Content uploaded by Ville-Petri Friman
Author content
All content in this area was uploaded by Ville-Petri Friman on Jan 16, 2025
Content may be subject to copyright.
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
1 of 15
MICROBIOLOGY
Siderophore synthetase- receptor gene coevolution
reveals habitat- and pathogen- specific bacterial iron
interaction networks
Shaohua Gu1,2,3, Zhengying Shao3, Zeyang Qu1, Shenyue Zhu3, Yuanzhe Shao2, Di Zhang1,
Richard Allen4, Ruolin He1, Jiqi Shao1, Guanyue Xiong1, Alexandre Jousset3, Ville- Petri Friman5,
Zhong Wei3*, Rolf Kümmerli4*, Zhiyuan Li1,2*
Bacterial social interactions play crucial roles in various ecological, medical, and biotechnological contexts. How-
ever, predicting these interactions from genome sequences is notoriously dicult. Here, we developed bioinfor-
matic tools to predict whether secreted iron- scavenging siderophores stimulate or inhibit the growth of community
members. Siderophores are chemically diverse and can be stimulatory or inhibitory depending on whether bacte-
ria have or lack corresponding uptake receptors. We focused on 1928 representative Pseudomonas genomes and
developed an experimentally validated coevolution algorithm to match encoded siderophore synthetases to cor-
responding receptor groups. We derived community- level iron interaction networks to show that siderophore-
mediated interactions dier across habitats and lifestyles. Specically, dense networks of siderophore sharing
and competition were observed among environmental and nonpathogenic species, while small, fragmented net-
works occurred among human- associated and pathogenic species. Together, our sequence- to- ecology approach
empowers the analyses of social interactions among thousands of bacterial strains and oers opportunities for
targeted intervention to microbial communities.
INTRODUCTION
Microbial communities populate all ecosystems on Earth from ter-
restrial to aquatic environments, aecting human health, agricul-
ture, and industry (1–3). e dynamics and functioning of these
communities are shaped by complex and unexplored interactions
between microorganisms (4,5). As the number of sequenced micro-
bial genomes continues to grow (6,7), there is enormous interest in
developing approaches to predicting microbial interaction networks
on the basis of genomic data. Such eorts are essential to obtain
complete insights into community functioning as many microor-
ganisms cannot be cultured in the laboratory (8), while their eco-
logical roles could still be inferred through sequence- to- interaction
mapping. Now, sequence- to- interaction mapping approaches pri-
marily focus on metabolic interactions, with genome- scale metabolic
models serving as the primary tool for establishing the pan- reactome
of microbial communities (9, 10). ese methods infer metabolic
reactions from the genome annotation of enzymes and then recon-
struct a ux model to understand how microorganisms take up
essential nutrients and release metabolic by- products into the envi-
ronment (11–13).
Despite the signicance of primary metabolism, there is in-
creasing evidence that also other secreted compounds synthesized
through secondary metabolism (14) play a major role in shaping mi-
crobial interactions (15,16). Nearly all microbes actively synthesize
and secrete compounds to fulll a diverse set of functions, including
communication, resource scavenging, motility, and attack of and de-
fense against competitors (17). Many of these secreted compounds
were previously considered nonessential for microbial growth in
laboratory settings but have since been shown to be critical for com-
petitiveness in natural environments (15,16). However, sequence-
to- interaction mapping has rarely been applied to secreted compounds,
particularly because the synthesis and mode of action of secondary
metabolites are challenging to predict.
Here, we developed a secondary metabolite sequence- to- interaction
approach focusing on iron- scavenging siderophores, one of the most
prevalent and diverse classes of microbial secondary metabolites (18).
Iron is critical for microbial growth and survival because of its impor-
tance as a catalytic group in enzymes guiding key biological processes
such as respiration and DNA replication (19). However, the concentra-
tion of bioavailable iron is typically below the required level in most
habitats (19–21), and upon iron limitation, nearly all bacteria produce
siderophores that eciently chelate iron from insoluble environmental
stocks (22,23). Siderophores are typically diusible and able to chelate
iron over a broad physical range (24). Once iron is bound, the complex
is recognized and taken up by specic receptors embedded into the
bacterial cell membrane (23). Diusible siderophores mediate several
types of social interactions. ey can cooperatively be shared among
contributing bacterial strains with matching receptors for the uptake of
the iron- siderophore complex (23,25). Siderophores can also be de-
ployed as competitive agents to limit access to iron if co- occurring
strains lack matching receptors (23,26). Last, siderophores can be ex-
ploited by cheater bacteria that have receptors for siderophore uptake
but do not pay the cost of producing siderophores themselves (22,23).
Consequently, while siderophore- mediated interactions have impor-
tant impacts on microbial community dynamics and functions (27–
30), we still poorly understand how these interactions scale up at
the network level in environmental and host- associated bacterial
populations.
1Center for Quantitative Biology, Academy for Advanced Interdisciplinary Studies,
Peking University, Beijing 100871, China. 2Peking- Tsinghua Center for Life Sciences,
Academy for Advanced Interdisciplinary Studies, Peking University, Beijing 100871,
China. 3Jiangsu Provincial Key Lab for Organic Solid Waste Utilization, Key Lab of
Organic- based Fertilizers of China, Nanjing Agricultural University, Nanjing, P. R. China.
4Department of Quantitative Biomedicine, University of Zurich, Winterthurerstr.
190, 8057 Zurich, Switzerland. 5Department of Microbiology, University of Helsinki,
00014 Helsinki, Finland.
*Corresponding author. Email: weizhong@ njau. edu. cn (Z.W.); rolf. kuemmerli@ uzh.
ch (R.K.); zhiyuanli@ pku. edu. cn (Z.L.)
Copyright © 2025 The
Authors, some rights
reserved; exclusive
licensee American
Association for the
Advancement of
Science. No claim to
original U.S.
Government Works.
Distributed under a
Creative Commons
Attribution License 4.0
(CC BY).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
2 of 15
e aim of our study was to infer how receptors and siderophores
have coevolved and use this information to develop algorithms that
identify matching siderophore- receptor pairs that predict interaction
networks in bacterial communities on the basis of sequence data.
We previously used the genome sequences of 1928 Pseudomonas
strains to develop bioinformatic pipelines that allowed us to predict
the chemical structure of 188 pyoverdines (the main siderophores of
this genus) and to identify 4547 FpvA- receptor genes (segregating
into 94 groups) involved in pyoverdine uptake (31). Here, we capi-
talize on this work to develop a coevolution pairing algorithm to
match the pyoverdines (key) and receptors (lock) into 47 unique
lock- key groups and over 90% of these predicted interactions could
be validated experimentally in vivo. Using the predicted lock- key
pairs, we then reconstructed siderophore- mediated iron interaction
networks among all Pseudomonas strains. We found that network
complexity was high among strains isolated from soil- , water- ,
and plant- derived habitats, whereas complexity was lower among
strains isolated from human- associated habitats. We further no-
ticed that interaction networks among pathogenic species were
small and loose with few pyoverdine interactions existing between
strains. e opposite was the case for strains from environmental
habitats. Together, the developed sequence- to- interaction mapping
tool can accurately predict social interaction networks mediated by
siderophores in complex bacterial communities. Our ndings sug-
gest that selection for social interactions varies across habitats and
lifestyles, thus providing valuable insights into community func-
tions and connectivity.
RESULTS
Three classes of pyoverdine strategies in Pseudomonas
strains and the lock- key (receptor- synthetase) principle
of coevolution
Our dataset consists of 1928 Pseudomonas strains, producing a total
of 188 chemically dierent pyoverdine types and featuring 94 dier-
ent receptor groups according to our recently developed bioinfor-
matic prediction tools (31). Our dataset contains 1928 unique
nonredundant strains as we previously deduplicated the dataset by
removing strains with high phylogenic similarity and high similari-
ty in pyoverdine synthetases. Of the 1928 nonredundant genomes,
403 are complete and 1525 are incomplete.
We rst explored the diversity of strains in terms of phylogeny,
ecological habitat, and pyoverdine strategies. At the phylogenetic
level, our dataset included a diverse set of Pseudomonas species,
where P. aeruginosa (28.7%), P. uorescens (7.0%), P. syringae (6.0%),
and P. putida (2.2%) were the most abundant ones (Fig. 1A). e
strains originated from diverse habitats, including human- derived
habitats (21.2%), soils (13.6%), plants (12.1%), and water (6.4%), al-
though the origin of many strains (39.5%) is unknown (Fig. 1A).
While our study focuses on pyoverdine, the widespread primary sid-
erophore of uorescent Pseudomonas spp., we next checked whether
secondary siderophores also occur among the 1928 strains. We found
that synthetase clusters for secondary siderophores were relatively
rare and occurred only in 392 strains (pyochelin: 15.1%; yersiniabac-
tin: 3.2%; pseudomonine: 1.8%; quinolobactin: 0.15%; desferriox-
amine: 0.05%; table S1). Given their rarity and the fact that secondary
siderophores have lower iron anity than pyoverdines (32,33), we
can reasonably exclude the possibility that secondary siderophores
have a meaningful inuence on the iron networks studied here.
To assess the diversity of pyoverdine production and uptake
strategies, we analyzed the absence or presence of pyoverdine syn-
thesis clusters and counted the number of FpvA pyoverdine re-
ceptors per strain. We found three types of pyoverdine- utilization
strategies (Fig. 1B). “Single- receptor producers” were the most com-
mon type (986 strains, 51.1%) and refer to strains with one pyover-
dine synthesis locus and one FpvA receptor gene. “Multireceptor
producers” were the second most common type (678 strains, 35.2%)
and include strains with one pyoverdine synthesis cluster but mul-
tiple FpvA receptor genes. “Nonproducers” were the least common
type (264 strains, 13.7%) and refer to strains that lack the pyover-
dine synthesis cluster but contain at least one receptor gene. While
strains can have multiple FpvA receptor genes, no strain carried
more than one pyoverdine synthesis cluster. is observation is in
line with the expected high costs of pyoverdine synthesis, which is
based on a series of gigantic modular enzymes known as nonribo-
somal peptide synthetases (34).
