The PhageExpressionAtlas reveals shared and unique transcriptional patterns across phage–host interactions

Abstract

Time-resolved transcriptomic profiling has been used to study phage–host interactions for more than a decade. However, the resulting datasets are not readily accessible for custom re-analysis, and resources are lacking that provide standardized processing, storage, and analysis of transcriptomes from phage infections. Here, we present the PhageExpressionAtlas, the first bioinformatics resource for storing time-resolved dual RNA-sequencing data from phage infections. This data was processed uniformly using a custom analysis pipeline and is presented for interactive exploration through visualization. The PhageExpressionAtlas currently hosts 42 datasets from 23 studies. Using the PhageExpressionAtlas, we replicate key findings from original publications and extend hypothesis testing across multiple phage–host systems. By systematically querying and analyzing the underlying database, we evaluate approaches to phage gene classification and find that uncharacterized phage genes dominate all infection phases, with their distribution depending strongly on the classification strategy. Moreover, we provide a comprehensive view of the expression dynamics of anti-phage defenses as well as host- and phage-encoded anti-defense systems in the infection context, indicating unique and conserved patterns of transcriptional regulation underlying bacterial anti-phage immunity and phage counter-strategies. Together, the PhageExpressionAtlas is a unifying resource that democratizes transcriptomics-driven analyses of phage–host interactions and supports integrative cross-study assessment.

Graphical Abstract

Graphical Abstract.

Introduction

Bacteriophages (phages) are viruses that infect bacteria and can lyse their hosts, thereby shaping microbial communities across diverse ecosystems. In recent years, phages have also gained increasing relevance in biotechnology and medicine. Phage-derived enzymes widely used in molecular biology [1], anti-phage defense systems such as CRISPR–Cas [2, 3], phage-encoded anti-defense strategies that counteract these systems [4], and additional mechanisms enabling host takeover [5, 6] have become central topics of investigation. In parallel, antibiotic resistance continues to rise as a major global health threat, motivating renewed interest in phages as alternatives or complements to antibiotics, including phage therapy [7].

To fully exploit the potential of phages and their genomic repertoires, a mechanistic understanding of phage–host interactions is essential. One key approach is dual RNA sequencing (dual RNA-seq), which simultaneously quantifies phage and host transcriptomes across infection stages [8]. Time-resolved dual RNA-seq studies across diverse phage infections have yielded fundamental insights into host responses, including anti-phage defense, as well as phage anti-defense and takeover programs and their interplay [9–13]. While these studies illuminate specific facets of individual phage–host systems, their primary analyses are typically framed by the hypotheses and functional annotations available at the time. As phage research continues to uncover new interaction mechanisms and gene functions, previously generated dual RNA-seq datasets become increasingly valuable for re-analysis and cross-study comparison, yet are rarely revisited in a systematic manner.

Notably, dual RNA-seq data are typically not processed in a standardized manner and are rarely made available through dedicated databases in a format that supports rapid re-analysis. Moreover, visual exploration options that are accessible to non-bioinformaticians remain limited, with only few exceptions [10]. As a consequence, comprehensive and systematic analyses—and especially comparisons across phage infections—are still lacking, despite their potential to substantially advance mechanistic understanding and to inform applications in biotechnology and therapy. It is therefore instrumental to provide a resource that enables access to, and interactive visualization and exploration of, transcriptome-resolved phage infections. While databases and tools for phage genomics and phenotypes are available [14, 15], comparable resources that facilitate gene expression-centric analyses are currently missing.

Here, we present the PhageExpressionAtlas, the first resource in phage research that enables interactive, visual exploration of phage–host interactions at the transcriptome level. To this end, we implemented a unified processing and analysis framework and uniformly processed 42 public time-resolved dual RNA-seq datasets spanning diverse phage–host systems across infection stages from 23 studies. The atlas includes multiple clinically relevant hosts, including pathogens such as Staphylococcus aureus and Pseudomonas aeruginosa, and covers both model and therapeutic phages, thereby representing a broad spectrum of infection modalities. Users can query the collection, download processed data, and visualize phage and host gene expression dynamics across infection phases for selected phage–host interactions using heatmaps and profile plots. In addition, PhageExpressionAtlas supports classification of phage genes into early, middle, and late expression classes, visualization of their temporal profiles, and inspection of their genomic context. This functionality provides insights into infection-stage-specific programs, supports functional annotation of phage genes, and facilitates investigation of genome organization. Leveraging the aggregated dataset collection, we systematically assess the transcriptional dynamics of bacterial anti-phage defense systems and identify programmed expression of phage-encoded anti-defense systems, extending current knowledge by placing defense and counter-defense modules into their native infection contexts.

Altogether, the PhageExpressionAtlas provides a foundation for a growing visualization and analysis resource that democratizes transcriptomics-driven phage research and enables integrative cross-study assessment. The PhageExpressionAtlas is available at https://phageexpressionatlas.cs.uni-tuebingen.de.

Materials and methods

Collection and curation of dual RNA-seq datasets of phage infections

Data collection

We searched PubMed, GEO, and Google Scholar for dual RNA-seq datasets of phage infections using the following keywords: “RNA-seq” or “dual RNA-seq” together with “phage,” “phage-host,” “phage infection,” “time-resolved.” Datasets were included in the PhageExpressionAtlas if they met the following criteria:

• the dual RNA-seq data cover at least two time points during phage infection,

• raw sequencing data are publicly available or, if not publicly available, were provided by the authors upon request,

• annotation of sequencing data is sufficient to assign samples to replicates and time points/conditions,

• strain identity, infection time points, replicate structure, raw-data accession, and compatible host/phage reference genomes were available.

Supplementary Table S1A summarizes all data accessions, associated study metadata, and the reference genome accessions used, which were retrieved from the respective repositories until August 2025. Overall, the collection spans a diverse range of phages—both temperate and virulent—and hosts (Fig. 1A and B). It is possible that other datasets may have been missed by the search and we welcome submissions—in form of notification by e-mail—by researchers about datasets that shall be included.

Figure 1.

