Embryo metabolite analysis and implantation potential prediction using chemiluminescent microfluidic chips with dielectric wetting valves

Abstract

Analysis of embryo metabolites offers a promising non-invasive strategy for assessing preimplantation developmental potential, yet the limited performance of current sensing platforms restricts physiological insights. Here we report a capillary-driven chemiluminescence microfluidic device equipped with three electrowetting valves, enabling simultaneous, offline quantification of glucose, lactate and pyruvate from as little as 3 μL of spent blastocyst culture medium (SBCM). Using a training set (n = 61) and validation set (n = 108) of human embryo transfers, we systematically recorded morphological development and clinical pregnancy outcomes. Metabolic flux analysis revealed that embryos with higher developmental potential consumed more glucose and pyruvate while producing more lactate. Integration of metabolic and morphological data yielded a predictive model of implantation potential with an area under the curve (AUC) of 92.0%, demonstrating robust performance. This platform establishes a powerful tool for embryo selection and may inform clinical decision-making in assisted reproduction.

Introduction

Non-invasive assessment of human embryo developmental potential before implantation predominantly relies on morphological analysis¹. It has been widely utilized by IVF clinics worldwide for over 40 years^2,3. Despite its widespread use, morphological grading is inherently subjective and often lacks predictive accuracy regarding the embryo’s developmental potential. To overcome the limitations of conventional approaches, advances in time-lapse imaging^4,5, comprehensive chromosomal screening⁶, proteomics⁷, and metabolomics⁸ have significantly enhanced the assessment of embryo viability. Among these, metabolic profiling offers critical physiological insights that complement morphological evaluation, providing an alternative way to study embryo’s developmental potential.

Techniques such as mass spectrometry⁹, high-performance liquid chromatography (HPLC)⁸, and Raman spectroscopy¹⁰ have been widely used to analyze metabolic changes in embryo culture media, including metabolites such as pyruvate, lactate, and alanine. Each method offers distinct advantages. For example, mass spectrometry provides high sensitivity and comprehensive analytical capabilities; HPLC excels in precise quantification and separation; and Raman spectroscopy enables non-destructive, rapid, and real-time analysis. However, despite their advantages, these technologies typically require complex sample preparation and costly instrumentation, and offer limited capability for manipulating and analyzing microliter-scale fluids, which has largely constrained their translation into routine clinical practice.

Microfluidic technology, with its advantages of minimal sample consumption, high integration, rapid analysis, and exceptional sensitivity, offers promising solutions to major challenges in assisted reproduction and holds substantial translational potential^{11,12,13,14,15,16}. In embryo culture, microfluidic systems can create microenvironments more conducive to development by finely regulating physicochemical and biochemical conditions, thereby directly or indirectly enhancing developmental competence^16,17,18,19. In the context of metabolic biomarker analysis, various microfluidic devices have been designed to quantify embryo metabolites and their key reactants^{20,21,22,23,24}. For instance, Urbanski et al. developed an automated microfluidic system for offline, non-invasive monitoring of glucose, lactate, and pyruvate in the culture medium of individual mouse embryos¹⁶. Studies by Heo and Obeidat further integrated embryo culture and metabolic monitoring within single devices, enabling long-term or continuous assessment of embryonic metabolic activity^21,22.

Microfluidic technology holds considerable promise for embryo metabolic analysis, yet its clinical application faces multiple challenges. First, it must enable sequential detection of multiple target metabolites in extremely small sample volumes²⁵, while remaining non-invasive. Second, sensors must combine low detection limits with a wide linear range to accommodate the large concentration differences among metabolites in embryo culture media—glucose and lactate are typically at mM levels, whereas pyruvate is often at μM concentrations²⁶. Moreover, because the total amount of metabolites from a single embryo over the entire culture period is limited, sensors must be capable of detecting small relative changes against high background concentrations, such as glucose²⁷. For microfluidic sensing to achieve broader clinical adoption, it is essential to minimize reliance on complex, bulky peripheral equipment and labor-intensive chip fabrication²⁸. Currently, most microfluidic systems for embryo metabolic sensing struggle to meet all these requirements simultaneously, representing a major bottleneck to their practical implementation in clinical embryo metabolic analysis.

Capillary-driven microfluidic chips offer a compelling alternative, leveraging capillary forces for autonomous fluid transport²⁹. This eliminates the need for external pumps, enhances portability, and enables precise fluid control. Moreover, capillary microfluidic platforms seamlessly integrate diverse analytical techniques—including colorimetry³⁰, fluorescence detection³¹, electrochemical analysis³², and chemiluminescence³³—facilitating rapid, high-sensitivity, and low-power metabolite detection. We developed a capillary microchip with integrated electrowetting valves for chemiluminescence-based metabolite detection. These valves enabled precise fluid regulation within the capillary network, triggering enzymatic oxidation and chemiluminescence reactions. The chip demonstrated excellent sensitivity and a broad linear detection range for multiple metabolites, enabling accurate quantification of glucose, lactate, and pyruvate in embryo culture media and linking metabolic activity to clinical pregnancy outcomes.

Results and discussion

The mechanism of metabolite detection

A capillary-driven chemiluminescence microchip incorporating dielectric wetting valves was designed to detect embryo metabolites to establish a multi-parameter predictive model for embryo developmental potential (Fig. 1a). The microfluidic chip consisted of an Ag electrode layer, a Polydimethylsiloxane (PDMS) dielectric layer, an ITO electrode layer, and a double-sided tape channel (see Supplementary Note 1 for details). The silver electrode (thickness: 6.78 ± 0.12 μm, Supplementary Fig. S9) was screen-printed on PET to serve as the excitation electrode. PDMS was spin-coated onto the Ag electrode layer to form a hydrophobic dielectric layer (~20 μm). The etched ITO electrode (thickness: 680 ± 50 nm) functioned as the ground electrode. The double-sided tape (thickness: 170 μm) was patterned to fabricate the microfluidic channel (12 mm × 2 mm). All components were subsequently assembled, with the detailed chip design and fabrication procedure described in the Supplementary Note 2. The ITO electrode (contact angle: 78.2° ± 1.52°, Supplementary Fig. S8a) and PDMS dielectric layer (contact angle: 118.5° ± 1.04°, Supplementary Fig. S8b) together constituted a hydrophobic valve within the capillary channel. Driven by capillary force on the glass surface (contact angle: 29.2° ± 0.55, Supplementary Fig. S8c), the sample (21 μL) spontaneously filled the microchannel and halted at the dielectric valve. Upon application of a 250 V potential, the valve surface became wetted, allowing flow to resume (Fig. 1b(i, ii)).