On the basis of these ndings, we hypothesized that in each
single- receptor producer, the sole receptor present should recognize
the self- produced pyoverdine to ensure tness benets when faced
with iron limitations. Consequently, synthetase and receptor pairs
should reect molecular coevolution, where mutational alterations
in the synthetase structure should select for corresponding changes
in the receptor sequences to preserve the lock- key relationship and
ecient iron uptake. To test this hypothesis, we focused on the 986
single- receptor producers and calculated the degree of covariation
between sequence distance matrices of the receptor, the synthesis
cluster, and 400 conserved genes. For the receptors (FpvA) and the
synthesis cluster, we used the feature sequences that are most pre-
dictive of receptor specicity and pyoverdine molecular structure,
as identied in our previous work (31). We found a strong correla-
tion between the distance matrices of the receptors and the synthe-
sis clusters (Pearson’s r= 0.88) and an intermediate correlation
between the receptor and the conserved genes (Pearson’s r=0.51)
(Fig. 1C). ese results provide strong evidence for an intimate co-
evolution between pyoverdine receptors and synthetases that greatly
exceeds the baseline phylogenetic eect. Notably, we observed strong
clustering patterns in the sequence space of the receptors, forming
distinctive blocks that closely match with the clustering patterns of
their corresponding pyoverdine synthesis clusters. Using our recep-
tor clustering pipeline (31), we identied 17 receptor groups among
the 986 single- receptor producers. ree out of the 17 receptor
groups represent the FpvA receptors found in the human pathogen
P. aeruginosa (labeled as GI, GII, and GIII + IV in Fig. 1C, le panel)
for which the selective uptake of the corresponding pyoverdines has
been demonstrated (22,35). ese analyses strongly indicate that
cognate receptors and synthesis genes have coevolved in single-
receptor producers, resulting in one- to- one “lock- key” relationships.
Coevolutionary lock- key groups cannot directly be inferred for
multireceptor producers because there is currently no method to
distinguish the “self- receptor” responsible for absorbing the self-
produced pyoverdine from the other FpvA receptors responsible for
the uptake of heterologous pyoverdines produced by other strains.
Moreover, receptor diversity seems to be much larger among multi-
receptor producers than among single- receptor producers. When
focusing on the 43 (of 94) receptor groups with more than 10 mem-
bers, we found that single- receptor producers covered only 14 of these
groups (32.6%), while multireceptor producers had a much more di-
verse receptor coverage (41 groups, 95.3%) (Fig. 1D). Nonproducers
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
3 of 15
Fig. 1. Classication of Pseudomonas strains and elucidation of the coevolution between pyoverdine synthetases and receptors. (A) Phylogenetic relationship
among the 1928 Pseudomonas strains based on the concatenated alignment of 400 single- copy conserved genes. Starting from inside, colors in the rst ring distinguish
the ve most prevalent species, with “Others” represent the remaining less abundant species. Colors in the second ring distinguish the four most prevalent sources of
isolation. In the third ring, claret and blank regions cover strains with complete pyoverdine synthetase clusters and strains without synthetase gene clusters, respectively.
In the fourth blue ring, the bar height indicates the number of FpvA receptors present in each strain. (B) Strains can be classied into three types by scoring the presence/
absence of a synthetase cluster and counting the number receptors in each genome: (i) single- receptor producers containing one pyoverdine synthetase cluster and one
FpvA receptor gene, (ii) multireceptor producers containing one pyoverdine synthetase cluster and several FpvA receptor genes, and (iii) nonproducers lacking a synthe-
tase gene but containing at least one receptor gene. (C) Heatmap visualizing distances between feature sequences of the FpvA receptors and pyoverdine synthetase
clusters and between FpvA feature sequences and phylogenetic genes among the 986 single- receptor producers. In all three heatmaps, the hierarchical clustering of the
strains follows the one used for the FpvA feature sequences (left panel). The black squares on the heatmaps denote the ve major FpvA groups. Three of these groups
correspond to the receptors found among P. aeruginosa strains and are labeled with black text. (D) Forty- three largest FpvA receptor groups with more than 10 members
(sorted by group size) and their frequency among single- receptor producers, multireceptor producers, and nonproducers.
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
4 of 15
had a similarly broad receptor coverage (34 groups, 79.1%). We also
found that single- receptor producers tend to connect more com-
pactly (mean silhouette index= 0.96 ± 0.16) than multireceptor
producers (mean silhouette index=0.78±0.19) and nonproducers
(mean silhouette index=0.79±0.20) in the sequence space (gs. S1
and S2), suggesting that receptors from single- receptor producers
might evolutionarily be more conserved than receptors from non-
producers and multireceptor producers. Together, these observa-
tions imply that single- and multireceptor producers take on dierent
roles in iron interaction networks.
Development of the coevolution pairing algorithm and
experimental validation for predicting iron
interaction networks
e aim of this section is to establish a lock- key receptor- pyoverdine
interaction map across all three strain types. A rst task in this pro-
cess is to identify receptors in multireceptor producers that are used
to take up the self- produced pyoverdine. A rst intuitive approach
was to check for receptors proximate to the pyoverdine synthetase
genes (solution 1), while an alternative approach was to use the
lock- key pairs identied for single- receptor producers and analyze
whether similar pairs occur in multireceptor strains (solution 2).
Even aer completing these two solutions, more than half of the re-
ceptor groups could not be paired with any pyoverdine synthetases.
Specically, solution 1 identies putative self- receptors in 87.1%
(591 of 678) of the multireceptor producers, while solution 2 could
be applied to only 68.7% (466 of 678) of multireceptor strains.
We thus developed an unsupervised learning algorithm, termed
the “coevolution pairing algorithm” (solution 3), which searches for
the set of synthetase- receptor combinations that maximizes coevo-
lutionary association on the basis of feature sequence distance be-
tween the synthetase and the receptor (Fig. 2, A and B; see the
“Using the coevolution relationship between synthetases and recep-
tors to identify self- receptors in producers” section for details). e
algorithm starts with a random association between synthetases and
receptors, resulting in a rst correlation between the two matrices
[for example, correlation coecient (cr)= 0.17]. is is then fol-
lowed by an iteration process that improves the association strength
of the two matrices until an optimized correlation is reached (for
our dataset, cr=0.85). Note that the cr value is expected to change
depending on the dataset used. We then checked for consistency in
self- receptor identication across the three solutions (Fig. 2C): (i)
the receptor is within 20,000–base pair (bp) proximity to the syn-
thetase; (ii) the lock- key pairs of single- receptor producers are ap-
plied to multireceptor producers; (iii) an unsupervised coevolution
pairing algorithm. Solution 1 and solution 2 yield high levels of con-
sistency (99.5% across 433 strains). e unsupervised solution 3
shares high consistency with solution 1 (93.7%, across 591 strains)
and solution 2 (94.4%, across 466 strains), indicating that all three
solutions are legitimate, with solution 3 having the advantage of be-
ing applicable to all strains.
Our coevolution pairing algorithm allocated the self- receptors of
the 1664 pyoverdine- producing strain into 47 distinct lock- key
groups on the basis of the receptor feature distance (Fig. 2D and g.
S3). Most self- receptors belonged to 17 lock- key groups (single- receptor
producers: 986=100%; multireceptor producers: 572=84.4%), while
the remaining self- receptors of multireceptor producers (106=15.6%)
segregated into 30 additional receptor groups (g. S3). Of the total
4547 FpvA genes detected, we identied 2883 receptors that are not
self- receptors and thus possibly serve as “cheating receptors” to take
up heterologous pyoverdines produced by other strains. Most of
these cheating receptors (2703=93.8%) also segregated into the 47
lock- key groups, conrming that they could be used to exploit at
least one of the 188 produced pyoverdines. e remaining cheating
receptors (180=6.2%) could not be linked to any of the 47 lock- key
groups, suggesting the existence of rare receptor groups that pre-
sumably match rare pyoverdine structures not covered by our data-
set (g. S3).
By combining the coevolution pairing algorithm and the lock-
key pairing, we predicted siderophore- mediated iron interaction
networks on the basis of pyoverdines that were produced and could
be taken up by corresponding receptors by community members.
We conducted two experiments in two dierent laboratories to vali-
date predicted interactions using the same set of methods to evaluate
siderophore- mediated microbial interactions. For the rst valida-
tion, we used a Pseudomonas community from the Laboratory of
Rhizosphere Microbial Ecology (LorMe) in Nanjing (China), which
was originally isolated from the tomato rhizosphere (29). We in-
cluded 24 independent strains and subjected their genomes to our
bioinformatic pipelines to predict pyoverdine structures, to nd all
FpvA receptors (31), to identify self- receptors, and to allocate py-
overdines and receptors into lock- key groups. We found that these
24 strains included 4 single- receptor producers, 16 multireceptor
producers, and 4 nonproducers (g. S4) and that their self- receptors
could be allocated to 13 lock- key groups (g. S5). With this infor-
mation, we predicted the pyoverdine- mediated iron interaction net-
work between the 24 strains (Fig. 2E). For experimental validation,
we conrmed the pyoverdine production status of the 20 putative
producers and 4 nonproducers (g. S4). We then followed a modi-
ed version of our previously established protocols to calculate the
net eect pyoverdine has on the growth of other strains (GEPyo)
while controlling for the eects of other metabolites in the superna-
tant (29). is approach allowed us to obtain an experimentally de-
rived pyoverdine- mediated iron interaction network (Fig. 2E and
g. S6). We found that 90% of the observed interactions (whether
strains are stimulated or inhibited by the pyoverdines of others) matched
the computationally predicted interactions from sequence data.
e second experimental validation involved strains from the
Zurich (Switzerland) collection, isolated from soil and freshwater
habitats (23). In this case, we used published experimental data from
the literature (36). e focus of this earlier study was to test whether
the opportunistic human pathogen P. aeruginosa PAO1 can invade
natural soil and pond communities on the basis of its ability to use
pyoverdine produced by the natural isolates. We used data from all
the strains for which genome sequences were available (PAO1 and
33 natural isolates) to establish pyoverdine- mediated iron interac-
tion networks (Fig. 2, F and G). We then applied our bioinformatic
pipelines as explained for the Nanjing collection and found a high
level of consistency (94%) between the predicted and observed
pyoverdine- mediated iron interactions (Fig. 2, F and G).
Together, the two validation experiments demonstrate that
siderophore- mediated microbial interactions can accurately be pre-
dicted on the basis of genome- sequence analysis using the lock- key
relationship between receptor and synthetase genes. A closer inves-
tigation of the confusion matrix for the Nanjing experiment showed
that the 10% inaccuracies (i.e., wrongly predicted interactions) pri-
marily consisted of false negatives (n=36, 78.3%) but few false
positives (n=10, 21.7%). is means that our algorithm has high
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
5 of 15
Fig. 2. Unsupervised coevolutionary algorithm to establish a lock- key pair map for pyoverdine synthetase and receptor groups and its experimental validation.
(A) The cartoon owchart depicts the coevolution pairing algorithm (solution 3) to match the synthetase in each strain to its “self- receptor” on the basis of an unsupervised
learning scheme that optimizes coevolutionary strength (cr values) between the feature sequence distance matrices of synthetases and matched receptors. (B) Progres-
sion of the cr across 3 × 104 iterations based on 50 independent runs. (C) Consistency of identifying the same receptors as self- receptors across three dierent solutions.
