Rapid and quantitative phage susceptibility test by ramanome Authors: Xiao Han, Xiaofu Wan, Yang Zhou, Xiaoting Fu, Xiaoshan Zheng, Bo Gao, Shi Huang, Anle Ge, Jiadong Huang, Hongzhou Lu, Jian Xu Published: 2026 ABSTRACT AbstractAntimicrobial resistance poses an escalating global threat, renewing interest in bacteriophage therapy as a precision alternative to antibiotics. However, clinical translation remains hindered by the lack of rapid and quantitative phage susceptibility testing (PST) platforms capable of evaluating host range, infection potency, and effective multiplicity of infection (MOI). Here, we present a ramanome‐based phage susceptibility test (RPST), a phenotypic platform that captures infection‐induced remodeling of bacterial macromolecular composition to unify these diagnostic requirements within a single workflow. RPST integrates four Raman biomarkers into a Composite Infection Index (CII), enabling rapid and lysis‐independent discrimination between susceptible and resistant bacterial populations within ~1 h, with 96.0% categorical concordance (24/25) to plaque assays. As a continuous population‐level metric, CII quantifies the proportion of infected cells, allowing quantitative ranking of phage potency against shared hosts. By resolving CII trajectories across the MOI and time, RPST further determines the minimal effective MOI, which is the lowest phage‐to‐bacterium ratio sustaining self‐propagating infection, thereby defining the lower boundary for therapeutic feasibility. Together, these capabilities transform PST from static compatibility assays into a dynamic and quantitative framework that bridges in vitro infectivity assessment and infection dynamics relevant to phage therapy. INTRODUCTION Antimicrobial resistance (AMR) remains a critical global health threat. In 2019, it caused over one million deaths, and the situation is worsening as effective antibiotic options continue to diminish 1 , 2 . The World Health Organization predicts that by 2050, AMR could lead to over 10 million deaths annually, surpassing cancer as the leading cause of mortality 1 . In this context, phage therapy has re‐emerged as a promising countermeasure. Bacteriophages, the natural viruses that target bacteria, have unique therapeutic advantages, including high specificity, self‐amplifying capacity, and minimal disruption to the host microbiome 3 , 4 . They directly eradicate target pathogens through lytic cycles and have shown remarkable efficacy against multidrug‐resistant bacterial strains and biofilm‐associated infections where antibiotics often fail 5 . This has spurred a surge of interest in phage therapy, with over 40 clinical trials currently registered on ClinicalTrials.gov that target a wide range of infections 6 , 7 . Despite its promise, clinical translation of phage therapy faces significant challenges, leading to inconsistent therapeutic outcomes 8 . This variability is largely due to three major obstacles: the intrinsically narrow host range of individual phages 8 , the inevitable emergence of bacterial resistance driven by the co‐evolutionary dynamics between phages and their hosts 3 , and the difficulty of achieving therapeutically relevant phage concentrations at infection sites, where effective clearance requires phage replication to outpace bacterial proliferation 9 . Addressing these barriers critically depends on the development of robust and clinically informative phage susceptibility testing (PST) platforms 10 . An ideal PST system must therefore (i) rapidly and accurately determine whether a phage can infect a given pathogen, thereby resolving the constraint of narrow host range; (ii) discriminate the relative potency of active candidates, enabling the selection of phages most capable of countering resistance; and (iii) evaluate whether infection dynamics, including phage replication and secondary infection, can sustain effective bacterial clearance under clinically relevant multiplicity of infection (MOI) conditions, ensuring therapeutic feasibility 11 . Current PST strategies can be broadly categorized as genotypic or phenotypic methods, yet neither satisfies the three essential diagnostic criteria. Genotypic methods, which quantify phage genome copies, offer high sensitivity and rapid single‐endpoint readouts. Although they can indirectly estimate infection potential, they do not measure actual infectivity, are prone to false positives and negatives, and lack the temporal resolution needed to capture infection dynamics, limiting their clinical relevance 12 . Among phenotypic methods, growth‐based assays such as plaque assays (the current gold standard), disk diffusion, broth microdilution, and optical density (OD) monitoring can reliably distinguish susceptible from resistant hosts, but are time‐consuming, offer only qualitative assessments of infection potency, and fail to resolve dynamic infection processes 8 , 10 , 13 . Metabolism‐based assays, including isothermal microcalorimetry and probe‐coupled flow cytometry, can shorten turnaround time and improve throughput, yet typically yield static, qualitative outputs that fail to simultaneously capture infectivity, relative potency, and phage replication dynamics 8 , 11 . Consequently, the development of a rapid, accurate, quantitative, and dynamic PST system remains a critical unmet need. To contextualize these unmet needs, representative PST strategies and their diagnostic capabilities are compared (Table 1), which suggests that no existing methods simultaneously resolve infectivity, relative lytic potency, and infection sustainability under defined MOI conditions. Table 1Comparison of representative phage susceptibility testing (PST) strategies and their diagnostic capabilities.Method/platformPrincipleKey advantagesMajor limitationsReferencesPlaque assayPlaque formationHigh accuracy; direct evidence of infectivity; providing plaque morphologyTime‐consuming; qualitative or semi‐quantitative; endpoint‐only 10 qPCRPhage genome copy numberRapid; highly sensitive; quantitativeUnable to measure infectivity; unable to distinguish productive from abortive infection; prone to false positives/negatives 12 Optical density (OD) monitoring/killing curvesBulk growth inhibitionDynamic monitoring; relatively rapidUnable to distinguish live/dead cells; confounded by