Fig. 1: Schematic of the metabolite detection process for embryo developmental potential assessment.

a The morphology and metabolite comprehensive analysis based on a chemiluminescence dielectric valve chip. b The sample detection process. (i) The sample flowed to the enzyme reaction zone and halted before the valve, where it reacted with the preloaded enzyme. (ii) The dielectric valve was activated, allowing fluid to resume flow and triggering the chemiluminescence reaction.

The enzyme and probe were preloaded and dried in designated areas of the PDMS dielectric layer. After the analyte was introduced at the chip inlet, the mixed solution immediately flowed to the front of the valve in each of the three branch channels. GO_X, LO_X, and PO_X oxidized their respective analytes, generating H₂O₂ in the corresponding channels. Subsequently, the flow control valves for glucose, lactate, and pyruvate were activated, allowing the H₂O₂-containing solution to enter the detection area, where H₂O₂ oxidized luminol to produce chemiluminescence.

Numerical and experimental studies of the dielectric wetting valves

The mechanism of the dielectric wetting valve has been reported extensively. In the numerical simulation model, the initial contact angles were assigned as 120° for the PDMS surface, 80° for the ITO surface, and 30° for the glass surface. The experimentally measured contact angles, presented in Supplementary Fig. S8, closely align with these values. After activation of the voltage of 250 V, although the bottom ITO surface had the same contact angle of 80°, the contact angle of the PDMS layer immediately dropped from 120° (Fig. 2a(i)) to 90.36° (Fig. 2a(ii, iii)). Meanwhile, the direction of capillary force shifted from upstream to downstream within the flow, thereby enabling the halted liquid to move through the valve as depicted by the gas/liquid interface line in Fig. 2a. The numerical model was employed to evaluate the impact of PDMS thickness (15–30 µm) and voltage (50–450 V) on the contact angle (Fig. 2b). For a given PDMS layer, the contact angle was inversely proportional to the voltage. While at the same voltage, the thinner PDMS led to the lower contact angle. To allow the liquid flow through the activated valve, the contact angle should be smaller than 100°³⁴. Therefore, the required applied voltage exceeded 200 V, 250 V, 300 V, and 300 V for the PDMS thickness of 15, 20, 25, and 30 µm respectively.

Fig. 2: Study of the dielectric valves.

a Simulation of fluid flowing through a dielectric valve upon voltage application. b Simulation results of the contact angle at the three-phase contact line in the valve at various PDMS thicknesses and voltages. c Schematic diagram of dielectric wetting: (i) a droplet on an electrode with a dielectric layer without an applied potential, and (ii) a droplet with an applied potential. (iii)The measured contact angle on the dielectric layer (20 µm thick) side of the PET-Ag-PDMS at various voltages, data were mean ± SD and representative of at three independent experiments. d Photos showing the stopped dye flowed through the relevant valves upon voltage application. e The CL intensity of H₂O₂ detected independently in three channels over 30 s.

In the absence of an applied potential, a droplet of conductive fluid resting on an electrode—coated with a both hydrophobic and dielectric layer—assumes a spherical cap shape at mechanical equilibrium. Upon the application of an electric potential, charges accumulate at the droplet-insulator interface, reducing interfacial tension and, consequently, the contact angle (θ), as illustrated in Fig. 2c(i, ii). This dielectric wetting phenomenon is governed by the Young-Lippmann equation (Eq. 1-1), which quantitatively describes the relationship between applied voltage and contact angle modification³⁵.

$$\cos \theta=\cos {\theta }_{0}+\frac{{\varepsilon }_{0}{\varepsilon }_{d}}{2d{\gamma }_{{LG}}}{V}^{2}$$

(1-1)

where θ and θ₀ represent the contact angles after and before voltage application, ε₀ and ε_d are the vacuum dielectric constant and the relative dielectric constant of the dielectric layer, respectively, d represents the thickness of the dielectric layer, γ_LG represents the gas-liquid surface tension, and V represents the applied potential.

The contact angle of 5 μL of PBS on a PDMS dielectric layer was measured at various voltages. The screen-printed silver electrode on the PET substrate was 6.78 ± 0.12 μm thick (Supplementary Fig. S9), and the PDMS film was 26.93 ± 0.35 μm thick (Supplementary Fig. S10), yielding a total dielectric layer with a thickness of approximately 20 μm. When voltages of 100 V, 150 V, 200 V, 250 V, 300 V, and 350 V were applied, the measured gas-liquid-solid three-phase contact angles were 116.2° ± 1.8°, 109.9° ± 2.7°, 101.4° ± 1.9°, 88.4° ± 1.6°, 76.1° ± 1.1°, and 57.0° ± 1.8°, respectively (Supplementary Fig. S11). The contact angle was 0.7°−2.1° larger than those obtained from the simulation (Fig. 2c(iii)). The error bars represent the standard deviations of three independent measurements unless otherwise stated. This suggested the accuracy of both simulation and experiments. Therefore, the voltage of 250 V and PDMS thickness of 27 μm were employed for all the rest studies unless stated otherwise.

Following PBS dye loading, spontaneous capillary flow initiated the chip inlet, halting in the hydrophobic region of the dielectric valve (Fig. 2d(i), and Supplementary Movie 1). After applying a 250 V voltage, the dye resumed flowing into the detection area (Fig. 2d(ii–iv), and Supplementary Movie 2). Subsequently, the detection capability of the device was assessed using 1 mM H₂O₂ solution (Fig. 2e). The valve controlled the flow as expected, and the chemiluminescence (CL) intensity in the detection area reached a maximum within 0.1 s before sharply decreasing.