Features, processing, and storage of dual RNA-seq datasets in the PhageExpressionAtlas. (A) Number of datasets collected for and represented in the PhageExpressionAtlas grouped by host species (summarizing different strains of the same species) and the phages infecting the hosts. Phages are denoted by letters from A to Y, bar height corresponds to the number of available datasets in the atlas. (B) Pie chart illustrating the fractions of temperate and virulent phages across all datasets. (C) Schematic illustration of key steps of dual RNA-seq data processing using the custom Nextflow pipeline and interactive downstream processing. Processing requires host and phage genome information alongside the raw dual RNA-seq data. (D) Schematic representation of the SQlite database, in which processed dual RNA-seq data are stored (“matrix_data”) in four different formats alongside phage and host information and the corresponding genome annotations.

Data processing

For unified processing of raw RNA-seq reads, we implemented a Nextflow [16] pipeline that performs all computational steps required to generate gene-level count tables (Fig. 1C). The pipeline comprises quality control with FastQC (v0.12.1) [17], adapter trimming with Cutadapt (v4.9) [18], and read mapping with HISAT2 [19] against a combined reference consisting of the host and phage genomes. Alignments were processed with Samtools (v1.6), and gene-level quantification was performed with featureCounts (v2.0.1). We retained primary alignments only, while allowing multi-mapping reads and reads overlapping multiple features to be counted. Finally, we employed MultiQC (v1.32) to summarize quality-control and processing metrics across samples.

To further process the resulting count tables, we implemented a Jupyter Notebook that applies a consistent post-processing workflow across datasets while permitting dataset-specific adaptations where required (Fig. 1C). Count tables were depleted of rRNA features retrieved from the genome annotation, followed by per-sample transcripts-per-million (TPM) normalization:

where the TPM of gene in a given sample is computed from its read count and gene length , and normalized by the sum of length-normalized counts across all genes in that sample. In addition to TPM, we generated a log+1-transformed representation of the data.

Replicate consistency was assessed by principal component analysis, and outlier replicates were removed in cases, where samples grouped well per time point but a single replicate clearly clustered with a different time point. This resulted in the removal of only one sample overall (sample t4_R1 in the Wolfram-Schauerte_2022 dataset). Replicates were then aggregated within each time point by computing the mean of TPM values (“TPM means”), which serves as the basis for the heatmaps displayed in the PhageExpressionAtlas. Importantly, we calculated a mean-stabilized variance by dividing the variance of each gene’s expression over time by its mean expression over time, which we use to rank order genes to display gene subsets in heatmaps.

To visualize temporal expression profiles, we further computed “fractional” expression from TPM means by dividing each gene’s TPM value at each time point by its maximal TPM value across all time points.

Phage gene classification assigns phage genes to early, middle, or late infection phases based on their temporal expression patterns. Following common practice in the field, we pre-computed dataset-specific expression classes during processing using study-derived phase boundaries on replicate aggregated TPM values (TPM means). Specifically, for each gene we determined whether its expression exceeded a fraction of its maximal expression across time points and then assigned the class according to the corresponding time bin:

 where is a relative expression threshold and , , and denote dataset-specific boundaries for the early, middle, and late infection phases, respectively. For pre-computed classes, we used (“Class Max”) and (“Class Threshold”). In most datasets, these settings recapitulated the phase-resolved gene class distributions reported in the original studies.

In addition, the atlas supports user-defined classification based on custom infection-phase boundaries (, , ) and a user-selected threshold relative to each gene’s maximal expression across time points. As a complementary, unsupervised approach, which is less dependent on choice of time points and thresholds, the atlas also allows users to cluster phage genes based on their temporal expression profiles (using fractional expression) with k-means (based on scikit-learn 1.7.2). For this, users can select the number of clusters k to be between one and five. The resulting clusters are sorted by the ascending time point of maximal expression of the cluster representative. Since the assignment of the clusters to temporal gene classes is non-trivial, we provide numeric cluster labels. Only for  clusters, clusters are labeled as “early-like,” “middle-like,” and “late-like.”

To enable efficient storage and querying of host and phage genomes, GFF files were restricted to gene features and reformatted. Where overlapping gene symbols occurred, we generated unique symbols; if no symbol was available, the gene identifier was used. We further improved the annotation of protein-coding phage genes by running Pharokka (v1.8.2) [20] with default settings on the protein FASTA files sorted by genomic coordinates, and we added resulting annotations and functional categories to the re-formatted GFF files. For annotated non-coding features (e.g. tRNAs), the corresponding biotype [transfer RNA (tRNA), non-coding RNA (ncRNA), etc.] was added to the category field.

Database creation

We implemented an SQLite3 database to store all data underlying the PhageExpressionAtlas (Fig. 1D). The database comprises five tables. The central table (“dataset”) stores dataset-level records together with associated study metadata and identifiers for the corresponding phage and host. For each dataset, we store four expression matrices to support different downstream analyses and visualizations: (i) the raw count matrix, (ii) the per-sample TPM-normalized matrix, (iii) mean TPM values aggregated per time point (“TPM means”), and (iv) the matrix of fractional expression values. Each matrix is stored as a pickled pandas DataFrame. Phage and host identifiers link to dedicated tables containing organism-level metadata, which in turn link to genome annotations derived from the corresponding GFF files and likewise stored as pickled pandas DataFrames.

We provide the database creation workflow as a Jupyter Notebook, which can be adapted to build a custom database and to run a local instance of the PhageExpressionAtlas. This enables private deployment, for example, to explore datasets that cannot yet be shared publicly.

To manage updates of external resources such as Pharokka [20] or DefenseFinder [2] for functional phage and/or host gene annotation, we will update the entire database of the PhageExpressionAtlas annually by systematically running these annotation tools across all present genomes. New datasets submitted to us by users will be added to the atlas on-demand.

PhageExpressionAtlas frontend and backend

Interface design and functionalities

Building on this foundation, we developed the PhageExpressionAtlas as an interactive web application to facilitate transcriptomics-based (re-)investigation of phage infections. The atlas enables users to search the database and visually explore selected datasets representing phage–host interactions of interest using common visualizations such as heatmaps and profile plots. In addition, expression-based classification of phage genes into infection phases supports hypothesis generation and functional interpretation of phage genes.

We structured the PhageExpressionAtlas into five pages. The landing page (“Home”) introduces the atlas, summarizes available exploration options of the website, and provides a brief background on the relevance of phage research and the role of dual transcriptomics for this (Supplementary Fig. S2).