aggregation and growth arrest; indirect and non‐quantitative infectivity 8 , 11 Disk diffusionGrowth inhibition zonesSimple; standardizedTime‐consuming; not quantitative; lacking kinetic information 13 Flow cytometry‐based assaysMembrane integrityRapid; high‐throughput; single‐cell resolutionNon‐specific signals (false positives); complex assay design; unable to distinguish productive from abortive infection 14 Isothermal microcalorimetryHeat flow from metabolismBroadly applicable; dynamicTime‐consuming; indirect metabolic readout via thermal signals 7 RPST (this work)Population‐level biochemical remodelingRapid; quantitative; simultaneously resolving infectivity, lytic potency, and minimal effective MOICurrently population‐averaged; requiring multi‐center validation‐MOI, multiplicity of infection; RPST, ramanome‐based phage susceptibility test.John Wiley & Sons, Ltd. Comparison of representative phage susceptibility testing (PST) strategies and their diagnostic capabilities. MOI, multiplicity of infection; RPST, ramanome‐based phage susceptibility test. Ramanomics has emerged as a new analytical platform in phage research 15 . For example, surface‐enhanced Raman spectroscopy (SERS)‐based probes have enabled the specific identification and discrimination of pathogenic microorganisms 12 , 16 , 17 , 18 , while spectral profiling of infected cells and biofilms has advanced the understanding of phage‐bacterium interactions 19 , 20 , 21 . The integration of tip‐enhanced Raman spectroscopy (TERS) further achieves nanoscale spatial resolution, allowing visualization of viral surface topology and bacterial adhesion proteins 22 . Owing to its rapid, label‐free, and non‐destructive molecular fingerprinting capability 23 , which captures a rich spectrum of cellular metabolic phenomes 24 , 25 , ramanomics holds strong potential for probing infection‐induced biochemical remodeling. Here, a ramanome 26 refers to a structured collection of spontaneous Raman spectra acquired from a certain number of cells within an isogenic population under a defined condition and time window, collectively representing a population‐level metabolic phenome. Each individual Raman spectrum encodes thousands of vibrational features associated with cellular macromolecules, such that the ensemble captures the composite biochemical state of the population rather than single‐molecule signatures. Importantly, intrinsic cell‐to‐cell heterogeneity is preserved within a ramanome and contributes biologically meaningful variation rather than experimental noise. In the context of phage infection, ramanomes primarily monitor host bacterial cells, capturing coordinated metabolic remodeling that emerges during the latent phase prior to overt lysis. Indeed, recent studies have shown that spontaneous Raman spectroscopy can distinguish infected from uninfected bacterial cells by detecting alterations in major macromolecular components within the fingerprint region 27 , 28 . However, these studies have largely focused on isolated phage‐host pairs, yielding case‐specific spectral changes rather than broadly applicable diagnostic principles. This fragmentation reflects the extensive genetic and physiological diversity among phages 29 , which drives heterogeneous host responses and hampers the establishment of a unified Raman‐based susceptibility testing framework. To overcome these limitations, a systematic and expandable ramanome‐based framework is needed to enable rapid, quantitative, and dynamic assessment of phage susceptibility across diverse phage‐host systems. In this study, we developed a ramanome‐based phage susceptibility test (RPST) that harnesses infection‐induced remodeling of bacterial macromolecular composition as a rapid phenotypic readout of phage‐host interactions. By capturing coordinated changes in nucleic acids, proteins, and lipids at the population level, RPST enables rapid and lysis‐independent discrimination between phage‐susceptible and phage‐resistant bacterial populations. Beyond binary discrimination, RPST provides a quantitative framework to evaluate infection efficacy across different MOI and to resolve differences in lytic potency among phages against the same host. In addition, by integrating single‐endpoint analysis with time‐resolved ramanome trajectories, RPST captures infection progression and the capacity of phage populations to sustain propagation over time. Together, these features position RPST as a rapid, quantitative, and dynamic phenotypic platform that addresses the core diagnostic requirements of PST, bridging static compatibility assessment and infection dynamics relevant to therapeutic feasibility. RESULTS Overview of the RPST workflow Both conventional plaque assays and RPST require prior preparation of host bacterial isolates from environmental or clinical samples, including strain purification and identification (Figure 1A). The conventional plaque assay, the clinical gold standard for phage susceptibility testing 8 , requires multiple labor‐intensive steps, including isolate expansion, double‐layer plating, incubation, and plaque enumeration, typically taking 11–21 h to determine host range (Figure 1B). In contrast, RPST leverages phage‐induced remodeling of bacterial macromolecular composition, captured by Raman spectroscopy, to classify infection outcomes within approximately 1 h (Figure 1C). Overview of ramanome‐based phage susceptibility test (RPST). (A) Sample preparation. Bacterial strains isolated from environmental or clinical samples were purified by streak plating and identified prior to downstream phage susceptibility testing (PST). (B) Conventional PST workflow. Isolated bacterial colonies were cultured in liquid medium to the logarithmic growth phase (4–8 h), followed by co‐incubation with phages and plaque assays using the double‐layer agar method. Plaques were visually inspected after an additional 6–12 h incubation, resulting in a total turnaround time of 11–21 h. TAT, turn around time. (C) RPST workflow. Bacterial colonies were suspended and co‐incubated with phages for 40 min, followed by centrifugal washing and transfer onto Raman‐compatible CaF₂ substrates. Raman spectra (≥60 spectra per sample) were subsequently acquired and analyzed using a standardized computational pipeline to generate the Composite Infection