Introduction

Microbial cell factories underlie numerous innovations for sustainable biofuels, biomaterials, biochemicals and biomedicines1,2,3. In the iterative design-build-test-learn (DBTL) cycle of microbial strain development, the “test” phase (i.e., phenotype-based strain screening) is frequently a rate-limiting and the most tedious step4,5,6. To improve speed and efficiency of the test phase, methods for high-throughput screening of phenotypic traits are pivotal.

Traditional strain screening methods, primarily colony-based plate assays, rely on macroscopic measurements such as colony size or metabolic indicators7,8. While effective for basic identification and purification, these methods lack the capacity for phenotypic screening and are limited by low throughput, delayed feedback, and an inability to address cellular heterogeneity. As a result, rare or superior strains with subtle phenotypic advantages often go undetected, limiting the potential of synthetic biology. On the other hand, laboratory automation has enhanced the efficiency of colony-based screening by reducing manual labor and increasing consistency9,10. However, these systems still rely on population-level evaluations, which fail to capture dynamic single-cell behaviors and the subsequent selection of candidate colonies with improved traits. This highlights the need for single-cell resolution screening that is capable of detecting rare phenotypes and enabling more precise strain optimization.

Droplet-based microfluidics offer a promising solution for high-precision single-cell screening by encapsulating individual cells in droplets4,11,12,13. This approach enables screening based on metabolic and growth phenotypes at the single-cell level14,15. However, the process of droplet generation, storage, and sorting requires multiple chips, leading to complex workflows and higher likelihood of contamination. Additionally, indexing droplets are usually difficult, thus real-time process monitoring in a droplet remains a major challenge16,17. Another limitation is the tendency for droplet fusion, which reduces system stability and sorting accuracy. While some of such fusion issues can be alleviated by dispersing the droplets onto plates, the open-system nature introduces contamination risks and oil-phase evaporation, which further complicates the overall process and reduces reliability18,19,20.

Zymomonas mobilis, a Gram-negative bacterial chassis cell of industrial value, has been extensively studied for its efficient sugar metabolism and remarkable ethanol production capabilities21. Moreover, its potential in producing biofuels and biochemical22,23,24,25,26, particularly ethanol and lactic acid27, has garnered attention due to its robust metabolic pathways and tolerance to toxic substrates and products28. These features, together with its genetic tractability, make Z. mobilis an ideal biorefinery chassis29,30. However, the inability to rapidly identify high-yielding and tolerant strains has hindered the development of superior strains.

In this work, we develop an AI-powered Digital Colony Picker (DCP) platform that automatically processes pre-engineered microbial clones, screens phenotypic signatures, and exports selected strains through integrated growth-metabolic profiling at single-cell resolution. In a microfluidic chip with 16,000 picoliter-scale microchambers, AI-driven image analysis dynamically monitors and screens single-cell morphology, proliferation, and metabolic activities with spatiotemporal resolution. Then clones of target phenomes are exported by the contact-free Laser-induced bubble (LIB) technique. Employing the screening of Z. mobilis mutants as a model, we apply DCP to identify and select strains with elevated lactate tolerance and production to accelerate strain optimization and functional gene discovery taking the DCP advantages of single-cell multi-modal phenotyping, capability of spatiotemporal monitoring, scalable throughput, and AI-guided contactless target-clone export.

Results

Overall design of the DCP system and workflow

To overcome the limitation of conventional droplet-based flow cytometry systems, which lack dynamic monitoring capabilities, we developed an innovative AI-powered, addressed static droplet system for automated high-throughput screening of strains. The system consists of four core modules (Fig. 1A), each serving a specific function to streamline the screening process. (i) Microfluidic chip module, which as the foundational component houses 16,000 identical, pL-scale microculture units designed for high-throughput single-cell isolation and cultivation in each unit. (ii) Optical module, which integrates microscopy and lasers for imaging and LIB-based selection. (iii) Droplet location module, which ensures precise positioning and traceability of microchambers, facilitating efficient target identification. (iv) Droplet export and collection module, which allows seamless transfer of selected monoclonal droplets (i.e., the ones with the target phenotype) through the microfluidic chip and into a 96-well collection plate for downstream analysis.

Fig. 1: The overall design of the DCP system.