The “Data Overview” page summarizes dataset characteristics as well as processing and storage in the database of the PhageExpressionAtlas (Fig. 1C and D; and Supplementary Fig. S3). It also provides descriptive statistics on the distribution of hosts and phages across included datasets (Supplementary Fig. S3A and B). A comprehensive, searchable table allows users to retrieve dataset details, including links to the original publications (Supplementary Fig. S3C). For each dataset, dedicated buttons link to the “Dataset Exploration” or “Genome Viewer” pages with the corresponding dataset pre-selected.

The “Dataset Exploration” page serves as the core interactive visualization interface of the PhageExpressionAtlas (Supplementary Figs S4 and  S5). Users can select a phage, host, and dataset of interest and obtain an overview of time-resolved expression dynamics (Supplementary Fig. S4A). In total, three visualization panels are displayed. The upper panel displays a heatmap of z-score-normalized expression (per gene) for all phage genes (left) based on mean TPM values per time point (Supplementary Fig. S4B). On the right, a corresponding heatmap shows the same number of the most variable host genes. As for all visualizations in the atlas, help buttons provide contextual information about the displayed plots, the underlying data, and available selection options.

The middle panel provides the results of phage gene classification functionality. Fractional expression profiles across the measured time points are shown for each phage gene as profile plots and colored according to user-defined classification criteria (Supplementary Fig. S4C). Users can select pre-computed classes (“Class Max” and “Class Threshold”) or perform custom classification by defining infection-phase boundaries via time-point thresholds and a relative expression threshold with respect to each gene’s maximal expression over the infection time course. The profile plots provide immediate visual feedback on the resulting classification, enabling rapid qualitative assessment given the observed expression trajectories.

The lower panel enables generation of z-score-normalized heatmaps and profile plots for user-selected phage and host genes (Supplementary Fig. S5). In addition, for phage genes, also a linear genome track is shown above the plots, allowing genes to be selected in their genomic context and highlighted to indicate the current selection. This track is also linked to the “Genome Viewer” page. This facilitates exploration of specific genomic loci and their transcriptional dynamics. An analogous selection workflow is available for host genes, but without an integrated genome visualization (Supplementary Fig. S5).

All plots on the “Dataset Exploration” page can be downloaded, and phage gene classes generated in the classification panel can be exported. Beyond these dedicated components, interactivity is supported throughout the page via linked selections.

The “Genome Viewer” page provides an additional perspective on time-resolved phage transcriptomes by integrating transcriptomic and genomic information (Supplementary Fig. S6A). Users can select a phage and the corresponding dataset, after which phage genes are classified using one of the three options also available on the “Dataset Exploration” page. The phage genome is then displayed as a circular genome plot, with an inner ring indicating functional categories (e.g. DNA, RNA, and nucleotide metabolism) and an outer ring highlighting the inferred temporal class (early, middle, late). Each ring consists of two tracks indicating the forward and reverse strand.

A dynamically adjustable selection window allows users to zoom into a genomic locus of interest and display it as a linear view on the right-hand side (Supplementary Fig. S6A). In this view, individual genes and their functional annotations and temporal classes can be interactively inspected, facilitating interpretation of transcriptional programs in their genomic context. For example, the T4 genome illustrates the organization of genes into loci with temporally coordinated transcription (Supplementary Fig. S6A), consistent with prior work [10, 21, 22].

To encourage researchers to contribute data to the PhageExpressionAtlas, we designed a “Data Submission” page, where we describe how to submit data to the atlas (Supplementary Fig. S6B). To ensure consistent processing and annotation, an Excel template is provided in which the user can add the relevant information on the raw sequencing data, as well as other information (e.g. metadata and reference genomes), so that the data can be processed for the atlas by us.

To support navigation of the PhageExpressionAtlas, we provide a comprehensive “Help” page that includes written instructions and video demonstrations for each page of the atlas.

Implementation

The PhageExpressionAtlas frontend was implemented in HTML, CSS, and JavaScript (JS). For UI elements (e.g. components and buttons), we used the open-source web component library Shoelace (v2.20.0).

The phage-shaped word cloud on the landing page was generated using the ECharts wordcloud extension (v2.1.0). Genome viewers on the “Dataset Exploration” and “Genome Viewer” pages were implemented with GoslingJS (v0.11.0). On the “Data Overview” page, the Sankey diagram and pie charts were created with Apache ECharts (v5.4.0), the sunburst plot with Plotly.js (v3.0.0), and the dataset table with Tabulator (v6.3.0). All interactive plots (profile plots and heatmaps) were implemented with Plotly.js. Profile plots use fractional expression values computed from mean TPM values per gene and time point (see above). Heatmaps are based on mean TPM values and are z-score-normalized using SciPy’s zscore function (v1.15.2), which is also used for subsequent gene-wise clustering. k-means-based clustering for phage gene classification is performed using the scikit-learn KMeans function (v1.7.2).

The backend was implemented in Python using Flask (v3.1.0) as a lightweight web application framework to handle API requests and URL routing. Database access is mediated via SQLAlchemy (v3.1.1), which connects the frontend to the SQLite database.

Automated accessibility checks were performed for every page according to WCAG 2.1 AA using the page scanner of the Chrome extension “Accessible Web Helper”.

The PhageExpressionAtlas runs in a Docker container hosted on the infrastructure of the Institute for Bioinformatics and Medical Informatics (IBMI) at the University of Tübingen and is integrated into the Tübingen Visualization Server (TueVis, https://tuevis.cs.uni-tuebingen.de).

Use cases for the PhageExpressionAtlas

We conducted two use cases with the PhageExpressionAtlas web interface. For the first use case, we analyzed the dataset from Wolfram-Schauerte et al. [10], which profiled the dual transcriptome of T4 phage infection of E. coli before infection () and at four time points post infection. We used the PhageExpressionAtlas to reproduce key findings from the original study, focusing on phage gene expression dynamics.

For the second use case, we analyzed Kayvirus phage infection of two S. aureus strains (SH1000 and Newman) from Finstrlova et al. [13] by studying host ncRNA expression during phage expression, an aspect not addressed in the original publication.