(D) Predicted lock- key pairs connected in sequence space: 1664 synthases (bottom left) linked to the 4547 receptors (top right). Arrows depict the 47 lock- key links be-
tween synthetase and receptor groups. The colored and black/gray- shaded lines represent groups with and without single- receptor producers, respectively. (E) Predicted
versus observed iron interaction network among the 24 experimental strains. Each circular node represents an experimental strain, and line colors stand for single- receptor
producers (green), multireceptor producers (yellow), and nonproducers (red). Hexagons represent the predicted 13 lock- key receptor- pyoverdine groups. Edges from
strain nodes to lock- key nodes represent pyoverdine production, while edges from lock- key nodes to strain nodes represent utilization. Green (nonusable pyoverdine) and
pink (usable pyoverdine) edges depict cases in which experimental observations match predicted interactions, while blue edges depict incorrect predictions. The pyover-
dine groups that are produced by at least one single- receptor producer are colored by the respective receptor group (Fig. 1C), whereas pyoverdine groups that are exclu-
sively produced by multireceptors are colored gray. (F to G) Predicted versus observed iron interaction networks based on data from a previous study carried out in the
Zurich lab. The predicted interactions were inferred by the algorithms presented in this study, while the experimental data are taken from table S2 of Figueiredo etal. (36).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
6 of 15
specicity (97.3%) but relatively low sensitivity (67.9%). Biological-
ly, this means that there might be additional siderophore receptor
groups present among our strains, which we did not capture.
Variation of pyoverdine- mediated iron interaction networks
across habitats
Following the successful validation, we applied the lock- key pairing
methodology to our full dataset to construct the pyoverdine- mediated
iron interaction network among all 1928 Pseudomonas strains (Fig. 3A).
To keep traceability in such an enormous network, we allocated
strains into microbial siderophore functional groups, defined as
strains that produce the same pyoverdine types and can use the same
repertoire of pyoverdines. Overall, the network featured 407 micro-
bial siderophore functional groups, 47 dierent lock- key receptor-
pyoverdine groups, 307 production edges, and 1788 utilization edges.
Our iron interaction network can be considered as a special ver-
sion of a bipartite network, characterized by two types of nodes (mi-
crobial functional groups and pyoverdine groups) and two types of
directional edges (utilization and production). In ecological bipar-
tite networks such as those associated with pollination and food
webs, topology plays a crucial role for ecological functions and com-
munity assemblies (37). e topological metrics of our network re-
veals considerable heterogeneity. Specically, the in- and out- degree
distribution of pyoverdine nodes is heavy tailed, indicating “hub”
siderophores that are either produced or used more extensively than
others (Fig. 3, B and C). is high degree of heterogeneity suggests
a nonrandom network structure, likely inuenced by specic eco-
evolutionary forces (e.g., erce competition for iron). We observed
higher than expected yet moderate modularity (Qb=0.51, com-
pared to Qb= 0.41 in a randomized network) and nestedness
(NODF =0.15, compared to NODF=0.04 in a randomized net-
work). ese values dier from mutualistic networks that typically
feature high nestedness (e.g., pollination networks), while antago-
nistic networks are typically characterized by high modularity (e.g.,
herbivory and host- parasite networks) (37,38). Together, the iron
interaction network may represent a distinct type of ecological bi-
partite network and stands out as one of the largest reported to date.
Next, we created separate networks for strains isolated from soil
(262 strains), plant (234), water (124), and human- derived (409)
habitats. We found that frequencies of the three pyoverdine strate-
gies and network topologies varied fundamentally between the four
habitats (Fig. 3, D to G). For example, among the soil- derived strains,
there were 56.9% multireceptor producers, 27.5% single- receptor
producers, and 15.7% nonproducers (table S2). In contrast, there
were only 10.0% multireceptor producers and 4.0% nonproducers
but 86.1% single- receptor producers among human- derived strains.
Regarding network topologies, we observed that the number of mi-
crobial functional groups was higher for soil (130, value scaled to
number of strains=0.50), plant (97, 0.41), and water (70, 0.56) hab-
itats than for human- related habitats (41, 0.10). Moreover, many
functional groups (60.7%) exclusively occurred in a single habitat:
soil (80, 23.7%), plant (56, 16.6%), water (43, 12.7%), and human
(26, 7.7%), whereas only 8 functional groups (g. S7A) and 11 py-
overdine groups (g. S7B) were present in all four habitats. One pos-
sible reason for why the networks may dier is that the phylogenetic
diversity (PD) varies across habitats. We found no evidence for such
a direct association. PD was not higher in habitats with more complex
networks (soil, PD=78.1; plant, PD=52.8; water, PD=39.1) than
in human- derived habitats (PD=61.2) featuring simpler networks.
Together, these results indicate that iron interaction networks seem
to evolve dierently across habitats, whereby the underlying factors
shaping these dierences need further investigation.
Variation of pyoverdine- mediated iron interaction networks
between pathogenic and nonpathogenic species
Pseudomonas spp. do not only populate a variety of habitats but can
also display diverse lifestyles. e most prominent division in life-
style occurs between pathogenic and nonpathogenic species. Here,
we explore whether iron interaction networks dier between these
two lifestyles. We allocated strains to pathogenic and nonpathogen-
ic species groups and predicted the iron interaction networks for all
species with more than ve strains (Fig. 4A and table S3). e most
abundant pathogenic species were the human pathogen P. aeruginosa
and the plant pathogen P. syringae, while the most abundant non-
pathogenic environmental species were P. uorescens and P. putida,
of which many are neutral or even benecial for hosts.
We observed multiple dierences in network topologies between
the two lifestyles (Fig. 4A). First, strains of pathogenic species were
mostly single- receptor producers or nonproducers (Fig. 4, A and B),
while strains of nonpathogenic species were primarily multireceptor
producers (Fig. 4, A and B). Second, the diversity of siderophore
functional groups was much lower in pathogenic species compared
to nonpathogenic species (Fig. 4C and table S2). Consequently, the
complexity of iron interaction networks (quantied by the entropy
of their functional groups) was lower in pathogenic species than in
nonpathogenic species (Fig. 4C). For example, P. aeruginosa (the
most abundant species in our dataset, 554 strains) has a simple inter-
action network with only three functional groups, while P. uorescens
(135 strains) has a complex interaction network with 13 functional
groups (table S2). Last, we calculate the synthetase count versus re-
ceptor count (termed as “Syn/Rec”) value dened as the ratio of syn-
thetase and receptor groups present in each lock- key group (Fig.
4E). A Syn/Rec ratio near one indicates that there are no cheating
receptors in this lock- key group, i.e., this pyoverdine is exclusive to
its producers and cannot be used by strains not producing it. Con-
versely, a Syn/Rec ratio near zero means that most receptors in this
lock- key group are cheating receptors, and the corresponding py-
overdine is more sharable and exploitable.
We found that the Syn/Rec ratio was much lower in nonpatho-
genic species than in pathogenic species (Fig. 4F). Furthermore, the
Syn/Rec ratio diered substantially across lock- key groups (Fig. 4E).
For example, the lock- key groups 39 (purple hexagon, present in
P. aeruginosa) and 35 (cyan, present in several plant pathogens, Fig.
4A) had high Syn/Rec ratios of 0.93 and 0.89, respectively. In contrast,
the lock- key group 94 (light green, present in P. uorescens) had a
much lower Syn/Rec ratio of 0.20. Collectively, the proportion of
multireceptor producers and the preference for producing shareable
or exclusive siderophores clearly dierentiated between pathogenic
and nonpathogenic species on the two- dimensional plane (Fig. 4G).
Together, these results show that nonpathogenic species form open
networks dominated by shareable pyoverdine lock- key groups, where-
as pathogenic species form more closed networks dominated by ex-
clusive pyoverdine lock- key groups. As in the section above, we tested
whether the simpler networks observed among pathogenic species is
explained by lower PD. Again, we found no evidence for such an as-
sociation (nonpathogenic species: P. uorescens, PD=31.5; P. putida,
PD=13.3/pathogenic species: P. aeruginosa, PD=71.5; P. syringae,
PD=18.3).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
7 of 15
Fig. 3. Pseudomonas iron interaction networks vary across habitats. (A) The predicted iron interaction networks mediated by pyoverdines among 1928 Pseudomonas
strains. Circular nodes represent functional groups (i.e., strains that produce the same pyoverdine type and use the same repertoire of pyoverdines) with node size being
proportional to the number of strains within this functional group. Line colors of circular nodes represent single- receptor producers (green), multireceptor producers
(yellow), and nonproducers (red). Hexagonal nodes represent lock- key pyoverdine groups with node size being proportional to the number of strains using the corre-
sponding pyoverdine. The hexagons of the 17 pyoverdine groups found among single- receptor producers are highlighted with the corresponding receptor group colors
(as shown in Fig. 1C), while the pyoverdine groups that are exclusively found among multireceptor producers are depicted by gray hexagons. Edges from circular to
hexagonal nodes represent pyoverdine production, while edges from hexagonal to circular nodes represent pyoverdine utilization (with edge color matching the color
of the functional group). (B) In- and (C) out- degree distribution of the pyoverdine nodes. The in- and out- degrees are dened by the number of edges pointing toward
(representing production) or originating from (representing utilization) a pyoverdine node. (D to G) Iron interaction networks of strains isolated from soil, plant, water and
human habitats. The color bars on the right of each panel show the proportion of single- receptor producers (green), multireceptor producers (yellow), and nonproducers
(red). Node symbols and colors are the same as in (A).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
8 of 15
Fig. 4. Network properties dier between pathogenic and nonpathogenic Pseudomonas spp. (A) Pyoverdine- mediated iron interaction network for pathogenic (top
row) and nonpathogenic (bottom row) species with more than ve strains in our database. Symbols and color codes match the ones in Fig. 3A. (B and C) Dierence of the
percentage of multireceptor producers and the number of the microbial functional groups between dierent species lifestyles. Pathogenic species are colored red, non-
pathogenic species are colored blue, and species that do not have a clear role in pathogenicity are colored gray. (D) Dierence of diversity quantied by the entropy of its
functional group frequencies between dierent species lifestyles. (E) Counts of annotated receptors (x axis) and annotated synthetases (y axis) for each lock- key group.
