Characterization of hypothetical proteins reveals function modules and virulence factors in Pseudomonas aeruginosa

Abstract

Hypothetical proteins (HPs), known as biological dark matter, are genetic sequences of which the biological functions and regulatory processes remain largely unclear. In Pseudomonas aeruginosa, HPs account for 36.4% of all coding genes, representing a significant gap in understanding their functions and contribution to the virulence network. To address these challenges, we present a deletion repertoire comprising 938 strains targeting 1165 HPs in P. aeruginosa. Starting with transcriptome-wide analysis, we inferred putative functional predictions for HP mutants through systematic profiling of differential expression and candidate transcriptional pathways. We integrated our expression profiles with public datasets (5707 genes × 2312 conditions). This allowed us to distinguish constitutively expressed genes from variable genes. Using independent component analysis (ICA), we identified 56 independent regulatory Modules, providing a comprehensive transcriptional-association framework. These contributed to the discovery and experimental validation of 20 virulence-related hypothetical proteins (VRHPs) and 3 antibiotic resistance-related hypothetical proteins (ARHPs). Evolutionary analysis across bacterial taxa provided additional insights into HP sequence conservation and evolutionary dynamics. In summary, our study presents a large-scale HP mutant library combined with integrated transcriptomic and machine learning approaches, establishing both a valuable resource and a systematic strategy for predicting and prioritizing candidate functions of HPs across bacterial species.

This study presents a deletion library of 938 hypothetical protein mutants in P. aeruginosa. Integrating transcriptomics and machine learning, it reveals comprehensive regulatory modules and discovers virulence and resistance factors.

Introduction

Genes with unknown functions, such as hypothetical proteins (HPs), present a major challenge in the postgenomic era¹. These HPs, also termed orphan genes, represent predicted coding sequences lacking functional classification due to the absence of recognizable protein sequence or structural motifs². Across both eukaryotic and prokaryotic domains, 25–50% of genes are considered HPs, remaining either completely uncharacterized or only partially understood^3–7. Among these, certain HPs demonstrate evolutionary conservation across multiple lineages, designated as conserved proteins⁸. Functional annotation of HPs can unveil biological pathways and identify potential therapeutic targets for drug development and screening⁹. Understanding gene interactions represents a fundamental level of biological organization, making the identification of functional modules and their components a key objective in systems biology^10–12. However, comprehensive functional classification of HPs remains limited, with few large-scale experimental validations confirming protein expression from predicted HP genes.

Pseudomonas aeruginosa is a Gram-negative, multidrug-resistant (MDR) opportunistic pathogen responsible for severe acute or chronic infections, such as chronic obstructive pulmonary disease, cystic fibrosis, cancer, traumas, burns, sepsis, and ventilator-associated pneumonia^13–17. Its genome encodes numerous virulence factors, including flagella, type IV pili, pyocyanin (PYO), lipopolysaccharides, and biofilms, which enable the bacteria to adapt to environmental challenges and complete their infection cycles^18–25. Conserved regulatory networks in Gram-negative bacteria demonstrate functions in virulence and antimicrobial resistance modulation, such as resistance nodulation division (RND) efflux pumps, quorum-sensing (QS) system, type III (T3SS) and type VI secretion system (T6SS), and c-di-GMP signaling^26–30.

The virulence network of P. aeruginosa is controlled by global regulatory factors, such as transcription factors (TFs), two-component system (TCS), and regulatory proteins that rely on protein-protein interactions^31–33. Our group has recently elucidated the mechanisms of a group of key virulence regulators, including VqsR, VqsM, AlgR, CdpR, RpoN, and AnvM^34–40. Our systematic analysis of P. aeruginosa transcriptional regulation encompassed multiple approaches: construction of a virulence-related regulatory network (PAGnet) with 20 interacting TF regulons, characterization of 182 TF DNA-binding patterns via HT-SELEX, and genome-wide mapping of 172 TF binding sites through ChIP-seq, establishing a comprehensive regulatory atlas^31,34. Within the transcriptional regulatory landscape, numerous HPs are present, though their functional roles and regulatory contributions remain undefined.

This critical gap persists in our understanding of multidrug-resistant P. aeruginosa’s virulence regulatory network. While 36.4% of its genes remain classified as HPs, over the past decade, only 83 of 2080 (4%) HP genes have been functionally characterized (Suppl. Data 1). Of these, 30 genes influence virulence and 10 affect antibiotic resistance, while others participate in metabolism, nutrient transport, stress responses, cellular processes, and enzymatic functions. For example, hyperexpression of PA3719-20 modulated MexR repressor activity, while deletion of PA2982 resulted in non-polar localization of both the flagellum and FlhF protein^41,42. The vast majority (1997 genes, 96%) remain uncharacterized. These findings underscore the critical importance of HPs and emphasize the need for systematic functional characterization to fully elucidate P. aeruginosa’s virulence and antibiotic resistance networks.

In this study, we addressed the substantial knowledge gap of HPs in P. aeruginosa, which comprises 36.4% of its coding sequences. Through the construction of a comprehensive 1165 HP mutant library and integration of transcriptomics with machine learning, we assigned potential functions to 1322 P. aeruginosa HPs. This analysis identified 20 virulence-related and 3 antibiotic resistance-related HPs, while revealing cross-species evolutionary patterns. Our work establishes both a valuable resource and a systematic framework for predicting and prioritizing candidate functions of biological dark matter across bacterial pathogens.

Results

938 deletion strains reveal 372 GO-restricted HPs

Initial genomic analysis of Pseudomonas aeruginosa PAO1, empowered by the Pseudomonas Genome Database annotations, identified 2080 HPs. Only 69 HPs harbored 113 significant functional domains as identified through SMART database analysis (Suppl. Fig. 1A). The absence of recognizable domains in 96.7% of HPs underscored significant knowledge gaps in bacterial protein functions. To further advance the interpretation of HPs, we constructed 938 deletion strains. This large-scale deletion repertoire encompassed 769 single HP gene deletions and 169 shared operon deletions covering 396 HP genes and 39 annotated genes (Fig. 1A) (Suppl. Data 1). In total, 1165 out of 2080 (56%) HP genes were included in the repertoire. We measured growth curves for all 938 deletion mutants in 96-well plates, 19 mutants (PA0267, PA0277, PA0285, PA0484, PA0522, PA0541, PA0557, PA4577, PA0348-49, PA2451-52, PA3248, PA0924, PA3819, PA0636-39, PA4573, PA0702, PA0834, PA0572, PA2660-61) exhibited impaired growth (Suppl. Data 2), indicating widespread transcriptional changes observed across HP deletions predominantly reflect specific regulatory roles rather than generalized fitness defects. We performed RNA-seq of 938 deletions at mid-log phase in LB conditions, and differentially expressed genes (DEGs) were calculated by comparing to the wild type under the same culture conditions. Two replicates across 938 mutant libraries and 22 WT libraries showed high consistency (Suppl. Fig. 1B). Globally, deletion strains exhibited a median of 101.5 DEGs, while individual genes were affected by a median of 12 different deletions, indicating extensive regulatory networks (Fig. 1B). Among 4937 genes analyzed, we identified 40 highly responsive genes that were differentially expressed in more than 300 mutants, with oxidoreductase-related genes (n = 8) and heme-binding (n = 6) representing significant functional prediction (Suppl. Fig. 1C).

Fig. 1. DEGs and GO enrichment analysis of 938 PAO1 HP mutants.

A The scheme shows the classification of the HP mutant library. B Distribution of differentially expressed gene (DEG) numbers per mutant versus the frequency of individual genes exhibiting candidate transcriptional associations across multiple mutants. The x-axis represents the log10 value of gene/mutant counts. n = 4937, n = 938. Centre line: median. Box bounds: 25th to 75th percentiles (interquartile range, IQR). Whiskers: extend to most extreme data points within 1.5 × IQR. C The Venn diagrams show the distribution of Gene Ontology (GO) terms among PAO1 HPs. The overlaps are between biological process, cellular component, and molecular function. D Circos plot illustrating the interconnections between different enriched GO functional categories associated with PAO1 HPs. The width of each segment corresponds to the number of HPs associated with that GO term. Colored ribbons connecting different GO terms represent shared HPs between functional categories. E Scatter plot displaying the topological parameters of PAO1 HPs. Each dot represents an HP, positioned according to its betweenness centrality (x-axis) and degree (y-axis). Betweenness centrality measures a protein’s importance as a bridge between different network clusters, while degree indicates the number of direct interactions. F The results of the minimum inhibitory concentrations (MICs) determination assays for wild-type (WT) PAO1, ΔPA3998, and the complemented strain (ΔPA3998/com) against meropenem, polymyxin B, tetracycline, and trimethoprim. Each group consists of three biological replicates. G Quantification of 3-oxo-C12-HSL (C12-HSL) production in WT, ΔPA0392-94, and ΔPA0392-94/com. Each group consists of four biological replicates. Error bars indicate the means ± SD of four independent experiments. The center line represents the median, the box bounds indicate the 25th and 75th percentiles, and the whiskers represent the minimum and maximum values. Statistical significance was calculated using a two-sided Student’s t-test, p-value = 0.0001/0.005 EV represents the empty vector pAK1900. H Congo red agar plate showing enhanced exopolysaccharide production in the ΔPA0392-94 mutant compared to the WT and complemented strain. Each group consists of three biological replicates. The same representative WT image is shown in Fig. 1H and Suppl. Fig. 6E, as both panels originate from the same experimental batch. Source data are provided as a Source Data file.