Index (CII). Total detection time was less than 1 h. The Raman spectral illustration and the CII discrimination schematic are derived from experimental data generated in this study. The image of the RAMS (Qingdao Single‐Cell Biotechnology Co. Ltd., China) was reproduced with permission from the manufacturer. Created in BioRender. Han X. (2026) https://BioRender.com/54mb1tf. The RPST workflow consists of three sequential steps: (i) phage‐bacteria incubation, involving co‐incubation of bacterial suspensions with phages for 40 min; (ii) ramanome acquisition, including washing, deposition onto Raman‐compatible substrates, and spectral collection; and (iii) data analysis, in which a standardized processing pipeline computes a Composite Infection Index (CII) for susceptibility determination (Figure 1C). Identification of phage‐induced biomarkers by Raman fingerprints to distinguish infected from uninfected bacterial cells Previous studies have demonstrated that Raman spectroscopy can capture infection‐induced alterations in cellular biomolecules 27 , 28 ; however, the biochemical diversity of phage‐host interactions complicates the identification of conserved diagnostic features. To establish a robust ramanome‐based framework for PST, we sought to identify spectral biomarkers that consistently distinguish infected from uninfected bacterial populations across representative systems. To this end, we analyzed three representative phage‐host systems (T1‐Escherichia coli ATCC11303, T4‐E. coli ATCC11303, and T4‐E. coli ATCC25922) chosen to sample both phage‐ and host‐dependent variation (T1 is FhuA‐dependent and relatively host‐restricted, whereas T4 is a broadly lytic Myoviridae phage 30 ). Ramanomes were profiled at 20–120 min (20‐min intervals; Figure S1A–C) to capture the trajectory from early adsorption and macromolecular remodeling to population‐level lysis, ensuring that candidate markers reflect sustained infection dynamics rather than transient noise. For each fingerprint peak, the area under the receiver operating characteristic curve (AUC) was computed between untreated and infected groups (MOI = 10). Peaks with mean AUC values > 0.80 at four or more time points across all three susceptible systems were retained, yielding 15 candidate spectral regions comprising a total of 86 individual Raman shifts (Figure 2A, top). Differential spectral analysis further refined these candidates to four conserved regions exhibiting consistent decreased or increased changes: 658–663 cm⁻¹ (C–S stretching in cystine, decreased), 713–719 cm⁻¹ (pyrimidine ring breathing in nucleic acids, decreased), 1430–1450 cm⁻¹ (CH₂ bending in lipids and protein, increased), and 1589–1595 cm⁻¹ (C═C skeletal vibration in aromatic amino acids, increased; Figure 2A, bottom; Table S1). These conserved signatures form the basis of subsequent RPST metric construction. Identification of infection‐associated Raman biomarkers. (A) Identification of biomarkers distinguishing infected from uninfected bacteria via applying receiver operating characteristic (ROC) analysis to Raman spectral fingerprint regions. Raman spectra were collected from three susceptible systems (T1‐Escherichia coli ATCC11303, T4‐E. coli ATCC11303, and T4‐E. coli ATCC25922) at six time points (20–120 min) at 20‐min intervals under MOI = 10. Peaks showing an average area under the ROC curve (AUC) > 0.80 in at least four of six time points across all three systems are considered candidate biomarkers (n = 15). Combined with difference spectrum analysis, only those showing consistent decreased or increased changes (either increased or decreased upon infection) across systems are retained, resulting in four universal candidate regions: 658–663 cm⁻¹ (decreased), 713–719 cm⁻¹ (decreased), 1430–1450 cm⁻¹ (increased), and 1589–1595 cm⁻¹ (increased). (B) Temporal dynamics of the four biomarkers compared between susceptible and resistant systems. In the susceptible T1‐ATCC11303 system, 658–663 cm⁻¹ and 713–719 cm⁻¹ gradually decreased, whereas 1430–1450 cm⁻¹ and 1589–1595 cm⁻¹ continuously increased over time. In contrast, in the resistant T1‐DH5α system, the initial decrease of 658–663 cm⁻¹ and 713–719 cm⁻¹ (≤40 min) was followed by recovery or stabilization, while 1430–1450 cm⁻¹ and 1589–1595 cm⁻¹ transiently decreased at early time points (≤60 min) and then remained unchanged or slightly declined. (C) Temporal dynamics of the four biomarkers in the weakly susceptible system under different MOI conditions. Under MOI = 10, 658–663 cm⁻¹ and 713–719 cm⁻¹ consistently decreased, while 1430–1450 cm⁻¹ and 1589–1595 cm⁻¹ increased. Under MOI = 1, responses fluctuate: the first two regions decreased at 40 min, increased at 80 min, and sharply dropped at 100 min; the latter two increased at 40–80 min but also declined at 100 min. Statistical analysis was performed using two‐sided Student's t‐test. ns, not significant; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Red asterisks indicate that the statistical value is higher in the experimental group than in the control group, while blue asterisks indicate lower. To evaluate whether these spectral regions can discriminate phage‐susceptible from resistant strains, we collected ramanomes from E. coli ATCC11303 (T1‐susceptible) and DH5α (T1‐resistant) after infection with phage T1 (MOI = 1) across 20–100 min, with uninfected cells (MOI = 0) as controls (Figures S1D,F). In E. coli ATCC11303, global spectral divergence from controls became evident at 40 min and intensified over time, whereas DH5α displayed only transient deviations (<60 min) that largely subsided by 80 min. Analysis of the four biomarker regions reinforced these trends (Figure 2B): in the susceptible strain, the 658–663 cm⁻¹ and 713–719 cm⁻¹ bands decreased progressively, whereas the 1430–1450 cm⁻¹ and 1589–1595 cm⁻¹ bands increased steadily (p < 0.0001). In resistant DH5α, early declines in the first two regions recovered or stabilized, while the latter two exhibited brief reductions that subsequently plateaued. Thus, both global spectra and biomarker dynamics enable robust discrimination between phage‐susceptible and resistant phenotypes. To assess detection sensitivity, we analyzed a weakly susceptible system: phage T4 infecting E. coli ATCC25922. Ramanomes were collected over 0–100 min under MOI = 0, 1, and 10, with MOI at 0 as the control (Figure S1C,E). In contrast, a strongly susceptible strain exhibited stable infection signatures even at MOI = 1 (Figure S2), precluding limit‐of‐detection analysis; subsequent tests therefore focused on ATCC25922. At the global spectral level, infection‐induced changes appeared by 20 min at MOI = 10, but only after 80 min at MOI = 1. Analysis of the four biomarker regions (0, 40, 80, and 100 min) confirmed this dependence (Figure 2C). At MOI = 10, the 658–663 and 713–719 cm⁻¹ bands decreased, whereas the 1430–1450 and 1589–1595 cm⁻¹ bands increased consistently (p < 0.0001). At MOI = 1, these responses were delayed and temporally variable: the first two bands declined at 40 min (p < 0.05), rose at 80 min (p < 0.05), and sharply decreased again by 100 min (p < 0.0001); the latter two increased at 40–80 min (p < 0.0001) but fell at 100 min (p < 0.0001). Thus, the diagnostic performance of individual Raman biomarkers depends on infection synchrony and population composition. Constructing a CII via multi‐feature integration for rapid and robust infection discrimination To achieve a robust RPST, the four Raman biomarkers were integrated into a unified diagnostic metric, called CII, and tested using a weakly susceptible system, the T4‐E. coli ATCC25922. Spectra from uninfected and infected populations were labeled accordingly, and standardized intensities from the four biomarker regions (658–663, 713–719, 1430–1450, and 1589–1595 cm⁻¹) were used as input features. To identify the optimal integration strategy, we compared six approaches: direct summation of normalized biomarker intensities, logistic regression, linear discriminant analysis (LDA), naïve Bayes, support vector machine (SVM), and random forest (RF). Receiver operating characteristic (ROC) and precision‐recall (PR) analyses showed that all machine‐learning approaches achieved high performance (AUC > 0.99, PR‑AUC > 0.99, accuracy > 0.94), whereas direct summation performed poorly (AUC = 0.439 ± 0.040, accuracy = 0.454 ± 0.026; Figure 3A,B). Among the classifiers, RF and LDA performed equally well in terms of ROC‐AUC and PR‐AUC; RF was chosen for subsequent analyses because of its better interpretability. Construction of a Composite Infection Index (CII) via multi‐feature integration to achieve rapid and robust infection discrimination. (A, B) Performance comparison of six machine learning models integrating four RPST‐derived biomarkers, evaluated by 5 × 5 cross‐validation: direct summation, logistic regression, linear discriminant analysis (LDA), naïve Bayes, support vector machine (SVM), and random forest (RF). Based on ROC (A) and precision–recall curves (B), both AUC and PR (precision‑recall)‐AUC follow RF = LDA > naïve Bayes ≥ SVM > logistic regression > direct summation, with RF achieving the best performance (AUC: 0.995 ± 0.003; PR‐AUC: 0.993 ± 0.003). (C, D) Comparison between single biomarkers and the RF‐based CII. Across repeated cross‐validation, CII consistently outperforms all individual features in both ROC (C) and precision–recall (D). (E) RF‐derived feature importance averaged across folds: 1430–1450 cm⁻¹ (0.516 ± 0.010), 1589–1595 cm⁻¹ (0.244 ± 0.012), 713–719 cm⁻¹ (0.089 ± 0.014), and 658–663 cm⁻¹ (0.151 ± 0.018). (F) Dual‐cutoff framework for CII‐based classification. The optimal cutoff (0.600) maximizing G‐mean defines infection status with 99% confidence bounds (low cutoff: 0.597; high cutoff: 0.603), categorizing populations as uninfected (CII < low cutoff), infected (CII > high cutoff), or uncertain (between cutoffs). (G, H) Temporal dynamics of CII. In T1‐E. coli ATCC11303 (susceptible), CII exceeded the cutoff by 20 min (0.730 ± 0.210), stabilizing by 40 min (0.991 ± 0.015). In T1‐E. coli DH5α (resistant), both the infected and control groups remained below the cutoff. (I) Validation across 25 phage‐host systems (E. coli, Pseudomonas aeruginosa, Salmonella enterica, and Klebsiella pneumoniae). CII‐based classification agreed with plaque assays in 24/25 cases (96.0% accuracy). To rigorously evaluate RF performance, we implemented a five‐time repeated five‐fold cross‐validation scheme (25 total folds), yielding a mean AUC of 0.995 ± 0.003 and a mean accuracy of 0.965 ± 0.014. However, when trained on individual biomarkers, RF models showed substantially lower performance (AUC = 0.80–0.96; PR‐AUC = 0.83–0.96; accuracy = 0.72–0.89) compared with the integrated model (Figure 3C,D), indicating that integration of multiple features improves classification robustness. To clarify the basis of the performance gain, we examined the weight of RF features. All four biomarkers contributed to classification with distinct weights: 0.516 ± 0.010 (1430–1450 cm⁻¹), 0.244 ± 0.012 (1589–1595 cm⁻¹), 0.151 ± 0.018 (658–663 cm−1), and 0.089 ± 0.014 (713–719 cm−1) (Figure 3E). The rankings remain consistent under various cross‐validation folds (standard deviation of ranks <0.2), confirming robustness of the selected biomarkers. To implement CII as a diagnostic readout, ROC analysis was used to define a dual‐cutoff framework for the CII. An optimal threshold (0.600) was identified by G‐mean maximization and refined into low (0.597) and high (0.603) cutoffs at the 99% confidence level, yielding three outcome categories: uninfected (< low cutoff), infected (> high cutoff), and uncertain (between cutoffs) (Figure 3F). To validate diagnostic performance, we applied the CII‐based PST to phage T1 infection of E. coli ATCC11303 (susceptible) and DH5α (resistant) at MOI = 1. In ATCC11303, the CII rapidly increased, exceeding the infected cutoff by 20 min and stabilizing by 40 min (20 min: 0.730 ± 0.210; 40 min: 0.991 ± 0.015; 60–100 min: ≥ 0.997 ± 0.008) (Figure 3G). In contrast, DH5α consistently remained within the uninfected zone (Figure 3H). To evaluate generalizability, the CII was validated across 25 phage‐host pairs encompassing E. coli, Pseudomonas aeruginosa, Salmonella enterica, and Klebsiella pneumoniae. After 40 min of co‐incubation at MOI = 10, CII‐based classification was compared with plaque assay results. The CII achieved concordant susceptibility calls in 24 of 25 systems (96.0% accuracy) (Figure 3I, Table S2). Quantitation of infection efficacy for rational selection of phage for therapy Beyond binary classification of