A Schematic of the DCP system. B Schematic of the chip. The chip consists of three layers, including a PDMS mold layer with microstructures, a metal film layer, and a glass layer. C DCP-based full process for single cell isolation, cell culture, monoclonal identification and sorting. Step 1: Vacuum-assisted single-cell loading and cultivation. Step 2: AI-powered identification and sorting of microscopic monoclones. D Optional step: the culture medium inside the microchamber can be replaced or replenished by injecting new liquid into the inlet. Scale bar: 60 μm.

Specifically, the microfluidic chip consists of three layers, including a PDMS mold layer with microstructures, a metal film layer, and a glass layer (Fig. 1B). The metal film, serving as a photoresponsive layer, is made of indium tin oxide (ITO), which facilitates the generation of microbubbles under rapid laser excitation. The ITO layer is uniformly deposited on the glass using magnetron sputtering, ensuring high-quality film coverage and uniformity. The ITO film exhibits a transparency of over 86%, enabling sufficient light transmission for clear visualization of single-cell-resolved aqueous bacterial colonies in the microchamber. Each microchamber is connected to a shared main channel via side channels, ensuring automatic efficient cell loading into the microchambers. A channel is designed around the perimeter of the microchamber on the chip, and water is introduced to reduce liquid evaporation from the edge microchamber. Gas-phase isolation between the microchambers prevents droplet fusion, supports stable incubation, and enables multiple media exchanges. A capillary tube is connected to the outlet to transfer the monoclonal droplets within the chip to a collection plate.

Based on this system, the high-throughput strain screening workflow proceeds in the following steps (Fig. 1C): Step 1: Vacuum-assisted single-cell loading and cultivation. The chip is pre-vacuumed, which allows rapid loading of a single-cell suspension in less than one minute. When the sample is subsequently introduced into the microchannels, residual air in the microchambers is absorbed by the PDMS layer, facilitating complete filling of the chambers without bubble entrapment. Gas-phase isolation between microchambers eliminates interference from oil phases and ensures stable droplet incubation, allowing for media replacement or addition as needed. Following loading, the chip is placed in a water-filled centrifuge tube and incubated in a high-precision temperature-controlled incubator, which allows individual cells within the microchamber to grow into independent microscopic monoclones (Supplementary Fig. 1). Step 2: AI-powered identification and sorting of microscopic monoclones. Following incubation, an oil phase is injected into the chip to facilitate droplet collection. Moreover, the introduction of the oil phase transforms the original gas intervals between each microchamber into oil intervals, ensuring that interference between the microchambers is avoided throughout the sorting process. The system automatically identifies the zero point of chip (upper-right corner by default) and uses AI-powered image recognition to detect microchambers containing monoclonal colonies. The motion platform positions the laser focus at the base of the identified microchamber. Using the LIB technique, microbubbles are generated at the chip membrane interface, propelling single-clone droplets toward the outlet. During the sorting process, multiple droplets may occasionally form, yet these droplets would remain encapsulated within the oil phase, thus contamination of adjacent microchambers is prevented. These droplets are collected at the capillary tip and transferred to a collection plate using a cross-surface microfluidic printing method31. The system also adjusts collection times in real-time based on droplet flow rates, ensuring precise collection of single clones. To mitigate the risk of environmental contamination, all experiments were performed within a biosafety cabinet under sterile conditions (Supplementary Fig. 2). Step 3 (optional): Liquid replacement. To optimize microbial colony growth, the system supports dynamic replacement of the liquid medium. Using gas gaps, culture media can be replenished or culture conditions changed through the chip inlet at any time (Fig. 1D). This capability enhances experimental flexibility and supports customized conditions for various research needs.

Microchamber-based bioreactors enhance single-cell cultivation and flexibility

The microchamber-based picoliter bioreactors effectively address the limitations inherent in traditional droplet-based systems by offering independent growth environments for each cell and by preventing droplet fusion (Fig. 2A). This feature not only prevents fusion of droplets but also allows each microchamber to provide sufficient space for microbial growth and metabolic activities, thus enhancing the overall performance of the system. These improvements enable a higher degree of precision in single-cell cultivation and subsequent analysis.

1. (i).

