Fast, Streamlined In Situ Genotyping of Bacterial Strains for Single-Cell and Subcellular Dynamics Profiling

Abstract

Temporal dynamics, such as cell-cycle oscillations, transcriptional bursting, and the regulated assembly and disassembly of phase-separated condensates, constitute an important layer of regulation in biology. Yet, existing technologies, including DNA sequencing, plate readers, and fluorescence microscopy, face a fundamental trade-off between temporal resolution, single-cell and subcellular information, and experimental throughput. This trade-off not only fragments our understanding of fundamental cellular processes but also impedes large-scale screening for dynamic phenotypes. Efforts to mitigate such limitations are beginning to emerge. One promising approach combines live-cell time-lapse imaging of pooled bacterial libraries in microfluidic chips with sequential fluorescence in situ hybridization (seqFISH) to connect dynamic phenotypes exhibited by individual cells with their corresponding genotypes. However, widespread adoption of this strategy has been constrained by two technical challenges: (1) compromised in situ genotyping efficiency due to nonspecific fluorescent probe aggregation in confined microfluidic growth chambers, and (2) the reliance on high-copy-number plasmids to ensure genotyping sensitivity, which often distorts dynamic phenotypes and perturbs normal physiology during live-cell imaging. Here, we address both limitations with an integrated and physiologically compatible solution. We introduce a simple surface-modification strategy to suppress fluorescent probe aggregation in microfluidic chips, enabling robust in situ genotyping to be completed overnight. In parallel, we implement barcoded copy-number-tunable plasmids to decouple the conflicting requirements of live-cell imaging and post hoc genotyping, thereby preserving delicate subcellular dynamics while maintaining genotyping sensitivity. Together, these improvements transform image-based bacterial screening into a faster and minimally perturbative workflow, expanding the scalability and versatility of high-throughput single-cell and subcellular dynamics screening for both fundamental and synthetic biology research.

Introduction

Time-resolved, single-cell characterizations of complex phenotypes in bacteria have enabled fascinating discoveries about the fundamental mechanisms governing temporal dynamics and subcellular organization in biology. For example, live-cell fluorescence imaging has been indispensable for uncovering stochastic and transient gene expression events in bacteria, which have far-reaching implications for various aspects of bacterial physiology, such as cell-fate transitions, , stress tolerance, and multidrug resistance. In synthetic biology, live-cell imaging has also been instrumental for prototyping synthetic gene circuits engineered to perform specific dynamic functions to assess their dynamic fidelity, precision, and long-term robustness. − However, despite its single-cell sensitivity and fine-grained temporal resolution, conventional optical microscopy is typically limited to tracking only two to three distinguishable strains simultaneously, substantially constraining its capacity for large-scale profiling of dynamic phenotypes in live bacterial populations.

To overcome the throughput limitations of conventional live-cell microscopy, recent years have witnessed a surge of new techniques which enable large-scale, single-cell characterization of bacterial populations across diverse genetic backgrounds. − However, among these, only a few preserve the ability to record dynamic phenotypes. These methods rely on powerful strategies that decouple phenotyping from genotyping, where dynamic phenotypes, such as cell growth, division, and fluorescence, are first recorded through time-lapse imaging, followed by post hoc genotyping to identify each cell’s strain identity. For example, in the SIFT (single-cell isolation following time-lapse imaging) platform, strains from a pooled bacterial library are cultured in parallel within individual growth traps of a microfluidic device known as the mother machine, , allowing their dynamics to be monitored at single-cell resolution for virtually unlimited cell generations. After live-cell imaging, cells displaying phenotypes of interest are isolated via optical trapping and transferred into designated containers, where they can be genotyped by sequencing. Because this retrieval process preserves cell viability and prevents cross-contamination, isolated cells can be used to reseed new cultures or subjected to further downstream analyses such as plate-based assays and next-generation sequencing.

A complementary approach, known as DuMPLING (dynamic u-fluidic microscopy-based phenotyping of a library before insitu genotyping), also relies on mother machine–based time-lapse imaging, but reveals each strain’s identity directly in cells fixed in situ within the growth traps (Figure ). In this approach, the bacterial strain library is constructed by fusing each genetic variantsuch as a sgRNA spacer, a protein mutant, or a fluorescently labeled subcellular componentto a unique barcode sequence. The one-to-one correspondence between each genetic variant and its barcode is known a priori via sequencing. Following live-cell imaging in the mother machine, cells are fixed within the chip, and their barcodes and thus strain identities are revealed by fluorescence in situ hybridization (FISH). This enables genotyping of the entire imaged library, not just a preselected subset, and has the potential to uncover critical genetic sequences and biochemical features that enable or abolish specific physiological traits in a high-throughput manner.

1.

Workflow of combinatorial seqFISH-based in situ genotyping experiments for screening single-cell and subcellular dynamics. (A) Library construction: each strain carries a dual-purpose plasmid that expresses both a genetic variant for dynamic phenotyping and its corresponding barcode for in situ genotyping. In this study, we employ a 12-digit ternary barcoding scheme, where each ternary digit (a “trit”) can take one of three possible values (0, 1, and 2). (B) Phenotyping: the pooled library is introduced into the mother machine, where individual “mother cells” positioned at the ends of growth traps generate isogenic progeny cells that eventually fill the growth traps. Time-lapse imaging is performed to capture single-cell features such as growth rate, morphology, and fluorescence, depending on the research question of interest. (C) Genotyping: combinatorial sequential FISH (seqFISH) is applied to determine the barcode sequence expressed by each strain in every mother-machine growth trap. In this study, digit values of 0, 1, and 2 are detected using readout probes labeled with Alexa Fluor 488, ATTO 550, and Alexa Fluor 647, respectively. As an example, barcode sequences for digit 1 are shown in black, with their corresponding readout probe sequences shown in green (for digit value 0), yellow (digit value 1), and red (digit value 2). (D) Barcode decoding: strain identities are determined by mapping barcode sequences to a predefined barcode-to-strain lookup table obtained from sequencing.