PST, the continuous nature of the CII enables quantitative assessment of infection responses. Unlike plaque‐based efficiency of plating (EOP), which yields discrete or qualitative outcomes, the CII provides a population‐level metric that varies continuously with infection conditions. To examine the dependence of CII on phage dosage, we analyzed the Ecp54‐E. coli DH5α pair, across five MOIs (0, 0.01, 0.1, 1, and 10) after 40 min of co‐incubation. CII values increase with MOI and exceed the infection cutoff (CII > 0.603) at MOI ≥ 0.1 (Figure 4A). Distribution analysis revealed that the fraction of infected cells rose progressively with MOI, exceeding uninfected cells at MOI = 0.1 and approaching complete infection at MOI = 10 (Figure 4B). Thus, CII‐based RPST can resolve continuous, MOI‐dependent infection responses. Quantitative assessment of relative phage potency by RPST. (A, B) Effect of MOI on the CII in the Ecp54‐E. coli DH5α system. Boxplots (A) show progressive CII elevation with increasing MOI after 40 min co‐incubation. Density plots (B) at 40 min reveal a gradual rise in the proportion of infected cells, exceeding uninfected populations at MOI = 0.1 and approaching complete infection at MOI = 10. (C, D) Evaluation of phage Pap12 against three P. aeruginosa strains. RPST‐based CII distributions (C) show MOI‐dependent increases for susceptible strains Pae1 and Pae307, surpassing the infection cutoff at MOI = 0.01 and 1, respectively, while the resistant strain Pae5 remained below the cutoff even at MOI = 10. Plaque‐based relative infection efficiency (RIE) assays (D) yielded RIE = 1 (Pae1), 0.01 (Pae307), and 0 (Pae5). (E, F) Evaluation of phage Kpap4 against three K. pneumoniae strains. RIE assays (E) show values of 1, 0.1, and 0 for Kpn25, Kpn30, and Kpn6, respectively. Corresponding RPST‐based CII distributions (F) demonstrate cutoff crossing at MOI = 0.1 for Kpn25, persistent sub‐cutoff levels for Kpn6 even at MOI = 10, and gradual but incomplete increases for Kpn30 (RIE > 0). (G) Comparative analysis of 12 E. coli phages (MOI = 10, 40 min). Nine systems show population CII above or close to the cutoff, with Ecp9 showing the highest CII. Density plots exhibit unimodal distributions for infective phages, while three noninfective ones display bimodal patterns with population CII remaining below the cutoff. NC, host bacteria not co‐incubated with phages. To determine whether RPST can resolve phage relative infectivity across multiple hosts, we tested phage Pap12 with P. aeruginosa isolates and phage Kpnp4 with K. pneumoniae isolates, each system including two plaque‐positive and one plaque‐negative strain. For P. aeruginosa, Pae1 entered the infectious state at MOI = 0.01, Pae307 at MOI = 1, while Pae5 remained below the threshold even at MOI = 10 (Figure 4C), consistent with relative infection efficiency (RIE) values of 1, 0.01, and 0, respectively (Figure 4D). For K. pneumoniae, Kpn25 crossed the infection threshold at MOI = 0.1, Kpn6 remained uninfected at MOI = 10, and Kpn30 showed gradual CII increases without crossing the infection threshold, despite being plaque‐positive (RIE = 0.1) (Figure 4E,F). To evaluate whether RPST can rank the relative efficacy of multiple phages against the same host, 12 E. coli lytic phages were tested against E. coli DH5α at MOI = 10 for 40 min. Nine systems exceeded or approached the infection threshold (Figure 4G), with Ecp9 showing the highest CII (p < 0.0001). Notably, systems below the threshold exhibited bimodal CII distributions, reflecting mixed infected and uninfected subpopulations (Figure 4F). Consistently, phages with CII above the threshold produced transparent plaques, whereas those below the threshold yielded turbid or no plaques (Figure S3). Determining the minimal effective MOI via dynamics of ramanomes during infection Beyond endpoint susceptibility classification, time‐resolved ramanome profiling enables assessment of whether infection can be maintained over time under different phage‐to‐bacterium ratios. Here, we define the minimal effective MOI operationally as the lowest initial MOI at which the CII rises above the infection threshold and remains stably elevated over the observation period. To assess whether RPST can resolve infection dynamics and determine the minimal effective MOI, we established a multidimensional experimental matrix incorporating five MOIs (0, 0.01, 0.1, 1, and 10) and seven time points (0–120 min at 20‐min intervals) across three phage‐bacterium systems: T1‐E. coli ATCC11303, T4‐E. coli ATCC11303, and T4‐E. coli ATCC25922. In the T1‐ATCC11303 system, CII increased rapidly with MOI (Figure 5A). At MOI = 10, the population surpassed the infection threshold within 20 min (CII = 0.968 ± 0.073 > cutoff). At lower MOI, CII rose more slowly but reached comparable infected levels within 60 min (CII = 0.983 ± 0.029–0.996 ± 0.007), indicating that infection can be sustained even under sparse initial phage input. Based on these dynamics, the minimal effective MOI for this system lies below 0.01. RPST enables quantitative assessment of effective MOI. (A–C) MOI‐ and time‐resolved trajectories of CII in three phage‐host systems. Populations were monitored under five MOIs (0, 0.01, 0.1, 1, and 10) across seven time points (0–120 min, 20‐min intervals). In the T1‐E. coli ATCC11303 system (A), populations at MOI = 0 remained below the infection cutoff throughout the experiment, whereas those at MOI = 10 cross the threshold within 20 min. Intermediate MOIs (0.01, 0.1, 1) exhibited delayed but complete transitions into the infected state within 60 min. In the T4‐E. coli ATCC11303 system (B), populations crossed the infection cutoff within ≤ 20 min at MOI = 10, 60 min at MOI = 1, 80 min at MOI = 0.1, and 120 min at MOI = 0.01, while remaining uninfected at MOI = 0. In the T4‐E. coli ATCC25922 system (C), only MOI = 10 led to a transition to the infected state. At lower MOIs, CII values rose transiently but subsequently declined, and population‐level CII never exceeded the infection cutoff. The T4‐ATCC11303 system exhibited a similar qualitative pattern, with all MOI > 0 conditions eventually reaching stable infection (Figure 5B). However, the transition kinetics differed markedly: higher MOI accelerated the onset of infection (≤20 min at MOI = 10), whereas