Study of the detection conditions

The concentrations of probes (luminol, PIP, and HRP) and enzymes (GOx, LOx, and POx) were evaluated to maximize the CL intensity. For luminol optimization, PIP (1 mM) and HRP (140 U/mL) concentrations were fixed. Increasing luminol from 0.5 to 2 mM produced no significant change in signal (Fig. 3a, P > 0.05), whereas further elevation to 8 mM reduced the intensity by 61.9%. This was also reported in previous studies and might be related to self-quenching at higher luminol concentrations^36,37. When optimizing PIP (p-iodophenol), luminol (2 mM) and HRP (140 U/mL) were held constant. As PIP concentration rose from 0.5 to 1.0 mM, the CL intensity dramatically increased to the maximum of 115.01 × 10³ counts/gate. It rapidly declined with the further increases in PIP concentration (Fig. 3a), potentially due to raised nonradiative transitions reducing luminescence³⁸. Finally, the CL intensity was directly proportional to the concentration of horseradish peroxidase (HRP), reaching a maximum once the concentration exceeded 100 U/mL (Fig. 3b), with luminol and PIP fixed at 2.0 mM and 1.0 mM, respectively. Therefore, luminol (2 mM), PIP (1 mM), and HRP (100 U/mL) were used as probes in the subsequent experiments. In the enzyme effect study, the CL intensity was proportional to the concentrations of GOx, LOx, and POx and reached maximal values at concentrations larger than 70, 20, and 25 U/mL, respectively (Fig. 3b, c). Reaction time was standardized at 5 min. Consequently, the above concentrations were employed in subsequent experiments.

Fig. 3: Study of detection conditions.

The influence of a the luminol and PIP concentrations, b the concentrations of GOx and HRP, c the concentrations of LOx and POx, and d sample incubation time on the CL intensity. e CL intensity comparison with and without sample incubation in the chip. f Variation in CL intensity when the dielectric film is not pre-embedded with enzymes and detection reagents. Data were mean ± SD and representative of at three independent experiments.

The incubation time of the sample with the enzyme has been reported to drastically impact chemiluminescence (CL) intensity. As shown in Fig. 3d, as the incubation time increased from 1 to 6 min, the chemiluminescence (CL) intensity of the samples gradually increased and stabilized. Compared to the unincubated samples, the CL intensity of incubated glucose (1 mM), lactate (1 mM), and pyruvate (1 mM) increased by 419-fold, 122-fold, and 62-fold, respectively (Fig. 3e). This suggested that the enzymatic oxidation of glucose, lactate, and pyruvate to H₂O₂ reached peak efficiency after incubation enabled by the valve.

To reduce device fabrication time and cost, the reusability of the combination of Ag electrode layer and PDMS dielectric layer was evaluated. After each detection process, the chip was thoroughly cleaned using alcohol (75%) and ultrapure water. In cases where enzymes and detection reagents were not pre-embedded in the dielectric layer, no CL signal appeared upon adding the same sample, confirming that the chip was effectively cleaned and ready for reuse (Supplementary Fig. S12). Furthermore, nearly identical detection signals (P > 0.05) were observed in the detection area when the chip was reused up to six times to detect 1 mM of glucose, lactate, and pyruvate (Fig. 3f).

Detection performance for glucose, lactate, and pyruvate

The microfluidic chip was employed to simultaneously detect glucose, lactate, and pyruvate, with optimized detection conditions (Supplementary Table S2). The real-time CL intensity reached peak values (referred to as CL intensity in the following section unless stated otherwise) within seconds after the fluid passed through the dielectric valve (Supplementary Figs. S13–S15). As shown in Fig. 4a, CL intensity increased from 59 counts/gate to 121.25 × 10³ counts/gate for glucose (1.0–2000 μM), from 53 counts/gate to 117.45 × 10³ counts/gate for lactate (0.5–2000 μM), and from 69 counts/gate to 111.49 × 10³ counts/gate for pyruvate (2.0–2000 μM). The CL intensity (y) exhibited a linear relationship with either concentration (C_g, C_l, C_p) or the logarithmic concentration (logC_g, logC_l, logC_p) of three analytes. For glucose, the linear regression models were y = 0.056C_g + 0.03 (R² = 0.974) for 1.0–100 μM and y = 91.39logC_g − 176.95 (R² = 0.993) for 100–2000 μM (Fig. 4b, c). For lactate, the models were y = 0.066C_l + 0.028 (R² = 0.983) for 0.5–100 μM and y = 86.24logC_l − 165.46 (R² = 0.996) for 100–2000 μM (Fig. 4b, c). For pyruvate, the models were y = 0.019C_p + 0.026 (R² = 0.999) for 2.0–100 μM and y = 81.47logC_p − 160.94 (R² = 0.991) for 100–2000 μM (Fig. 4b, c). The limits of detection (LOD), calculated as 3STD/S (with S as the slope of the linear regression models and STD as the standard deviation of blank samples), were 0.16, 0.15, and 0.55 μM for glucose, pyruvate, and lactate, respectively.

Fig. 4: Detection of glucose, lactate, and pyruvate.

a The CL intensity as analyte concentration varied. Linear correlation between CL intensity and b the concentration or c the logarithmic concentration of three analytes. d Simulation of H₂O₂ diffusion in three rectangular channels. Dr. Siwei Bai from the Department of Electrical and Computer Engineering, Technical University of Munich, provided assistance with the numerical simulations. e Concentration distribution of H₂O₂ in a rectangular channel at different times. f CL response for different analytes using the related reagents for glucose, lactate, and pyruvate to demonstrate the detection specificity. g Chemiluminescence intensities of glucose (i), lactate (ii), and pyruvate (iii) were measured on chips stored at 4 °C for varying durations. Data were mean ± SD and representative of at three independent experiments.