The FISH-based genotyping strategy offers a powerful means to comprehensively map genotypes to phenotypes at single-cell resolution. In principle, using a ternary (0/1/2) barcoding scheme and N cycles of combinatorial sequential FISH (seqFISH), , up to 3 ^N distinct strains can be genotyped within a single experiment. The feasibility of this approach has been demonstrated in a macroscopic flow chamber, where Escherichia coli cells are chemically tethered to a glass coverslip. However, extending combinatorial seqFISH to the microfluidic channels of the mother machine, which is essential for long-term dynamic phenotyping, has proven more challenging. The small dimensions of mother machine growth traps, together with chemical properties of poly­(dimethylsiloxane) (PDMS) used in chip fabrication, promote nonspecific aggregation of barcode-targeting FISH probes during the genotyping step, which in turn significantly deteriorates signal quality. For example, in the original DuMPLING study, a single round of FISH hybridization required approximately 16 h, presumably to allow sufficient signal accumulation over background fluorescence arising from nonspecifically adsorbed probes. A subsequent study estimated that a complete FISH cycle took roughly 24 h, and consequently, forwent combinatorial seqFISH in favor of sequential hybridization without probe stripping between rounds, which markedly reduced the number of strains that could be profiled within a single experiment.

In addition to challenges associated with slow and inefficient probe hybridization and stripping, image-based in situ genotyping experiments have thus far relied almost exclusively on high-copy-number plasmids to express barcode mRNA, , suggesting the necessity of sufficient barcode expression levels for successful execution. However, such requirement significantly limits the range of experiments that can be performed. For instance, faithful recording of promoter activities and protein dynamics typically requires fluorescent reporters to be either integrated into the chromosome or expressed from low-copy-number plasmids, to avoid overexpression artifacts that can lead to spurious phenotypes. − In synthetic biology, high-copy-number plasmids not only reduce production yields due to increased metabolic burden , but can also undermine the precision and stability of synthetic gene circuits.

In this study, we present an image-based screening workflow compatible with high-throughput profiling of single-cell and subcellular dynamics using pooled bacterial libraries. Building on the established paradigm of live-cell phenotyping followed by in situ genotyping, , our approach improves the signal quality and speed of combinatorial seqFISH-based in situ genotyping, and is now compatible with screening single-cell and intracellular dynamic phenotypes that are sensitive to plasmid copy number induced artifacts.

These advances are achieved by addressing critical bottlenecks that have limited previous implementations, including slow genotyping efficiency within confined microfluidic environments and the incompatibility between faithful phenotyping and high plasmid copy numbers. First, we implement a simple surface modification strategy to overcome FISH probe aggregation in microfluidic growth traps, enabling robust overnight (<24 h) in situ genotyping across 12 seqFISH rounds. Second, by incorporating a plasmid amplification step between the live-cell imaging and genotyping, we temporally decouple phenotypic measurement from barcode signal amplification. This extra step resolves the conflicting plasmid copy number requirements between the two experimental stages, allowing single-cell and subcellular dynamics to be faithfully recorded at high throughput with reduced physiological perturbation or metabolic burden. Finally, using the bacterial MinCDE system as a case study, we demonstrate that this phenotyping-amplification-genotyping workflow supports systematic profiling of subcellular dynamics using bacterial mutagenesis libraries.

Together, these advances significantly improve seqFISH-based in situ genotyping efficiency in microfluidic chips, while preserving delicate intracellular dynamics during live-cell imaging. The resulting framework provides a generalizable route for linking single-cell dynamic phenotypes to genotypes in microfluidic systems, paving the way for broader adoption in systems and synthetic biology.

Results and Discussion

Surface Modification of Microfluidic Chips Improves Signal-to-Background Ratio for In Situ Genotyping

We reasoned that the duration of in situ genotyping could be reduced by either increasing the true signal amplitude or suppressing background fluorescence in each seqFISH cycle. We therefore begin our study by tackling the previously reported background issues arising from probes nonspecifically adsorbed to the mother machine, with the goal of simultaneously improving both the signal-to-background ratio and accelerating in situ genotyping. To quantitatively assess the extent of nonspecific probe aggregation, we simultaneously introduced a mixture of three commonly used fluorescent probes, Alexa Fluor 488, ATTO 550, and Alexa Fluor 647 (Figure A), into empty mother machine chips without cells. The chemical properties of these dyes, including hydrophobicity and charge, encompass a range that is representative of most commercially available fluorescent probes. , The dimensions of the mother-machine growth traps (average width = 1.4 μm; average vertical height = 1.29 μm; length range = 20–30 μm) closely match those used in prior studies, − including a design optimized for E. coli MG1655, the bacterial strain examined in the present study. The feeding trench was deliberately made with a large height (100 μm), allowing the use of a relatively high flow rate (∼35 μm/min), near the upper end of previously reported values, to promote efficient removal of probes adsorbed to the chip. Nonetheless, in chips fabricated with standard PDMS curing protocols, we still observed substantial nonspecific probe retention following stripping buffer perfusion, particularly for ATTO 550 and Alexa Fluor 647, which strongly adhered to the microfluidic growth traps (Figure B, upper panel).

2.