lower MOI (0.01–1) required extended incubation (60–120 min) before achieving comparable infection levels. Therefore, although both systems support infection propagation at low MOI, the temporal dynamics of establishment vary between phage‐host pairs. In contrast, T4‐ATCC25922 system displayed a distinct dynamic profile. Only MOI = 10 led to sustained infection (CII = 1 from 20 to 120 min; Figure 5C). At lower MOI (0.01–1), CII exhibited transient increases followed by declines to baseline. For example, at MOI = 1, CII rose to 0.463 ± 0.410 by 40 min, fell to 0.041 ± 0.124 at 60 min, rose again to 0.319 ± 0.342 at 80 min, and returned to baseline thereafter, indicating that initial infection events have failed to propagate effectively within the population. Thus, initial infection events did not lead to stable population‐level infection, placing the minimal effective MOI for this system above 1. Together, time‐resolved CII trajectories delineate system‐specific thresholds for sustained infection and reveal pronounced differences in infection establishment kinetics across phage‐host combinations. DISCUSSION Despite considerable progress in PST, current strategies remain unable to simultaneously meet the key diagnostic requirements of accurately determining host susceptibility, ranking lytic potency, and evaluating the effective MOI that can sustain infection. RPST addresses this gap by leveraging population‐level Raman phenotyping as a direct, proliferation‐independent readout of infection‐induced biochemical remodeling. At the population level, productive lytic phage infection entails a profound reallocation of host cellular resources, redirecting biosynthetic capacity from host maintenance toward viral genome replication and virion assembly 31 , 32 , 33 . Consistent with the diagnostic objective of RPST, ramanome‐based readouts capture relative, population‐averaged biochemical remodeling rather than absolute molecular abundances. In fact, as Raman spectra are normalized by the total peak area across the fingerprint region, changes in individual bands reflect shifts in the relative composition of cellular constituents within infected populations. The four conserved Raman biomarkers map onto core biochemical remodeling processes, nucleic acid metabolism, protein biosynthesis, and CH₂‐rich macromolecular accumulation, which have been repeatedly identified as dominant host responses to lytic phage infection by independent omics studies across diverse systems 31 , 32 , 33 . Specifically, the infection‑induced increase in the 1430–1450 cm⁻¹ band, often broadly assigned to lipids 21 , more plausibly reflects enhanced protein‐associated CH₂ vibrations arising from viral structural and replication‐associated proteins rather than de novo lipid synthesis 34 . Conversely, the infection‑associated decrease in the nucleic acid‐associated 713–719 cm⁻¹ band does not indicate reduced nucleic acid synthesis, but rather that protein accumulation outpaces nucleic acid expansion during productive infection. Given the distinct Raman scattering cross‐sections of different functional groups 35 , individual spectral features are context‐susceptible, underscoring the necessity of integrative, multi‐biomarker strategies for robust and generalizable PST assessment. However, infection‐induced biochemical remodeling does not unfold synchronously across bacterial populations, particularly under low MOI or during early infection stages. Under these conditions, infected and uninfected cells coexist, and the fraction of productively infected cells evolves dynamically over time. As a result, even biologically conserved Raman features are vulnerable to dilution by uninfected or transiently perturbed subpopulations, leading to unstable or context‐dependent diagnostic performance when considered in isolation. Integrating multiple complementary Raman biomarkers into a CII therefore provides a more stable and generalizable diagnostic signal. By aggregating biochemical dimensions associated with nucleic acids, proteins, and CH₂‐containing macromolecules, CII amplifies coherent infection‐associated remodeling while suppressing noise arising from population heterogeneity. This integrative strategy enables robust discrimination across diverse phage‐host systems, even under challenging conditions such as low MOI or asynchronous infection. Crucially, CII‐based RPST achieves high concordance with plaque‐assay‐based susceptibility calls across diverse phage‐host systems and greatly reduces turnaround time. This establishes RPST as a reliable phenotypic PST readout capable of rapidly and accurately distinguishing susceptible from resistant bacterial strains at the population level. Beyond binary susceptibility classification, the continuous nature of CII enables quantitative assessment of infection efficacy. Because CII reflects the fraction of cells undergoing sustained biochemical remodeling, its magnitude provides a population‐level proxy for phage lytic potency under defined conditions. This property allows RPST to rank infective phages targeting the same host and to resolve differences that remain indistinguishable by plaque morphology alone. In contrast to plaque assays, which provide only qualitative or categorical readouts, RPST reframes phage susceptibility testing from a static compatibility assessment into a quantitative evaluation of infection strength, thereby supporting the rational selection of the most effective phage candidates. Importantly, infection efficacy alone does not guarantee persistence. Effective phage therapy requires that infection dynamics be sustained under clinically realistic phage‐to‐bacterium ratios, such that phage amplification and secondary infection can outpace bacterial proliferation. By resolving CII trajectories across MOI and time, RPST quantifies the lowest phage‐to‐bacterium ratio that supports self‐propagating infection at the population level, thereby defining the minimal effective MOI. Unlike plaque assays, which provide only endpoint compatibility, RPST captures both the initiation and persistence of infection, linking early phenotypic susceptibility with dynamic infection outcomes relevant to in vivo efficacy. Moreover, time‐resolved RPST enables longitudinal monitoring of infection trajectories, revealing whether bacterial