Supplementary Table S3 summarizes chemiluminescence-based methods for glucose and lactate detection, with paper- and cloth-based microfluidic chips widely explored due to their capillary-driven transport and low sample requirements. Li et al. reported a paper device with microchannels of varying lengths for sequential glucose and lactate detection using 2 μL of sample, achieving limits of detection (LODs) of 8 and 15 μM, respectively³³. Zhang et al. developed a cloth-based platform that combined gravity and capillarity for glucose detection with an LOD of 9.74 μM³⁹. Our group further advanced sensitivity with a spin-disc paper device, achieving LODs of 0.34 μM (glucose) and 0.30 μM (lactate), representing 23- and 50-fold improvements⁴⁰.The chip developed in this study achieved an additional 53% and 50% reduction in LODs and a 100% and 700% expansion in detection ranges for glucose and lactate, respectively. These improvements stem from using glass as the lower substrate in the dielectric wetting valve chip, which enhances photon collection compared with paper. Supplementary Table S4 summarizes optical microfluidic platforms for pyruvate detection, most relying on flow injection. For instance, Urbanski et al. achieved a 10 μM LOD with automated loading and mixing. In contrast, our chip reduced the pyruvate LOD 17-fold²⁰. Overall, this capillary-driven microfluidic chip enables precise flow control through dielectric valves, prolongs multistep reaction times, and significantly enhances detection performance. Its structural flexibility further supports multi-channel integration, offering strong potential for large-scale applications.

As illustrated in the design schematics, hydrogen peroxide (H₂O₂) is generated only within the enzymatic reaction zones. For H₂O₂ to influence another channel, it would need to diffuse over a sufficiently long distance during the incubation period. To address this, we conducted simulations of 1 mM H₂O₂ diffusion within the rectangular channels to monitor concentration changes across the three enzymatic reaction regions. Fig. 4d shows the concentration profiles of 1 mM H₂O₂ after 10, 120, 240, and 360 s of diffusion, and the corresponding concentration–position distributions over time are presented in Fig. 4e. The results indicate that after 6 min, H₂O₂ at the outermost edge of a channel had diffused only 2.5 mm, leaving a substantial distance from adjacent channels, thereby confirming the independence of the three measurements.

The detection specificity of the chip is shown in Fig. 4f. As expected, the target analytes—glucose, lactate, and pyruvate—and their mixtures displayed extremely high CL intensities, whereas the blank sample and non-target analytes showed almost no CL signal. These results indicate that the CL biosensor integrated into the dielectric valve chip possesses high selectivity, a property primarily attributed to the specificity of the enzymatic reactions involved.

Furthermore, the storage stability of the chip with preloaded enzymes was evaluated (Fig. 4g(i–iii)). After 15-day storage in a 4 °C fridge, the device retained 97.4%, 95.3%, and 96.3% of its initial CL intensity for detecting glucose (1 mM), lactate (1 mM), and pyruvate (1 mM), respectively. There was no significant difference in CL intensity of any analyte over the storage period (P > 0.05), making it potentially suitable for mass production and transport for off-site use.

Detection of metabolites in single blastocyst spent culture medium

The initial concentrations of glucose, lactate, and pyruvate in the fresh media were measured at 2.42, 4.21, and 0.25 mM, respectively (Supplementary Table S5). As the concentrations of glucose, lactate, and pyruvate were separately adjusted to 1.92 mM, 5.21 mM, and 0.15 mM, the experimental error ranges detected using the chip were −1.2 to 5.0%, −4.2 to 1.7%, and −4.0 to 0.6%, respectively. Notably, the relative standard deviations (RSDs) for triplicate measurements remained consistently below 3.22%. Both the experimental error ranges and RSDs demonstrate that the chip exhibits high accuracy in quantitative analysis of analytes within complex media. Subsequently, metabolites in SBCM were quantified using a microfluidic chip, with embryos cultured individually. Each assay required only 3 μL of spent culture medium. Fig. 5a presents the metabolic profiles of 61 human embryos, with positive and negative values indicating production and consumption, respectively.

Fig. 5: Detection of the embryonic metabolites.

a Metabolites consumption or production rates obtained from 61 embryos, data were mean ± SD and representative of at three independent experiments. b The glucose consumption rate, c the lactate production rate, and d the pyruvate consumption rate of embryos with different pregnancy outcomes (The pregnancy and non-pregnancy groups comprised 33 and 28 embryos, respectively. Statistical significance was determined by one-way ANOVA). The box depth represents the interquartile range (50% of the data), whiskers represent the 5 and 95% quartiles. The horizontal line within the box indicates the median glucose consumption. Single (*), double (**), and triple asterisks (***) indicate significant differences with p value of <0.05, <0.01, and <0.001, respectively, compared to gestated blastocysts.

Supplementary Table S6 summarizes major advances in embryo metabolic biomarker detection. Early studies used 96-well plates or glass slides with ultrasensitive fluorescence methods^27,41, while LC–MS, NMR, and HPLC enabled multiplex detection but lacked precise quantification^42,43,44,45. With the emergence of microfluidics, embryo culture and metabolic analysis have become more precise and efficient. Urbanski et al. reported an automated microfluidic system with laser-induced fluorescence, enabling non-invasive monitoring of glucose, lactate, and pyruvate in 0.5 μL of spent medium from single mouse embryos (0–1000 μM, LOD: 10 μM)²⁰. Heo et al. developed a computer-controlled platform capable of fluid handling and quantitative glucose detection in the 1–100 μM range²¹. Obeidat et al. further integrated multi-electrochemical sensors to measure dissolved oxygen, glucose, and lactate in microchambers, achieving real-time metabolic evaluation with LODs of 0.5, 0.4, and 1.22 μM, respectively²². Compared to existing systems, the chip developed in this study exhibits significantly enhanced sensitivity, with detection limits of 1.0 μM for glucose, 0.5 μM for lactate, and 2.0 μM for pyruvate. It also features a substantially wider dynamic detection range, with upper-to-lower limit ratios of 2000, 4000, and 1000, respectively. These characteristics align well with the intrinsic concentration differences of metabolites in embryo culture media—millimolar levels for glucose and lactate, and micromolar levels for pyruvate—enabling the platform to overcome the challenge of low metabolic flux in individual embryos. In summary, this chemiluminescence-based microfluidic platform allows for sensitive, accurate, and non-invasive detection of key embryonic metabolites, offering a powerful tool for embryo metabolomic analysis.