Surface modification of mother machine chips improves signal-to-background ratio (S/B) in in situ genotyping. (A) Chemical structures (in NHS-ester forms) of the three fluorescent dyes used in this study for genotyping are shown, along with their excitation and emission maxima, octanol–water distribution coefficients (log D, more negative values indicate greater hydrophilicity at the bulk level), and charge at pH 7.4. (B) Representative fluorescence images of 15 empty growth traps, illustrating dye retention after 1 h of continuous stripping-buffer perfusion in untreated versus PDMS-BCP–treated chips. For each of the three fluorescent color channels, images are displayed on the same intensity scale for direct comparison of probe aggregation. (C) Schematic diagram of PDMS-BCP treatment on a mother machine microfluidic chip. Side chains of incorporated BCP are colored in blue. The schematic is not drawn to scale and does not quantitatively represent the fraction of surface-modified sites. (D) Quantification of raw fluorescence intensities for three fluorescent dyes in untreated and treated chips (over 200 growth traps were analyzed for each fluorescent dye; error bars are standard errors of the mean from 2 independent experiments; *p < 0.05, **p < 0.01, ***p < 0.001 by two-tailed t-test). (E) Representative images from three seqFISH rounds with Test Strain 1, showing hybridization results in untreated and treated chips. Vertically projected (averaged) intensities for each dye are shown below the corresponding growth trap, normalized to the maximum intensity across all dyes and traps. Combinatorial seqFISH results from all 12 digits are shown in Figure S1. (F) Comparison of S/B values between untreated and treated chips. (G) Percentage of incorrectly genotyped growth traps (those with S/B < 1) in untreated and treated chips. (H) Representative filmstrips and time traces of signal amplitudes from individual genotyping rounds, illustrating that each seqFISH round, including hybridization and probe quenching, can be completed within 70 min. Filmstrips for each color channel are shown on the same intensity scale for direct comparison. Solid and open triangles mark the entry of probes and quenching buffer, respectively. Thin lines in the time traces represent single growth traps, while thick lines show the average. Single-cell amplitudes are normalized to the maximum value within the corresponding color channel. Scale bars = 5 μm.

The interaction between fluorescent dyes and microfluidic surfaces is governed by a combination of hydrophobic adsorption and specific interfacial trapping mechanisms, rendering many commonly used dyes susceptible to nonspecific adsorption on the hydrophobic PDMS surfaces used in microfluidic devices. For instance, ATTO 550 exhibits strong hydrophobic affinity for PDMS, whereas Alexa Fluor 647, despite being more hydrophilic in bulk solution, is prone to surface trapping due to its tetra-anionic character and highly polarizable cyanine core, which facilitate multivalent electrostatic interactions and π–π stacking at PDMS interfaces (Figure A). In principle, surface coatings such as detergents or bovine serum albumin can partially mitigate nonspecific probe binding; however, these layers can be easily removed during the harsh stripping steps required between consecutive hybridization rounds in combinatorial seqFISH, leaving subsequent cycles susceptible to renewed probe aggregation. To overcome this limitation, we sought to modify the surface property of PDMS chips in a more robust manner.

Conventional approaches of increasing the surface hydrophilicity, such as plasma treatment or silanization, either have short-lived effects (typically lasting only hours to a few days) or require complicated and labor-intensive chip fabrication protocols. Instead, we resort to a simpler approach by doping the PDMS prepolymer with a trace amount of PDMS–poly­(ethylene glycol) (PEG) block copolymer (referred to as BCP hereafter) during chip curing (Figure C). This strategy requires no additional fabrication or sample preparation steps, is compatible with optical microscopy, and has been reported to confer stable surface hydrophilicity to PDMS devices for over two months.

Comparing dye intensities in empty growth traps confirmed that BCP surface modification markedly reduces probe retention across all three dyes (Figure B, lower panel; Figure D). These results suggest that BCP treatment is a promising strategy for reducing background issues caused by nonspecific probe aggregation in the mother machine, thereby enhancing both the accuracy and efficiency of in situ genotyping.

To directly test this, we performed combinatorial seqFISH-based in situ genotyping on an E. coli strain (hereafter referred to as Test Strain 1) expressing a 12-digit ternary barcode of known sequence: 221221110021 (Figure E). The barcode was placed under a strong phage-derived T7 promoter, with expression induced by arabinose through production of genomically integrated T7 RNA polymerase under a pBAD promoter.

In each of the 12 genotyping rounds, three single-stranded DNA readout probesconjugated with Alexa Fluor 488 (for digit 0), ATTO 550 (for digit 1), or Alexa Fluor 647 (for digit 2)were introduced into the device. Since all cells carry the same barcode sequence, by design, only one of the three probes was expected to remain bound during each hybridization cycle for all growth traps. However, across multiple cycles we observed substantial interference from readout probes other than those corresponding the ground truth value (e.g., Alexa Fluor 488 and ATTO 550 at digit 1, Alexa Fluor 488 and Alexa Fluor 647 at digit 3, etc.) (Figure E). These nonspecifically adsorbed probes produced elevated background fluorescence that obscured the true signals from specifically hybridized probes, leading to comparable intensities between true and spurious signals, and in some cases even incorrect barcode calls. On the other hand, we were able to obtain visibly cleaner genotyping results with BCP-treated chips, with minimal interference from nonspecifically bound probes across all 12 cycles (Figures E, S1, and S2).

Because raw fluorescence intensities can vary across dyes and hybridization rounds, to quantitatively evaluate seqFISH signal qualities with and without BCP-treatment, we quantified the signal-to-background ratio (S/B) for each growth trap by first averaging raw pixel intensities along the longitudinal axis of the growth trap (Figure S3). The resulting one-dimensional (1D) lateral intensity profile of each of the three fluorescent dyes was then fit to a Gaussian function to extract the signal amplitude above background. For every digit in each growth trap, the S/B was defined as the ratio between the amplitude of the ground-truth dye (e.g., Alexa Fluor 647 for digit 1 of Test Strain 1) and the amplitude of the strongest non–ground-truth dye. A digit can be correctly genotyped if the amplitude of the ground-truth dye exceeds that of the competing dye, i.e., if S/B > 1.

Comparing the S/B between untreated and surface-modified chips revealed that BCP-treatment indeed significantly enhanced signal quality across all three ground truth digit values (and thus all three fluorescent dyes) (Figure F). Due to the improved S/B, the probability of barcode misidentification was markedly reduced (Figure G). Time-lapse imaging of the genotyping procedure further reveals that, with the improved S/B, a single hybridization cycle (probe binding, washing and stripping) can be efficiently and realizably accomplished within about 1 h (Figure H). This enhancement enables the practical implementation of combinatorial FISH with probe stripping between consecutive hybridization rounds within the mother machine. For example, using our 12-digit barcode design, a complete 12-round in situ genotyping procedure can be completed overnight.