Subsequently, Gene Ontology (GO) terms were enriched for each deletion strain. A total of 265 mutants had GO terms in all three main categories of GO terms, Biological Process, Molecular Function, and Cellular Component (Fig. 1C). Notably, the complete absence of GO terms for 287 mutants (including 372 HPs) highlighted a significant annotation gap, reflecting either condition-specific genetic roles or the inadequacy of current GO classification systems in classifying bacterial proteins (Fig. 1C). Application of the aPEAR clustering algorithm revealed 35 distinct communities of biological themes (Suppl. Data 3). The largest functional communities identified were associated with ribosomal structure (164 mutants), peptide binding (70 mutants), and translation (39 mutants), reflecting that these transcriptionally responsive clusters are sensitive to global physiological perturbations upon genetic disruption (Suppl. Fig. 1D). In conclusion, we systematically established a large-scale mutant resource encompassing 1165 HPs, providing a comprehensive transcriptomic profile.

Non-random distribution of HPs across functional communities

To investigate beyond a single annotation, we examined co-perturbation patterns among HPs and identified functional communities. Remarkably, HP-mediated community connections revealed structured bridges (Fig. 1D). The strongest connections were built between the pyoverdine biosynthetic process and the siderophore transmembrane transporter activity (n = 23), suggesting coordinated regulation by distinct sets of HPs across these two communities. The observed HP-mediated co-enrichment of translational and electron transfer processes (n = 9) indicated previously unrecognized regulatory linkage between ribosomal and respiratory functions, transcending basic metabolic correlations. To systematically verify that the functional connectivity mediated by these mutants was statistically non-random, we performed a gene-level Monte Carlo permutation test across 10,000 independent iterations. We computed empirical p-values for all functional connections to assess the probability that they occurred by chance. Out of the 65 total observed connections, 35 were validated as statistically robust (Suppl. Fig. 1E; Suppl. Data 3). Notably, nearly all pairwise functional combinations with an observed co-occurrence frequency greater than 2 in our dataset were confirmed to be statistically significant, except for “siderophore transmembrane transporter activity + structural constituent of ribosome” (Suppl. Fig. 1F). These results demonstrated the high reliability of our identified core functional axes.

Building on this, we constructed an HP-GO undirected network (a network connecting HPs to their associated GO terms). We calculated the topological parameters of degree, betweenness centrality, and degree centrality, identifying two “cross-GO hub” HPs with high betweenness centrality and high degree as key HPs (Fig. 1E). Experimental tests supported the practical efficacy of this strategy in prioritizing functionally relevant candidates. GO analysis of ΔPA3998 revealed significant enrichment in antibiotic response pathways (p-value of 1.90 × 10⁻⁴; Suppl. Fig. 1G). Subsequent minimum inhibitory concentration (MIC) assays indicated reduced resistance to multiple antibiotics, including meropenem, polymyxin B, tetracycline, and trimethoprim, supporting the potential role of HP PA3998 in antimicrobial resistance (Fig. 1F). ΔPA3998 exhibited impaired PAO1 growth prior to the stationary phase, but grew comparably to the wild type thereafter (Suppl. Fig. 1H). Undirected network analysis highlighted another HP PA0392-94, with GO enrichment revealing its multifaceted regulatory roles across phenazine biosynthesis, cytochrome c oxidase activity, copper ion binding, and multiple secretion systems (Type III secretion system (T3SS), Type VI secretion system (T6SS), and Sec). Crucially, subsequent phenotypic assays experimentally validated the dual regulatory nature of PA0392-94, confirming its role in the positive regulation of the las system (Fig. 1G) coupled with negative control of exopolysaccharide (EPS) production (Fig. 1H). ΔPA0392-94 may modestly slow growth (Suppl. Fig. 1H), thereby attenuating the expression of pqsBE, and lasB (Suppl. Fig. 1I), as well as 3OC₁₂-HSL production. In sum, our results revealed organized network communities of co-perturbed pathways across the mutant library, while further targeted analysis led to the identification and experimental validation of two HPs in P. aeruginosa PAO1.

Sixty-seven HP mutants master seven virulence pathways

P. aeruginosa employs diverse virulence pathways and secretes multiple virulence factors to establish successful host infection and promote disease progression⁴³. To investigate the role of HPs in virulence regulation, we focused on 298 genes in seven major virulence pathways in P. aeruginosa (Suppl. Data 4): Type VI secretion system (T6SS), Type III secretion system (T3SS), quorum sensing (QS), pyocyanin production (PYO), motility, antibiotic resistance, and exopolysaccharide & biofilm formation (EPS&Biofilm) (Suppl. Fig. 2A). For each HP deletion, we quantified the association of each HP mutant with each pathway by calculating the ratio of pathway-specific genes in DEGs among all genes in the pathway. This quantitative analysis enabled pathway assignment of HPs and revealed broad-spectrum putative virulence modulators with exceptional pathway-specific influence (Suppl. Fig. 2A). To identify the virulence pathway-associated candidate HPs, we modeled the response ratios across all mutants using a log-normal distribution for each target pathway. We applied a 95th percentile right-tail threshold as an initial effect-size cutoff. Under this primary filter, 123 distinct mutants (encompassing 166 HPs) were captured as potential candidates exhibiting high-ranking biological effects (Suppl. Fig. 2B). To further control for variations in mutant-specific transcriptomic perturbation scales and eliminate global transcriptional noise, we subsequently subjected the mutants to a rigorous Hypergeometric Distribution Test (Suppl. Fig. 2C). A mutant was formally prioritized as a genuinely significant “Pathway-Associated Candidate” only if it satisfied the p-value < 0.05 within this hypergeometric framework and passed 95 percentile at the corresponding pathway.

With this two-tier analytical strategy, we defined 67 high-confidence mutants with broad transcriptional effects spanning all seven virulence pathways (Fig. 2A). By observing the overlapping and distinct regulatory profiles of these pathway-associated candidate HPs, we observed that the vast majority exhibit highly specialized, single-pathway specificity (Fig. 2B). Notably, only a single mutant, PA1478, demonstrated significant multi-targeting capacity across three distinct virulence frameworks (T3SS, T6SS, and Motility). Additionally, five mutants demonstrated a dual-pathway transcriptional modulation capacity. Specifically, PA0319 exhibited candidate transcriptional associations with both QS and T6SS; PA1112 modulated Antibiotic Resistance and QS; PA1203 was functionally associated with Antibiotic Resistance and Motility; the PA1305-PA1308 cluster showed concurrent candidate connections with PYO and T6SS; and PA4963 was tied to EPS & Biofilm alongside T6SS.

Fig. 2. In-depth analysis of HPs exhibiting candidate transcriptional associations with PAO1 virulence pathways.

A Violin plots display the distribution of -log10(p-value) of individual HPs on seven virulence pathways. The p-values are derived from the one-side hypergeometric enrichment analysis for all HP mutants. The red vertical dashed line denotes the standard statistical significance threshold (p-value = 0.05). For each functional pathway, the sample size (n) represents the total number of HPs. 67 orange dots are “Pathway-Associated Candidates”, represent the mutants satisfied the p-value < 0.05 within this hypergeometric framework and exceeding the 95th percentile (in the log-normal distribution fitting) within corresponding pathway. Grey dots represent non-significant mutants. B UpSet plot showing the distinct HPs or intersection of HPs that exceed the 95th percentile in/among different virulence pathways. The horizontal bars (bottom left) denote the cumulative number of significant pathway-associated candidate HPs linked to each individual pathway, while the vertical bars (top) represent the exact size of distinct intersecting cohorts or single-pathway specific regulators. C Functional association network of 20 pathway-associated candidate HPs connected to multiple virulence systems. For each HP, a radar plot shows the proportion of HP that affects the pathway. The genes in each pathway are displayed. Purple nodes representing multi-targeting HP. Circular layouts display the genes within each of the core virulence pathways. Radar charts project the multi-dimensional functional association profiles for selected top candidate HPs. Each axis of the radar plot denotes the response ratio to a specific pathway.

Prioritizing HPs by their pathway-associated response ratios, we drew a detailed functional association network featuring the top three regulators for each pathway, encompassing 20 HP mutants (Fig. 2C). This analysis revealed specific patterns. For example, the dual-pathway-associated candidate PA0319 coordinates both las and rhl quorum-sensing system signal synthesis and export. The ΔPA0319 mutant exhibited attenuated quorum sensing, evidenced by reduced 3OC12-HSL production (Suppl. Fig. 2D). The growth curve showed that ΔPA0319 does not affect the growth of PAO1 (Suppl. Fig. 2E). In sum, we provided putative functional predictions for 67 mutants (including 79 HPs) to seven pathways, revealing both pathway-specific candidate functions and cross-pathway transcriptional hubs among virulence-associated HPs (Fig. 2).