The metabolite profile of each embryo was further correlated to its clinical pregnancy outcome post-implantation. The mean glucose consumption for pregnant embryos was 118.92 ± 23.78 pmol/embryo/h, which was 17.67% higher than that of non-pregnant embryos (101.06 ± 26.41 pmol/embryo/h, p = 0.007, Fig. 5b). The mean lactate production for pregnant embryos was 231.16 ± 51.24 pmol/embryo/h, while non-pregnant embryos produced 34.96% less lactate (171.28 ± 46.61 pmol/embryo/h, p < 0.001, Fig. 5c). Additionally, pyruvate consumption rate of pregnant embryos was 38.21 ± 7.24 pmol/embryo/h and 11.99% higher than that of non-pregnant embryos (Fig. 5d). Despite the small difference in average values, a significant difference (34.12 ± 6.69 pmol/embryo/h, p = 0.027) was still observed. These results demonstrate that pregnant and non-pregnant embryos already display significant differences in glucose and pyruvate consumption as well as lactate production during culture, suggesting that metabolomic analysis of embryo culture media may serve as an effective tool for predicting pregnancy outcomes.

Multi-parameter predictive model for embryo selection

To evaluate the predictive value of key metabolic parameters—glucose consumption, lactate production, and pyruvate consumption—we generated receiver operating characteristic (ROC) curves (Supplementary Fig. S16). Based on the Youden index, the optimal thresholds for predicting clinical pregnancy were 101.73 pmol/embryo/h for glucose consumption, 187.39 pmol/embryo/h for lactate production, and 31.45 pmol/embryo/h for pyruvate consumption (Fig. 6a). Embryos were then assigned metabolic scores of 3, 2, 1, or 0 based on their overall metabolic status. In parallel, the Gardner scoring system was employed to evaluate embryonic development morphology. Blastocysts were categorized into four grades—Ⅰ, Ⅱ, Ⅲ, and Ⅳ(Fig. 6b). Morphological assessments were conducted on 61 embryos, and the distribution of their scores is presented in Supplementary Fig. S17.

Fig. 6: Multi-parameter prediction model for embryo selection.

a Embryo metabolic screening for pregnancy potential prediction. b Representative images of embryos across different grades used for pregnancy potential prediction. c The metabolite production/consumption rates of embryos with various morphological scores (Data were mean ± SD, embryos were classified into grades I–IV, with 6, 19, 14, and 22 embryos in each grade, respectively). d The number of embryos resulting in pregnancy, live births, and the total number of embryos were recorded for metabolism scores of 3, 2, 1, and 0. e The number of embryos resulting in pregnancy, live births, and the total number of embryos with morphology scores of Ⅰ–Ⅳ. f The percentage of various metabolic scores in clinical pregnant and non-pregnant embryos. g The percentage of distinct morphological scores in clinical pregnant and non-pregnant embryos. h The predictive outcomes for different embryo pregnancy potential predictive models. These included the morphological model (blue line), the metabolomics model (green line), and the combined model of morphology and metabolomics (M&M model, red line). Comparisons between ROC curves were conducted using DeLong’s test.

The relationship between metabolite concentrations and their morphological scores was further analyzed (Fig. 6c). The glucose consumption rates for embryos with morphology scores of Ⅰ–Ⅳ were 125.21 ± 25.26, 113.36 ± 29.99, 111.56 ± 31.36, and 103.96 ± 18.50 pmol/embryo/h, respectively. The corresponding lactate production rates were 232.61 ± 36.09, 219.88 ± 63.42, 195.88 ± 68.37, and 186.75 ± 43.72 pmol/embryo/h, while the corresponding pyruvate consumption rates were 34.55 ± 10.98, 37.80 ± 7.22, 38.13 ± 6.75, and 34.41 ± 6.25 pmol/embryo/h. Moreover, a one-way ANOVA revealed no significant differences across morphology scores for each analyte (p = 0.33, p = 0.15, p = 0.32). These findings suggest that metabolic activity is largely independent of morphological grading.

As shown in Fig. 6d, the numbers of clinical pregnant/total embryos and live births/pregnancies were 22/24 and 20/22 for metabolic score 3, 9/14 and 8/9 for score 2, 2/17 and 2/2 for score 1, and 0/6 and 0/0 for score 0. Similarly, Fig. 6e shows that for morphological grades I, the numbers of clinical pregnant/total embryos and live births/pregnancies were 5/6 and 5/5, 11/19 and 9/11 for grade II, 5/14 and 5/5 for grade III, and 12/22 and 11/12 for grade IV.

Notably, a cumulative metabolic score of 3-0 correlated with clinical pregnancy rates of 66.67, 27.27, 6.06, and 0%, respectively (Fig. 6f). Furthermore, embryos with a cumulative metabolic score above 1 were identified with a clinical pregnancy rate of 93.94% (Supplementary Table S7). Embryos with morphological scores of Ⅰ– Ⅳ achieved clinical pregnancy rates of 15.15, 33.34, 15.15, and 36.36%, respectively (Fig. 6g). Notably, 48.49% of embryos resulting in pregnancy had high scores (Ⅰ and Ⅱ). Importantly, the metabolic score outperformed morphological assessment, enabling a 45.45% higher detection rate for embryos that became pregnant.