Integrative analysis with the PhageExpressionAtlas database

Beyond the use cases of the PhageExpressionAtlas web interface, we systematically queried its database allowing for integrative analysis across different phage infections.

Probing phage gene classification

To evaluate the classification of phage genes into early, middle, and late w.r.t. infection time across diverse infections, we selected 18 datasets from 12 studies that include a pre-infection measurement and sample time points spanning the three major infection phases (Supplementary Table S1B).

For benchmarking, we defined a proxy ground truth for late gene expression based on the assumption that genes involved in virion assembly and host lysis are expressed late during infection. Accordingly, genes annotated by Pharokka as “connector,” “lysis,” “head and packaging,” “tail,” and “integration and excision” were treated as ground-truth late genes. These functions are typically enriched among late-expressed genes [8, 21]. While this proxy ground truth is necessarily imperfect and incomplete, it provides a practical benchmark for large-scale comparison across datasets. Using these labels, we assessed classification performance with the following three standard metrics.

Recall quantifies the fraction of ground-truth late genes that are predicted as late:

where true positives (TP) denote ground-truth late genes classified as late, and false negatives (FN) denote ground-truth late genes not classified as late. Under the assumption that these late labels are correct, recall represents the primary metric.

Precision measures the fraction of classified late genes that are labeled as ground-truth late:

where false positives (FP) are genes classified as late but not labeled as ground-truth late. Because the annotation-based ground truth is likely incomplete and may vary across phage taxa, we did not expect precision to be consistently high.

As a third measure, we computed the overall fraction of genes classified as late among all non-ground truth late genes to monitor how parameter choices affect the class distribution.

We defined a grid search over (i) expression thresholds relative to each gene’s maximal expression across time points () and (ii) relative infection-time boundaries used to bin measurements into early, middle, and late phases. For each dataset, we iterated over all parameter combinations and performed classifications using the criterion defined above. Our analysis focused primarily on the pre-computed “Class Max” and “Class Threshold” settings.

Further, we performed k-means clustering to classify phage genes into early, middle, and late genes using . To assign the clusters to the respective classes, the cluster centroids were sorted by the time point of their maximal expression. For example genes in the cluster with earliest maximal expression are assigned to the “early” class.

For the systematic assessment of phage gene class distributions across the 18 datasets, some phages were represented by multiple datasets. To ensure that each phage gene contributed only once to the cross-phage summary, we computed a consensus class per gene across datasets. Classes were encoded numerically as early = 0, middle = 1, and late = 2. For genes present in multiple datasets, the mean class value was calculated and converted back to a discrete class using nearest-integer rounding, with half values assigned to the later class. Thus, mean values below 0.5 were assigned to early, values from 0.5 to below 1.5 to middle, and values of 1.5 or higher to late.

Expression of phage and host orthogroups and (anti-)defense systems

For the integrative analysis across infections, we selected 18 datasets from 12 studies that include a pre-infection measurement and exhibit clear transcriptional signatures of phage takeover (Supplementary Table S1B).

For all phage and host genomes, we used OrthoFinder (v3.1.2) [23] to infer orthogroups from the amino acid sequences of annotated protein-coding genes. In addition, we ran DefenseFinder (v2.0.1) [2] on the protein sequences sorted by genomic location to identify anti-phage defense systems in hosts as well as phage-encoded anti-defense systems.

For each dataset, we extracted gene expression values (fractional expression or z-score-normalized TPM means per measured time point) for genes assigned to orthogroups or to defense/anti-defense systems. Because datasets span heterogeneous time series, we harmonized infection progression by binning each time course into the following relative intervals: uninfected (pre-infection time point), immediate early (first 10%), early (subsequent 15%), middle (subsequent 25%), and late (final 50%) of the sampled infection duration. We visualized expression dynamics across datasets using boxplots (fractional expression and z-score-normalized TPM means per bin), accompanied by bar plots indicating per-dataset gene contributions. In addition, we used profile plots to display fractional expression trajectories of individual genes within each dataset.

Results

Uniform processing and storage of dual RNA-seq data

By searching PubMed, GEO, and Google Scholar, we found and collected 42 time-resolved dual RNA-seq datasets from 23 peer-reviewed studies, which met our selection criteria (see the “Materials and methods” section) and together span a broad range of phage–host interactions (Fig. 1A and Supplementary Table S1A). Datasets from the same study often assess the same phage infecting different host strains, the same host infected by distinct phages, or infections with phage cocktails. The largest fraction of datasets covers infections of P. aeruginosa, followed by S. aureus and E. coli (Fig. 1A). Overall, the collection covers transcriptomes from infections by 25 bacteriophages, predominantly virulent phages (84%) (Fig. 1B).

For some datasets, pre-computed count tables are available via GEO; however, for most studies only raw sequencing reads are accessible. We therefore established a Nextflow pipeline to process raw RNA-seq data uniformly and reproducibly into gene-level count tables under a dual-mapping setup in which reads are mapped simultaneously to the host and phage genomes (Fig. 1C). We validated the pipeline on the T4–E. coli dataset (comprising five time points, three biological replicates each) [10] by comparing our TPM-normalized expression matrix to the TPM-normalized matrix provided via GEO. As expected, we observed high concordance () for matched samples (Supplementary Fig. S1B).

We then standardized interactive downstream processing of the count tables, which involves replicate assessment, outlier removal, and in silico rRNA depletion. This workflow not only provides a consistent processing strategy across studies, but also formats the data for database storage. Based on the normalized, aggregated counts tables, phage genes are assigned to the three commonly used temporal classes (“early,” “middle,” “late”) using two strategies: “Class Max” and “Class Threshold.” Replicates are aggregated for visualization within the PhageExpressionAtlas, and variance-stabilized expression values are used to select gene subsets for heatmaps. All normalized and summarized versions of the counts tables can be downloaded via the atlas.

In addition, we improved functional annotation of phage genes using Pharokka [20]. This rarely increased functional gene-level annotation, but consistently provided standardized functional categories suitable for visualization in the atlas. Finally, all processed matrices (see the “Materials and methods” section), together with phage and host metadata and genome annotations, were stored in an SQLite database that constitutes the core data structure of the atlas backend (Fig. 1D).