Integration of baseline expression and cross-condition plasticity

While DEG and virulence pathway-level analyses identified key pathway-associated HPs, our comprehensive transcriptomic dataset enabled putative functional insights. Beyond the absolute baseline expression level of individual genes, we examined the cross-condition expression plasticity, quantified by expression variance across multiple environmental contexts. Leveraging our transcriptomic data for a higher-resolution investigation, we evaluated replicate consistency using transcripts per million (TPM-normalized) read counts. Pearson correlation analysis revealed exceptional reproducibility (third quartile r = 0.9999) between replicates, validating the reliability of our sequencing data (Suppl. Fig. 3A). Unsupervised hierarchical clustering based on Euclidean distances of log2-transformed mean TPM between replicates segregated HP mutants into two major transcriptomic groups (Fig. 3A). Analysis of expression consistency revealed 24 genes with persistently low expression (log2 TPM < 2 in > 99% of HP mutants; Suppl. Fig. 3B). These consistently low-expressed genes predominantly comprised small RNAs (sRNAs) and phz genes, with notably GO terms of phenazine biosynthetic process (Suppl. Fig. 3C). Forty-eight genes maintained low expression levels (log2 TPM < 3) across over 99% HP mutants, including five HPs (PA1531, PA1918, PA1923, PA2091, PA2315; Fig. 3B). Interestingly, GO analysis revealed enrichment of pathogenesis and cell adhesion terms under LB growth conditions, consistent with expected physiological responses (Fig. 3C). To understand the context-dependent expression of these low-expressed HPs, we performed comparative analysis incorporating 566 published transcriptomes and our wild-type RNA-seq data (Suppl. Data 5). These incorporated published transcriptional profiles spanned multiple experimental paradigms, such as key transcription factor mutations, virulence factor deletions, antibiotic treatment and environmental challenges to wild-type PAO1, strengthening analytical reliability. While Euclidean distance analysis clearly distinguished between different experimental conditions (Suppl. Fig. 3D), these HPs maintained consistently low expression across all datasets (Fig. 3B). Such loci may remain completely dormant under in vitro laboratory growth parameters, potentially requiring specific host-interactive signals, or distinct environmental niches to trigger their induction^5,44. Alternatively, they may function at low transcript copies that are biologically potent but present minimal transcript abundance⁴⁵. Analysis identified 29 highly expressed genes (log2 TPM > 10 in > 95% of HP mutants), including four HPs (PA1592, PA3309, PA4352, PA4463; Fig. 3D; Suppl. Fig. 3E). While these HPs showed lower expression in wild-type and published datasets, the majority of genes in this group were associated with ribosomal function and translation (Fig. 3E).

Fig. 3. Baseline expression and cross-condition plasticity analysis of HPs.

A Unsupervised hierarchical clustering of the expression matrix of 938 HP mutant strains based on log2-transformed TPM values. The heatmap shows gene expression patterns (blue to red gradient representing low to high expression, log2(TPM + 1) scale) that segregated mutants into two major regulatory clusters: C1 (n = 812) and C2 (n = 126). B The box plots show the distribution of 5 HP genes’ log2(TPM + 1) values across the HP mutants expression matrix, public datasets, and WT expression matrix, showing the expression pattern consistency. For (B, D, F), the sample size of five genes/HPs are same: HP mutants (n = 938), Public dataset (n = 566), WT (n = 44). Centre line: median. Box bounds: 25th–75th percentiles (interquartile range, IQR). Whiskers: extend to most extreme data points within 1.5 × IQR. C GO enrichment analysis of 48 consistently low-expressed genes (log2(TPM) < 3 in > 99% of HP mutants). The p-values are calculated using a one-sided Fisher’s Exact Test. D Comparative analysis of 4 HP gene expression levels (log2(TPM + 1)) between the HP mutants expression matrix, public datasets, and WT expression matrix. E GO enrichment analysis of 29 highly expressed genes (log2(TPM) > 10 in > 95% of HP mutants). The p-values are calculated using a one-sided Fisher’s Exact Test. F Variance analysis of gene expression across different genetic backgrounds (HP mutants expression matrix, WT expression matrix, and public dataset). The p-values are calculated using a two-tailed Mann-Whitney U test. G GO enrichment analysis of genes with the highest and lowest variance. H Expression variance value of phenazine biosynthesis genes. I Expression variance value of xisF4 and pfsE. J Base mean value of RNA-seq data and RT-qPCR results show the relative mRNA expression levels. Left: RNA-seq base mean values representing individual data points (n = 1). Right RT-qPCR validation presented as ΔCT values; data are shown as mean ± SD with individual biological replicates plotted (n = 3). Concordance between RT-qPCR and RNA-seq data validated the expression trends. No statistical hypothesis testing or p-value adjustment was performed for this qualitative trend comparison.

To investigate expression dynamics, we calculated gene-specific variance across log2 TPM matrices from HP mutants, wild-type, and public datasets. HP mutants exhibited significantly higher expression variance compared to wild-type and published data (Fig. 3F). This enhanced variability suggested broader transcriptomic perturbation of HP⁴⁶. Putative functional analysis of expression extremes (top and bottom 30 genes by variance) revealed an intriguing split in the regulation of phenazine biosynthesis (Fig. 3G). While phzC1 and phzS showed high expression variability, other pathway components (phzD1, phzD2, phzE1, phzE2, phzF1, phzF2) maintained stable expression, suggesting differential regulatory control within the phenazine biosynthetic pathway (Fig. 3H; Suppl. Fig. 3F). Notably, phzD1, phzD2, phzE1, phzE2, phzF1, and phzF2 also ranked among the lowest expressed genes. The phage-associated genes xisF4 and pfsE, organized in a single operon, demonstrated significant expression variance across HP mutant strains (Fig. 3I). Concordance between RT-qPCR and RNA-seq data validated these findings (Fig. 3J). This variance likely captured a combination of direct downstream target shifts and widespread indirect effects, including secondary stress responses, growth-deceleration signatures, or global physiological adaptations^47–49.

ICA successfully captured signals from the transcriptome data of HP mutants

While initial analysis revealed distinct expression patterns, the complex matrix suggested further underlying regulatory structure governing bacterial physiology and virulence. To extract higher-order relations from the established gene expression matrix and infer potential HP roles, we applied Independent Component Analysis (ICA), enabling the distinction of biological signals and identification of individual Modules^50,51 (Suppl. Fig. 4A). Following dataset integration, we implemented rigorous quality control measures, retaining samples with high biological replicate correlation (Pearson’s r > 0.8 for log2(TPM); Suppl. Fig. 4B). After normalizing data to the corresponding wild-type, dimensionality reduction via principal component analysis (PCA) retained 1017 components, capturing > 95% of data variance (Suppl. Fig. 4C). The final matrix contained expression data for 5707 genes across 2312 RNA-seq conditions, comprising 1746 conditions from our HP mutant compendium and 566 conditions from published PAO1 RNA-seq studies representing 117 distinct conditions (Suppl. Data 5). Robust ICA employed 120 iterations with strict convergence criteria (tolerance = 1e-7), identifying 56 regulatory Modules (Suppl. Fig. 4D). The Modules derived from Matrix M represented cohesive relationships between genes, dynamic expression patterns, and signals, while the Matrix A measured signal strength of individual Modules under different conditions⁵⁰. Regulatory Modules were annotated by comparison to an established transcriptional regulatory network (TRN)⁵¹ comprising 1346 regulator–target relationships involving 146 regulators (Suppl. Data 6). Each Module was named after the corresponding regulator(s) based on overlap with TRN target sets. These regulatory Modules were further grouped into 12 functional categories based on the known functions of the regulator(s) and the profiles of genes within each Module (Fig. 4A). These comprehensive Modules encompassed 1377 non-HP genes and 430 HPs (Suppl. Fig. 4E; Suppl. Data 7). HPs structurally embedded within ICA Modules are genes that co-vary with other genes within the same Module, suggesting shared transcriptional regulation by a common leading transcription factor⁵¹.

Fig. 4. Independent component analysis reveals regulatory Modules.

A Treemap visualization of 56 regulatory Modules organized by 12 functional categories. Each square represents a specific Module with a size proportional to the number of genes. Major functional categories are highlighted with labeled boxes, including Amino Acid Metabolism, Metal Homeostasis, Carbon Metabolism, Defense Mechanism, Energy Metabolism, Secretion Systems, Structural Components, Quorum Sensing, Stress Response, Prophages, Transcription, and Translation. B Activity distribution of TagR1 Module across different conditions. Six bins are classified based on the range of activity scores. Box plots showing activity scores grouped into six bins ranging from maximally repressed (Bin 1) to maximally activated (Bin 6), with sample sizes indicated above each box. Centre line: median. Box bounds: 25th to 75th percentiles (interquartile range, IQR). Whiskers: extend to most extreme data points within 1.5 × IQR. C Table showing activity scores of specific conditions affecting TagR1 Module. D Venn diagram showing overlap between gene composition of quorum-sensing Modules. Overlapping genes are indicated below. E Manhattan plot showing gene weights for MvfR Module, with predicted functional categories. The y-axis represents gene weights, and the x-axis shows gene position on the genome. F Correlation matrix showing the relationships of activities among three QS Modules LasR, RhlR, and MvfR.