The above analysis revealed that both metabolic and morphological scores independently act as positive predictors of pregnancy outcomes. The ROC curves were generated for the morphology model, metabolomics model, and a combined model (M & M model). The study results demonstrated that the AUC of the pregnancy assessment model based on embryonic morphology was 61.8%, while the AUC of the metabolomics model significantly increased to 89.4% (Fig. 6h). The AUC of the combined model reached 92.0%, significantly higher than that of the morphology model (p < 0.001). Metabolite profiling has been applied to predict embryo implantation potential (Supplementary Table S8). Fernandes et al. employed ESI–MS to measure lactate and pyruvate, achieving predictive AUCs of 72.7 and 83.0%, respectively⁴⁶. Liang et al. used LC–MS to analyze five metabolites, yielding an AUC of 88%⁹, while Huo et al. identified amino acid markers by HPLC, raising the AUC to 90.4%⁸. Although informative, these approaches rely on complex preprocessing and costly instrumentation. By contrast, our microfluidic platform offers a simple, rapid, and highly sensitive assay capable of multiplexed metabolite detection, providing a more practical tool for embryo assessment.

Clinical validation of the embryo selection prediction model

Metabolites in the validation set of 108 human embryo culture media were analyzed using our self-developed dielectric valve-based detection chip. As shown in Fig. 7a, the metabolic profiles of individually cultured embryos were obtained, with positive values indicating metabolite production and negative values indicating consumption. In the validation set data, the average consumption rates of glucose and pyruvate by the embryos were 126.87 ± 40.46 pmol/embryo/h and 33.74 ± 12.81 pmol/embryo/h, respectively, while the lactate production rate was 259.79 ± 129.04 pmol/embryo/h (Supplementary Table S9). Gardner et al. revealed that human embryos consume glucose and pyruvate at rates of 131.0 ± 11.4 pmol/embryo/h⁴¹ and 37.7 ± 1.8 pmol/embryo/h⁴⁷, respectively—values that align closely with those observed in this study.

Fig. 7: Clinical validation of the multi-parameter prediction model for embryo selection.

a Metabolites consumption or production rates obtained from 108 embryos, data were mean ± SD and representative of at three independent experiments. b The glucose consumption rate, c the lactate production rate, and d the pyruvate consumption rate of embryos with different pregnancy outcomes in the validation set (The pregnancy and non-pregnancy groups comprised 69 and 39 embryos, respectively. Statistical significance was determined by one-way ANOVA). The box depth represents the interquartile range (50% of the data), whiskers represent the 5 and 95% quartiles. The horizontal line within the box indicates the median glucose consumption. Single (*) and triple asterisks (***) indicate significant differences with p value of <0.05 and <0.001, respectively, compared to gestated blastocysts. e The metabolite production/consumption rates of embryos with various morphological scores (Data were mean ± SD, embryos were classified into grades I–IV, with 23, 53, 13, and 19 embryos in each grade, respectively). f The number of embryos resulting in pregnancy, live births, and the total number of embryos were recorded for metabolism scores of 3, 2, 1, and 0. g The number of embryos resulting in pregnancy, live births, and the total number of embryos with morphology scores of Ⅰ–Ⅳ. h The percentage of various metabolic scores in clinical pregnant and non-pregnant embryos. i The percentage of distinct morphological scores in clinical pregnant and non-pregnant embryos. j The predictive outcomes for different embryo pregnancy potential predictive models in the validation set. These included the morphological model (yellow line), the metabolomics model (blue line), and the combined model of morphology and metabolomics (M&M model, red line). Comparisons between ROC curves were conducted using DeLong’s test.

In the validation set, pregnant embryos exhibited a mean glucose consumption rate of 143.57 ± 29.61 pmol/embryo/h, which was 47.53% higher than that of non-pregnant embryos (p < 0.001; Fig. 7b), demonstrating strong statistical significance. This trend was consistent with the training set, where a highly significant difference was also observed (p = 0.007). For lactate production, pregnant embryos in the validation set produced 303.17 ± 120.06 pmol/embryo/h, representing a 65.63% increase over non-pregnant embryos (p < 0.001; Fig. 7c). Similarly, in the training set, lactate production was significantly lower in non-pregnant embryos, with a 35.08% reduction relative to pregnant embryos (p < 0.001), indicating that decreased lactate production consistently characterized non-pregnant embryos across both datasets. In addition, pyruvate consumption in the validation set was 35.79 ± 13.22 pmol/embryo/h for pregnant embryos, 19.3% higher than in non-pregnant embryos (p = 0.025; Fig. 7d), a statistically significant difference that was in line with the training set (p = 0.027). Although embryos resulting in clinical pregnancy exhibited differences in certain metabolic fluxes compared with non-pregnancy embryos, the distributions of individual metabolic parameters still showed considerable overlap when analyzed in isolation, which may lead to misclassification of implantation potential. This observation underscores that flux changes of a single metabolite are insufficient to reliably discriminate clinical pregnancy outcomes. Instead, a combined analysis of multiple key metabolic fluxes in the culture medium is required to achieve a more robust prediction of embryo implantation potential.

We next analyzed the relationship between metabolite concentrations and embryo morphology scores (Fig. 7e). The distribution of morphological scores was presented in Supplementary Fig. S18. The glucose consumption rates for embryos with morphology scores I–IV were 138.32 ± 31.23, 126.78 ± 43.02, 140.52 ± 31.52, and 103.91 ± 40.86 pmol per embryo per hour, respectively. No significant differences were observed among embryos with scores I–III (p = 0.27). The corresponding lactate production rates were 279.05 ± 113.66, 252.27 ± 131.74, 268.31 ± 153.59, and 251.61 ± 128.74 pmol per embryo per hour, while pyruvate consumption rates were 35.28 ± 12.09, 33.27 ± 13.67, 34.58 ± 12.02, and 32.59 ± 12.47 pmol per embryo per hour, respectively. No significant differences were detected in lactate or pyruvate metabolism among the morphology groups (p = 0.85 and p = 0.89). These findings were consistent with the training set and further support that embryo metabolic activity is largely independent of morphology scores.