Postphenotyping Plasmid Amplification Enables Minimally Perturbative Live-Cell Imaging

Having resolved the issue of nonspecific probe aggregation in mother machine growth traps, we next seek to extend the applicability of this platform by enabling compatibility with low-copy-number plasmids, which are often preferred in systems and synthetic biology applications.

To achieve robust barcode expression while maintaining low plasmid copy number during live-cell imaging, we turned to copy-number-tunable plasmids. ,, Among available options, the copy number-inducible trigger plasmid (pTrig) offers one of the widest dynamic ranges. − pTrig harbors two replication originsmini-F origin for stable plasmid maintenance and oriL, derived from P1 phage, for plasmid copy number amplification (Figure A). Upon IPTG induction, the replication protein RepL, expressed from the lac promoter, binds to oriL to initiate plasmid amplification, increasing the plasmid copy number from ∼1–3 copies to several hundred copies per cell.

3.

Postphenotyping plasmid amplification permits in situ genotyping without maintaining high plasmid copy numbers during live-cell imaging. (A) Schematic of the mechanism of plasmid copy number amplification using dual-origin, barcoded vectors. Cells are first phenotyped in mother machine growth traps in low-copy plasmid state to minimize perturbation, followed by IPTG-induced plasmid amplification prior to combinatorial seqFISH. (B) Representative filmstrips of a single growth trap with cells expressing sfGFP under a constitutive promoter on a pTrig plasmid, with or without IPTG. (C) sfGFP intensity time traces from individual growth traps in the presence or absence of IPTG. (D) Distribution of sfGFP intensities after 6 h with or without IPTG. More than 80% of growth traps exposed to IPTG reached an intensity of 100 (vertical dashed line), indicating successful plasmid copy number amplification. (E) Combinatorial seqFISH images of uninduced and IPTG-induced Test Strain 2 cells, showing that reliable genotyping requires both high plasmid copy number (via IPTG induction) and a strong promoter. All fluorescence images are shown on the same intensity scale. Complete 12-digit combinatorial seqFISH results for Test Strain 2 are shown in Figure S4C. (F) Representative combinatorial seqFISH images of a pooled barcode library expressed from the T7 promoter on the pTrig plasmid. Growth traps marked with • indicate those that could not be reliably genotyped due to low signal amplitudes. Complete 12-digit combinatorial seqFISH results are shown in Figure S5. (G) Distribution of S/B values at the 67th percentile (across all 12 digits per growth trap) for barcodes expressed from the pTrig plasmid. In ∼23% of growth traps, the 67th percentile S/B value fell between 0 and 1, meaning at least 8 of 12 digits (67%) of a growth trap were incorrectly genotyped, comparable to or worse than random guessing. (H) Box plots comparing the signal amplitudes between incorrectly (S/B ≤ 1) versus correctly (S/B > 1) genotyped digits. Scale bars = 5 μm.

We cloned the same T7 promoter and 12-digit barcode sequences from Test Strain 1 into the pTrig backbone to generate Test Strain 2. The pTrig backbone also encodes a constitutively expressed fluorescent protein (sfGFP), providing a means for visualizing the plasmid copy-number amplification process in individual cells. Tracking the sfGFP time course reveals that postphenotyping plasmid amplification can be reliably accomplished within about 6 h following IPTG induction (Figures B–C and S4A,B), for at least 80% of the cells (Figure D). Additionally, the strong sfGFP signal generated during this process can be efficiently quenched by methanol (MeOH) fixation prior to genotyping, ensuring that sfGFP and presumably related fluorescent proteins are suitable reporters for live-cell phenotyping without interfering with subsequent in situ genotyping.

Combinatorial seqFISH results demonstrates that the strategy of postphenotyping plasmid amplification indeed yielded sufficiently strong fluorescence signals for both Test Strain 2 (Figures E,F and S4C) as well as a pooled library of known barcode sequences (Figure S5). For most growth traps, barcode calling can be reliably performed. However, in about 23% of growth traps, at least 8 of 12 digits (≥67%) were misclassified (S/B < 1) (Figure G), comparable to random guessing. Closer inspection revealed that incorrectly identified digits had maximum signal amplitudes (across all three probes) significantly lower than those of correctly identified digits and close to background, indicating insufficient barcode production. This finding is consistent with the ∼20% of growth traps that failed to show plasmid amplification in the sfGFP experiments. In practice, such growth traps can be safely excluded from analysis using an intensity threshold.

As an additional check, we asked whether expression of barcode mRNAs under a strong T7 promoter would be sufficient for reliable in situ genotyping. However, in the absence of IPTG-induced plasmid amplification, none of the readout probes produced a detectable signal above background, indicating that low-copy-number plasmids alone do not yield enough barcode mRNA for seqFISH-based genotyping (Figure S6).

A Streamlined Workflow for Screening Subcellular Spatiotemporal Dynamics

Having verified the feasibility of postphenotyping plasmid amplification for in situ genotyping, we next sought to demonstrate the streamlined workflow integrating both dynamic phenotyping and in situ genotyping. A unique strength of image-based screening, compared with mainstream techniques such as plate reader assays or DNA sequencing, is that it preserves both single-cell dynamics as well as subcellular information. The three-protein MinCDE system in E. coli provides an ideal test case: this three-protein network exhibits not only oscillatory temporal dynamics but also striking spatial patterns that regulate cell division-site placement. Specifically, interactions between MinD and MinE proteins generate pole-to-pole oscillations of all three proteins within the cell. As a result, MinC, the cell division inhibitor carried along by MinD, displays a time-averaged concentration that is highest at cell poles and lowest at midcell, allowing a typical rod-shaped bacterial cell to undergo binary fission and divide symmetrically, so that cellular resources are partitioned equally among the two progenies. Recent work has shown that the dynamic patterns formed by Min proteins are highly robust against fluctuations of the external growth condition. In contrast, earlier experimental and computational studies have demonstrated that these spatiotemporal patterns critically depend on the reaction kinetics between MinD and MinE, − highlighting an inherent fragility or tunability in dynamic pattern formation governed by their biochemical properties.