Single cell loading and distribution. The precise loading and distribution of single cells into the microchambers is critical for ensuring reliable colony formation and screening outcomes. To optimize this process, we performed Poisson distribution calculations (λ = 0.3), which suggested that a cell concentration of 1 × 10⁶ cells/mL would be ideal for single-cell loading into 300 pL microchambers32. This concentration was chosen based on the goal of minimizing the likelihood of multiple cells being loaded into a single microchamber, which could interfere with colony formation. Statistical analysis of E. coli cells labeled with green fluorescent protein confirmed that, at this concentration, approximately 30% of the microchambers contained a single cell, while only 5% contained multiple cells (Supplementary Fig. 3). By fine-tuning the cell concentration, we were able to minimize multi-cell occupancy, ensuring that monoclonal colonies formed reliably. This optimization is critical for the success of high-throughput screening, as it enables the isolation of genetically uniform colonies, facilitating more accurate downstream analysis (Fig. 2B).

1. (ii).

Mitigation of evaporation. Microchambers, due to their small volume, are particularly sensitive to liquid evaporation, which can alter the concentration of nutrients and metabolites in the cultivation environment. To tackle this challenge, we typically placed the chip within a 50 mL centrifuge tube that was 10% filled with water, which ensures a saturated vapor environment around the chip, maintaining high humidity throughout the incubation process (Supplementary Fig. 1). This approach effectively maintained a stable liquid volume, allowing for consistent environmental conditions over the duration of the cultivation period. Fluorescent sodium solution was used to monitor the liquid areas within the microchambers, and the results showed a liquid loss rate of approximately 6% after 24 h (Fig. 2C). However, for shorter-term cultivation (e.g., less than six hours for E. coli), the liquid loss was negligible, allowing for reliable monoclonal colony formation under stable conditions.

1. (iii).

Rapid and efficient replacement of the liquid medium. To support long-term cultivation and flexible experimental conditions, the microchambers facilitated rapid liquid exchange through molecular diffusion. Visualization experiments with fluorescent sodium and rhodamine B solutions showed that complete liquid replacement occurred within 5 min (Fig. 2D). The near-zero flow velocity within microchambers, as confirmed by COMSOL simulations, minimized microbial loss during liquid exchange (Fig. 2E). Cell counts before and after replacement showed no significant differences, and no cross-contamination was observed among microchambers, even for motile species such as E. coli (Fig. 2F).

Fig. 2: Evaluation of an innovative microchamber-based bioreactor.

A The innovative method of gas-phase isolation ensures that each microchamber becomes a separate culture unit. Individual cells can form a microscopic monoclone. Scale bar: 60 μm At least three independent experiments were performed, and similar results were obtained. B The number of cells within a microchamber follows a Poisson distribution probability model. Cells were counted in the microchamber and classified as 1 or >1. n = 3 biological replicates. Error bars: standard deviation (SD). Data are presented as mean +/− SD. Source data are provided as a Source Data file. C The proportion of microchambers that lost liquid in the microchamber array on the chip was detected each day. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file. D A visualization solution was employed to illustrate fluid exchange and monitor fluorescence intensity changes. Scale bar: 160 μm. At least three independent experiments were performed, and similar results were obtained. E COMSOL simulations indicate that while the main channel has a rapid flow rate, the flow within the microchamber is nearly zero, ensuring minimal impact on the liquid exchange process. F Monitoring changes in the number of microscopic monoclones in chips before and after the liquid exchange process. n = 3 biological replicates, no significant difference were determined by two-tailed Student’s t test. Data are presented as mean +/− SD. Source data are provided as a Source Data file.

AI-powered identification and automated export of microchamber with target phenotypes based on images

Automated identification and localization of microchambers harboring target microbial phenotypes are pivotal for downstream retrieval and screening, particularly given the high density of microchamber arrays on a DCP chip. To address this issue, we developed an AI-assisted image processing procedure to enable fully automated detection and analysis of microchambers (Fig. 3A). Specifically, an object detection model based on Faster R-CNN was implemented to accurately identify microchambers within images and output bounding box coordinates, which are precisely mapped to physical locations via perspective transformation, ensuring accurate localization and indexing33,34. Subsequently, a semantic segmentation model based on SegNet was employed to delineate the liquid regions within each microchamber, effectively mitigating background noise and interference from non-target areas, and generating region-of-interest (ROI) masks35. The images were then processed and analyzed using OpenCV, extracting quantitative metrics such as colony area and grayscale intensity features, thereby providing reliable indicators of microbial growth state and metabolic activity.