We applied the metabolomics-based scoring system to evaluate implantation potential in the validation cohort (Supplementary Fig. S19). As shown in Fig. 7f, the numbers of clinical pregnant/total embryos and live births/pregnancies were 33/40 and 29/33 for metabolic score 3, 30/33 and 24/30 for score 2, 6/27 and 3/6 for score 1, and 0/8 and 0/0 for score 0. Similarly, Fig. 7g shows that for morphological grades I, the numbers of clinical pregnant/total embryos and live births/pregnancies were 18/23 and 13/18, 34/53 and 27/34 for grade II, 10/13 and 9/10 for grade III, and 7/19 and 7/7 for grade IV. Notably, the cumulative clinical pregnancy rates for embryos with metabolic scores of 3, 2, 1, and 0 were 47.83, 43.48, 8.69, and 0%, respectively (Fig. 7h). Embryos with scores >1 achieved a cumulative pregnancy rate of 91.39% (Supplementary Table S10), closely matching the rate observed in the training set (93.94%). By contrast, clinical pregnancy rates for morphology grades I–IV were 26.09, 49.28, 14.49, and 10.14%, respectively (Fig. 7i). Notably, 75.37% of pregnant embryos were derived from high-grade (I or II), a proportion markedly higher than that observed in the training cohort (48.49%). The data in Supplementary Table S11 indicated that, in both the training and validation cohorts, the effects of potential confounding factors—including maternal age, body mass index (BMI), duration of infertility, endometrial thickness, and oocyte number—could be excluded in the analysis of clinical pregnancy outcomes. These findings indicate that the metabolic scoring system provides greater accuracy and reliability than morphology alone in predicting implantation potential.

Receiver operating characteristic (ROC) curve analysis further assessed model performance. As shown in Fig. 7j, in the validation set, the AUC of the morphology-based model was 68.2%, which was significantly improved by the metabolomics model (86.5%). The combined morphology–metabolomics (M&M) model achieved an AUC of 87.3%, markedly higher than the morphology model (DeLong test, p < 0.001), consistent with the training set results.

The aforementioned research demonstrates that a metabolomic evaluation model can be established to predict embryonic implantation potential by leveraging the metabolic differences between pregnant and non-pregnant embryos. This approach is fully compatible with existing embryo morphological assessment and time-lapse imaging techniques, thereby enabling a multidimensional screening strategy for selecting embryos with high developmental competence. As the method only utilizes spent culture medium without altering the standard in vitro fertilization (IVF) protocol, it holds significant promise as an auxiliary decision-making tool for identifying embryos with the best developmental potential.

Limitations of the study

The primary limitations of this study include the limited sample size, and larger cohort studies will be required to improve model robustness. Culture medium collection followed routine clinical workflows and may therefore not fully capture metabolic dynamics at other developmental stages, such as the eight-cell or blastocyst stages. In addition, the experimental design did not distinguish morphologically similar embryos derived from the same patient, which represents a further limitation. Future work could involve large-scale, prospective multicentre cohort studies integrated within standard clinical workflows to enable metabolic analyses of morphologically similar embryos and embryos across developmental stages. Moreover, integrating metabolomic assessment with emerging non-invasive embryo evaluation approaches, including time-lapse imaging, proteomic profiling, and microRNA analysis, may help establish a multidimensional assessment framework, providing a powerful tool for embryo selection.

Methods

Study participants and samples

This study was approved by the Research Ethics Committee of Shenzhen Zhongshan Obstetrics and Gynecology Hospital (formerly known as Shenzhen Zhongshan Urology Hospital); approval number: SZZSECHU-20180021. The participants were infertility patients who underwent in vitro fertilization (IVF) or intracytoplasmic sperm injection (ICSI) followed by vitrified single-blastocyst transfer at our center between January 2021 and July 2023. In total, 169 samples of spent blastocyst culture medium (SBCM) were collected and allocated to a training set (n = 61) and a validation set (n = 108). The training set was constructed retrospectively with the specific aim of building a predictive model. To ensure a balanced distribution for model development, we selectively included 34 samples from embryos resulting in clinical pregnancy and 27 samples from embryos that did not, with collection dates spanning non-consecutive periods across the study timeframe. In contrast, the validation set was designed prospectively to assess the effectiveness of the model in practical applications. It comprised all SBCM samples that met the inclusion criteria from a defined period (June to July 2023).

Embryo Culture and Sample Collection

Oocytes retrieved by ultrasound-guided aspiration underwent fertilization by either IVF or ICSI. Fertilization was assessed 17 ± 1 h after insemination, and the presence of two pronuclei was taken as evidence of normal fertilization. Zygotes were cultured in Quinn’s Advantage Cleavage Medium (SAGE BioPharma, USA) supplemented with 10% (v/v) serum protein substitute until day 3. Embryo culture was maintained in a Labtech C200 incubator under stable conditions (6% CO₂, 5% O₂, 37 °C, 95% humidity), which were applied consistently. Cleaving embryos were transferred to Quinn’s Advantage Blastocyst Medium, supplemented with 10% (v/v) serum protein substitute, and cultured individually in 25 μL droplets until blastocyst formation. Following vitrification, spent blastocyst culture medium (SBCM) was collected and stored at −80 °C for subsequent analysis.

Embryo transfer and confirmation of clinical pregnancy

All patients underwent vitrified–warmed single blastocyst transfer. Endometrial preparation followed either natural or artificial protocols. Serum hCG was measured 11 days after transfer, with concentrations >5 IU/L indicating biochemical pregnancy. Clinical pregnancy was confirmed by the presence of an intrauterine gestational sac on transvaginal ultrasound 30 days after transfer.