The TagR1 Module emerged as the largest regulatory Module, comprising 74 genes and 49 HPs that exceeded the gene weight threshold, indicating their significant contribution to the Module (Suppl. Fig. 5A). Analysis revealed four key clusters. Besides the tag cluster that contains tagR1 itself, gene cluster tss and hsi are both associated with the T6SS, where mag cluster is involved in metal transport and virulence^52–54. This transcriptional coherence validated ICA’s ability to link expression dynamics with biological function, while demonstrating the coordinated regulation of physically clustered genes. The TagR1 Module activity was classified into six bins, ranging from maximally repressed (bin 1) to maximally activated (bin 6) (Fig. 4B). Notably, ΔmutS, ΔrpoN, and ΔgacA mutants exhibited strong repression, clustering in bins 1 and 2 with activity scores of −29.50, −23.03, and −16.42, respectively (Fig. 4C). This repression aligned with known molecular functions. For instance, MutS influences virulence by controlling mutation rate, while RpoN and GacA are established T6SS regulators^35,55,56. The observed repression in these mutants supported TagR1’s role in coordinating T6SS and virulence processes. Additionally, previous studies showing enhanced biofilm formation in wspE, pelA, and pslBCD mutants under doxorubicin treatment suggested potential activation of TagR1 under similar conditions⁵⁷ (Fig. 4C).

Three quorum-sensing Modules show dynamic response to different conditions via distinct gene sets

The quorum-sensing system in P. aeruginosa is highly complex, especially in its mechanisms for regulating virulence factors, biofilm formation, and environmental adaptation⁵⁸. The application of ICA to P. aeruginosa revealed extensive Modules for quorum-sensing regulators LasR, RhlR, and MvfR. These Modules showed distinct gene compositions, with an overlap of nine genes observed between LasR and RhlR Modules (Fig. 4D; Suppl. Fig. 5B, C). The MvfR Module comprised nine genes, including the pqs and phn operons, further supporting the coordinated expression dynamics of physically clustered genes (Fig. 4E). Correlation analysis of LasR, RhlR, and MvfR Module activities demonstrated moderate concordance, reflecting their overlapping yet distinct roles in quorum-sensing regulation (Fig. 4F). This pattern suggested both coordinated and independent activation mechanisms responding to different environmental conditions. Comparative analysis of MvfR-Module activity across conditions showed that this Module is predominantly activated, evidenced by 362 datasets in bin 6 (Suppl. Fig. 6A). Notably, we observed reductions in Module activity of corresponding TF knockout strains, aligning with expected regulatory behavior (Suppl. Fig. 6B). The ΔPA2430 mutant exhibited markedly low LasR Module activity and reduced production of QS signals (Suppl. Fig. 6B; C). The ΔPA4852 mutant exhibited the second-lowest MvfR Module activity, surpassed only by ΔmvfR, with subsequent validation showing reduced pyocyanin production (Suppl. Fig. 6B; D). The ΔPA4952 mutant showed the strongest reduction in RhlR Module activity among HP mutants, supporting PA4952’s putative role as a PYO-associated modulator. Distinctly, both LasR and RhlR Modules were highly suppressed in the ΔPA2229-30 mutant, which was classified into the QS pathway based on DEG pathway enrichment, correlating with experimental validation of decreased pyocyanin production, EPS production, and impaired Las and Rhl systems function (Suppl. Data 1; Suppl. Fig. 6B; E, F). Furthermore, mouse infection models confirmed reduced virulence phenotypes (Suppl. Fig. 6G). The growth curve showed that mutation of these HPs did not affect the growth of PAO1 (Suppl. Fig. 6H). In sum, we present an ICA-based network that integrated HP mutant data in P. aeruginosa, which identified 56 regulatory Modules with their constituent genes and condition-specific activities.

Two perspectives of activity matrix analysis for inferring the potential roles of 951 HPs

Module activity (Matrix A) reflects coordinated expression patterns of functionally linked genes, with condition-specific activation profiles correlating with individual Module responses to environmental signals⁵⁹. Following the above observations, this relationship implied the alignment between condition-specific activity and underlying biological function. To track collective regulatory responses, we proposed a dual-perspective approach to analyze Matrix A for HP function inference (Fig. 5A). From our first analytical perspective, we categorized Module responses to HP mutations by examining activity patterns. ICA analysis revealed a predominant trend toward Module repression rather than activation, especially in Module LasR (p-value = 3.1e-29; Suppl. Fig. 7A, B). Through examining the significant activated and repressed Module activity (depicted in Methods), 244 HP mutants demonstrated potential transcriptomic impact, each affecting at least one Module (Fig. 5B; Suppl. Fig. 7C). Condition ∆PA0319 exhibited the strongest repressive effect, decreasing the activities of 7 Modules (Fig. 5C). This broad suppression aligned with previously observed reductions in quorum sensing (Suppl. Fig. 2D). In sum, this functional prioritization linked 244 HP mutants (including 310 HPs) to 30 distinct Modules, revealing that 93 HPs showed pleiotropy by affecting more than 2 Modules (Suppl. Fig. 7C, D; Suppl. Data 1).

Fig. 5. Putative functional predictions of HPs by computing correlations between expression patterns and 56 regulatory Module activities.

A Conceptual framework showing dual-perspective analysis of Module activity matrix (Matrix A). The Left panel (Perspective 1) shows the activity matrix with rows as Modules activity score and columns as HP mutation conditions. The right panel (Perspective 2) illustrates correlation analysis between HP expression and Module activity across conditions. B Bubble plot showing HP mutants’ impact. The X-axis shows the number of activated Modules, the Y-axis shows the number of repressed Modules, with bubble size indicating the total number of affected Modules. Dark color represents multiple HP mutant conditions. C Bar chart for the ΔPA0319 condition, which shows strong repressive effects across multiple Modules. D Number of HP genes that have correlation coefficients greater than or equal to 0.4 with regulatory Modules. E Bubble plot showing average correlation scores of 27 Modules that have both positive and negative correlations with HPs. The X-axis represents the average correlation score of negatively associated HPs, the Y-axis shows the average correlation score of positively associated HPs, with bubble size indicating the total number of correlated HP genes. F Venn diagram showing overlap between HPs with putative functional predictions generated by Perspective 1 and Perspective 2, with 105 HPs prioritized by both approaches. G Venn diagram of 36 HPs associated with quorum-sensing Modules by both perspectives. H Bar chart showing 3OC12-HSL production in wild-type (WT), mutant, and complemented (COM) strains. Each group consists of four biological replicates. I Correlation networks showing the association of three HP genes (PA3764, PA4463-65, and PA4128) with different regulatory Modules. J Antibiotic sensitivity assay results. Each group consists of three biological replicates. K, L Phenotypic validation showing colony morphology and pigment production. Panel (K) shows swarming motility images comparing WT/EV and mutant strains. Panel (L) shows quantification of pyocyanin production across different mutant strains. Each group consists of three biological replicates. Error bars indicate the means ± SD of three independent experiments. Statistical significance was calculated using a two-sided Student’s t-test. *p-value < 0.05. **p-value < 0.01. ***p-value < 0.001. EV represents the empty vector pAK1900. The same representative WT image is shown in Fig. 5L and Suppl. Fig. 6D, as both panels originate from the same experimental batch. Source data are provided as a Source Data file.

While the first approach successfully prioritized 310 of 2080 HPs through Module activity regulation, it was limited by mutant dataset availability. Our second, broader approach correlated expression dynamics of all 2080 HPs with 56 regulatory Module activities across experimental conditions, enabling comprehensive putative functional predictions of the complete HP set (Fig. 5A). Of 116,480 analyzed correlations, we examined the distribution of all coefficients and used a normal distribution as a background model. We applied a stringent tail criterion and selected the extreme 0.05% quantile of the fitted distribution as the cutoff (r ~ 0.4). Only 1758 (1.5%) demonstrated strong associations with Module activities, prioritizing 746 HPs with significant candidate transcriptional connections to 49 Modules’ dynamic activity patterns (absolute Pearson correlation ≥ 0.4; Supplementary Fig. 7E, F). Four Modules (CpxR, LiuR, Sfa3, and CueR) showed the highest HP correlation frequency, dominated by negative transcriptional correlations, whereas FleN distinctly exhibited positive correlations with 82 HPs (Fig. 5D; Suppl. Fig. 7G). While 27 Modules showed both positive and negative HP correlations, EutR emerged with the highest mean positive correlation score, specifically across four HPs (Fig. 5E). Our second analytical perspective, based on Module activity correlations, generated functional predictions for 746 HPs, building upon and extending our perspective 1 analysis (Fig. 5F).

Among 951 HPs prioritized for putative roles, 36 HPs showed associations with quorum-sensing Modules (Fig. 5G). Fascinatingly, four mutants (PA4132, PA3335, PA1211-15, and PA0586-88) all demonstrated markedly reduced levels of 3OC₁₂-HSL compared with WT, while the complemented strains restored production (Fig. 5H). Intriguingly, HPs frequently demonstrated connections to multiple regulatory Modules (Suppl. Fig. 7C, H). For validation, we selected HPs showing associations with phenotype-linked Modules (Fig. 5I). We found that PA4463 is associated with the FleQ Module, while PA4465 is linked to the FleN Module (Fig. 5I). These associations are consistent with our experimental results showing that the PA4463–65 mutant significantly reduces P. aeruginosa swarming motility and PYO production (Fig. 5K–L). Moreover, the ΔPA4463–65 strain exhibited decreased tolerance to meropenem and trimethoprim (Fig. 5J), consistent with our earlier finding that the PA4463–65 is part of the antibiotic resistance-related pathway. In addition, PA4128, which was classified into the PYO pathway, showed a strong correlation with the RhlR and RsmA Modules, and PA3764 correlated with the QscR-1 Module (Fig. 5I); notably, both were experimentally confirmed to positively regulate PYO production (Fig. 5L). Collectively, these results indicated that the high HP-regulatory Module activity correlations (absolute Pearson correlation ≥ 0.4) can serve as an effective screening framework for prioritizing and inferring the putative roles of HPs.