Lock- key groups are colored the same as those in Fig. 3A. The gray dotted line represents 0.1 and 0.9 Syn/Rec ratios. (F) Dierence of the weighted Syn/Rec ratio for each
species between dierent species lifestyles. (G) The percentage of multireceptor producers (y axis) and the weighted Syn/Rec ratio (x axis) for each species clearly separate
pathogenic and nonpathogenic strains (segregation illustrated by the dashed line). The size of the dot represents its diversity quantied by the entropy of its functional
group frequencies. Same color coding as that in (B).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
9 of 15
Modeling the relationship between pyoverdine utilization
strategies and community dynamics
To explore the connection between pyoverdine utilization strategies
of individual strains and the resulting community dynamics, we
built simple ecological models of siderophore- mediated ecological
competition (Fig. 5A; see the Supplementary Materials for details)
(39). Our chemostat- type model is an extension of classical resource-
consumer models (40–42). It simulates the process during which
various types of siderophore producers secrete their specic sidero-
phores into the environment, which may then be used by the mi-
crobes in the system to uptake iron for their growth. We randomized
key microbial parameters, such as growth budgets and resource par-
titioning strategies, among the individual members of the commu-
nity to avoid dependencies on specic parameters. To achieve model
reliability, we ran repeated simulations to gather statistical insights
into community composition and dynamics. Note that our model
considers bacterial agents that dier in their siderophore strategies.
e strategies and situations modeled could apply to any microbes
and are not limited to pseudomonads.
Our model revealed that dierent single- receptor producers can-
not coexist, consistent with the competitive exclusion principle (39)
and matching our observation that a high proportion of single-
receptor producers was associated with simple network structures
(Fig. 4A). In a well- mixed environment, competitive exclusion dic-
tates that the number of stably coexisting species cannot exceed the
number of limiting nutrients. Single- receptor producers only ab-
sorb the siderophores they produce, and their interactions with
other single- receptor producers are mediated by competition for
Fig. 5. Mathematical model exploring the relationship between pyoverdine utilization strategies and community dynamics. (A) Schematic diagram of the iron
competition model. Each microbe
i
can produce siderophores of type
j
with resource budget
αi
and obtain iron by absorbing siderophore- iron complexes through cor-
responding receptors (fraction of receptors denoted as
vij
). Growth rate is proportional to the total amount of absorbed iron and the fraction of resources allocated to
primary metabolism,
1−α
i
. In each simulation, 20 random strains are assembled into communities and then compete in a chemostat- like model until a steady state is
reached. (B) Distribution of the number of species at the steady state for communities initiated with a single strategy type: single- receptor producers (green), multirecep-
tor producers (yellow), and nonproducers (red). (C) Schematic diagram depicting the invasion scenario together with the dierent possible outcomes. Putative invaders
are introduced at low frequency into a local community at the steady state. (D) Examples illustrating common invasion dynamics. (Top) A local community with three
species (various blue shades) is invaded by a single- receptor producer (purple). (Middle) A local community is invaded by a multireceptor producer. (Bottom) Summary
statistics showing the proportion of cases, in which invaders completely wiped out the local community as a function of their pyoverdine utilization strategy. a.u., arbitrary
units. (E) Likelihood of a single- receptor producer to invade a local community of multireceptor producers for four dierent iron interaction networks. Scenarios from top
to bottom show the following: (i) no overlap in siderophore utilization between the local community and the invader; (ii) the multireceptor producer from the local com-
munity can cheat on the invader’s siderophore; (iii) the invader and one multireceptor producer share the same siderophore; (iv) scenario combining (ii) and (iii). Schemat-
ics are simplications and do not reect the actual number of species in the local community.
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
10 of 15
free iron. Without cross- utilization, iron remains the sole limiting
nutrient, thereby allowing only a single species to persist in the com-
munity. In contrast, multireceptor producers were more likely to
coexist on the basis of our model (Fig. 5B), mirroring our observa-
tion that a high proportion of multireceptor producers correlated
with increased network diversity and complexity (Figs. 3 and 4A).
Last, nonproducers could not exist in the absence of producers in
our models.
To delve deeper into the connection between pathogenicity and
pyoverdine utilization strategies, we simulated the consequences of
invasion by introducing a strain (“invader”) with a higher resource
budget (available for intrinsic growth and siderophore production)
into a steady- state community (“local community”) (Fig. 5C). Across
numerous parameters sets, we consistently observed that invasions
by single- receptor producers and nonproducers tended to disrupt local
communities (Fig. 5D). Successful single- receptor producers typically
wiped out the local community. Conversely, successful multireceptor
producers integrated themselves into and coexisted with the local com-
munity (Fig. 5D). Briey, the more severe consequences of invasion
provide a plausible explanation for the association between certain
strategies and pathogenicity.
Invasion probability further depended on the pyoverdine lock-
key relationship between the invader (here modeled as a single-
receptor producer, e.g., matching a pathogenic lifestyle) and the
local community (Fig. 5E). e probability of successful invasion
was close to zero if the local community could cheat on the invader’s
siderophore, which is more likely when the invader produces sider-
ophores with low Syn/Rec ratio. Conversely, invasion probability
was high if the invader could produce an exclusive siderophore or if
one of the local community members produced the same sidero-
phore as the invader (i.e., enabling cheating of the invader). Together,
our modeling results reveal that the relative proportion of single-
and multireceptor producers and their interactions via siderophores
has fundamental consequences for community diversity and net-
work complexity.
DISCUSSION
Predicting interactions between microbes from sequence data oers
exciting opportunities for understanding microbiome assembly and
stability and may lay the foundation of biotechnological and medical
microbiome interventions. While sequence- to- interaction mapping
has predominantly been carried out for primary metabolism involving
resource consumption, conversion, and cross- feeding, there are few
approaches to reconstructing microbial interactions on the basis of
secondary metabolites (antibiotics, toxins, siderophores, and surfac-
tants) (43–45). In our paper, we oer such an approach by developing
a coevolution- inspired computational approach to infer iron interac-
tion networks mediated by pyoverdines (a class of iron- scavenging
siderophores) within communities of Pseudomonas bacteria. Pyover-
dines can both promote (through molecule sharing) and inhibit
(through iron blocking) the growth of other strains depending on a
molecular lock- key (receptor- pyoverdine) mechanism (23,29). Our
coevolution pairing algorithm managed to pair 188 pyoverdine types
and 4547 receptors into 47 lock- key groups. Our experimentally vali-
dated approach allowed us to reconstruct the iron interaction network
of 1928 Pseudomonas strains. We found intriguing network dierenc-
es between habitats (soil, plant, water, and human- derived habitats)
and between microbial lifestyles (pathogenic and nonpathogenic).
Large and highly connected siderophore- mediated iron interaction
networks occurred among nonpathogenic environmental strains,
whereas small and disconnected network dominated among patho-
genic strains. ese results suggest that selection pressures shaping
bacterial interaction networks may dier fundamentally between hab-
itats and lifestyles.
Our sequence- to- ecology approach underscores several challeng-
es associated with the reconstruction of interactions driven by sec-
ondary metabolites. e rst challenge is that the chemical structures
of secondary metabolites are oen dicult to infer from sequencing
data because the metabolites are produced by nonribosomal peptide
and polyketide synthesis pathways. e second challenge is to iden-
tify pyoverdine receptor genes among the many dierent types of
siderophore receptor genes each strain has. We solved these chal-
lenges in a previous paper (31), where we developed approaches on
the basis of feature sequences that allowed us to infer the chemical
structure of 188 pyoverdines and identify 4547 pyoverdine receptor
genes. e third challenge was to pair pyoverdines to matching re-
ceptors within and across strains. While coevolution analyses are a
widely used computational tool, used in diverse areas ranging from
ab initio protein structure to host- pathogen interaction predictions
(46), we could not use existing algorithms, such as DCA, SCA, and
Evoformer (47–49). e reason is that these classical site- based co-
evolution methods depend on paired sequences between which the
degree of covariation is quantied, yet the existence of multireceptor
producers impeded direct assignments of synthetase- receptor pairs.
For this purpose, we developed an unsupervised learning algorithm
(called the coevolution paring algorithm), which yielded 47 receptor-
synthetase lock- key pairs. is step was essential to reconstruct iron
interaction networks. Our pipeline has the potential to be applied to
several other microbial traits. For example, microbial membrane re-
ceptors coevolve with phages (50,51), and pairing phages with the
receptor they use for infection could provide insights into host-
pathogen coevolution and, thus, bacteria- host interaction networks
in natural communities.
Our sequence- to- interaction mapping together with the math-
ematical model yielded several important biological insights into
the ecology and evolution of microbial iron interaction networks.
First, multireceptor producers seem to be the glue of iron interac-
tion networks. They harbor a large diversity of pyoverdine and
receptor types, which can connect many other single- and multire-
ceptor producers and foster the formation of large and highly con-
nected networks. Second, networks of nonpathogenic species and
communities in natural soil and water habitats were large and high-
ly interconnected. ese networks were dominated by multirecep-
tor producers and were more open to invasion. In other words, new
strains could easily integrate themselves into existing communities
without much disturbance to the local community. is suggests
that selection for siderophore and receptor diversity is particularly
high in species- rich habitats. ird and in contrast to the above,
networks of pathogenic species and communities in human- derived
habitats were small and scattered and had low complexity. ey
were dominated by single- receptor producers with a compromised
repertoire of pyoverdine and receptor types. Single- receptor pro-
ducers were predicted to be the invaders with severe consequence,
having a high potential to disrupt and displace the local com-
munity. Utilization of exclusive siderophore groups increases the
chances of successful invasion. ese results suggest that selection
favors exclusive siderophores in pathogens as successful invasion is
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
11 of 15
a key aspect of their lifestyle. Last, strains with a strict cheating
strategy (nonproducers) occur but are relatively rare (13.6%). However,
our approach yields a conservative nonproducer estimate because
regulatory nonproducers that have a synthetase cluster but do not
express it also occur (22) but cannot be detected with our approach.
Given that nonproducers are fully dependent on siderophore pro-
ducers, it is intuitive to understand that they are (i) more common
in environmental habitats featuring many multireceptor producers
with a rich pyoverdine repertoire and (ii) the best at invading mul-
tireceptor communities.
Our insights on pyoverdine utilization strategies and their conse-
quences for community dynamics reveal ways of how siderophore-
mediated interactions could be leveraged for biotechnological ap-
plications (52). In this context, there is great interest in using probiotic
strains in agriculture to protect crops from infections by bacterial plant
pathogens. ere is increasing evidence that siderophore- mediated in-
teractions play a key role in this process (29,30). For example, it was
shown that plant- benecial Chryseobacterium strains use their sidero-
phores to suppress the plant pathogen Ralstonia solanacearum (53). e
approach has recently been extended to human pathogens, which were
found to be suppressed by exclusive siderophores from environmental
Pseudomomnas spp. (52). While these studies explored siderophore in-
teractions across the species boundaries without clear knowledge on
the specic receptor setup, we here propose a strategic two- pronged
design approach. First, single- receptor producers with exclusive sidero-
phores should be designed and used as probiotics to competitively
exclude pathogens in simple communities. Second, multireceptor pro-
ducers with an exclusive siderophore against the pathogen and
cheating receptors using the pathogen’s siderophores should be de-
signed and used as probiotics in more complex communities. ey
could integrate themselves without major disruption of the local
community yet still competitively exclude the pathogen via the
exclusive siderophore.
