we ordered them according to the sum of their ranks in each organ ( Figure 5B ). The top 5 pathways included a sulfur-containing compound, two terpenoids, a phenylpropanoid, and a hormone synthesis pathway. There are some pathways that show organ-specific expression (e.g. young-flower-specific nitrogen-containing secondary compound, PWY-6442).

Since SM pathways expand and diversify by gene duplications (Hofberger et al., 2013), the predicted SM pathways might likely represent several separate, related pathways biosynthesizing related metabolites. For example, several enzymes implicated in flavonoid biosynthesis (PWY-7897, Figure 5C ) form an aerial root-specific cluster (green clade). This suggests that these genes biosynthesize a flavonoid in aerial roots, which might differ from flavonoids produced in flowers (first five enzymes, pink and brown clade, Figure 5C ).

To investigate whether these functional sub-clusters are frequently found within the predicted pathways, we calculated the gene expression correlation (captured by Pearson Correlation Coefficient, PCC) within a pathway. We assumed that the median PCC should be high if the enzymes within a pathway are involved in the biosynthesis of the same metabolite. Surprisingly, while some of the pathways showed a high average PCC (blue line, Figure S6 ), many pathways showed average PCC values like randomly chosen gene pairs (orange line, average PCC ~ 0.25). Conversely, using tissue-specific gene expression data to identify the clusters containing >75% of annotated enzymatic reactions within a pathway resulted in a much higher average PCC value for nearly all pathways (green line, Figure S6D ). This result shows that sequence similarity-based approaches such as PathwayTools can be supported by gene expression analyses to identify genes from tissue-specific pathways.

M.dodecandrum might express over 100 SM pathways ( Table S12 ) and metabolites ( Figure 4A ). Since these pathways should have specific functions, we expect the functionally-related pathways to be expressed at the same time and place. To identify such patterns, we investigate which pathways are more connected in the co-expression network than expected by chance. We first set a PCC threshold of ≥0.6, as the PCC values obtained from the randomized pathway assignments were all<0.6 ( Figure S5 , randomized pathway assignment confidence interval is shown in orange). We then counted the number of co-expression edges connecting any two pathways and calculated which pathways were more connected than expected by chance (see methods). Interestingly, we observed high connectivity between the different pathways, where, e.g., most phenylpropanoid, terpenoid, and hormone biosynthesis pathways were significantly connected, indicating transcriptional coordination between the pathways ( Figure 6A ). Between 93 pathways, there were 251 significant connections.

To take a closer look at the connections that are found within and across the pathways, we investigated the gene co-expression networks of two connected pathways. The first pathway involves the biosynthesis of the plant hormone ethylene, which has numerous roles in plant development and stress responses (Chang, 2016). Ethylene is synthesized from L-methionine by three enzymatic steps comprising methionine adenosyltransferase, 1-aminocyclopropane-1-carboxylate synthase (synthase), and 1-aminocyclopropane-1-carboxylate oxidase (oxidase) (Chang, 2016) Interestingly, the co-expression network of the ethylene biosynthesis pathway shows several groups of co-expressed synthases and oxidases ( Figure 6B , left pathway, blue edges connect co-expressed genes), suggesting that M.dodecandrum contains several ethylene biosynthesis pathways. The other pathway represents the biosynthesis of phytosterol terpenoids, which comprise multiple steps that convert cycloartenol to stigmasterol, crinosterol, and brassicasterol in Arabidopsis thaliana. This pathway in M.dodecandrum contains several dehydrogenases, methyltransferases, isomerases, demethylases, and desaturases that modify the cycloartenol (or a related compound) to other sterols ( Figure 6B , right pathway). While PWY-2541 has been characterized in Arabidopsis, it is likely that this pathway produces other, related compounds in M.dodecandrum. Within each pathway, we also observed subclusters of genes that are co-expressed with the same gene(s) in the other pathway. While the analysis does not reveal any clear organ-specific clusters like in Figure 4C , this could suggest separate clusters of SM pathway expansion where subclustering occurs not only within a pathway but with other pathways too. The co-expression edges found between the ethylene and phytosterol pathways indicate that the two pathways are transcriptionally positively coordinated. This suggests that an increase in ethylene might result in increased biosynthesis of phytosterols. This is in line with a recent study that showed that exogenous ethylene led to a profound change in the ratio of stigmasterol to sitosterol (Markowski et al., 2022).