Fig. 3: AI-powered image identification and sorting of microbial monoclones.

A AI-powered image analysis. The captured images can be used to identify the microchamber in the image by object detection algorithm and the region of interest (ROI) can be segmented based on OpenCV. Subsequently, a semantic segmentation model identifies bacteria inside the microchamber and provides relevant information such as the number of bacteria, gray scale area of the bacterial region. B Different numbers of bacterial samples were used to load into the microchamber. The images were acquired automatically by AI, which also identified the number of bacteria within the microchamber. At least three independent experiments were performed, and similar results were obtained. C The number of bacteria identified by AI was found to be linearly correlated with the number of bacteria loaded into the microchamber, with R2 of 0.993. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file. D AI-powered panoramic image analysis. Starting from one end of the chip, panoramic identification of the chip can be achieved using S-shaped window shifting. The images from each screen were analyzed to form the panoramic image of the chip, and the image and position information of each microchamber was collected and stored. In addition, fluorescence information corresponding to each microchamber of the chip can be obtained based on the coordinate information obtained from the bright field. Through regular monitoring, the image change information of each chamber can be obtained, providing technical support for obtaining the single-cell scale growth phenotype. E Using LIB technology, microscopic monoclones were exported as droplets, which were collected into 96-well plates for incubation with the assistance of the capillary. F The accuracy of droplet recovery was evaluated, and the probability of recovered droplets being re-culturable was assessed based on E. coli monoclonal droplets. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file.

To assess the quantitative ability of the AI-assisted image recognition algorithm in identifying bacterial numbers within microchambers, we loaded E. coli at various concentrations into the chip and captured images for analysis (Fig. 3B). The bacterial counts within the microchambers were accurately identified and quantified, showing a linear correlation with the actual input concentrations of E. coli (R² = 0.998) (Fig. 3C). This high accuracy in identifying bacterial numbers demonstrates the robustness of our AI-powered system in processing and analyzing microchamber images.

Given the limited field of view in conventional microscopic imaging, multiple imaging sessions are often necessary to capture all the microchambers across the entire chip. To enhance the monitoring and identification process at the chip scale, we introduced the “From Point to Face” strategy, which systematically expands the imaging field through a dynamic, step-by-step process (Supplementary Movie 1). The “window shifting” method used for image acquisition enables the transition from single field-of-view images to a panoramic overview of the chip (Fig. 3D). By orchestrating precise platform movement and integrating automatic focusing, calibration, and deep learning techniques, high-quality images can be obtained even in cases of limited image clarity. Once the entire chip is captured, all microchambers undergo image analysis using OpenCV, and each chamber is tagged with positional data. This process allows the system to identify ~800 microchambers per minute (Supplementary Movie 1).

With microchamber positions already recorded, fluorescent field images can be easily captured by moving the platform without the need for re-identifying the microchambers (Supplementary Movie 2). Phenotype recognition under the fluorescent field, which includes distinguishing bacterial growth and metabolic activity, can be completed in a matter of minutes. The images obtained under the fluorescent field are then cross-referenced with those from the bright field, facilitating the concurrent retrieval of both growth and metabolic phenotype information for bacterial strains within the microchambers.

Once the AI-assisted system identifies the desired monoclones, the next step is to export the microscopic single-clones to a macroscopic system for further cultivation or analysis. The sorting process is carried out in two steps: (i) the target monoclones are exported from the microchamber into the microchannel as droplets; (ii) the droplets are “printed” into the macroscopic system, such as a 96-well plate (Fig. 3E). Specifically, the AI-powered image recognition system automatically identifies and quantifies each microchamber on the chip and indexes them according to set thresholds to locate the target droplets. LIB technology is then used to propel liquid from the microchambers into the main channel, which is filled with oil (Supplementary Movie 3). This allows the microscopic single-clones to be exported from the microchambers in droplet form and flow toward the outlet.