In conclusion, we succeeded in developing a sequence- to- interaction
mapping approach for siderophores that has a high potential to de-
liver deeper insights into microbial ecology. Given that iron is a key
trace element that is limited in most environments, siderophore-
mediated interactions are an ideal entry point for secondary metab-
olite analysis from sequence data. While we focused on Pseudomonas
strains, we know that siderophore- mediated interactions occur across
the genus boundaries. For example. P. aeruginosa has receptors to
take up enterobactin produced by Enterobacteriaceae spp. and schizo-
kinen produced by. R. solanacearum (33). us, the next step would
be to apply our concepts to more diverse bacterial communities to
derive microbiome- level iron- interaction maps that could further
guide rational designs for biotechnological microbiota interventions.
MATERIALS AND METHODS
Construction of phylogeny tree
e phylogenetic tree depicted in Fig. 1A was constructed using the
PhyloPhlAn3 pipeline (54). PhyloPhlAn is a comprehensive pipe-
line that encompasses the identication of phylogenetic markers,
multiple sequence alignment, and the inference of phylogenetic
trees. In this analysis, we used over 400 universal genes dened by
PhyloPhlAn as our selected phylogeny markers. Subsequently, the
taxonomic cladogram was generated using the iTOL web tool
(http://huttenhower.sph.harvard.edu/galaxy/).
Using the coevolution relationship between synthetases and
receptors to identify self- receptors in producers
To establish all lock- key relationships between synthetases and re-
ceptors in their sequence spaces (see the “Variation of pyoverdine-
mediated iron interaction networks across habitats” section for
results), it is necessary to identify the self- receptors in each produc-
er strain. Assuming that the strongest coevolution occurs between
the synthetase and its cognate self- receptor, the pipeline of identify-
ing self- receptors consists of following three key parts:
Calculation of the indel distance matrix between pyoverdine
synthetase sequences
To precisly quantify the evolutionary distance between synthetases,
a more accurate method than the full sequence alignment is re-
quired. In the case of module- or domain- level duplication, dele-
tion, or insertion event, the p- distance [dened as the number of
positions with dierent bases (or amino acids) in the two sequences
divided by the total number of positions] between two closely re-
lated sequences can become markedly high, even for single- receptor
producers whose receptor sequences belong to the same classica-
tion group. is phenomenon is common and can potentially lead
to the erroneous clustering of synthetase genes. Consequently, we
undertook a two- step approach to enhancing the accuracy of calcu-
lating sequence distances between synthetase genes.
Step 1. To address this issue, we initially conducted a global se-
quence alignment between any two synthetase feature sequences
using the Needleman- Wunsch algorithm. Using the BLOSUM50
scoring matrix, we categorized all loci as “matched,” “similar,” or
“unmatched” on the basis of their sequence similarity. Subsequently,
we eliminated consecutive unmatched loci that extended for more
than
Lunmatch
amino acids, as these were deemed fragment mis-
matches resulting from module- or domain- level indels. e pri-
mary objective of this step was to mitigate the sequence distance
between synthetase sequences belonging to strains within the same
group. In our algorithm, the threshold
Lunmatch
was set at 10 ami-
no acids.
Step 2. For the remaining sequence segment, the ultimate dis-
tance, denoted as
Distancesyn
, was determined as
Here,
pdistance
represents the p- distance of the remaining se-
quence, while
pmatch
signies the proportion of consecutive matched
loci exceeding
Lmatch
amino acids in length. Given the fewer con-
secutive matched loci observed between strains belonging to dier-
ent groups, this step was implemented to further accentuate the
disparities in synthetase sequence distances between strains within
the same group and those in dierent groups. In our algorithm, we
set the threshold for
Lmatch
at 10 amino acids.
Calculation of the correlation between synthetase and
receptor sequence distance matrices
To eectively quantify the coevolutionary relationship between syn-
thetase and receptor sequences, we used the correlation between the
distance matrices of synthetic and receptor sequences. We deter-
mined self- receptor pairs by checking if the grouping of synthetase
sequences matched the grouping of receptor sequences. Recogniz-
ing that the Pearson correlation coecients between synthetase and
receptor sequence distance matrices could be sensitive to minor
changes in sequence distances, we executed the following two steps
Distancesyn
=pdistance ×
(
1−p
match)
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
12 of 15
to robustly quantify the correlation between synthetase and receptor
sequence distance matrices:
Step 1. We applied binarization to both the synthetase and recep-
tor sequence distance matrices using respective thresholds,
Tsyn
and
Trec
. e distance matrix for synthetase was the indel distance ma-
trix described in the preceding “Calculation of the indel distance
matrix between pyoverdine synthetase sequences” section, and the
distance matrix for receptors was calculated by the alignment of the
FpvA feature sequence (Pro168 to Ala295) (31). Elements exceeding
the threshold were assigned a value of 1, while elements falling be-
low the threshold were assigned a value of 0.
Step 2. Following binarization, it is possible that indirect connec-
tions could exist, involving intermediate strains, in addition to the
direct connections between strains. To distinctly separate multiple
groups, we connected all connected components that are linked to
each other, directly or indirectly, as one group. en, the distance
between strains belonging to dierent groups was standardized to 1,
while the distance between strains within the same group was con-
sistently set to 0. is connected component approach created an
unweighted network encompassing all strains. is network pre-
served the connection information between strains while eectively
mitigating any disturbances caused by minor uctuations in se-
quence distances. In our algorithm, we designated both
Tsyn
and
Trec
as 0.3.
Unsupervised coevolution pairing algorithm for identifying
the self- receptor
Using the correlation calculated in the “Calculation of the correla-
tion between synthetase and receptor sequence distance matrices”
section, we developed an unsupervised algorithm that leverages
random sampling and simulated annealing to identify self- receptors
from a multitude of receptors in each multireceptor producer. An
overview of our algorithm is presented as follows (as depicted
in Fig. 3A):
Step 1. Initial processing.
1) In the case of
N
multireceptor producer strains, we initiated the
process by computing the synthetase indel distances between all pair-
wise strains, as described in the “Calculation of the indel distance ma-
trix between pyoverdine synthetase sequences” section. is yields a
synthetase sequence distance matrix with dimensions
N×N
.
2) We randomly chose one receptor for each strain to construct
the initial receptor list. Subsequently, we calculate the receptor se-
quence distance matrix, which is also of size
N×N
, on the basis of
this initial receptor list.
3) We executed the connected component clustering procedure
for both the synthetase and receptor sequence matrices, as described
in the “Calculation of the correlation between synthetase and recep-
tor sequence distance matrices” section. Following this, we comput-
ed the Pearson correlation coecient between these two matrices.
Step 2. Random sampling. We randomly selected
Nbatch
strains
and introduced random perturbations to the receptor numbers as-
sociated with these strains, thus generating a perturbed receptor list.
Step 3. Calculation of the correlation coecient.
1) We recalculated the receptor sequence distance matrix by the
perturbed receptor list.
2) e connected component clustering procedure in the “Calcu-
lation of the correlation between synthetase and receptor sequence
distance matrices” section was applied to the receptor distance ma-
trix. e correlation coecient between the synthetase and receptor
distance matrices was calculated by methods in the “Calculation of
the correlation between synthetase and receptor sequence distance
matrices” section.
Step 4. Simulated annealing. According to the correlation coe-
cient calculated in Step 3:
1) If the correlation coecient is smaller than the previous value,
we would accept the perturbed receptor list.
2) If the correlation coecient failed to decrease, we would ac-
cept the change of the receptor list with a probability denoted as
Paccept
.
Step 5. We returned to Step 2 and continued the interaction, until
the correlation coecients converge or a maximial number of itera-
tions is reached.
In each iteration, the number of randomly selected strains was
adaptive. We used a smaller
Nbatch
when the number of iterations
was limited, aiming to achieve a faster rate of correlation coecient
improvement. Conversely, when the number of iterations was ex-
tensive, we opted for a relatively larger
Nbatch
to introduce a wider
range of perturbations. In our algorithm, the value for
Nbatch
was set
as follows
Furthermore, the inclusion of the simulated annealing step was
instrumental in preventing the algorithm from getting stuck in local
optimal solutions. In our algorithm, the acceptance probability, de-
noted as
Paccept
, was congured as follows
Here, in the acceptance probability calculation,
cr′
represents the
correlation coecient derived from the perturbed receptor list,
cr
denotes the correlation coecient from the original receptor list,
and
Siteration
signies the current iteration count. We discarded itera-
tions that decreased cr values and continued with those that in-
creased cr values until an optimization plateau was reached. We
predicted the self- receptor of all multireceptor producers on the
basis of the nal assignment. We used Pearson’s correlation coe-
cient for all our analyses. It is more suitable for measuring func-
tional groups than Spearman’s rank correlation. e former is more
sensitive to clustered data to measure functional coupling instead of
phylogenetic associations within a functional group.
DNA extract of Pseudomonas strains
We used 24 Pseudomonas strains that were originally isolated from
tomato rhizosphere (29) to test the eects of pyoverdine on interac-
tions between strains. e genomic DNA of Pseudomonas strains
was initially extracted using the Invitrogen PureLink Genomic
DNA kit. e DNA quantity and quality were tested by a NanoDrop
ND- 1000 Spectrophotometer (ermo Fisher Scientic). e DNA
was puried further using the Quick- DNA Miniprep Plus kit.
Illumina HiSeq sequencing
For Illumina paired- end sequencing of each strain (55), at least 3 μg
of genomic DNA was used for sequencing library construction. Paired-
end libraries with insert sizes of ~400 bp were prepared following
Illumina’s standard genomic DNA library preparation procedure.
Puried genomic DNA is sheared into smaller fragments with a de-
sired size by Covaris, and blunt ends are generated by using T4
DNA polymerase. Aer adding an “A” base to the 3′ end of the blunt
N
batch =
{
Random integer from [1, 5], iterations <20,000
Random integer from [1, 10], iterations
>
20,000
P
accept =exp
(
cr�−cr
1
∕S
iteration )
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
13 of 15
phosphorylated DNA fragments, adapters are ligated to the ends of
the DNA fragments. e desired fragments were then puried through
gel electrophoresis, then selectively enriched, and amplied by poly-
merase chain reaction. e index tag was introduced into the adapter
at the polymerase chain reaction stage as appropriate, and we did
a library quality test. Last, the qualied Illumina paired- end library
was used for Illumina NovaSeq 6000 sequencing (150 bp*2, Shanghai
BIOZERON Co., Ltd.).