populations have a trend toward infection dominance or recovery. Although further validation is required, the minimal effective MOI derived from RPST provides a rational, quantitative metric to guide phage formulation and dosing under realistic infection conditions. Nevertheless, there are several limitations. RPST prioritizes speed and throughput by profiling population‐level Raman signals, which inevitably masks single‐cell heterogeneity. Under low MOI or early infection stages, infected and uninfected cells may coexist, and infection‐induced biochemical remodeling may fail to propagate across the population. In such cases, transient or intermediate CII values reflect non‐sustained, population‐level infection outcomes rather than definitive cellular fates. RPST therefore does not distinguish specific intracellular failure modes, such as abortive or pseudolysogenic infections 36 , but instead reports whether infection‐associated remodeling is self‐sustaining at the population level. Future integration with single‐cell Raman acquisition strategies, such as Raman flow cytometry 37 , may enable more refined classification of infection states by resolving cell‐to‐cell heterogeneity that is inherently masked in population‐level ramanome measurements. In addition, the generalizability of ramanome‐based susceptibility metrics across instruments and laboratories remains an important consideration for clinical translation. RPST mitigates instrument‐specific variability by relying on multiple conserved spectral bands spanning relatively broad wavenumber ranges, rather than single‐wavenumber features, and by integrating these biochemical dimensions into the CII. Standard ramanome calibration procedures, including silicon‐based wavenumber referencing and multi‐peak molecular standards, can further support inter‐platform reproducibility 38 . While large‐scale multi‐center validation remains a future goal, these design choices provide a practical foundation for extending RPST across different Raman platforms. Moreover, although our validation has focused primarily on Gram‐negative bacteria, the RPST biomarkers capture relative remodeling of core macromolecular categories rather than Gram‐negative‐specific structures. Consequently, the underlying diagnostic principle is, in principle, also applicable to Gram‐positive organisms. On the other hand, differences in cell envelope architecture and baseline biochemical composition between Gram‐negative and Gram‐positive bacteria are expected to influence spectral baselines and infection dynamics. Accordingly, extension of RPST to Gram‐positive species will require additional calibration and validation, rather than direct transfer of thresholds established here. Extending the spectral framework to Gram‐positive organisms, as well as to polymicrobial samples and biofilm‐associated infections, will therefore be an important step towards broader clinical translation. Besides, complex clinical matrices may introduce spectral interference and compositional variability, underscoring the need for systematic validation under increasingly realistic infection contexts. With further optimization, RPST can substantially accelerate phage therapy development and clinical decision‐making. Its rapid, label‐free, and quantitative nature makes it suitable for not just preclinical phage screening, susceptibility testing, and potency ranking, but real‐time monitoring of treatment response in personalized therapy and the rapid detection of emerging resistance. Looking forward, the integration of RPST with flow‐mode 37 , 39 , 40 or multi‐modal single‐cell Raman‐activated sorting 41 , 42 and mass‐paralleled, artificial intelligence (AI)‐directed single‐cell culturomics 43 , 44 can bridge the gap between phenotypic screening and the isolation of genetically unique phages or resistant bacterial clones, fueling discovery and mechanistic studies. More broadly, the RPST framework provides a generalizable paradigm for probing virus–microbe interactions through quantitative phenotypic fingerprints. By building expansive ramanomic databases, we can envision an AI‐driven future where infection outcomes and optimal phage matches can be predicted directly from a bacterial isolate's ramanome. Collectively, this work demonstrates that population‐level Raman phenotyping can unify infectivity detection, potency quantification, and infection sustainability assessment within a single framework, paving the way to a more predictive and personalized approach to antimicrobial therapy. MATERIALS AND METHODS Bacterial strains, media, and culture conditions Reference strains E. coli ATCC11303 and ATCC25922 were obtained from the American Type Culture Collection (ATCC, USA). The laboratory‐stored E. coli DH5α and clinical isolates (S. enterica Sal2, Sal8, Sal28; P. aeruginosa Pae1, Pae5, Pae307; and K. pneumoniae Kpn6, Kpn25, Kpn30) were provided by Shenzhen Third People's Hospital (Shenzhen, China). All strains were cryopreserved at – 80°C until use. E. coli strains were cultured on Luria‐Bertani (LB) agar and in LB broth (Hope Bio). S. enterica and P. aeruginosa strains were cultured on tryptic soy agar (TSA, Aobox) and in tryptic soy broth (TSB, Solarbio). K. pneumoniae strains were cultured on blood agar plates (Hope Bio) and in Mueller‐Hinton broth (MHB, Hope Bio). Before experiments, frozen strains were thawed, streaked onto respective agar plates, and incubated overnight at 37°C. Single colonies were inoculated into liquid media for subsequent phage‐bacteria interaction experiments. Phage types and propagation Lytic phages targeting E. coli (T1, T4, Ecp2, Ecp5, Ecp9, Ecp10, Ecp11, Ecp12, Ecp19, Ecp32, Ecp44, Ecp54, Ecp91, and Ecp101), S. enterica (Salp6), P. aeruginosa (Pap12), and K. pneumoniae (Kpnp4) were obtained from Shenzhen Third People's Hospital and stored at 4°C in SM buffer (50 mM Tris‐HCl, 100 mM NaCl, 8 mM MgSO₄, pH 7.5). Phages were propagated on their respective laboratory hosts in liquid culture. Briefly, 100 µl of phage suspension (109–1010 plaque‐forming units per milliliter [PFU/ml]) was added to 500 ml logarithmically growing bacteria (OD600 = 0.5) at a MOI of 0.01–0.1. After incubation at 37°C with shaking for 5 h, lysates were centrifuged (10,000 g for 30 min) and filtered (0.22 µm). Phage titers were determined by