The outlet of chip is coupled with a capillary tube for droplet collection. The time required for droplets to travel from various microchamber locations to the capillary tube varies due to the driving force provided by the oil phase. To account for this, the system dynamically adjusts the droplet transfer time. We evaluated the droplet collection efficiency by using an array of fluorescent sodium droplets. The strong fluorescence signal emitted by these droplets made it easy to visualize their collection into a 96-well plate under the microscope. By using LIB technology, the target single-clones were successfully transferred to the collection plate. The 96-well plate was pre-loaded with 200 μL of mineral oil to prevent evaporation. The droplet recovery efficiency was found to be ~97% (Fig. 3F).

After droplet collection, the recovered clones need to be further cultured to allow for the growth of sufficient bacterial populations for downstream analysis. To demonstrate this, we generated a single-cell droplet array of E. coli and incubated it at 37 °C for 6 h. Subsequently, 60 single-clone droplets were automatically picked from the chip via laser and cultured on solid agar plates. Notably, 96% of the microscopic single-clones successfully grew on the plates (Fig. 3F). This high success rate highlights the reliability of the sorting and obtaining live cells after the laser picking for downstream scale-up cultivation process.

High-throughput single-cell phenotypic analysis and sort using the DCP system

Microfluidic bioreactors enable the capture and analysis of single cells to investigate key biological parameters, such as growth, morphology, and productivity36. However, conventional systems are limited to passive observation and lack the capability to precisely isolate phenotype-specific cells37. The DCP system overcomes this limitation by integrating single-cell dispersion, microchamber cultivation, and microscopic monoclone extraction into a seamless workflow. Moreover, thousands of bacterial monoclonal microchambers are phenotyped in parallel, while those with target phenotypes identified and exported in an automatic, sequential fashion.

Dynamic monitoring of single-cell growth

Using E. coli as a model, we tracked the complete, dynamic process of monoclonal formation within microchambers using DCP (Fig. 4A). Images of 500 microchambers were dynamically monitored from cell loading to after 48 h of culture and found that there was no cell growth caused by mutual contamination between microchambers. This result showed that the microchambers provided independent growth space for single cells. Leveraging an AI-based image recognition algorithm, we successfully traced the proliferation trajectories of 70 single-cell-containing microchambers, generating their respective time-resolved growth curves (Fig. 4B). Comparison of these curves revealed an initial exponential growth phase lasting approximately five hours, followed by a plateau phase due to physical space saturation within the microchambers. Notably, some microchambers exhibited a prolonged lag phase, suggesting potential heterogeneity in microenvironmental adaptation mechanisms. Thus, DCP allows the comparison and identification of such growth-based phenotypes for in-depth mechanistic investigations.

Fig. 4: Single-cell scale growth and metabolic phenotype analysis and sorting.

A The combination of bright field and fluorescence field allows precise identification of strain growth (bright field plaque area) as well as target substance expression (fluorescence signal) of monoclones within different microchamber chambers in the chip. B The number of cells within the microchamber can be monitored over time using a DCP-based system, which allows the formation of a growth curve at the single-cell level. C The process of gradual proliferation of individual E. coli to form microscopic monoclones was monitored. The absence of contamination observed in the microchamber lacking cells demonstrates the independence of this microchamber. Scale bar: 60 μm. D The accuracy and stability of the sorting procedure were evaluated by sorting fluorescent E. coli from mixtures with varying ratios (E. coli expressing green fluorescent protein and E. coli not expressing green fluorescent protein, ratios of 1:10, 100, 1000). Following one round of screening, the target cells were successfully obtained, demonstrating high stability and accuracy. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file.

Validation of fluorescent phenotype sorting efficiency

To demonstrate the capability of system for profiling and sorting metabolic phenotypes, we constructed gradient mixtures of GFP-expressing E. coli strains ATCC 35218 and DH5α at target-to-non-target ratios of 1:10, 1:100, and 1:1000. A multimodal detection strategy enabled highly efficient sorting: (i) bright-field imaging combined with AI-based recognition facilitated rapid localization of monoclonal microchambers, and (ii) fluorescence imaging quantitatively assessed GFP expression levels (Fig. 4C). Utilizing LIB, the DCP system achieved a sorting accuracy exceeding 95% when the target-to-background ratios were 1:10 and 1:100. At a more demanding ratio of 1:1000, the post-sorting proportion of target cells increased to nearly 80% after a single round of sorting (Fig. 4D), demonstrating exceptional performance in high-throughput screening of rare metabolic phenotypes.