PacBio sequencing
Whole- genome sequencing was performed by Pacic Biosciences
Sequel II technology (PacBio). e DNA was made into SMRTbell
libraries using the Express Template Prep Kit 2.0 from PacBio ac-
cording to the manufacturer’s protocol. Samples were pooled into a
single multiplexed library, and size was selected using Sage Sciences’
BluePippin, which uses the 0.75% DF Marker S1 High- Pass 6 kb–10 kb
v3 run protocol and S1 marker. A size selection cutoff of 8000
(BPstart value) was used. e size- selected SMRTbell library was
annealed and bound according to the SMRT Link setup and se-
quenced on a Sequel II.
Genome assembly
e raw paired- end reads were trimmed and quality controlled by
Trimmomatic (version 0.36, www.usadellab.org/cms/index.php?page=
trimmomatic) with parameters (SLIDINGWINDOW:4:15 MINLEN:75).
Clean data obtained by the above quality control processes were used
for further analysis.
Raw PacBio reads were converted to fasta format with Samtools
Fasta (www.htslib.org/doc/samtools.html). e Illumina data were
used to evaluate the complexity of the genome and to correct the
PacBio long reads. First, we used Unicycler (https://github.com/rrwick/
Unicycler) to perform genome assembly with default parameters
and received the optimal results of the assembly. GC depth and
genome size information was calculated by custom perl scripts,
which helped us to judge whether DNA samples were contami-
nated or not. Last, the strain genome was circularized with Circlator
(http://sanger- pathogens.github.io/circlator/).
Genome annotation
For the prokaryotic organism, we used the ab initio prediction
method to get gene models for every strain. Gene models were iden-
tied using GeneMark. en, all gene models were blastp against
nonredundant (NR in NCBI) database, SwissProt (http://uniprot.
org), Kyoto Encyclopedia of Genes and Genomes (www.genome.jp/
kegg/), and COG (www.ncbi.nlm.nih.gov/COG) to do functional
annotation by the blastp module. In addition, tRNAs were identied
using tRNAscan- SE (version 1.23, http://lowelab.ucsc.edu/tRNAscan-
SE), and ribosomal RNAs were determined using RNAmmer (version
1.2, https://services.healthtech.dtu.dk/services/RNAmmer-1.2/).
Measuring the growth of Pseudomonas strains and their
pyoverdine production
All Pseudomonas strains were stored at −80°C. Before the experi-
ments, a single colony of each strain was selected randomly, grown
overnight in lysogenic broth, washed three times in 0.85% NaCl, and
adjusted to an optical density at 600 nm (OD600) of 0.5 using a spec-
trophotometer (SpectraMax M5, Sunnyvale, CA). To quantify the
growth and siderophore production of each pseudomonad strain un-
der iron- limited conditions, we transferred 2 ml of overnight cultures
into a new 250- ml glass ask containing 150 ml of MKB medium
[K2HPO4 (2.5 gl−1), MgSO4·7H2O (2.5 gl−1), glycerin (15 ml l−1),
casamino acids (5.0 gl−1), pH 7.2] in threefold replication. Following
24 h of incubation at 30°C with shaking (rotary shaker set at 170 rpm),
we measured growth (OD600) with a spectrophotometer at room
temperature (SpectraMax M5, Sunnyvale, CA) and then harvested
the cell- free supernatant from bacterial cultures by centrifugation
(8000 rpm, 8 min at 4°C) and ltration (using a 0.22- μm lter).
e supernatant was then divided into two parts for (i) measuring
pyoverdine production and (ii) testing the eects of pyoverdine on
interactions between Pseudomonas strains. Briey, pyoverdine pro-
duction levels were measured (relative uorescence units with excita-
tion at 400 nm and emission at 460 nm) with a spectrophotometer at
room temperature (SpectraMax M5, Sunnyvale, CA).
Testing the eects of pyoverdine on interactions between
Pseudomonas strains
To avoid interference from other metabolites as much as possible,
we adapted the method of Butaitė etal. (22) to crudely purify py-
overdine from the supernatants of 20 producers collected through
the above steps. For the cross- feeding assay, we suspended each pu-
ried pyoverdine in 2 ml of Milli- Q water and passed the solution
through a 0.22- μm lte r.
Following the above steps to obtain 24 strains of bacterial uid
(OD600=0.5), we diluted invader precultures 100- fold into new 96-
well plates and subjected them to the following three experimental
conditions in threefold replication. (i)
SNlimited
: each strain individu-
ally growing in 180 μl of 10% MKB medium supplemented with 20 μl
of aqueous solution of pyoverdine crude extract. (ii)
SNreplenished
:
each strain individually growing in 180 μl of iron- rich 10% MKB
medium supplemented with 20 μl of aqueous solution of pyoverdine
crude extract (removes the eect of pyoverdine but retains the eect
of other metabolites). e iron- rich condition was achieved by add-
ing iron(III) solution (1 mM FeCl3·6H2O and 10 mM HCl) into
MKB medium (nal concentration equaling 50 μM). (iii)
SNcontrol
:
each strain individually growing in 180 μl of iron- limited 10% MKB
medium supplemented with 20 μl of 0.85% (w/v) NaCl instead of
supernatant (control mimicking the addition of spent medium). We
measured each pseudomonad strain growth (OD600) of each repli-
cate aer 24 h of incubation at 30°C under static conditions.
Subsequently, we calculated the eect of each producer’s pyover-
dine crude extract on each pseudomonad strain growth as the growth
eect, denoted as
GEtreatment
, was determined as
where
SNtreatment
is
SNlimited
or
SNreplenished
. For this calculation, we
took the average eects across the three replicates. From these measures,
the net GE of pyoverdine can be measured as
GEpyo =GEli −GEre
.
is is possible because we used the exact same supernatants for
SNlimited
and
SNreplenished
, but pyoverdines are only important for
growth in the former condition and not in the latter condition,
where iron is available in excess (29). In principle, GEpyo>0 indi-
cates pyoverdine- mediated facilitation. However, because there is
substantial experimental variation between experimental replicates,
we increased the threshold value of GEPyo>0.05 and classied
values above this threshold as positive interactions, where the re-
ceiving strain can use the respective pyoverdine for iron acquisition
(interaction type 1). Conversely, GEPyo≤0.05 values were classied
GEtreatment
=
[(
SN
treatment
∕SN
control)
−1
]
×
100
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
14 of 15
as neutral or negative interactions, where the receiving strain
cannot use the respective pyoverdine for iron acquisition (interac-
tion type 0).
Supplementary Materials
This PDF le includes:
Supplementary Text
Figs. S1 to S7
Tables S1 to S3
References
REFERENCES AND NOTES
1. R. Cavicchioli, W. J. Ripple, K. N. Timmis, F. Azam, L. R. Bakken, M. Baylis, M. J. Behrenfeld,
A. Boetius, P. W. Boyd, A. T. Classen, T. W. Crowther, R. Danovaro, C. M. Foreman,
J. Huisman, D. A. Hutchins, J. K. Jansson, D. M. Karl, B. Koskella, D. B. Mark Welch,
J. B. H. Martiny, M. A. Moran, V. J. Orphan, D. S. Reay, J. V. Remais, V. I. Rich, B. K. Singh,
L. Y. Stein, F. J. Stewart, M. B. Sullivan, M. J. H. van Oppen, S. C. Weaver, E. A. Webb,
N. S. Webster, Scientists’ warning to humanity: Microorganisms and climate change. Nat.
Rev. Microbiol. 17, 569–586 (2019).
2. X. Wang, Z. Wei, K. Yang, J. Wang, A. Jousset, Y. Xu, Q. Shen, V. P. Friman, Phage
combination therapies for bacterial wilt disease in tomato. Nat. Biotechnol. 37,
1513–1520 (2019).
3. N. D. Sonnert, C. E. Rosen, A. R. Ghazi, E. A. Franzosa, B. Duncan- Lowey,
J. A. González- Hernández, J. D. Huck, Y. Yang, Y. Dai, T. A. Rice, M. T. Nguyen, D. Song,
Y. Cao, A. L. Martin, A. A. Bielecka, S. Fischer, C. Guan, J. Oh, C. Huttenhower, A. M. Ring,
N. W. Palm, A host–microbiota interactome reveals extensive transkingdom connectivity.
Nature 628, 171–179 (2024).
4. H. K. Kuramitsu, X. He, R. Lux, M. H. Anderson, W. Shi, Interspecies interactions within oral
microbial communities. Microbiol. Mol. Biol. Rev. 71, 653–670 (2007).
5. A. Konopka, S. Lindemann, J. Fredrickson, Dynamics in microbial communities:
Unraveling mechanisms to identify principles. ISME J. 9, 1488–1495 (2015).
6. J. Handelsman, Metagenomics: Application of genomics to uncultured microorganisms.
Microbiol. Mol. Biol. Rev. 68, 669–685 (2004).
7. A. Almeida, A. L. Mitchell, M. Boland, S. C. Forster, G. B. Gloor, A. Tarkowska, T. D. Lawley,
R. D. Finn, A new genomic blueprint of the human gut microbiota. Nature 568, 499–504
(2019).
8. P. D. Schloss, J. Handelsman, Metagenomics for studying unculturable microorganisms:
Cutting the Gordian knot. Genome Biol. 6, 229 (2005).
9. C. Gu, G. B. Kim, W. J. Kim, H. U. Kim, S. Y. Lee, Current status and applications of
genome- scale metabolic models. Genome Biol. 20, 121 (2019).
10. A. V. Colarusso, I. Goodchild- Michelman, M. Rayle, A. R. Zomorrodi, Computational
modeling of metabolism in microbial communities on a genome- scale. Curr. Opin. Syst.
Biol. 26, 46–57 (2021).
11. A. R. Pacheco, M. Moel, D. Segrè, Costless metabolic secretions as drivers of interspecies
interactions in microbial ecosystems. Nat. Commun. 10, 103 (2019).
12. S. Gude, M. E. Taga, Multi- faceted approaches to discovering and predicting microbial
nutritional interactions. Curr. Opin. Biotechnol. 62, 58–64 (2020).
13. S. Louca, M. Doebeli, Calibration and analysis of genome- based models for microbial
ecology. eLife 4, e08208 (2015).
14. K. A. J. Bozhüyük, L. Präve, C. Kegler, L. Schenk, S. Kaiser, C. Schelhas, Y.- N. Shi,
W. Kuttenlochner, M. Schreiber, J. Kandler, M. Alanjary, T. M. Mohiuddin, M. Groll,
G. K. A. Hochberg, H. B. Bode, Evolution- inspired engineering of nonribosomal peptide
synthetases. Science 383, eadg4320 (2024).