To provide a proof-of-concept demonstration of the full experimental pipeline, and to illustrate the potential of this platform for screening dynamic subcellular patterns, we performed live-cell imaging followed by in situ genotyping on a small, pooled library of four E. coli strains expressing either wild-type (WT) or mutant variants of MinE. Point mutations were introduced within the membrane-targeting sequence (MTS) of MinE, a region known to stimulate MinD’s ATPase activity and to be essential for proper subcellular pattern formation. , MinC translationally fused to mKate2 was used as a proxy for visualizing the dynamic output of this protein network. All three proteins, including MinE variants and mKate2-MinC, are expressed from the native promoter of the minCDE operon on the pTrig plasmid (Figure A). The endogenous minCDE operon on the E. coli genome was deleted to eliminate interference from natively expressed Min proteins.

4.

Phenotype-genotype mapping of copy-number-sensitive intracellular spatiotemporal dynamics. (A) Construct of the dual-purpose plasmid for phenotyping and genotyping MinE variants. (B) Live-cell dynamics of mKate2-MinC (red) in four strains, each expressing a different MinE variant, were recorded by time-lapse imaging and subsequently resolved by in situ genotyping to reveal strain identities. Live-cell phenotyping images are cropped around the ends of mother machine growth traps to highlight the dynamic positioning of mKate2-MinC within a single bacterial cell (white contours). Scale bars = 3 μm (upper panel) or 5 μm (lower panel).

Live-cell fluorescence imaging revealed different spatiotemporal dynamics of mKate2-MinC, ranging from the characteristic pole-to-pole oscillation to nearly immobile puncta. Following postphenotyping plasmid amplification and in situ genotyping, the identity of each growth trap was resolved, establishing a one-to-one correspondence between mKate2-MinC dynamics and the minE variant (Figure B). As expected, cells expressing WT minE displayed robust pole-to-pole oscillations, traversing the cell length roughly every 30 s, confirming both the integrity of the genetic construct and the reliability of the imaging system. The MinE­(L8Y) mutant retained oscillatory behavior but with reduced speed, whereas the MinE­(F6C) and MinE­(N13Q) mutants severely impaired mKate2-MinC mobility, suggesting these residues may serve as functional hotspots for tuning MinCDE dynamics.

As with many other oscillatory systems in biology, such as the KaiABC circadian clock in cyanobacteria or the synthetic repressilator, the fidelity of Min dynamics is highly sensitive to plasmid copy number. , Therefore, visualization of the Min system typically relies on fluorescently tagged proteins expressed from low-copy plasmids. − In line with this, we observed that Min proteins expressed from medium- or high-copy-number plasmids completely abolished the characteristic pole-to-pole oscillations in wild-type cells (Figure S7), underscoring the need for distinct plasmid copy-number regimes between the phenotyping and genotyping phases. Looking ahead, our approach of leveraging copy-number-tunable plasmids for barcode expression could be used to systematically map functional hotspots in MinD and MinE using a scanning mutagenesis library, offering insights into the robustness and tunability of spatiotemporal pattern formation.

Conclusion

Image-based phenotype–genotype screening is an emerging approach with the potential to overcome key limitations of prevailing methods by simultaneously offering high-throughput capacity, single-cell and subcellular sensitivity, and fine-grained temporal resolution. ,,, Yet, despite its promise, this approach has remained largely underutilized and has not seen widespread adoption in either basic or applied research. Here, by addressing the major issues of nonspecific probe aggregation and the reliance on high-copy-number plasmids for barcode expression, we extend its utility to better accommodate a broader range of experimental needs, including fast in situ genotyping (∼1 h per barcode digit) within the mother machine using combinatorial seqFISH, as well as the need to maintain low plasmid copy numbers during the live-cell phenotyping stage.

It should be noted that the strategies introduced here are unlikely to represent the only viable solutions. For instance, alternative surface modification techniques, such as poly­(vinyl alcohol) (PVA) deposition, may offer a viable strategy to mitigate nonspecific probe aggregation. While our primary aim has been to enhance genotyping sensitivity through direct hybridization of readout probes to barcode mRNAs, the improved surface properties of the PDMS device may also support a broader range of nucleic acid and protein-based biochemistry within the confined geometry of the mother machine. For example, combinatorial FISH employing both nonfluorescent primary probes and fluorescently labeled secondary probes, which has previously been hampered by probe aggregation, may similarly benefit. In addition, beyond copy-number-tunable plasmids, recent studies have demonstrated that chromosomally encoded barcodes can be amplified via enzyme-catalyzed rolling circle amplification (RCA), , offering another feasible route for signal amplification and barcode readout. A key advantage of RCA-based approaches is their compatibility with chromosomally integrated genetic constructs, which can be particularly important for applications requiring precise quantification of gene expression noise, as chromosomal copies are known to reduce extrinsic variability relative to plasmid-borne systems. At the same time, RCA involves multiple nucleic acid priming and enzymatic reaction steps, which increase experimental complexity, cost, and susceptibility to false-positive signals due to amplification of misprimed templates. In contrast, our plasmid-copy-number–based strategy offers a simple, modular, and enzyme-free means of signal amplification that integrates seamlessly with live-cell imaging workflows while minimizing phenotypic perturbations for most dynamic screening applications.

Taken together, this new workflow enhances the technical versatility and procedural flexibility of image-based, single-cell and subcellular dynamics screening, facilitating broader adoption for the discovery and characterization of a wide range of dynamic phenomena in fundamental and synthetic biology, from gene expression and cell-cycle progression to phase-separation and subcellular pattern formation.

Methods

Host Strains

To systematically assess the efficacy of microfluidic chip surface treatments in enhancing the signal-to-background ratio of in situ genotyping, as well as the efficiency of postphenotyping plasmid copy-number amplification upon IPTG induction, we first conducted experiments using a single strain carrying a known barcode, rather than a mixed strain library. This design minimized potential confounding effects from factors such as differential barcode expression (e.g., arising from variations in mRNA folding kinetics) that could occur in a pooled strain library. These strains are referred to as Test Strain 1 (high-copy plasmid backbone) and Test Strain 2 (copy-number–tunable backbone) throughout this study.