We thus provide a standardized processing workflow for dual RNA-seq data of phage infections creating a unified data structure. Through application of our pipeline to 42 datasets and their storage in a database, we created the fundamental basis for the PhageExpressionAtlas.

The PhageExpressionAtlas enables transcriptome-based re-analysis of phage–host interactions

The PhageExpressionAtlas provides several functionalities for interactive interrogation of dual transcriptomes during phage infection. We use the term “atlas” by analogy to widely used single-cell transcriptomics atlases [24], which provide access to uniformly processed and integrated single-cell RNA-seq datasets and thereby serve as community “lookup” resources. Depending on the research question, users can obtain an overview of host responses and phage transcriptional dynamics (Supplementary Figs S4 and  S5), classify phage genes and inspect their genomic organization (Supplementary Figs S4 and  S6), or explore specific phage and host genes of interest, for example genes belonging to defined pathways or functional categories (Supplementary Fig. S5).

To illustrate these applications, we first reproduced key findings from a study characterizing the transcriptome of T4 phage infecting E. coli [10]. We immediately captured the dynamic transcriptional program of T4 genes over the infection time course (Fig. 2A), showing a clear separation into early, middle, and late expression classes, as reported previously [10, 22]. Inspection of phage gene classes in the genome viewer recapitulated the known clustering of temporally co-regulated genes in operon-like genomic organization as reported previously [21] (Supplementary Fig. S7A). Finally, using the “Gene Selection” panel, we revisited a canonical transcriptional cascade (Fig. 2B): three middle genes encode the proteins gp33, gp45, and gp55, which assemble into a complex that activates late transcription [25], thereby inducing late genes such as gene 68, which encodes the major capsid protein of bacteriophage T4 [26]. Together, this example highlights how the atlas supports exploration from global transcriptome overview and expression-based gene classification to targeted interrogation of specific regulatory modules.

Figure 2.

Targeted assessment of the transcriptomes of phage infections with the PhageExpressionAtlas. (A, B) Analysis of T4 phage gene expression based on the dataset from Wolfram-Schauerte et al. [10]. The overall trend of phage gene expression over the time course of infection is illustrated in a z-score-normalized heatmap (A). Profile plots of the relative expression of the four T4 phage genes 33, 45, 55, and 68, where each gene’s expression is normalized to its maximal TPM value over the time course of infection (B). Expression of five genes (ssrA, rnpB, clpP, clpX, cls) in S. aureus strain Newman (C) and SH1000 (D) during Kayvirus phage infection [13].

The atlas provides analogous functionality for host transcriptome analysis. To demonstrate this, we revisited the observation that certain ncRNAs, including the transfer-messenger RNA (tmRNA) SsrA and the small RNA RnpB, remain at relatively constant levels during T4 infection, suggesting that they may be host factors required for infection or otherwise maintained by the phage [10]. To assess whether this pattern generalizes, we analyzed Kayvirus infection of two S. aureus strains (Newman and SH1000) [13]. Using heatmaps, we recapitulated the reported global decrease in relative host transcript abundance during infection in both strains [13] and observed strain-dependent shifts in the distribution of phage gene classes (Supplementary Fig. S7B and C). While late phage genes were consistently classified as late in both hosts, infection of strain SH1000 shifted several genes classified as middle in Newman into the early class, indicating host-dependent timing differences in early and middle phage transcription.

We then examined expression profiles of the host ncRNAs ssrA and rnpB alongside essential protein-coding genes involved in proteolytic quality control (clpP, clpX) and lipid synthesis (cls). Whereas the protein-coding genes’ transcripts decreased over the infection time course in both strains, ssrA and rnpB remained comparatively abundant throughout infection (Fig. 2C and D). This observation is consistent with the hypothesis that these ncRNAs are either required during infection or are less susceptible to phage- and/or host-induced RNA turnover. Overall, these use cases illustrate how the atlas enables rapid, cross-system hypothesis testing on both the phage and host sides.

Assessing approaches to phage gene classification

Phage gene classification can provide a complementary route to inference of gene function, since temporal expression programs are linked to the major stages of infection progression. In the PhageExpressionAtlas, we provide two default classification schemes, “Class Max” and “Class Threshold.” “Class Max” assigns each gene to the infection phase in which it reaches its maximal expression, whereas “Class Threshold” assigns the earliest phase in which expression reaches at least 20% of the gene’s maximal value, using infection phase boundaries derived from the original studies. Such heuristic classifications are commonly used in the literature, but are rarely evaluated systematically [10, 13, 22]. We therefore leveraged the PhageExpressionAtlas database to benchmark parameter choices for expression-based phage gene classification.

We selected datasets from the PhageExpressionAtlas spanning diverse phage–host interactions, which cover transcriptomic measurements before infection and across the three major infection phases (early, middle, late) that amount to a total of 18 datasets (Supplementary Table S1B). To quantify classification accuracy, we defined proxy ground-truth labels for a subset of late phage genes—here referred to as ground-truth late genes. To quantitatively measure the classification outcome, we defined three metrics. Recall defines how many of the ground-truth late genes are indeed classified as such, while precision measures how many of late classified genes are indeed ground-truth late genes. Finally, we also quantified the fraction of classified late genes among non-ground-truth late genes.

First, we applied this analysis to the two default schemes, “Class Max” and “Class Threshold.” “Class Max” captured a larger fraction of ground-truth late genes across most datasets, whereas “Class Threshold” showed reduced recall in the majority of cases (Fig. 3A). At the same time, precision was higher for “Class Max,” indicating that a substantial fraction of genes classified as late are represented among the ground-truth late genes. When examining the fraction of additional classified late genes (i.e. those not contained in the ground-truth late genes), we observed low values for most datasets. Notably, the highest fractions occurred in phages with small ( genes) or particularly large ( genes) genomes (Supplementary Fig. S8A). This is consistent with sparse functional annotation in many phage genomes. We performed an additional assessment using different threshold values and time boundaries beyond our predefined classes indicating that “Class Max” represents a solid approach to phage gene classification (Supplementary Figs S9 and  S10). Also, we tested an unsupervised clustering-based approach for phage gene classification, which is independent from the choice of thresholds and infection phase boundaries. For this, we employed k-means using clusters, which we assigned to early, middle, and late classes based on the timing of maximal expression of the cluster representative. We achieved similar recall and precision as for “Class Max” (Supplementary Fig. S11A). In the PhageExpressionAtlas, k-means-based classification is enabled using between one and five clusters, while only for three clusters early, middle, and late labels can be assigned computationally.