15. K. Penn, C. Jenkins, M. Nett, D. W. Udwary, E. A. Gontang, R. McGlinchey, B. Foster,
A. Lapidus, S. Podell, E. E. Allen, B. S. Moore, P. R. Jensen, Genomic islands link secondary
metabolism to functional adaptation in marine Actinobacteria. ISME J. 3, 1193–1203
(2009).
16. K. Scherlach, C. Her tweck, Mediators of mutualistic microbe–microbe interactions. Nat.
Prod. Rep. 35, 303–308 (2018).
17. L. C. Vining, in Ciba Foundation Symposium 171 - Secondary Metabolites: Their Function and
Evolution (Wiley, 2007), pp. 184–198.
18. R. He, J. Zhang, Y. Shao, S. Gu, C. Song, L. Qian, W.- B. Yin, Z. Li, Knowledge- guided data
mining on the standardized architecture of NRPS: Subtypes, novel motifs, and sequence
entanglements. PLoS Comput. Biol. 19, e1011100 (2023).
19. S. C. Andrews, A. K. Robinson, F. Rodríguez- Quiñones, Bacterial iron homeostasis. FEMS
Microbiol. Rev. 27, 215–237 (2003).
20. P. W. Boyd, M. J. Ellwood, The biogeochemical cycle of iron in the ocean. Nat. Geosci. 3,
675–682 (2010).
21. D. Emerson, E. Roden, B. S. Twining, The microbial ferrous wheel: Iron cycling in terrestrial,
freshwater, and marine environments. Front. Microbiol. 3, e00383 (2012).
22. E. Butaitė, M. Baumgartner, S. Wyder, R. Kümmerli, Siderophore cheating and cheating
resistance shape competition for iron in soil and freshwater Pseudomonas communities.
Nat. Commun. 8, 414 (2017).
23. J. Kramer, Ö. Özkaya, R. Kümmerli, Bacterial siderophores in community and host
interactions. Nat. Rev. Microbiol. 18, 152–163 (2020).
24. G. E. Leventhal, M. Ackermann, K. T. Schiessl, Why microbes secrete molecules to modify
their environment: the case of iron- chelating siderophores. J. R. Soc. Interface 16,
20180674 (2019).
25. S. A. West, S. P. Diggle, A. Buckling, A. Gardner, A. S. Grin, The social lives of microbes.
Annu. Rev. Ecol. Evol. Syst. 38, 53–77 (2007).
26. M. Miethke, M. A. Marahiel, Siderophore- based iron acquisition and pathogen control.
Microbiol. Mol. Biol. Rev. 71, 413–451 (2007).
27. Y. Seyoum, K. Baye, C. Humblot, Iron homeostasis in host and gut bacteria – A complex
interrelationship. Gut Microbes 13, 1–19 (2021).
28. U. E. Schaible, S. H. E. Kaufmann, Iron and microbial infection. Nat. Rev. Microbiol. 2,
946–953 (2004).
29. S. Gu, Z. Wei, Z. Shao, V. P. Friman, K. Cao, T. Yang, J. Kramer, X. Wang, M. Li, X. Mei, Y. Xu,
Q. Shen, R. Kümmerli, A. Jousset, Competition for iron drives phytopathogen control by
natural rhizosphere microbiomes. Nat. Microbiol. 5, 1002–1010 (2020).
30. S. Gu, T. Yang, Z. Shao, T. Wang, K. Cao, A. Jousset, V. P. Friman, C. Mallon, X. Mei, Z. Wei,
Y. Xu, Q. Shen, T. Pommier, Siderophore- mediated interactions determine the disease
suppressiveness of microbial consortia. mSystems 5, e00811- 19 (2020).
31. S. Gu, Y. Shao, K. Rehm, L. Bigler, D. Zhang, R. He, J. Shao, A. Jousset, V.- P. Friman, Z. Wei,
R. Kümmerli, Z. Li, From sequence to molecules: Feature sequence- based genome mining
uncovers the hidden diversity of bacterial siderophore pathways. eLife 13, RP96719
(2024).
32. P. Cornelis, Iron uptake and metabolism in pseudomonads. Appl. Microbiol. Biotechnol. 86,
1637–1645 (2010).
33. R. Kümmerli, Iron acquisition strategies in pseudomonads: Mechanisms, ecology, and
evolution. Biometals 36, 777–797 (2023).
34. K. A. J. Bozhüyük, F. Fleischhacker, A. Linck, F. Wesche, A. Tietze, C. P. Nieser t, H. B. Bode,
De novo design and engineering of non- ribosomal peptide synthetases. Nat. Chem. 10,
275–281 (2018).
35. J. M. Meyer, A. Stintzi, D. de Vos, P. Cornelis, R. Tappe, K. Taraz, H. Budzikiewicz, Use of
siderophores to type pseudomonads: The three Pseudomonas aeruginosa pyoverdine
systems. Microbiology 143, 35–43 (1997).
36. A. R. T. Figueiredo, Ö. Özkaya, R. Kümmerli, J. Kramer, Siderophores drive invasion
dynamics in bacterial communities through their dual role as public good versus public
bad. Ecol. Lett. 25, 138–150 (2022).
37. E. Thébault, C. Fontaine, Stability of ecological communities and the architecture of
mutualistic and trophic networks. Science 329, 853–856 (2010).
38. J. Bascompte, P. Jordano, C. J. Melián, J. M. Olesen, The nested assembly of plant–animal
mutualistic networks. Proc. Natl. Acad. Sci. U.S.A. 100, 9383–9387 (2003).
39. S. Jiqi, L. Yinxiang, L. Jingyuan, G. Shaohua, L. Zhiyuan, Siderophore piracy promotes
dynamical coexistence in microbial community. bioRxiv [Preprint]. 2023. https://doi.
org/10.1101/2023.11.21.568182.
40. D. Tilman, Resource Competition and Community Structure (Princeton University Press, 1982).
41. V. Dubinkina, Y. Fridman, P. P. Pandey, S. Maslov, Multistability and regime shifts in microbial
communities explained by competition for essential nutrients. eLife 8, e49720 (2019).
42. T. Taillefumier, A. Posfai, Y. Meir, N. S. Wingreen, Microbial consortia at steady supply. eLife
6, e22644 (2017).
43. M. C. B. Tsilimigras, A. A. Fodor, Compositional data analysis of the microbiome:
Fundamentals, tools, and challenges. Ann. Epidemiol. 26, 330–335 (2016).
44. S. Weiss, W. van Treuren, C. Lozupone, K. Faust, J. Friedman, Y. D eng, L. C. Xia, Z. Z. Xu,
L. Ursell, E. J. Alm, A. Birmingham, J. A. Cram, J. A. Fuhrman, J. Raes, F. Sun, J. Zhou,
R. Knight, Correlation detection strategies in microbial data sets vary widely in sensitivity
and precision. ISME J. 10, 1669–1681 (2016).
45. K. Faust, J. Raes, Microbial interactions: From networks to models. Nat. Rev. Microbiol. 10,
538–550 (2012).
46. H. M. Chen, W. Guo, J. Shen, L. Wang, J. N. Song, Structural principles analysis of
host- pathogen protein- protein interactions: A structural bioinformatics survey. IEEE
Access 6, 11760–11771 (2018).
47. J. Greensmith, J. Feyereisl, U. Aickelin, The DCA: SOMe comparison. Evol. Intell. 1, 85–112
(2008).
48. S. Mirjalili, SCA: A sine cosine algorithm for solving optimization problems. Knowl.- Based
Syst. 96, 120–133 (2016).
49. J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool,
R. Bates, A. Žídek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie,
B. Romera- Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy,
M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver,
O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, D. Hassabis, Highly accurate protein
structure prediction with AlphaFold. Nature 596, 583–589 (2021).
Downloaded from https://www.science.org on January 16, 2025
Gu et al., Sci. Adv. 11, eadq5038 (2025) 15 January 2025
SCIENCE ADVANCES | RESEARCH ARTICLE
15 of 15
50. H. G. Hampton, B. N. J. Watson, P. C. Fineran, The arms race between bacteria and their
phage foes. Nature 577, 327–336 (2020).
51. K. E. Kortright, B. K. Chan, J. L. Ko, P. E. Turner, Phage Therapy: A renewed approach to
combat antibiotic- resistant bacteria. Cell Host Microbe 25, 219–232 (2019).
52. V. Vollenweider, K. Rehm, C. Chepkirui, M. Pérez- Berlanga, M. Polymenidou, J. Piel,
L. Bigler, R. Kümmerli, Antimicrobial activity of iron- depriving pyoverdines against human
opportunistic pathogens. eLife 13, RP92493 (2024).
53. K. Rehm, V. Vollenweider, S. Gu, V. P. Friman, R. Kümmerli, Z. Wei, L. Bigler,
Chryseochelins—Structural characterization of novel citrate- based siderophores
produced by plant protecting Chryseobacterium spp. Metallomics 15, mfad008
(2023).
54. F. Asnicar, A. M. Thomas, F. Beghini, C. Mengoni, S. Manara, P. Manghi, Q. Zhu, M. Bolzan,
F. Cumbo, U. May, J. G. Sanders, M. Zolfo, E. Kopylova, E. Pasolli, R. Knight, S. Mirarab,
C. Huttenhower, N. Segata, Precise phylogenetic analysis of microbial isolates and
genomes from metagenomes using PhyloPhlAn 3.0. Nat. Commun. 11, 2500 (2020).
55. Y. Wang, Z. Sha, X. Ren, A new species of Orchomenella (Amphipoda, Tryphosidae)
described from hydrothermal vent in the Okinawa Trough, Northwest Pacic. ZooKeys
1184, 261–271 (2023).
56. S. P. Diggle, M. Whiteley, Microbe Prole: Pseudomonas aeruginosa: Opportunistic
pathogen and lab rat. Microbiology 166, 30–33 (2020).
57. B. S. Scales, R. P. Dickson, J. J. LiPuma, G. B. Hunagle, Microbiology, genomics, and
clinical signicance of the Pseudomonas uorescens species complex, an unappreciated
colonizer of humans. Clin. Microbiol. Rev. 27, 927–948 (2014).
58. D. L. Arnold, G. M. Preston, Pseudomonas syringae: Enterprising epiphyte and stealthy
parasite. Microbiology 165, 251–253 (2019).
59. C. L. Patten, B. R. Glick, Role of Pseudomonas putida indoleacetic acid in development of
the host plant root system. Appl. Environ. Microbiol. 68, 3795–3801 (2002).
60. M. Espinosa- Urgel, A. Salido, J.- L. Ramos, Genetic analysis of functions involved in
adhesion of Pseudomonas putida to seeds. J. Bacteriol. 182, 2363–2369 (2000).