Pan-Genome of 1295 P. aeruginosa strains reveals Module-specific evolution patterns

We expanded our analysis to the species level by examining 1295 P. aeruginosa strains with complete or chromosome assemblies available in the NCBI database. Pan-genome analysis of these strains identified 72,131 total genes, with core genes comprising 2.09% of the pan-genome (Fig. 6A; Suppl. Fig. 8A). Pan-genome analysis included 2079 PAO1 HPs, excluding PA1414 (recently characterized as an sRNA regulator of infection). Over 50% of HPs constituted core genes, with only 139 HPs present in fewer than 200 strains (Fig. 6B). With pan‑genome results, we found no significant difference in presence frequency between HPs and non‑HP genes, indicating a broadly similar evolutionary tempo (Fig. 6C).

Fig. 6. Pan-genome Analysis of 1295 P. aeruginosa Strains.

A Pie chart showing gene distribution in the P. aeruginosa pan-genome. Core genes (present in 99% ∼ 100% strains), Soft core genes (present in 95% ∼ 99% strains), Shell genes (present in 15% ∼ 95% strains), and Cloud genes (present in 0% ∼ 15% strains). B Density plot showing the distribution of HPs based on strain occurrence. C The violin plot with box plot shows the comparison of HP and non-HP genes’ occurrence frequency in pan-genomes. Sample size: HP (n = 2079), non-HP genes (3502). For (C–E), Centre line: median. Box bounds: 25th to 75th percentiles (interquartile range, IQR). Whiskers: extend to most extreme data points within 1.5 × IQR. The p-value was calculated using a two-tailed Mann-Whitney U test to compare two groups. D The violin plot with box plot shows the SNP frequencies in HPs and non-HP genes. Sample size: HP (n = 923), non-HP genes (1639). The p-values are calculated using a two-tailed Mann-Whitney U test to compare two groups. E The violin plot with box plot shows the dN/dS fraction in HPs and non-HP genes. Sample size: HP (n = 1755), non-HP genes (2818). The p-values are calculated using a two-tailed Mann-Whitney U test to compare two groups. F Scatter plot showing conservation frequency of Module-correlated HPs across the pan-genome. The X-axis represents -log10(p-value) and the Y-axis shows the mean frequency of occurrence. G Genomic organization and frequency distribution of prophage-associated HP genes. H Venn diagram illustrating the overlap and uniqueness of HPs among three predominant sequence types. Important HPs are highlighted. I Venn diagram illustrating the overlap and uniqueness of HPs among the three predominant regions. Distinct and important HPs are highlighted. J Bar graph showing HP conservation across 24 bacterial strains. The upper panel displays the number of conserved HPs in each strain, while the lower panel shows the distribution of sequence identity percentages. K Hierarchical clustering of HP conservation patterns across different bacterial strains.

To further delineate evolutionary differences between HPs and non‑HP genes, we conducted more in‑depth analyses. SNP analysis comparing 1294 strains against the PAO1 reference genome revealed that while HPs and other genes showed similar strain distribution patterns, HPs exhibited significantly higher SNP frequencies (Fig. 6D). Using SNPGenie, we performed a within‑group dN/dS analysis based on whole‑genome alignments of all isolates to the PAO1 reference. We found that HPs exhibit significantly higher dN/dS than non‑HP genes, suggesting an elevated rate of amino acid substitution (Fig. 6E). Twenty HPs showed dN/dS > 1, and only two of the non-HP genes showed dN/dS > 1 (PntAA and PA3067, a transcription factor). Conversely, several proteins involved in translation and core cellular processes (e.g., RpmF, RpmG, RpmH, RplA, and PilT) displayed very low dN/dS, consistent with purifying selection removing deleterious nonsynonymous changes to preserve function (Suppl. Data 8). This discovery implied distinct evolutionary constraints on HP sequences despite similar conservation patterns across P. aeruginosa. While such signatures can occasionally indicate adaptive positive selection, they more broadly reflect relaxed purifying selection or lower functional constraints, allowing higher mutational tolerance without compromising cellular viability⁶⁰.

Moreover, we analyzed HP conservation patterns across highly correlated HP–Module groups. While average HP occurrence frequency across the pan-genome was 0.876, distinct conservation patterns emerged among HPs highly correlated with regulatory Modules (Fig. 6F). Sixteen HPs associated with the Anr Module, putatively involved in anaerobic respiration, showed exceptional conservation (mean frequency = 0.999, p-value = 7.8e-87). In contrast, five HPs (PA0717, PA0718, PA0719, PA0722, PA0726) linked to the XisF4 Module demonstrated significantly lower conservation (mean frequency = 0.07), representing the least conserved HP group. These five HPs are in the same gene cluster after XisF4 in PAO1, annotated to encode bacteriophage (Fig. 6G). HP conservation patterns associated with XisF4 aligned with typical mobile genetic element evolution rather than core metabolic Modules, supporting their predicted bacteriophage origins.

Inter and intra-species evolutionary analysis predicts conserved HP functions

Multi-locus sequence typing (MLST) analysis of 1265 strains using seven housekeeping genes (acsA, aroE, guaA, mutL, nuoD, ppsA, trpE) identified ST235, ST111, and ST253 as predominant sequence types in this pan-genome database (n = 70, 33, and 32, respectively; Suppl. Fig. 8B). Minimum spanning tree analysis based on HP repertoire similarity (Jaccard index) revealed distinct phylogenetic distributions of these prevalent STs (Suppl. Fig. 8C). ST235, ST111, and ST253 resided at distant branches, suggesting ST-specific HP acquisition or retention (Fig. 6H). Consistent with the ST-specific separation in HP repertoires, several ST-specific loci with predicted functional links reinforced tentative lineage-specific adaptations. In ST111, PA2430 was associated with quorum-sensing systems. ST235 and ST253 both harbored an HP PA3764, which was transcriptionally associated with the QscR-1Module and pyocyanin production. These ST-specific genes likely contribute to strain-specific variations in virulence, stress responses, and antibiotic resistance.

Geographical analysis of ST distribution further links HP evolution to the epidemiological context (Suppl. Fig. 9A–C). Upon further investigation, we observed that P. aeruginosa isolates from China, the USA, and Germany harbored the majority of identified HPs (Suppl. Fig. 9B). Notably, despite ranking fifth in terms of isolate number, Germany ranked third in HP diversity, suggesting strain diversity may drive HP repertoire richness in certain geographical regions. PA2430, an HP involved in the regulation of quorum-sensing systems and previously validated through mutant phenotyping in our study, was found to be absent in isolates from the USA (Fig. 6I).

To assess the evolutionary conservation of HPs across the bacterial kingdom, we analyzed sequence identity patterns of 2079 HPs (excluding PA1414, which was reclassified as sRNA) across 24 bacterial strains (including Acinetobacter baumannii, Bacillus cereus, etc.) using BLASTP. The analysis revealed a strong phylogenetic signal in HP conservation, with retention rates declining sharply with evolutionary distance from P. aeruginosa (Fig. 6J). Within the P. aeruginosa strains, conservation remained high in P. aeruginosa PAK and P. aeruginosa PA96 (retained 1958 HPs), with only several specific losses. For instance, PA1088-1090, which were composed of the flagellin glycan, were lost in these two strains⁶¹. However, even within the Pseudomonas genus, conservation decreased substantially in more distant species. Phytopathogen Pseudomonas syringae strains maintained fewer than 800 HPs, demonstrating extensive gene content divergence despite genus-level relatedness. Correlation-based clustering of sequence identity patterns across all strains revealed distinct conservation groups (Fig. 6K). Among these, Cluster 4 emerged as a highly restricted group containing only 30 HPs. This cohort included PA0561 and PA0625 (previously associated with antibiotic tolerance in PAO1⁶²), alongside PA2004 (enriched in antibiotic resistance-related GO terms in this study). Rather than definitively establishing functional equivalence, the persistent maintenance of this conserved set across diverse bacterial lineages suggests that these genes are under evolutionary constraints to preserve their sequence integrity⁶³. In summary, HPs demonstrated a wide spectrum of genetic variability along evolutionary trajectories, marked by distinct patterns of lineage-specific conservation and divergence. Our systematic analysis of HP transcriptomic profiles and functional association networks provides a valuable reference for inferring the roles of uncharacterized homologs across bacterial species, serving as a useful resource to accelerate functionally directed HP research.