the double‐layer agar method and expressed as PFU/ml. Crude lysates were stored at 4°C for short‐term use or at – 80°C with 10% glycerol for long‐term storage. Phage sensitivity assay (double‐layer agar assay) A 200 μl aliquot of mid‐log‐phase bacterial suspension (0.5 McFarland units) was mixed with 2 μl of phage suspension (108 PFU/ml) in 5 ml of 0.7% semi‐solid agar and overlaid onto respective agar plates. Plates were incubated at 37°C for 8 h. All assays were performed in triplicate. RIE assay Phage suspensions were serially diluted 10‐fold in SM buffer, and 2 µl of each dilution was spotted on the bacterial lawn and cultured at 37°C overnight. The lowest phage titer producing visible plaques was recorded for each phage‐host pair. For each host, the phage requiring the lowest titer was defined as the reference (RIE = 1). The RIE of other phages was calculated as the ratio of the reference titer to the minimal plaque‐forming titer of each test phage, reflecting their relative ability to initiate infection under identical conditions. Sample preparation for ramanome detection Bacterial colonies were suspended in respective liquid media and adjusted to 0.5 McFarland units (5 × 107–108 CFU/ml) using a densitometer (530 nm). Phage suspensions were diluted to target MOIs and mixed with bacterial suspensions in equal volumes (n = 3 per group). Mixtures were incubated in 48‐well plates at 37°C with shaking (200 rpm). At designated time points (0–120 min, 20‐min intervals), 1 ml aliquots were centrifuged (10,000 g, 2 min), washed thrice with ultrapure water, and resuspended in 10 μl sterile water. Approximately 3 μl of the suspension was placed onto quartz substrates for ramanome analysis. Raman spectrometry Spectra were collected using a confocal Raman microscope (RAMS, Qingdao Single‐Cell Biotechnology Co. Ltd.) equipped with a 532 nm laser, 1200 lines/mm grating, and a 100× objective (NA = 1.25; Olympus). Prior to spectral acquisition, wavenumber calibration was performed using a silicon standard with a laser power of 20 mW, a pinhole size of 25 μm, and an integration time of 0.5 s. Acetaminophen was additionally measured as a molecular reference under a laser power of 20 mW, a pinhole size of 125 μm, and an integration time of 1 s. Raman spectra of samples were then acquired with a laser power of 30 mW and an integration time of 1 s. Approximately 60 spectra were acquired per sample over the 400–3500 cm⁻¹ region, and the sampling depth is sufficient to accurately quantify metabolic features of the sample 28 , 29 . Ramanomic data preprocessing and analysis Raw spectra were precessed using the RamEx package (version 1.0.0) 45 , which integrates multiple processing modules, including cosmic‐ray removal, spectral smoothing (Savitzky‐Golay), baseline correction (polynomial fitting), spectral‐range selection (fingerprint region), and peak area normalization. The find_markers_roc function (RamEx package: https://github.com/qibebt‐bioinfo/RamEx) identified infection‐specific biomarkers based on ROC curves (AUC > 0.80). Statistical significance was determined using Student's t‐test, with p < 0.05 considered significant. Random forest model training and CII calculation An RF classifier was implemented in Python (RRID:SCR_008394) 3.09 using the scikit‐learn library (v 1.3.2). The model was trained on four Raman biomarkers from labeled infected and uninfected bacterial populations, with the predicted infection probability defined as the CII for each sample. The model utilized the following fixed hyperparameters to ensure reproducibility: n_estimators = 100, max_depth = none, min_samples_split = 2, min_samples_leaf = 1, criterion = ‘gini’, bootstrap = true, and random_state = 42. Model performance and generalizability were evaluated using a five‐time repeated five‐fold cross‐validation scheme (25 total folds). In each iteration, four folds were used for training and one for testing, ensuring that every sample was included in a test fold. The classifier's performance was assessed by AUC and accuracy computed across all test folds. Feature importance was computed for each fold to assess the stability of biomarker contributions. ROC‐based dual cutoffs were applied to classify samples as uninfected, infected, or uncertain. This framework enables reproducible CII computation for downstream analyses, including MOI‐dependent infection dynamics and comparative phage potency. Additional data analysis and visualization Statistical analysis was performed using SPSS 18 (IBM, RRID: SCR_002865). For group comparisons, Student's t‐test was employed with a nominal significance level of p < 0.05. All statistical tests were two‐sided. Comparisons were defined a priori between specified experimental conditions at each time point, and different time points were treated as independent experimental conditions rather than repeated measures. Accordingly, no multiple comparisons were conducted on the same dataset, and no correction for multiple testing was applied. Data visualization was implemented with OriginPro 2021 (OriginLab Corporation, RRID:SCR_014212), custom Python (RRID:SCR_008394) 3.09 scripts, and schematic illustrations were created via BioRender.com. AUTHOR CONTRIBUTIONS Xiao Han: Conceptualization; data curation; formal analysis; methodology; supervision; visualization; writing—original draft; writing—review and editing. Xiaofu Wan: Investigation; methodology. Yang Zhou: Investigation; resources; writing—review and editing. Xiaoting Fu: Investigation; methodology. Xiaoshan Zheng: Writing—review and editing. Bo Gao: Investigation; methodology. Shi Huang: Writing—review and editing. Anle Ge: Funding acquisition; Writing—review and editing. Jiadong Huang: Writing—review and editing. Hongzhou Lu: Investigation; resources; writing—review and editing. Jian Xu: Conceptualization; funding acquisition; project administration; supervision; writing—original draft; writing—review and editing. ETHICS STATEMENT As all bacterial strains and phages were from residual samples used in clinical diagnosis or strains from their subcultures, the criteria for exemption were met. CONFLICTS OF INTEREST Prof. Jian Xu is among the founders of Single‐Cell Biotech Co. Ltd. All other authors declare no competing interests. Supporting information Supplementary Materials 20260512.