61. A. Ramette, M. Frapolli, M. F. L. Saux, C. Gruaz, J. M. Meyer, G. Défago, L. Sutra,
Y. Moënne- Loccoz, Pseudomonas protegens sp. nov., widespread plant- protecting
bacteria producing the biocontrol compounds 2,4- diacetylphloroglucinol and
pyoluteorin. Syst. Appl. Microbiol. 34, 180–188 (2011).
62. D. Ercolini, A. Casaburi, A. Nasi, I. Ferrocino, R. di Monaco, P. Ferranti, G. Mauriello, F. Villani,
Dierent molecular types of Pseudomonas fragi have the same overall behaviour as meat
spoilers. Int. J. Food Microbiol. 142, 120–131 (2010).
63. J. Lalucat, A. Bennasar, R. Bosch, E. García- Valdés, J. Palleroni Norberto, Biology of
Pseudomonas stutzeri. Microbiol. Mol. Biol. Rev. 70, 510–547 (2006).
64. U. Tattawasart, J. Y. Maillard, J. R. Furr, A. D. Russell, Development of resistance to
chlorhexidine diacetate and cetylpyridinium chloride in Pseudomonas stutzeri and
changes in antibiotic susceptibility. J. Hosp. Infect. 42, 219–229 (1999).
65. L. Gardan, H. Shak, S. Belouin, R. Broch, F. Grimont, P. A. Grimont, DNA relatedness
among the pathovars of Pseudomonas syringae and description of Pseudomonas tremae
sp. nov. and Pseudomonas cannabina sp. nov. (ex Sutic and Dowson 1959). Int. J. Syst.
Bacteriol. 49, 469–478 (1999).
66. H. Y. Yun, H. Kim, First report of Pseudomonas amygdali causing bacterial leaf spot of
Mallotus japonicus in South Korea. For. Pathol. 51, e12707 (2021).
67. A. W. T. F. Chin, G. V. Bloemberg, I. H. Mulders, L. C. Dekkers, B. J. Lugtenberg, Root
colonization by phenazine- 1- carboxamide- producing bac terium Pseudomonas
chlororaphis PCL1391 is essential for biocontrol of tomato foot and root rot. Mol. Plant
Microbe Interact. 13, 1340–1345 (2000).
68. J. M. Yu, D. Wang, L. S. Pierson III, E. A. Pierson, Eect of producing dierent phenazines
on bacterial tness and biological control in Pseudomonas chlororaphis 30- 84. Plant
Pathol. J. 34, 44–58 (2018).
69. S. W. Kwon, J. S. Kim, I. C. Park, S. H. Yoon, D. H. Park, C. K. Lim, S. J. Go, Pseudomonas
koreensis sp. nov., Pseudomonas umsongensis sp. nov. and Pseudomonas jinjuensis sp.
nov., novel species from farm soils in Korea. Int. J. Syst. Evol. Microbiol. 53, 21–27 (2003).
70. G. F. Rakova, T. Y. Korshunova, L. F. Minnebaev, S. P. Chetverikov, O. N. Loginov, A new
bacterial strain, Pseudomonas koreensis IB- 4, as a promising agent for plant pathogen
biological control. Microbiology 85, 333–341 (2016).
71. S. M. K ang, S. Asaf, A. L. Khan, Lubna, A. Khan, B. G. Mun, M. A. Khan, H. Gul, I. J. Lee,
Complete genome sequence of Pseudomonas psychrotolerans CS51, a plant growth-
promoting bacterium, under heavy metal stress conditions. Microorganisms 8, 382 (2020).
72. Y. Li, D. Wang, S. Cao, X. Wang, G. Ren, First report of bacterial leaf spot on tobacco caused
by Pseudomonas psychrotolerans in China. Plant Dis. 107, 935 (2023).
73. A. A. Belimov, I. C. Dodd, V. I. Safronova, N. Hontzeas, W. J. Davies, Pseudomonas
brassicacearum strain Am3 containing 1- aminocyclopropane- 1- carboxylate deaminase
can show both pathogenic and growth- promoting properties in its interaction with
tomato. J. Exp. Bot. 58, 1485–1495 (2007).
74. B. Dutta, R. Gitaitis, G. Agarwal, T. Coutinho, D. Langston, Pseudomonas coronafaciens sp.
nov., a new phytobacterial species diverse from Pseudomonas syringae. PLOS ONE 13,
e0208271 (2018).
75. S. M. Bhatawadekar, Community- acquired urinar y tract infection by pseudomonas
oryzihabitans. J. Glob. Infect. Dis. 5, 82–84 (2013).
76. Y. Hou, Y. Zhang, L. Yu, X. Ding, L. Liu, L. Wang, S. Huang, First report of Pseudomonas
oryzihabitans causing rice panicle blight and grain discoloration in China. Plant Dis. 104,
3055 (2020).
77. W. Adam, F. Heckel, C. R. Saha- M öller, M. Taupp, P. Schreier, A highly enantioselective
biocatalytic sulfoxidation by the topsoil bacterium Pseudomonas frederiksbergensis.
Tetrahedron: Asymmetry 15, 983–985 (2004).
78. C. Ramos, I. M. Matas, L. Bardaji, I. M. Aragón, J. Murillo, Pseudomonas savastanoi pv.
savastanoi: Some like it knot. Mol. Plant Pathol. 13, 998–1009 (2012).
79. L. Tvrzová, P. Schumann, C. Spröer, I. Sedláček, Z. Páčová, O. Šedo, Z. Zdráhal, M. Steen,
E. Lang, Pseudomonas moraviensis sp. nov. and Pseudomonas vranovensis sp. nov., soil
bacteria isolated on nitroaromatic compounds, and emended description of
Pseudomonas asplenii. Int. J. Syst. Evol. Microbiol. 56, 2657–2663 (2006).
80. M. Gennari, F. Dragotto, A study of the incidence of dierent uorescent Pseudomonas
species and biovars in the microora of fresh and spoiled meat and sh, raw milk, cheese,
soil and water. J. Appl. Bacteriol. 72, 281–288 (1992).
81. S. Verhille, N. Baida, F. Dabboussi, D. Izard, H. Leclerc, Taxonomic study of bacteria isolated
from natural mineral waters: Proposal of Pseudomonas jessenii sp. nov. and
Pseudomonas mandelii sp. nov. Syst. Appl. Microbiol. 22, 45–58 (1999).
82. R. Li, Y. Jiang, X. Wang, J. Yang, Y. G ao, X. Zi, X. Zhang, H. Gao, N. Hu, Psychrotrophic
Pseudomonas mandelii CBS- 1 produces high levels of poly- β- hydroxybutyrate.
Springerplus 2, 335 (2013).
83. M. Elomari, L. Coroler, S. Verhille, D. Izard, H. Leclerc, Pseudomonas monteilii sp. nov.,
isolated from clinical specimens. Int. J. Syst. Bacteriol. 47, 846–852 (1997).
84. M. Tohya, S. Watanabe, K. Teramoto, K. Uechi, T. Tada, K. Kuwahara- Arai, T. Kinjo, S. Maeda,
I. Nakasone, N. N. Zaw, S. Mya, K. N. Zan, H. H. Tin, J. Fujita, T. Kirikae, Pseudomonas
asiatica sp. nov., isolated from hospitalized patients in Japan and Myanmar. Int. J. Syst.
Evol. Microbiol. 69, 1361–1368 (2019).
85. D. Kakembo, Y. H. Lee, Analysis of traits for biocontrol performance of Pseudomonas parafulva
JBCS1880 against bacterial pustule in soybean plants. Biol. Control 134, 72–81 (2019).
86. U. Behrendt, A. Ulrich, P. Schumann, Fluorescent pseudomonads associated with the
phyllosphere of grasses; Pseudomonas trivialis sp. nov., Pseudomonas poae sp. nov. and
Pseudomonas congelans sp. nov. Int. J. Syst. Evol. Microbiol. 53, 1461–1469 (2003).
87. M. von Neubeck, C. Huptas, C. Glück, M. Krewinkel, M. Stoeckel, T. Stressler, L. Fischer,
J. Hinrichs, S. Scherer, M. Wenning, Pseudomonas helleri sp. nov. and Pseudomonas
weihenstephanensis sp. nov., isolated from raw cow's milk. Int. J. Syst. Evol. Microbiol. 66,
1163–1173 (2016).
88. P. Ioannou, G. Vougiouklakis, A systematic review of human infections by Pseudomonas
mendocina. Trop. Med. Infect. Dis. 5, 71 (2020).
89. C. O. Onwosi, F. J. Odibo, Eects of carbon and nitrogen sources on rhamnolipid
biosurfactant production by Pseudomonas nitroreducens isolated from soil. World J.
Microbiol. Biotechnol. 28, 937–942 (2012).
90. A. Novinscak, M. Filion, Long term comparison of talc- and peat- based phytobenecial
Pseudomonas uorescens and Pseudomonas synxantha bioformulations for promoting
plant growth. Front. Sustain. Food Syst. 4, e602911 (2020).
Acknowledgments: We thank X. Huang for the insight on pyoverdine receptor analysis.
Funding: This work was supported by the National Key Research and Development Program of
China (nos. 2021YFF1200500 and 2021YFA0910700), National Natural Science Foundation of
China (nos. 42107140, 32071255, 41922053, and T2321001), and National Postdoctoral Program
for Innovative Talents (no. BX2021012). R.K. is supported jointly by a grant from the Swiss
National Science Foundation no. 310030_212266. V.- P.F. is supported jointly by Research Council
of Finland, Novo Nordisk Fonden, and The Finnish Research Impact Foundation. Author
contributions: S.G. performed most computational and experimental data analyses in this
research and drafted the manuscript. Z.S. and S.Z. performed the experiment of testing
pyoverdine- mediated interaction between Pseudomonas strains. Z.Q. performed the
mathematical model exploring the relationship between pyoverdine utilization strategies and
community dynamics. D.Z. developed the unsupervised coevolution pairing algorithm for
identifying the self- receptor, and R.A. oered the insight on this algorithm. Y.S., R.H., J.S., and G.X.
assisted in cleaning up the codes. A.J. and V.- P.F. oered insightful comments and assisted in
revising and writing of the manuscript. R.K. and Z.W. oversaw the project, designed experiments,
and revised the manuscript. Z.L. conceptualized and oversaw the project and revised the
manuscript. Competing interests: The authors declare that they have no competing interests.
Data and materials availability: All data needed to evaluate the conclusions in the paper are
present in the paper and/or the Supplementary Materials. The origin data and code can be
accessed at the Dryad repository: https://doi.org/10.5061/dryad.z08kprrpg.
Submitted 16 May 2024
Accepted 11 December 2024
Published 15 January 2025
10.1126/sciadv.adq5038
Downloaded from https://www.science.org on January 16, 2025