Collectively, by seamlessly integrating bright-field dynamic tracking with fluorescence-based metabolic analysis, the DCP system establishes a complementary analytical strategy: the former enables real-time monitoring of cellular proliferation, while the latter provides precise quantification of target-metabolite levels. Such a single-cell, multi-modal phenotyping approach significantly enhances the efficiency of high-throughput screening and offers an methodological framework for profiling the comprehensive phenomic changes in bioprocesses that are mediated by an isogenic population or a consortium.

DCP-based high-throughput screening of Z. mobilis strains with high lactic acid production and tolerance

Z. mobilis, known for its efficient sugar metabolism and ethanol production, has gained broad interest for industrial production of biofuels and biochemicals, especially ethanol and lactic acid22,23,24,25,26. Its robust metabolic pathways, tolerance to high substrate and product concentrations, and genetic tractability make it an ideal industrial host for metabolic engineering. However, current screening technologies, such as those relying on macroscopic phenotypic assessments, are unable to support high-throughput, high-precision strain optimization, since capturing dynamic single-cell behaviors or subtle phenotypic differences of mutant cells remains a challenge.

Therefore, we developed a high-throughput workflow using the DCP system to screen Z. mobilis strains with high lactic acid production (metabolic phenotype) and tolerance to potassium lactate (growth phenotype) (Fig. 5A). A D-lactic acid biosensor derived from Pseudomonas fluorescens was employed to monitor intracellular D-lactic acid concentrations through fluorescence intensity38. The initial strain of Z. mobilis was engineered by introducing the heterologous LmldhA gene (encoding D-lactate dehydrogenase) from Lactiplantibacillus mesenteroides subsp. mesenteroides ATCC 8293 to enable lactic acid biosynthesis. This genetic modification was necessary because native Z. mobilis produces minimal lactate under anaerobic conditions39. The engineered strain additionally harbors a gene repression system designed to implement genome-wide CRISPR interference and a LldR-based lactate-inducible system, and was subsequently loaded onto the DCP chip for single-cell cultivation.

Fig. 5: Process and results of high throughput screening of strains using DCP.

A Schematic diagram of application of DCP to screen for Z. mobilis strains with enhanced lactic acid production and tolerance. B Bright field images can detect microscopic monoclones formed by single cells. Combined with fluorescence images can directly index microscopic monoclones that grow fast and have high lactate production. LIB technology enables the target microscopic monoclones to be exported and collected as droplets. At least three independent experiments were performed, and similar results were obtained. C Following one round of screening, five strains were obtained and incubated in shake flasks for 24 h. Lactic acid production by the strains was determined by HPLC. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file.

Screening for high lactic acid production

After 8 h of incubation, the low bacterial count of 10 in the microchamber suggests they are still in the growth phase and require more time. Extending the incubation to 24 h resulted in a significant increase to 200 in bacterial numbers, making them suitable for further analysis (Supplementary Fig. 4). In an efficient, fully visualized process, bright-field and fluorescence imaging identified fluorescent monoclones that correspond to the high lactate yield traits, and the target monoclones were automatically exported via LIB into individual wells in a 96-well plate as droplets (Fig. 5B).

After a single round of DCP-based screening, five potential target strains were directly isolated. To further validate the acid production capacity of the strains obtained from the sorting, they were transferred to shake flasks for further incubation. Lactic acid production was quantified using high-performance liquid chromatography (HPLC). After 24 h of cultivation, the selected strains showed enhanced lactic acid production, with strain S-3 exhibiting a 17.3% improvement compared to the control strain (Fig. 5C). These results highlight the capability of the DCP system to efficiently isolate high-yield strains in a streamlined workflow.

Screening for high lactic acid production and tolerance

To further enhance stress tolerance, the high-producing strain S-3 was subjected to another round of DCP-based screening under lactate stress conditions (30 g/L potassium lactate). Using bright-field imaging, we monitored cellular growth and viability under stress, while fluorescence imaging was used to capture lactate-associated metabolic activity. By integrating multimodal data on growth, metabolism, and stress tolerance from both imaging modalities, we found that most microchambers exhibited severe growth inhibition under lactate stress, with only a few chambers showing active cell proliferation (Supplementary Fig. 5). In some microchambers, strong fluorescence signals were observed despite the presence of only a few cells in the corresponding bright-field images (Supplementary Fig. 5). This discrepancy likely resulted from stress-induced cell lysis and the subsequent release of intracellular fluorescent proteins. To identify candidate strains with both high productivity and enhanced tolerance, we selected microchambers that exhibited both clear cell proliferation in bright-field images and strong fluorescence in the corresponding fluorescent field. These target microchambers were identified and isolated using DCP, and the recovered strain was further cultivated and designated as S-3-1.