Figure 3.

Assessment of phage gene classification methodologies. (A) Metrics of phage gene classification for the atlas’ default settings “Class Max” and “Class Threshold” for 18 datasets. Recall defines the fraction of the ground-truth late genes in the categories “connector,” “lysis,” “head and packaging,” “tail,” and “integration and excision” annotated using Pharokka [20], which were classified as late genes. Precision describes the fraction of classified late genes among the ground-truth late genes. Fraction other late describes the percentage of classified late genes outside the ground-truth late genes. (B) Distribution of early, middle, and late gene fractions after classification using “Class Max” and “Class Threshold” criteria across datasets. The dataset legend serves for panels (A) and (B). (C) Total class count using the “Class Max” criterion across Pharokka-annotated gene categories across the 18 datasets (11 phages) presented in panels (A) and (B). Each gene is represented only once and for genes present in different datasets, average class voting was used (see the “Materials and methods” section).

We next compared the overall distribution of early, middle, and late gene fractions across datasets for both classification schemes. “Class Threshold” displayed a bias toward early classifications, whereas “Class Max” tended to yield a larger fraction of late calls (Fig. 3B and Supplementary Fig. S8B). Importantly, these distributions depend on both the sampled time points and the specific phage–host system under investigation. Notably, the same phage profiled under different growth conditions, during host manipulation (e.g. knockdown), or in distinct host strains can yield different class distributions, highlighting the sensitivity of phage transcriptional programs to environmental and host-dependent factors. We note, however, that our approach does not directly distinguish gene-specific timing shifts from global differences in infection progression between conditions. For example, infection of one host strain may proceed faster than infection of another, causing the same transcriptional program to occur at different absolute sampling times. In future implementations, such effects could be addressed by aligning time courses to lifecycle-associated transcriptional landmarks, for example the onset or peak expression of late-gene modules comprising lysis-associated and structural genes. Such lifecycle-normalized alignment could help separate global acceleration or delay of infection from genes whose expression timing changes specifically between conditions. This concept was not implemented in the present study.

Using “Class Max,” we then assessed how predicted temporal classes distribute across functional gene groups. For the 18 datasets, which comprise a total of 11 different phages including model phages T4, T7, and LUZ19 as well as less well annotated phages phiKZ, vB_PaeM_PA5oct, and phiCbK, we calculated the average class for each gene, when more than one dataset included the respective phage. Despite the use of dedicated phage annotation tools [20], most phage genes remain uncharacterized (between 40% and 80 % per phage) and dominate the early, middle, and late classes (Fig. 3C). Unexpectedly, “Class Max” classified a large fraction of genes of unknown function as late. Since this classification is purely based on the peak of expression over the time course of infection, we also employed k-means clustering, which produced the inverse trend that uncharacterized genes are mostly expressed early during infection (Supplementary Fig. S11B). As expected, genes annotated as head and packaging, tail, and lysis were predominantly classified as late. Interestingly, phage-encoded tRNAs and long non-coding RNAs, which are present only in a subset of phages, also tended to be classified as late, suggesting that such phage-encoded RNA-based regulatory functions may be particularly relevant at later infection stages. This underscores the utility of transcriptomics-derived temporal classes as an additional layer of evidence for functional inference across the infection cycle.

Together, these analyses constitute a first cross-dataset benchmark of expression-based phage gene classification, revealing condition- and host-dependent shifts in class distributions for the same phages, as well as reproducible trends across functional gene groups.

The PhageExpressionAtlas identifies recurring signatures of phage takeover, gene-function timing, and defense/anti-defense expression

Next, we exploited the collection of time-resolved infection transcriptomes to assess recurring trends in phage and host gene expression across datasets. To this end, we selected 18 datasets that comprise at least three post-infection time points and include a transcriptome measurement of the bacterial culture prior to infection. Consistent with prior reports, the relative phage transcript abundance increases over the course of infection (Fig. 4A). Notably, the final phage transcriptome fraction varies widely between phages and datasets, ranging from below 20% to above 80% at the end of the sampled infection period, highlighting the diversity of phage–host interactions captured in the atlas.

Figure 4.

Trends of phage and host gene expression. (A) Host and phage transcriptome abundance over the relative time course of infection. Lines represent means across datasets, the interval is determined by standard deviation from the mean (n = 17 datasets). (B) Boxplot indicating expression of orthologous genes relative to the E. coli house-keeping genes rpoD, hcaT, gyrA, recA, and adk summarized into uninfected, immediate early, early, middle, and late infection phases for n = 17 datasets. Data points represent fractional expression values. Full figure in Supplementary Fig. S12A. (C) Same boxplot as in panel (B) for orthologous groups of five characteristic T4 phage genes. Full figure in Supplementary Fig. S13A. (D) Boxplot indicating relative expression of anti-phage defense systems in diverse bacterial hosts binned into five infection stages as in panels (B) and (C). Full figure including dataset (n = 17) contributions is available in Supplementary Fig. S14A. (E) Profile plot indicating relative expression of the Rst_HelicaseDUF2290 defense system in P. aeruginosa PAO1 infected by LUZ19 (Brandao datasets) and vB_PaeM_PA5oct phage (Lood). (F) Boxplot showing fractional expression of host-encoded anti-defense systems across infection phases for n = 9 datasets. Full figure available in Supplementary Fig. S14A. (G) Boxplot showing fractional expression of phage-encoded anti-defense systems across immediate early, early, middle, and late infection phase for n = 10 datasets. Full plot available in Supplementary Fig. S15B. (H) Profile plots illustrating relative expression of two anti-CBASS genes encoded in T4 phage across n = 3 datasets. All profile plots show relative infection time on the x-axis bringing datasets to a comparable scale. Relative time scales in all plots relate to infection phases as follows: 0–0.1 (immediate early), 0.1–0.25 (early), 0.25–0.5 (middle), 0.5–1.0 (late).