Discussion

In this study, we aimed to establish a systemic framework for the putative functional inference and prioritization of the genome-wide HP sets of P. aeruginosa. Integration of transcriptional profiling across 938 HP mutants with statistical tools and machine learning analysis enabled us to generate candidate pathway inferences for 1322 HPs, significantly expanding our understanding of previously uncharacterized genes of P. aeruginosa (Fig. 7A). Specifically, our multi-faceted analytical framework encompassed GO analysis, virulence pathway mapping, and two distinct ICA-based approaches. We acknowledged that DEG-based functional assignment cannot distinguish direct regulation from indirect transcriptional effects. Therefore, the pathway/Module links reported here should be interpreted as predicted transcriptional associations, and mechanistic causality will require follow-up experimental validation. Based on evidence tiering across these methods, 868 HPs were supported by only one approach, whereas 454 HPs were supported by two or more independent approaches (Fig. 7B). All four approaches provided putative functional linkages for 23 HPs (Fig. 7B). Notably, perspective 2 of our ICA analysis, which examined correlations between HP expression dynamics and Module activities, provided the broadest functional predictions across the HP repertoire (Fig. 7B). Intriguingly, despite integrating both our log-phase LB culture data and diverse expression profiles from published datasets, 758 HPs remained functionally uncharacterized across all inference methods. This observation presented two possibilities. Either these HPs represent truly non-functional genetic elements in an evolved strain, or their functions remain undetected due to the inherent limitations of laboratory-based approaches. While we leveraged extensive in vitro datasets, the inability to infer potential roles for these HPs suggested that additional contexts may be necessary to fully understand their potential biological roles.

Fig. 7. Summary of putative functional prediction for P. aeruginosa hypothetical proteins.

A Table summarizing the comprehensive putative functional predictions for P. aeruginosa PAO1 hypothetical proteins (HPs) in this study. The table shows different predictive pipelines and their coverage. B Intersection analysis of the four computational approaches. C Identified ARHPs and VRHPs in this study. Relative phenotype results are labeled.

Previous studies have successfully applied ICA to transcriptomic data across diverse species, including Saccharomyces cerevisiae, Arabidopsis thaliana, P. aeruginosa, and human cancer cells, enabling effective identification and characterization of functionally related gene Modules^59,64–66. However, the effectiveness of ICA analysis relies heavily on the dataset quality, abundance, and completeness of transcriptional regulatory network (TRN) information. Here, our study leveraged the first P. aeruginosa expression dataset that included HP knockout data and current TRN annotations, maximizing analytical power. Our comprehensive HP mutation dataset also provided valuable insights into HP-specific regulatory effects. Remarkably, the observed relationships between HP functions and Module activities enabled the development of two complementary analytical perspectives for matrix A analysis. Experimental validation through phenotypic and RT-qPCR studies confirmed the robustness of these approaches, offering a valuable reference for data-driven functional prioritization. When performing ICA analysis, the first step was to determine an appropriate dimensionality for decomposition. At dimension 250, we observed relatively low background noise, with 57 single-gene Modules and 37 regulatory Modules identified (Suppl. Fig. 10A). However, at this setting, the boundaries between the LasR and RhlR regulons were not well defined, as both were merged into the Sfa2 Module (Suppl. Fig. 10). To achieve a higher-resolution and more precise functional inference of HP genes, we therefore selected dimension 1000 as an optimal balance between noise reduction and Module resolution.

Virulence mechanisms and antibiotic resistance represent the primary research priorities in P. aeruginosa pathobiology. Through systematic gene deletion, complementation, and wild-type comparisons, we identified 20 virulence-related HPs (VAHPs) and 3 antibiotic resistance-related HPs (ARHPs) (Fig. 7C). While these experiments established direct causality between HP deletions and phenotypic changes, the underlying regulatory mechanisms remain to be elucidated. For example, the PA2229–30 loci demonstrated a potential role as crucial modulators of pathogenicity in our phenotypic assays. Specifically, ICA analysis revealed suppression of LasR and RhlR Modules in ΔPA2229-30, while pathway enrichment identified QS pathway association. This reminded us of a previous result that RsaL binds to the PA2228-PA2229 intergenic region, suggesting a potential regulatory connection between these components. While multiple analyses confirmed PA2229-30’s role in pathogenicity and inferred its potential regulatory linkage, its precise and detailed role and related pathway warranted further investigation. In particular, follow-up studies under infection-relevant or stress conditions may be required to fully elucidate the context-dependent regulation implied by these QS-associated signals.

In addition to the ARHPs and VRHPs experimentally identified in this study, the utility of our analytical methods was further supported by evidence from published research. For example, PA0096–0101 has been previously identified as a T6SS vgrG1b cluster. And from our ICA-based analysis, PA0098 and PA0100 were both found to be strongly associated with the TagR1 Module⁶⁷. Similarly, PA1088–1090, which have been shown to constitute components of the flagellin glycan, exhibited close correlation with the FleN Module in our analysis⁶¹. PA1895 was associated with the QscR-1 and RhlR Modules, while PA1897 correlated with the QscR-1 Module; consistent with these associations, the PA1895–1897 cluster has been experimentally validated to participate in quorum-sensing regulation⁶⁸. Furthermore, mutations in PA3518 and PA3519 were previously shown to reduce P. aeruginosa copper sensitivity, and both genes displayed strong associations with the CueR Module in our study⁶⁹. Together, these observations reinforced the consistency between our Module-based inferences and HP functions reported in the literature, underscoring the robustness of this approach for functional inference.

HPs exhibit dynamic evolutionary patterns across both P. aeruginosa strains and diverse bacterial species. The newly identified 22 ARHPs and VRHPs showed different conservation patterns among 24 bacterial strains. Six HPs (PA2430, PA3335, PA1211, PA1214, PA1213, PA1215) were exclusive to P. aeruginosa, while four HPs (PA0319, PA0392, PA3998, PA3764) were found only within the Pseudomonas genus. The remaining 11 HPs had homologous proteins in non-Pseudomonas bacteria. Among these more distantly related species, Neisseria meningitidis Z2491 contained a homolog of PA4852, and Bordetella bronchiseptica 253 possessed homologs of 6 HPs (PA4852, PA0588, PA4128, PA2230, PA0586, PA4132). Notably, PA4852, PA0394, and PA0586-88 were highly conserved across more than 10 bacterial species, highlighting them as promising candidates for further investigation in diverse bacterial systems. We believe that the sequence conservation of these HPs with putative functional predictions provides clues for potentially conserved roles across bacterial species, establishing a reference database for subsequent research. Crucially, while all experiment-related observations are managed at the mutant level and rigorously attributed to whole operons rather than individual genes. We projected these collective functional inferences onto each individual constituent gene within our final database to facilitate genome-wide cross-referencing. To ensure maximum data transparency, the exact identities of the specific HPs involved in each multi-gene deletion have been explicitly annotated and itemized in the Suppl. Data 1. Overall, we provided a foundational resource and analytical framework for prioritizing and generating functional hypotheses for unannotated genes, not only in P. aeruginosa but also across the broader bacterial phylogeny. Given the single-condition nature of the primary mutant transcriptomic profiling, we caution that the present predictions are not comprehensive across all environments or infection states, and future multi-condition datasets will be essential to extend this framework to broader physiological contexts. A longstanding reason why many HPs remain unannotated is that some of them may require specific inducing signals to become functional, or may act primarily under extreme or specialized conditions (stress-responsive, niche-specific, or host-induced)^5,70. Notably, the sequencing and experimental conditions examined in this study were limited, which likely explains why we captured only a portion of the complete functional landscape. All HP putative functional prediction results are accessible through an interactive graphical user interface (https://jiadhuang0417.shinyapps.io/PA_HP/).

Methods

Bacterial strains, plasmids, primers, and growth conditions

All research complies with all relevant ethical and biosafety regulations of City University of Hong Kong. All animal procedures were approved by the Department of Health in Hong Kong. Pseudomonas aeruginosa PAO1 (PAO1) and Escherichia coli MG1655 (E. coli) strains were cultured at 37 °C in Luria-Bertani (LB) medium, either on agar plates or in broth with agitation at 220 rpm. Antibiotic selection was applied as follows: PAO1 cultures contained gentamicin (60 μg/mL), trimethoprim (300 μg/mL), or carbenicillin (150 μg/mL), while E. coli cultures were supplemented with kanamycin (50 μg/mL), gentamicin (15 μg/mL), or carbenicillin (60 μg/mL). Given the high intrinsic resistance of PAO1 (MIC > 900 μg/mL), carbenicillin served as the β-lactam selective agent. Complete strain information, plasmid constructs, and primer sequences are provided in Suppl. Data 9.

Generation of deletion mutants and complemented strains

Initially, we planned to build a deletion library for all 2080 annotated HP genes in PAO1. We then excluded HPs located exclusively in the middle of operons from the deletion library list. Thus, we planned to construct 1372 mutants covering 1679 HPs (deposited in Zenodo). Ultimately, due to practical constraints, we obtained RNA-seq profiles for 938 HP mutants (including 1165 HPs). We compared our sets with published Tn‑seq essentiality annotations⁷¹. Of 352 reported essential genes, 29 map to our 2080‑HP universe, and 21 were included in our planned mutant list. Only 3 of these 21 (PA0022, PA0404, PA0442) were successfully deleted, whereas 18 failed during construction—consistent with essential or growth‑critical functions impeding knockout.