The background strain SYL2, used for making Test Strain 1, Test Strain 2, and the barcoded libraries, was derived from E. coli MG1655, a K-12 substrain commonly regarded as the wild type of this model organism. In SYL2, the coding sequence of the araB gene was replaced with a pBAD promoter driving T7 RNA polymerase (T7RNAP), together with a tetracycline resistance gene (tetB) as a selectable marker. The entire pBAD-T7RNAP-TetB construct was PCR-amplified as a single fragment from the commercially available BL21-AI strain (ThermoFisher, C607003). In addition, the motA gene, which encodes flagellar motility protein A, was deleted to prevent cells from swimming out of mother machine growth traps during live-cell imaging. For experiments involving coexpression of Min proteins and barcodes, the background strain SYL4 was generated by deleting the minCDE operon from the genome of SYL2. All genomic deletions and insertions were carried out using λ Red–mediated recombineering. Transformations were performed by electroporating 2 μL of plasmid solution into 30 μL of cell suspension prepared in 10% glycerol. A list of strains used in this study is provided in Table S1.

Construction of Test Strains 1 and 2

All plasmids used in this study were constructed by Gibson assembly using NEBuilder HiFi DNA Assembly Master Mix (New England BioLabs, E2621L). A list of plasmids used in this study is tabulated in the Table S2.

The high-copy plasmid used to generate Test Strain 1 (plasmid pYC01) was built by inserting the T7 promoter, the testing barcode sequence (synthesized by Twist Bioscience), rrnB T1 terminator, and a spectinomycin resistance gene into the pUC19 vector (NEB N3041S). DNA fragments for Gibson assembly were amplified by PCR with Phanta polymerase (Vazyme P510–02), and the vector backbone was linearized by PCR.

The copy-number-tunable plasmid used to generate Test Strain 2 (plasmid pYC13), which expresses superfolder GFP (sfGFP) along with a barcode sequence, was assembled by inserting two sets of elements into the pTrig vector (Addgene plasmid #104576): (i) the fluorescent protein module, including the constitutive infB promoter, sfGFP coding sequence, and rrnB terminator, and (ii) the barcode module, including the T7 promoter, the barcode sequence, and rrnB T1 terminator. The pTrig backbone harbors the repL gene under control of the Plac promoter; upon IPTG induction, repL expression is activated, enabling RepL to bind its phage-derived replication origin within the coding region and initiate plasmid copy-number amplification.

Barcoded Library Generation

To construct the high-copy and the copy-number-tunable plasmids expressing various barcode sequences, the original single barcode sequences from plasmids pYC01 and pYC13 were replaced with a library of randomized barcodes. Each barcode is a 252-bp DNA sequence composed of 12 subsequences, termed “digits”. Each digit can take one of three possible sequences, denoted as digit values 0, 1, or 2. These 12 (digits) × 3 (values) = 36 barcode sequences were selected from a subset of published sequences previously validated for compatibility with combinatorial seqFISH and are listed in Table S3. Each 20-bp barcode digit consists solely of the nucleotides A, T, and G, and is separated from adjacent digits by a single C to minimize the formation of secondary structures in the barcode mRNA.

The barcode library was assembled by combinatorial ligation following the protocol as described in Emanuel et al. Briefly, 42-bp oligonucleotides representing all possible adjacent digit pairs were synthesized (BGI Genomics). For example, digits 1 and 2 can be combined in 3 × 3 = 9 ways (digit 1–0 with digit 2–0, digit 1–1 with digit 2–0, digit 1–2 with digit 2–0, digit 1–0 with digit 2–1, digit 1–1 with digit 2–1, digit 1–2 with digit 2–1, digit 1–0 with digit 2–2, digit 1–1 with digit 2–2, digit 1–2 with digit 2–2), and accordingly 9 oligos were synthesized and purchased for this digit pair. Oligo sets for all other digit pairs (digits 3 and 4, digits 5 and 6, digits 7 and 8, digits 9 and 10, and digits 11 and 12) were synthesized similarly. To enable combinatorial ligation between consecutive digit-pairs, complementary oligo sets representing digit pairs on the antisense strand were also synthesized (for digit pairs 2 and 3, digit pairs 4 and 5, digit pairs 6 and 7, digit pairs 8 and 9, and digit pairs 10 and 11). To allow subsequent PCR amplification of all possible barcode sequences, a universal 20-bp primer-binding site was appended to both the 5′ and 3′ ends of the 12-digit barcode. Ligation between the forward primer-binding site and digit 1, or between the reverse primer-binding site and digit 12, was facilitated by three additional oligonucleotides, each overlapping with the primer-binding site and one of the three possible sequences (digit values) of the first or last digit. All oligonucleotides were dissolved and combined in TE buffer at 100 nM, assembled by T4 ligation and overlap PCR. The resulting random barcode library was then inserted into the pUC19 and pTrig backbones using the same cloning strategy as for plasmids pYC01 and pYC13. For this study, where knowing the ground truth identities of individual barcode sequences is crucial for quantifying S/B, colonies from the transformed plate were picked and sequenced individually via Sanger sequencing. One day before the experiment, freezer stocks prepared from these sequenced colonies were pooled by inoculating into a single culture tube, which was then loaded into the mother machine for imaging, following the same procedure as for single strains.

Construction of the Barcoded MinE-Variant Strains

pYC3_x plasmids (PminC-mKate2:minC minD minE, where the subscript x denotes different MinE variants) encoding Min proteins and their mutants were constructed by multifragment Gibson assembly. The Min module comprises both the native promoter for the minCDE operon together with the coding sequences of the Min proteins. The promoter sequence and coding sequences of MinC, MinD and wild-type MinE were PCR-amplified from E. coli MG1655 genomic DNA. An mKate2 tag was fused to the N-terminus of MinC to monitor its dynamics. DNA sequences encoding MinE variants with random single amino acid substitutions were synthesized by Twist Bioscience. The Min modules carrying these MinE variants were Gibson-assembled into the existing pTrig library containing different barcodes. A few colonies from the transformed plates were then picked for imaging to demonstrate the full dynamic phenotyping and genotyping pipeline.