Expression of host and phage orthologous genes

We then used OrthoFinder [23] to infer orthogroups across bacterial hosts and across phages, respectively, to assess the expression dynamics of homologous host and phage functions across diverse infections. First, we examined orthogroups containing E. coli housekeeping genes, including rpoD (encoding the subunit of RNA polymerase), hcaT (a putative 3-phenylpropionate transporter), gyrA (DNA gyrase subunit A, essential for DNA replication), recA (recombinase A involved in DNA repair and the SOS response), and adk (adenylate kinase, central to adenylate metabolism and energy homeostasis). As expected, the relative transcript abundance of the corresponding orthologs decreases upon phage infection across a broad range of phage–host systems (Fig. 4B and Supplementary Fig. S12A). We observed a similar trend for orthogroups encoding ribosomal proteins (Supplementary Fig. S12B). Together, these patterns are consistent with a core feature of many lytic infections: progressive remodeling of transcriptome composition such that host transcripts decline in relative abundance as phage transcription increases, driven by host transcript degradation and/or dominance of the phage transcriptional program [8].

Next, we investigated orthogroups of phage genes with comparatively well-defined functions using T4 as a reference. nrdA encodes a ribonucleotide reductase subunit required for deoxyribonucleotide synthesis to support phage DNA replication, typically during the middle phase of infection [27, 28]. Accordingly, transcription of nrdA orthologs peaked predominantly in the middle phase across diverse phages infecting, among others, P. aeruginosa, S. aureus, and E. coli (Fig. 4C and Supplementary Fig. S13A). Similarly, T4 gene 47, which encodes a subunit of a DNA exonuclease involved in DNA repair, replication, and recombination [29], showed orthologs in multiple phages with expression maxima again concentrated in the middle phase, coinciding with the period, in which phage DNA replication typically occurs.

For canonical early genes, we found that the ADP-ribosyltransferases ModA and ModB are consistently most highly expressed in the early infection phase across datasets (Fig. 4C and Supplementary Fig. S13A). In addition, T4 frd transcription—with an ortholog detected in the P. aeruginosa phage vB_PaeM_PA5oct—peaked early, consistent with its role as a dihydrofolate reductase generating precursors for deoxythymidylate synthesis [30]. Conversely, T4 gene 20, encoding the portal protein that forms the capsid vertex required for docking of the DNA-packaging motor [31], behaved as a strictly late gene in both T4 and vB_PaeM_PA5oct, as expected for a structural assembly component.

Finally, screening the largest orthogroups revealed a DNA polymerase orthogroup present in six phages, with expression extending from early through middle and into early late phases depending on the phage and study (Supplementary Fig. S13B). This variability suggests that phage-encoded DNA polymerases can be deployed with different temporal programs while contributing to the shared goal of progeny production.

Insights into the role of defense and anti-defense systems via dual transcriptomes

Finally, we assessed the transcriptional dynamics of host-encoded defense and phage-encoded anti-defense systems across diverse phage–host interactions. This area is rapidly expanding due to the continued discovery of bacterial defense systems [2] and phage-encoded countermeasures [4]. However, their expression dynamics in time-resolved dual transcriptomes have not yet been systematically compared across “native” phage–host systems.

To this end, we employed DefenseFinder [2] to annotate host- and phage-encoded (anti-)defense systems and extracted their expression profiles from the corresponding datasets. Across multiple studies, we observed an overall decrease in relative transcript abundance for many host defense systems over the infection time course, including CRISPR–Cas, restriction-modification (RM) systems, and toxin–antitoxin (TA) modules (Fig. 4D and Supplementary Fig. S14A and B). Notably, while within-system expression profiles change over time, the rank of these systems within the host transcriptome (z-score) remains comparatively stable in many cases (Supplementary Fig. S14A and B), consistent with broad host transcriptome remodeling during lytic infection.

For a subset of systems, we observed transiently elevated expression immediately after infection. This includes the abortive infection system AbiD in a clinical S. aureus isolate and the Ceres defense system in Mycobacterium smegmatis (Fig. 4D and Supplementary Fig. S14C). These patterns suggest that some defenses may be transcriptionally induced upon infection, although transcriptional induction alone may be insufficient for effective clearance in the assayed systems. Possible explanations include limited induction magnitude, a mismatch between transcriptional and protein-level dynamics, and/or the presence of phage countermeasures.

We also observed system- and phage-dependent heterogeneity. For example, the two-gene defense system Rst_HelicaseDUF2290 showed different temporal dynamics upon infection by distinct phages. During infection of P. aeruginosa with the large phage vB_PaeM_PA5oct, fractional transcript abundance decreased early, whereas during infection with the smaller phage LUZ19 (~50 genes) the decrease was most apparent in the middle phase (Fig. 4E). For both phages, we did not detect annotated anti-defense systems using DefenseFinder. Given the large fraction of uncharacterized genes in vB_PaeM_PA5oct, additional anti-defense functions may remain unannotated, and broad RNA degradation mechanisms could also contribute to the observed dynamics, as described for T4 [32, 33].

Intriguingly, several bacterial hosts (e.g. E. coli and S. aureus) that encode multiple first-line defense systems (including RM and CRISPR–Cas) also encode corresponding anti-defense modules according to DefenseFinder predictions. In the analyzed datasets, these host-encoded anti-defense transcripts generally decreased in relative abundance over infection (Fig. 4F and Supplementary Fig. S15A), while their transcriptome rank (z-score) was often only modestly affected. This pattern suggests that substantial de novo transcription of these modules is not a universal hallmark of infection in the profiled conditions; functional contributions may instead depend on pre-existing protein pools, translation dynamics, or condition-specific regulation.

Finally, among the five phages in which DefenseFinder identified anti-defense systems, we examined their temporal expression programs (Fig. 4G and Supplementary Fig. S15B). Anti-RM and anti-TA modules were expressed across multiple phases, with a tendency toward earlier expression in several datasets, consistent with the expectation that nucleic-acid-targeting defenses can act rapidly after genome injection. In contrast, anti-CBASS mechanisms [34] encoded by bacteriophage T4 (e.g. 57B/acb1 and vs.4/acb2) showed predominantly middle/late expression (Fig. 4G and H), aligning with the timing of CBASS activation [35]. CBASS is triggered by cyclic nucleotides [35], which anti-CBASS proteins can degrade to prevent defense activation [34]. Notably, although CBASS is not encoded in the E. coli strains represented in our dataset collection, anti-CBASS expression in T4 appears as a conserved transcriptional program across conditions, potentially reflecting a broadly protective strategy rather than induction solely by the presence of CBASS.