Gene deletion mutants were constructed using a SacB-based allelic exchange strategy as previously described in refs. ^39,72. Briefly, flanking regions (~1500 bp upstream and ~1000 bp downstream) of the target gene were PCR-amplified and subsequently cloned into the suicide vector pEX18Ap, generating pEX18Ap-HP. A 0.9 kb gentamicin resistance cassette, excised from pPS858 using XbaI, was inserted into pEX18Ap-HP to yield pEX18Ap-HP-Gm. The resultant constructs were introduced into PAO1 via electroporation, with primary selection for gentamicin resistance. Double-crossover recombinants were identified by their resistance to gentamicin and ability to grow on 5% sucrose, indicating successful gene replacement. Previous studies have demonstrated that retention of the Gm cassette does not exert polar effects on downstream gene expression^34,40,73. To further validate the phenotypes observed in mutant strains, complementation experiments were performed by electroporating pAK1900 plasmids harboring the respective HP genes into the corresponding mutants, with selection for carbenicillin resistance.

Growth curve measurement

Growth phenotypes were assessed by optical density measurements at 600 nm (OD₆₀₀). For the systematic screen of 938 deletion mutants, strains were cultured directly in 96-well microtiter plates, and growth was monitored in a plate reader. Briefly, strains were recovered from glycerol stocks on LB agar, single colonies were inoculated into LB and grown overnight at 37 °C with shaking. Overnight cultures were diluted 1:100 into fresh LB and dispensed into 96-well plates (200 μl per well). Each screening batch included a wild-type (WT) control measured on the same plate to control for batch-to-batch variation. We used the area under the growth curve (AUC) as an integrative measure of growth fitness^74–77. To reduce systematic batch-to-batch variation, we normalized each mutant’s AUC to the wild-type (WT) control measured in the same batch, yielding a relative AUC (AUC_mutant/AUC_WT). We then calculated Z-scores based on the relative AUC distribution, and classified strains with ∣Z∣>3 as having a significant growth defect relative to WT^78,79. For growth curves shown in the main text for ARHPs and VRHPs, strains were instead cultured in 15 mL shaking tubes (LB, 37 °C, aeration by orbital shaking) to better match standard flask growth conditions. At the indicated time points, aliquots were transferred to a 96-well plate and OD₆₀₀ was measured using the same plate reader settings.

RNA-seq library preparation

P. aeruginosa PAO1 wild-type and HP mutant strains were cultured overnight in LB medium containing appropriate antibiotics at 37 °C. These cultures were subsequently diluted 100-fold in fresh LB and grown at 37 °C with agitation (220 rpm) until reaching mid-logarithmic phase (OD₆₀₀ of 0.6, mid-log phase). Bacterial cells were harvested from 2 mL culture aliquots by centrifugation (13,800 x g, 4 °C). Total RNA was isolated using the TianGen RNA extraction kit following the manufacturer’s protocol, including the recommended DNase treatment to eliminate genomic DNA contamination. Ribosomal RNA was depleted according to the RiboRid protocol⁸⁰. mRNA-enriched samples were used to generate cDNA libraries using the NEBNext Ultra™ II RNA Library Prep Kit (New England Biolabs) per manufacturer’s instructions. Libraries were sequenced on an Illumina HiSeq 2000 platform.

RNA-seq data analysis

Public PAO1 RNA-seq datasets were systematically curated from the NCBI Sequence Read Archive (SRA) database, prioritizing studies with a minimum of two biological replicates per condition to ensure statistical robustness across the broadest possible range of experimental contexts. Raw sequencing data underwent comprehensive preprocessing: public datasets were processed using trim_galore (v0.6.10) with automatic adapter detection algorithms and stringent quality filtering parameters (-q 20 --phred33 --stringency 3), while HP library sequences were trimmed using Illumina TruSeq adapter sequences under identical quality control thresholds. Post-trimming data quality was rigorously assessed using FastQC (v0.12.1) to evaluate per-base sequence quality, GC content distribution, and potential contamination, with aggregate quality metrics compiled using MultiQC (v1.17) for batch visualization.

Processed reads were mapped to the P. aeruginosa PAO1 reference genome (RefSeq assembly GCF_000006765.1) using STAR aligner (v2.7.11b) with prokaryote-optimized parameters, specifically setting --alignIntronMax 1 to account for the compact gene architecture typical of bacterial genomes. Gene-level read quantification was performed using featureCounts (v2.0.6), with read counting configured according to sequencing type (paired-end or single-end) and with multi-mapping reads excluded. The resulting count matrices underwent additional filtering to remove low-expression genes before differential expression analysis using DESeq2 (v1.48.2) within the R statistical environment, employing variance stabilizing transformations and Wald statistical tests with Benjamini-Hochberg multiple testing correction.

GO enrichment and functional community definition

Differential expression data were categorized by individual mutant names. Functional over-representation analysis (ORA) for the DEGs of each mutant was performed using the enricher function within the clusterProfiler package (v4.18.4). PA-specific GO annotations were downloaded from the Pseudomonas Genome Database (https://www.pseudomonas.com/). Enriched GO terms were filtered at a significance threshold of p-value < 0.01. To mitigate redundant GO terminology and synthesize overarching biological themes, the enriched GO terms for each mutant were clustered using the aPEAR package (v1.0) via default parameters. We defined the clusters as functional communities. The chord diagram was constructed and visualised using the circlize R package (v0.4.18). Duplicate communities within the same mutant were merged to yield a set of distinct functional attributes per mutant. Mutants significantly enriched in more than one distinct functional community (n_communities > 1) were retained, and all possible pairwise combinations (n = 2) of their enriched communities were generated. The connections (links) within the Circos plot delineate the co-occurrence of two functional categories within the same mutant genotype. The counts (chord thickness) thus reflect the exact number of distinct HP mutants that share that specific pairwise functional footprint, revealing underlying functional axes mediated by these mutants.

Permutation-based statistical validation of the functional communities connectivity

We implemented a gene-level Monte Carlo permutation framework through 10,000 independent sampling iterations. For each individual biological mutant, the precise count of experimentally identified DEGs was recorded. A corresponding randomized pseudo-mutant gene set was then generated by stochastically sampling the same number of genes from the total Pseudomonas aeruginosa PAO1 genomic background. Within each iteration, the complete ensemble of 938 randomized pseudo-mutant gene sets sequentially underwent identical downstream analysis, including GO enrichment and aPEAR clustering pipeline. The pairwise co-occurrence frequencies across all identified functional communities were computed for each run to construct an empirical null distribution for every interconnected functional axis. The empirical p-value for each observed functional link was calculated using the following formula.

$$P_{emp}=\frac{{\sum }_{m=1}^{M}I\left(F_{rand,m}\ge F_{obs}\right)}{M}$$ 1

Where $F_{obs}$ represents the experimentally observed co-occurrence frequency of a given functional link (our real data), $F_{rand,m}$ represents the co-occurrence frequency of the corresponding link in the randomized simulation, $M$ denotes the total number of independent sampling iterations ($M$ = 10,000), and $I$ is an indicator function that equals 1 if the condition is satisfied and 0 otherwise. Inter-community functional connections exhibiting a $P_{emp}$ < 0.05 were considered statistically robust.

RNA extraction and quantitative PCR analysis

Total RNA was isolated using the RNA extraction kit (Tiangen Biotech) according to the manufacturer’s protocol. RNA concentration and purity were quantified using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific). First-strand cDNA synthesis was performed using the FastKing RT Kit (Tiangen Biotech) following the manufacturer’s instructions.

Quantitative PCR (RT-qPCR) was conducted using the SuperReal Premix Plus (SYBR Green) Kit (Tiangen Biotech). Each 10 μL reaction contained 100 ng of cDNA template and 200 nM of each primer. All reactions were performed in technical triplicates. The 16S rRNA gene served as an endogenous reference control for normalization. Amplification was performed on a QuantStudio 3 Real-Time PCR System (Thermo Fisher Scientific) using the following thermal cycling conditions: initial denaturation at 95 °C for 10 min, followed by 40 cycles of denaturation at 95 °C for 10 s and combined annealing/extension at 60 °C for 30 s. Relative gene expression was calculated using the 2^(-ΔΔCt) method.

Antibiotic susceptibility testing

Minimum inhibitory concentrations (MICs) were determined for P. aeruginosa PAO1 wild-type and HP mutant strains against five clinically relevant antibiotics: meropenem, polymyxin B, ciprofloxacin, tetracycline, and trimethoprim. Bacteria were cultured overnight in LB medium at 37 °C with agitation (220 rpm) and subsequently diluted to approximately 10^5-10^6 CFU/mL. Serial two-fold dilutions of antibiotics were prepared in 96-well plates, and 5 μL of the diluted cultures were inoculated into 120 μL of fresh LB containing the respective antibiotic concentrations. After 24 h of incubation at 37 °C, the MIC was defined as the lowest antibiotic concentration that completely inhibited visible bacterial growth.