Microfluidic Chip Fabrication

Both the master molds and the PDMS chips for microfluidic device were designed and produced in-house. The chip comprises two layers: growth traps (layer 1) and feeding trenches (layer 2). Each mother machine contains two feeding trenches, with 85 growth traps groups situated on each side of the feeding trench. The growth traps tested had an average vertical height of 1.29 μm (as measured by KLA-Tencor P-7 Surface Profiler), and an average width of 1.4 μm (measured using optical microscopy), ensuring that each growth trap can only be seeded by one mother cell. The feeding trench is 100 μm in height and 200 μm in width.

The master mold of the microfluidic chip was fabricated at the Nanosystem Fabrication Facility, HKUST. Layer 1 was made with a mixture of SU-8 2000:5 and SU-8 3005 (2:1 v/v mixed) and layer 2 with 100% SU-8 2050 at a height of 100 μm. All UV exposures were performed using the SUSS MA6 (for layer 1) and the AB-M (for layer 2) mask aligners with mercury I-line light filtered through a 360 nm long-pass filter (Omega Optical, PL-360LP). Hard baking was conducted at 160 °C for 2 h for each layer. Detailed photolithography parameters are given in Table S4.

The microfluidic chips were fabricated using PDMS and glass slides. A mixture of 64 g Sylgard 184 elastomer, 8 g curing agent and 0.1 g PDMS–PEG BCP (Gelest, DBE-712) was poured onto the master mold. The mixture was degassed for 3 h, and then cured overnight on a hot plate at 90 °C. The cured PDMS was carefully peeled from the mold, and holes were punched at the inlet and outlet ports using 17G blunt needles. The PDMS was then treated with plasma for 3 min. After the plasma treatment, the PDMS was immediately bonded to a glass 40 mm coverslip.

To reduce cell adhesion and prevent bacterial biofilm formation, the mother machine chip was surface-passivated on the day of the experiment by sequentially injecting 100 μL 5% (w/v) Tween-80 and 100 μL 5% (w/v) bovine serum albumin (BSA) diluted or dissolved in DNase/RNase-free distilled water (Invitrogen, 10977035), with a 10 min incubation after each step. After passivation, the concentrated cell culture was loaded into the chip.

Cell Loading

On the day prior to the experiment, a single colony (for Test Strain 1 or Test Strain 2) or multiple colonies (for barcoded libraries) were picked from LB agar plates and inoculated into 5 mL of LB broth in a 50 mL conical tube. Cultures were incubated overnight at 37 °C with shaking. The following day, the overnight culture was diluted 1:100 in LB supplemented with the appropriate antibiotics (50 μg/mL kanamycin for strains carrying the high-copy-number pUC19 plasmid, and 100 μg/mL carbenicillin for those carrying the pTrig plasmid) and grown at 37 °C with shaking at 200 rpm until reaching the exponential phase (OD₆₀₀ ≈ 0.5). Cells were harvested by centrifugation at 500g for 3 min, resuspended in 0.1 mL LB, and loaded into the mother machine.

Before imaging, the inlet and outlet ports were temporarily sealed with a narrow strip of 3M Scotch Magic Tape. The chip was then incubated at 37 °C for 2 h to allow cells to adapt to the microfluidic environment and diffuse into growth traps. The tape was subsequently removed, and the inlet and outlet were connected to the automated fluidics system to initiate imaging.

Microscopy

All imaging experiments were performed using a Nikon Ti2-E motorized inverted fluorescence microscope equipped with a CFI Plan Apochromat Lambda DM 100X Oil immersion phase contrast objective and a Photometrics Prime BSI Express sCMOS camera. Automated fluid exchange was controlled using a custom-built system originally designed for MERFISH experiments. , This setup included a peristaltic pump (Gilson Minipuls 3) and three daisy-chained valve positioners (Hamilton MVP 36798), each connected to an eight-way valve (Hamilton HVXM 8–5, 36798). Due to the small dimensions of the microfluidic chip, the maximal permissible flow rate is intrinsically limited to ∼150 μL/min. This, combined with the long fluidic path (∼1 m), hinders rapid switching of buffers at the mother machine inlet. To overcome this, we installed a solenoid valve (FluidicLab) near the chip inlet. During reagent switching, this valve reroutes the flow to a secondary waste container, bypassing the microfluidic chip and thereby enabling faster buffer exchange without risking chip damage. Flow rates were measured by the FS3Microfluidic Flow Sensor (FluidicLab), and all fluidics components were controlled via serial communication by a desktop computer according to manufacturers’ specifications.

During live-cell imaging, cells were supplied with LB growth medium containing the appropriate antibiotics, delivered either via a KD Scientific Legato 100 syringe pump or through the integrated fluidics system with a flow rate between 25–40 μL/min. For genotyping experiments involving the pTrig plasmid, to alleviate catabolite repression which could inhibit the transcription of repL gene and thereby limit plasmid copy number amplification, the mother machine was subsequently infused with M9 medium supplemented with 0.5% (v/v) glycerol for 2 h to elevate intracellular cAMP levels. Following this preconditioning step, pTrig plasmid amplification and barcode expression were simultaneously induced by continuous infusion of a barcode expression buffer consisting of M9 medium supplemented with 0.5% (v/v) glycerol, 2 mM IPTG, 2% (w/v) l-arabinose, and the appropriate antibiotics, for 4 h. Throughout live-cell imaging, including the plasmid amplification and barcode induction stages, temperature was held at 37 °C using a Tokaihit thermal box. After barcode induction, the heating unit was turned off to slow the degradation of barcode mRNA. Cells were then fixed on chip with pure methanol for 1 h and permeabilized with 70% ethanol for 1.5 h. Next, 90% (v/v) formamide in 2×SSC was flowed through the chip for 30 min to facilitate probe binding by relaxing the secondary structures of barcode mRNA.