The S-3-1 strain was reloaded onto the microfluidic chip, and an increased number of microchambers supporting active bacterial growth was observed. Iterative screening was performed by selecting microchambers exhibiting robust growth based on bright-field imaging and further refining the selection using fluorescence intensity as an indicator. Strains were preserved after each round of screening. By the fourth screening cycle, the strain displayed a significant improvement in potassium lactate tolerance, evidenced by a substantial increase in the number of microchambers supporting normal proliferation. The final strain, which exhibited both high cell proliferation and strong fluorescence signals, was designated as S-3-4.

To further validate the performance of the selected strains, fermentation experiments were conducted using both the enriched strains and the control strain in RMG5 medium supplemented with 30 g/L potassium lactate. This concentration was selected based on our tolerance assays showing that wild-type Z. mobilis exhibits growth inhibition near 30 g/L potassium lactate (Supplementary Fig. 6), thereby enabling effective discrimination of strain tolerance phenotypes. Strains S-3-3 and S-3-4 exhibited superior growth performance under lactate stress (Table 1, Supplementary Fig. 7), with growth rates increased by approximately 12.64% and 77.01% from the control strains, respectively (Table 1). Moreover, S-3-4 was the only strain capable of completely consuming glucose under lactate stress, with a 19.71% increase in net lactate production (calculated by subtracting the initial 30 g/L supplemented lactate) over the control (Table 1, Supplementary Fig. 7). These findings suggest a significant enhancement in lactate tolerance and production, especially in strain S-3-4.

Discovery of ZMOp39x027 enhances lactate tolerance and growth rate

To elucidate the molecular basis of enhanced lactate tolerance in S-3-3 and S-3-4, the CRISPRi plasmid was cured to generate derivative strains S-3-3t and S-3-4t respectively. Quantitative comparison under lactate stress (RMG5-30LA) revealed that while S-3-3t maintained similar growth kinetics to its parental strain, it showed significant reductions in both specific growth rate (0.072 ± 0.002 vs 0.080 ± 0.001 h−1; p < 0.001 by Student’s t test) and maximum biomass (0.941 ± 0.08 vs 1.221 ± 0.06 OD600nm; p < 0.001), representing 10% and 22.9% decreases respectively (Fig. 6A). Notably, these parameters aligned with baseline levels observed in the control strain (0.073 ± 0.005 h−1, 0.948 ± 0.027 OD600nm), suggesting CRISPRi-mediated repression of ZMO1323 confers plasmid-dependent lactate tolerance, which is consistent with previous results38. In contrast to this CRISPRi plasmid-dependent phenotype, the engineered strain S-3-4t retained comparable growth characteristics post-plasmid excision, with no significant differences in growth rate (0.153 ± 0.001 vs 0.152 ± 0.001 h−1; p å 0.1) or maximum biomass (2.617 ± 0.029 vs 2.657 ± 0.055 OD600nm; p å 0.1) relative to the plasmid-bearing strain (Fig. 6B). This phenotypic stability strongly implies that genomic adaptation, rather than transient transcriptional modulation, drives the acquired lactate tolerance in this genetic background.

Fig. 6: Fermentation experiments to probe the mechanism of enhanced lactate tolerance.

A, B Fermentation tests of mutant strains S-3-3, S-3-3t, S-3-4 and S-3-4t. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file. C Growth curves of the mutants that overexpress ZMO0008, ZMO0566, ZMO0733, ZMOp32x018 and ZMOp39x027. n = 3 biological replicates. Data are presented as mean +/− SD. Source data are provided as a Source Data file.

To test this hypothesis, we performed whole-genome sequencing of strain S-3-4, revealing 44 nucleotide mutations, 18 of which were synonymous (Supplementary Table 1). Further validation through Sanger sequencing confirmed nine non-synonymous mutations, four of which were located in non-coding regions. The remaining mutations were identified in the