Discussion and conclusion

Here, we introduce the PhageExpressionAtlas as an interactive visualization resource for exploring phage–host interactions at the transcriptional level. We established a unified processing framework to process and analyze 42 datasets from 23 studies spanning a broad range of phage–host systems. We demonstrate that users can move from whole-transcriptome heatmaps to selected gene profiles, temporal gene classes, and genome-context views without local preprocessing using the PhageExpressionAtlas. Included datasets can be re-analyzed to confirm published findings and to formulate and test new, user-defined hypotheses. To the best of our knowledge, the PhageExpressionAtlas is the first dedicated resource that provides centralized access to uniformly processed time-resolved transcriptomic data of phage infections.

Using the PhageExpressionAtlas database, we benchmarked approaches to expression-based phage gene classification. We note that this evaluation is constrained by the lack of comprehensive ground-truth labels across phages and gene classes, reflecting the large fraction of phage genes with unknown function. Nonetheless, using a proxy late-gene set enriched for structural, assembly, and lysis functions, we found that assigning temporal classes based on maximal expression (“Class Max”) correctly classifies most of these genes as late across diverse phages. Moreover, transcriptional timing was sensitive to experimental context, including growth conditions and host strain, underscoring the importance of standardized processing and cross-study comparison. Consistent with the sparse state of phage annotation, the majority of genes classified as early, middle, or late remained of unknown function, highlighting the need for continued functional characterization across infection stages and functional modules.

Future work could leverage supervised machine learning to predict phage gene classes directly from sequence, rather than from transcriptomic profiles. Such models could assist functional annotation and mechanistic investigation, as the infection phase in which a gene is expressed can be informative about the broader processes it contributes to [8]. The PhageExpressionAtlas provides a foundational training and benchmarking resource for these efforts and may complement state-of-the-art phage annotation and defense/anti-defense discovery pipelines such as Pharokka [20], Phold [36], and DefenseFinder [2].

To enable cross-system comparisons, we assessed expression of orthologous gene groups across hosts and across phages. As expected, we observed the global decline of relative host transcript abundance during lytic infection, consistent with decades of prior work [8–11, 21, 37]. We recorded similar dynamics for many host-encoded defense systems. In some cases—depending on the underlying time-series resolution—defense transcripts showed an early increase immediately after infection followed by a rapid decline. Overall, these patterns suggest that defense-related transcripts are generally not selectively stabilized and may be affected by the same phage-triggered processes that remodel or degrade the host transcriptome. Previous work reported that, despite pronounced host transcriptome degradation during T4 infection, the host proteome remains comparatively stable [10]. Together, these observations are consistent with a model in which baseline (pre-infection) expression levels contribute substantially to the cell’s defense capacity, whereas de novo transcription during infection may play a more limited role.

In this context, a recent transcriptional survey across bacterial species and strains under diverse physiological conditions highlighted condition-dependent priming of the defense arsenal [38]. Accordingly, one plausible explanation is that many atlas datasets—often generated under near-optimal laboratory growth conditions—may capture states in which defense systems are expressed at lower levels than in more variable environmental settings, potentially lowering the barrier for successful infection even in the absence of annotated anti-defense mechanisms. At the same time, given that anti-defense discovery is still emerging, currently unannotated anti-defense modules may be identified in the datasets contained in the atlas. Moreover, a single-microbe RNA-seq-based study recently uncovered the heterogeneous expression levels of defense systems in phage–host interactions [39], which indicate that the bulk average may not accurately reflect the “native” interaction between phage and host immunity on the transcriptional level. Overall, these analyses illustrate how the PhageExpressionAtlas enables systematic, cross-dataset investigation of transcriptional dynamics of bacterial anti-phage immunity and phage counter-defense strategies, providing an entry point for mechanistic follow-up studies of this co-evolutionary arms race.

Looking forward, we will prioritize three extensions of the PhageExpressionAtlas: integration of newly submitted dual RNA-seq datasets, cross-system comparison modules, and additional omics layers such as proteomics and transcriptome architecture. First, we will continuously integrate additional time-resolved dual RNA-seq datasets as they become available, including datasets contributed by authors directly. To encourage data contributions, we implemented a dedicated “Data Submission” page with instructions for researchers to make us aware of public data that is not yet included in the atlas. Second, we will implement additional modules to further enhance interactivity and visualization, including more integrative cross-phage and cross-host analyses to enable systematic comparisons across infection systems. Finally, we aim to incorporate additional omics modalities, including proteomics [10] and transcriptome architecture data generated by differential RNA-seq [40, 41] and ONT-cappable-seq [42]. Together, these extensions would enable interrogation of multiple regulatory layers during phage infection and support a more holistic view of phage–host interactions.

Supplementary data

Supplementary data is available at NAR Genomics & Bioinformatics online.

Conflict of interest

None declared.

Funding

This work was supported by funding from the Wilhelm Schuler Stiftung (University of Tübingen) awarded to M.W.-S. for the development of the PhageExpressionAtlas. Also, this project was funded by the Federal Ministry of Research Technology and Space (BMFTR) and the Baden-Württemberg Ministry of Science as part of the Excellence Strategy of the German Federal and State Governments. We acknowledge support from the Open Access Publication Fund of the University of Tübingen.

Data availability

The PhageExpressionAtlas is available at phageexpressionatlas.cs.uni-tuebingen.de. No novel datasets were generated in this work. All dataset accessions are documented in Supplementary Table S1A. The database is available at https://github.com/Integrative-Transcriptomics/PhageExpressionAtlas/tree/main/PhageExpressionAtlas/instance.

Code availability

The code underlying the PhageExpressionAtlas, the creation of the database, and custom analyses presented in this work are available at https://github.com/Integrative-Transcriptomics/PhageExpressionAtlas and archived in zenodo at https://doi.org/10.5281/zenodo.20797245. The processing pipeline is available at https://github.com/Integrative-Transcriptomics/phage_host_dual_transcriptomics and archived in zenodo at https://doi.org/10.5281/zenodo.19249672.