The permeabilized cells were subsequently probed with 12 rounds of combinatorial seqFISH. FISH probes were synthesized by conjugating one of three fluorescent dyes (Alexa Fluor 488, ATTO 550, or Alexa Fluor 647) to oligos via NHS ester chemistry (BGI genomics). The oligo sequences were complementary to those representing individual barcode digits. Each seqFISH round consists of three steps: probe entry (hybridization), washing, and quenching. First, hybridization buffer [1% (w/v) dextran sulfate, 5% (w/v) ethylene carbonate, and 2×SSC, along with the probe sets for the corresponding digits] was continuously flowed into the chip for 30 min. Next, the washing buffer ([40% (v/v) formamide with 2×SSC]) was continuously introduced for 20–30 min to remove nonspecifically bound probes, during which microscopy images for genotyping were acquired. Finally, quenching buffer ([90% (v/v) formamide in 2×SSC]) was flowed for 10–20 min to extinguish fluorescent signals, allowing the subsequent seqFISH round to proceed. A complete list of chemicals used for chip passivation and combinatorial seqFISH imaging are tabulated in Table S5.

The fluorescence excitation light source was provided by the Lumencor SPECTRA light engine using the following excitation color channels: GFP (457/28 nm), Cy3 (555/28 nm), mCherry (575/25 nm) and Cy5 (637/12 nm). Chroma fluorescence filter sets were used to further filter the excitation wavelengths and separate emission from excitation. The excitation, dichroic, and emission filter wavelengths are as follows: GFP (ex. 470/40, di. 495, em. 525/50), Cy3 (550/20, 570, 585/20), mCherry (560/40, 585, 630/75), and Cy5 (640/30, 660, 690/50). Image acquisition parameters, in the format of (fluorophore name, Lumencor SPECTRA light engine color channel name, % of Lumencor SPECTRA light engine power, exposure time), are as follows: (sfGFP, GFP channel, 10%, 300 ms), (Alexa Fluor 488, GFP channel, 10%, 300 ms), (ATTO 550, Cy3 channel, 10%, 500 ms), (mKate2, mCherry channel, 10%, 200 ms), (Alexa Fluor 647, Cy5 channel, 10%, 300 ms). Vertical drifts of focal plane during the experiment were compensated using the Nikon Perfect Focusing System.

Data Analysis

Raw images acquired using the NIS-Elements image acquisition program were exported as TIFF files and analyzed using a custom-written MATLAB software suite. First, frame drifts between consecutive time points were corrected using cross-correlation, and images were rotated to align the mother machine growth traps vertically within each frame. Image segmentation was performed using custom-trained deep neural networks based on the U-Net architecture with an encoder depth of three. Because pixels corresponding to individual growth traps constitute only a small fraction of the entire image, model training faces a severe foreground–background class imbalance. To address this issue, we adopt a two-stage training strategy. In the first stage, a network model is trained to detect the entire growth-trap array across the horizontal axis of the entire image, including the spaces between adjacent traps. In the second stage, a separate network is trained to identify individual growth traps within the restricted regions detected by the first model. The first network performs binary pixel classification (trap array versus background), whereas the second network performs three-class pixel classification (individual traps, intertrap spacing, and background). The resolution of input images is flexible as long as they are resized to 128 × 128 pixels to match the input requirements of both networks. Training was performed using stochastic gradient descent, with a mini-batch size of 8 and an initial learning rate of 0.005. With mild data augmentation (horizontal and vertical flipping and small-angle rotations), we found that a relatively small training data set (approximately 30 images) spanning diverse trap orientations, cell-occupancy states, and illumination conditions was sufficient to achieve reliable segmentation performance after 50–100 epochs, although larger training data sets are expected to further improve robustness. Scripts for bright-image preprocessing, model training and growth trap segmentation are available at GitLab (https://gitlab.com/liaogroup.23/zhang-cai-et-al-2026). Following segmentation, filmstrips of detected individual growth traps were generated by cropping a 50 (width) × 285 (length) pixel region of interest (ROI) centered on each detected growth trap. For genotyping, fluorescence images were first processed with a 3 × 3 median filter to reduce camera shot noise. Pixel intensities from each of the three fluorescent channels were then averaged along the longitudinal axis of each growth trap, yielding a one-dimensional lateral intensity profile across the trap width, which was subsequently fitted to a symmetric Gaussian function of $I=b+A·\exp \left[\frac{{-(x-x_0)}^2}{2{\sigma }^2}\right]$ to extract the signal amplitude A above the background b (fitting to an asymmetric Gaussian function does not significantly change the results). The identity of each barcode digit was assigned according to the color channel with the highest amplitude (Alexa Fluor 488 for digit value 0, ATTO 550 for digit value 1, and Alexa Fluor 647 for digit value 2). The signal-to-background ratio (S/B) of each growth trap was quantified by dividing the amplitude of the ground-truth color channel (known from sequencing for Test Strains 1 and 2 and for the pooled libraries in this study) by that of the nonground-truth color channel with the highest amplitude. Occasionally, in cases where inefficient probe binding or incomplete stripping resulted in most growth traps exhibiting amplitudes below 20 (arbitrary fluorescence unit) in all three fluorescent color channels, additional FISH rounds were selectively performed for the affected barcode digit, and the round yielding the highest amplitude (among all rounds done for that digit) was used to assign the digit value.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/cbmi.5c00220.

• In situ genotyping of Test Strain 1 across 12 rounds of seqFISH; plasmid copy number amplification followed by in situ genotyping for Test Strain 2 (pTrig backbone), and genetic constructs and photolithography parameters (PDF)

§.

W.Z. and Y.C. contributed equally to this work. Y.L., W.Z., and Y.C. designed the study. W.Z. performed microfluidic chip fabrication, fluidics system setup, and microscopy. Y.C. performed molecular biology experiments. M.H.C. assisted with molecular cloning and imaging. Y.X. and A.R.W. assisted with photolithography. W.Z., Y.C. and Y.L. performed data analysis and wrote the manuscript. All authors assisted with manuscript preparation.

The authors declare no competing financial interest.