C4-HSL detection

C4-HSL activity was measured using an E. coli-based reporter system as previously described in ref. ⁸¹. The reporter strain harboring both pKD-rhlI-lux (rhlI promoter-lux fusion) and pMCSG19-rhlRI (encoding RhlR and RhlI) was cultured overnight in LB containing kanamycin (50 μg/mL) and carbenicillin (60 μg/mL) at 37 °C. The culture was diluted to an OD₆₀₀ of 0.1 in fresh selective LB medium. P. aeruginosa wild-type, mutant and complementary strains were grown overnight in LB at 37 °C, and cell-free supernatants were obtained by centrifugation (18,800 × g 15 min). For bioassays, 120 μL of diluted reporter culture was combined with 30 μL of P. aeruginosa supernatant in a 96-well microplate. Luminescence was quantified after 4 hours of incubation at 37 °C with agitation using a Synergy 2 Plate Reader (BioTek).

C12-HSL detection

The bioassay for C12-HSL was conducted similarly to the C4-HSL assay⁸¹, using an E. coli reporter strain containing pMSG19-lasR (expressing LasR) and a lasI promoter-lux fusion. The experimental protocol was identical to that described for C4-HSL detection.

PQS detection

PQS signal detection was performed according to a modified version of a previously described method⁸². The reporter strain was constructed by transforming pKD-pqsA-lux (pqsA promoter-lux fusion) into a P. aeruginosa ΔpqsA mutant (deficient in PQS production). The reporter was cultured in LB containing trimethoprim (400 μg/mL) at 37 °C for 16 h and subsequently diluted to an OD₆₀₀ of 0.1 in fresh LB. Cell-free supernatants from P. aeruginosa strains were obtained as described above. Bioassays were performed by combining 25 μL of supernatant with 125 μL of diluted reporter culture in a 96-well microplate, and luminescence was measured as described for the C4-HSL and C12-HSL bioassays.

Measurement of pyocyanin production

Pyocyanin assay was performed according to previous study with modification⁸³. The mutants of PAO1 strains were grown overnight in LB liquid medium at 37 °C with shaking at 250 rpm. Cultures were transformed into fresh PYO medium containing 6 g/L alanine, 0.1 g/L K₂HPO₄, 2 g/L MgSO₄, 1% glycerol and 0.1 g/L ferric citrate. After 14 h cultivated, the cells were harvested by centrifugation at 18,800 × g for 15 min. The supernatants were transferred into a 96-well plate, on which the OD₆₉₅ of each sample was measured.

Swarming motility

Swarming motility was assessed on plates containing 0.4% agar, 8 g/L nutrient broth, and 5 g/L glucose, as previously described in ref. ⁸⁴. Aliquots (2.5 μL) of overnight LB cultures were spotted onto the center of swarming plates. After 18 hours of incubation at 37 °C, the diameter of the swarming zone was measured. Images were captured using a Bio-Rad imaging system.

Congo red binding assay

Biofilm matrix production was evaluated using a Congo red binding assay⁸⁵. LB-based agar plates (1%) were supplemented with 0.05% (v/v) glutamate, Congo red (80 μg/mL), 2 mM MgSO₄, and 0.2% (v/v) glucose. Aliquots (2.5 μL) of overnight cultures were spotted onto these plates and incubated at 37 °C overnight followed by 3–4 days at room temperature.

Mouse infection assay

Mouse infection assays were performed according to a previously established protocol⁴⁰. HP mutants, wild-type strains, and complementary strains were grown overnight in Luria-Bertani broth at 37 °C with shaking at 220 rpm. Bacteria were subsequently inoculated into fresh LB broth and cultured to mid-logarithmic phase. Wild-type C57BL/6 female mice (4–6 weeks), which are susceptible to P. aeruginosa infection, were used for all experiments. Mice were housed in a specific pathogen-free (SPF) facility under a regulated 12 h light/12 h dark cycle, with the ambient temperature maintained at 22 ± 2 °C. 10⁷ colony-forming units (CFUs) of P. aeruginosa were intranasally instilled into each mouse (three mice per group). The mice were euthanized 18 hours post-infection to quantify P. aeruginosa CFUs in the lungs. Excised lungs were homogenized in phosphate-buffered saline, and serial dilutions were plated on Pseudomonas isolation agar (PIA). After overnight incubation at 37 °C, P. aeruginosa colonies were counted. The intranasal infection model employed is a well-established method for studying P. aeruginosa respiratory infections. This model enables bacteria to progress naturally from the nasal cavity to the lungs, closely mimicking natural infection routes^40,86. All animal procedures were approved by the Department of Health in Hong Kong and the ethical approval number is (24-12) in DH/HT&A/8/2/5 Pt.14.

Virulence pathway analysis

Pathway transcriptional response proportions were calculated by normalizing the number of pathway-specific DEGs to the total gene count in each pathway. For each HP mutant and pathway combination, the response proportion (p) was computed as the ratio of pathway-associated DEGs (d) to the total pathway gene count (P). i: Virulence pathway index.

$$p=\frac{d_i}{P_i}$$ 2

We fitted a normal distribution to the p-values of HPs within each pathway and defined pathway-specific significance thresholds based on the 95th percentile of the right tail, ensuring that the cutoff is tailored to the underlying distribution of each individual pathway. HPs passed thresholds were considered as virulence pathway-associated candidate HPs.

Robust independent components calculation

To derive robust independent components from RNA-seq datasets, log-transformed transcripts per million (log(TPM)) values were calculated and quality control was applied by retaining only samples with Pearson correlation coefficients ≥ 0.8 between biological replicates for both public and HP datasets. Each dataset was normalized relative to its corresponding reference condition to eliminate technical variation and mitigate batch effects.

Independent component analysis (ICA) was performed following the iModulonMiner⁸⁷ framework. FastICA (v1.7.2) was run 120 times for each tested dimensionality with a convergence tolerance of 10⁻⁷ with random seeds. Pairwise distance between components was calculated as one minus the absolute Pearson correlation coefficient. Robust independent components were identified using DBSCAN clustering (scikit-learn v1.7.2) with an epsilon of 0.1 and a minimum cluster size of 50. All components meeting these clustering criteria were retained as robust Modules. We performed ICA at 1000 dimensions to prevent under-decomposition, given the exceptionally large size of our dataset. For Module annotation, we used PyModulon (0.2.1) together with the latest TRN.

Two perspectives for Matrix A

For perspective 1, we identified outlier responses by analyzing Module activities across HP mutant conditions. Activity scores were grouped by Module under different HP mutant conditions. We calculated the first (Q1) and fifth (Q5) quartiles of activities and their interquartile range (IQR). Outliers were defined as observations falling outside Q1 - 2×IQR or Q5 + 2×IQR thresholds, representing significant repression or activation, respectively. For perspective 2, we calculated the Pearson correlation coefficients between condition-specific HP expression patterns and Module activities under different conditions. We modeled the global distribution of Pearson correlation coefficients between all HPs and Modules using a normal background distribution. Significant associations were defined using the top 0.05% tail of the fitted distribution, corresponding to a cutoff of r = 0.396, which was rounded to 0.4 for simplicity and interpretability. HPs with an absolute Pearson correlation coefficient of at least 0.4 were considered significant in perspective 2.

Pan-genome analysis, SNP calling, dN/dS calculation and MST tree

A total of 1295 P. aeruginosa strains were downloaded from the NCBI database. Pan-genome analysis was performed using Roary⁸⁸. SNPs were called using PAO1 as a reference genome via Snippy (v4.6.0). dN/dS were calculated using SNPGenie⁸⁹. The Minimum Spanning Tree (MST) was constructed to visualize the relationships between TF sets based on their Jaccard distances. All analysis and plot codes were uploaded to GitHub.

Statistical analysis

Different statistical tests for specific experiments and analyses, labeled in the figure legends. RNA-seq experiments have two replicates, and other experiments were performed at least three times. For GO enrichment analysis, adjusted p values were calculated with Benjamini-Hochberg methods. No statistical method was used to predetermine the sample size. No data were excluded from the analyses.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

X.D. and X.S. conceived the project and designed experiments. C.Y., J.L., X.S., and Y.Z. carried out major experiments. C.Y., B.L., L.X. and Y.D. analyzed data and prepared figures and tables. C.Y., B.L., L.X., X.S. and X.D. wrote and revised the manuscript. All authors read and approved the final manuscript. J.H., Y.S., N.L., F.C., S.L., T.L., T.F., Y.L., G.Q. and A.Y. were involved in the project. All authors read and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Matthew Wolfgang and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This study was funded by the Guangdong Major Project of Basic and Applied Basic Research (2020B0301030005), National Natural Science Foundation of China (32272619 and 32470112 to X. S.; 32570202 to X.D.), General Research Funds of Hong Kong (11102223, and 11103225 to X.D.), Collaborative Research Fund (C7033-20G to A.Y.). The funders had no role in study design, data collection, interpretation, or the decision to submit the work for publication.

Data availability

RNA-seq data are available in the National Center for Biotechnology Information Gene Expression Omnibus database under accession codes GSE256222. DEGs, M matrix, and A matrix are deposited to the Zenodo repository (https://doi.org/10.5281/zenodo.20453783)⁹⁰. A comprehensive PA-HP database allows researchers to efficiently explore their HPs (input PA locus) of interest, thereby facilitating further functional studies in P. aeruginosa. The website is https://jiadhuang0417.shinyapps.io/PA_HP/. Source data are provided with this paper.

Code availability

All related codes were uploaded to GitHub (https://github.com/LuBeifang/PA-HP)⁹¹.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77512-1.