Characterization of the dielectric wetting valves

The dielectric wetting mechanism within the capillary channel was studied numerically. The level set method was employed to accurately capture the dynamic behavior of the gas-liquid interface. Under the influence of an electric field, the contact angle of the gas-liquid interface was corrected using the Young-Lippmann equation, allowing the precise description of the electric field’s modulation of interface behavior. By systematically simulating the variation in contact angle under different voltage and dielectric layer thickness conditions, the study further determined the optimal parameter configuration for the recovery of fluid flow. Detailed simulations are provided in the Supplementary Note 3. The mechanism and performance of the valves were investigated experimentally. Contact angle measurements were performed on a silver electrode–PDMS dielectric layer using a goniometer (Drop Analyzer, Dalue Technologies, Zhuhai, China), with 5 μL deionized water under applied voltages of 0 or 150–400 V. Valve-controlled fluid manipulation was validated using a diluted dye solution under 250 V, with the process recorded by camera. Sequential chemiluminescence detection was further assessed using 1 mM H₂O₂ as the analyte. A 5 μL mixture of 2.0 mM luminol, 1.0 mM PIP, and 140 U/mL HRP was preloaded into the detection zone, after which sample introduction and sequential valve actuation generated chemiluminescence signals recorded by the instrument.

Study of the detection conditions

The chemiluminescence (CL) intensity of the detection using the microfluidic chip is influenced by the concentrations of reagents and the enzyme incubation time. To optimize luminescence intensity, the CL intensity of detecting 1 mM H₂O₂ using probes with various concentrations (luminol: 0.5–8 mM, PIP: 0–4 mM, and HRP: 20–140 U/mL). Additionally, the impact of varying concentrations of GO_X (10–120 U/mL), LO_X (5–30 U/mL), or PO_X (10–35 U/mL) was evaluated by detecting 1.0 mM glucose, lactate, or pyruvate, respectively. The dielectric valve’s ability to stop the fluid flow was then used to explore the effects of incubation time (0–6 min) on the detection signal. The CL intensity optimization process is detailed in the Supplementary Note 4.

The reusability of the dielectric valve chip was exploited to save detection costs. After each detection, the PDMS dielectric layer was detached from the double-sided tape and cleaned by alcohol (75%) and ultrapure water before being reassembled to a chemiluminescence microfluidic chip. After embedding the enzyme and detection substrates, the chip was reused to detect analytes (1 mM glucose, lactate, and pyruvate). The chip’s performance during six reusing cycles was assessed by detecting analytes on the film without pre-embedded reagents.

Detection performance for glucose, lactate, and pyruvate

The performance of the chip in detecting a mixture of glucose, lactate, and pyruvate diluted in PBS was evaluated. Five microliters of each enzyme solution (70 U/mL GO_X, 20 U/mL LO_X, 25 U/mL PO_X) and 5 μL of substrate solution containing luminol (2.0 mM), PIP (1.0 mM), and HRP (100 U/mL) were respectively pre-embedded to in the enzyme reaction area before the valve and the detection area behind the valve on the PDMS dielectric layer. After assembly of the whole device, 21 μL of the analyte mixture was added to the chip inlet. The liquid mixture immediately flowed to the front of the valves in the three branch channels. After incubating the oxidation reaction for 4, 5, and 6 min, the valves controlling the flow of glucose, lactate, and pyruvate were respectively activated at the voltage of 250 V. Meanwhile, the CL signal was monitored using a photomultiplier tube (PMT, Hamamatsu H7421) with a gate time of 10 ms to record the signal variation over 60 s after each valve activation.

The selectivity of the dielectric valve chip for glucose (1.0 mM), lactate (1.0 mM), and pyruvate (1.0 mM) in the presence of common interferents was evaluated. Additionally, the prepared detection chip was stored at 4 °C, and samples containing 1.0 mM glucose, lactate, and pyruvate were tested on days 1, 3, 5, 8, and 15 to assess chemiluminescence (CL) response and stability. Furthermore, the generation and diffusion of 1 mM H₂O₂ within a rectangular microchannel were simulated to evaluate potential cross-interference among the three reaction zones. Further details are provided in the Supplementary Notes 5 and 6.

Detection of metabolites in single blastocyst spent culture medium

The chip was first employed to measure the baseline concentrations of glucose, lactate, and pyruvate in a fresh embryo culture medium (Quinn’s Advantage Blastocyst Medium). Briefly, 3 μL medium was mixed with 18 μL PBS buffer (pH 7.4) before being introduced into the chip inlet. Then the CL signal for three analytes was recorded in the detection area. Afterward, the medium was either added with 1.0 mM lactate or diluted to reduce the glucose and pyruvate concentrations by 0.5 mM and 0.1 mM, respectively. This step aimed to simulate the embryo-conditioned medium with lactate production and glucose/pyruvate consumption. The samples were subsequently tested to determine the concentration of the three analytes separately. Finally, the chip was applied to analyze spent culture medium from individual blastocysts in the training cohort (n = 61) to validate its performance.

Multi-parameter predictive model for embryo selection

The embryo metabolic scoring system was established using glucose, lactate, and pyruvate as key indicators of metabolic activity. We quantified glucose and pyruvate consumption rates together with lactate production rates during embryo culture, and correlated these with clinical pregnancy outcomes. Based on optimized metabolic thresholds predictive of pregnancy, embryos were classified such that values above the threshold received a score of 1 and those below received 0. The cumulative scores across the three metabolites stratified embryos into four metabolic grades: 3, 2, 1, or 0. Blastocysts were graded according to the Gardner system, evaluating expansion (scale 1–6), inner cell mass quality (A–C), and trophectoderm cohesiveness (A–C). Based on inner cell mass and trophectoderm scores, blastocysts were classified into four categories: grade Ⅰ (AA), grade Ⅱ (AB, BA), grade Ⅲ (BB), and grade Ⅳ (AC, CA, BC, CB). Finally, this study integrated the embryo metabolic score, clinical morphological score, and transfer experiment results to construct a metabolic model, a morphological model, and a combined metabolism-morphology model to evaluate embryo implantation potential. For further details, please refer to the Supplementary Note 7.

Clinical validation of the embryo selection prediction model

Metabolites in the culture medium of the 108 validation cohort embryos were analyzed using our dielectric valve-based chip. We calculated and benchmarked the consumption/production rates of key metabolites (glucose, pyruvate, lactate) against existing literature. Each embryo then received integrated metabolic and morphological scores via our metabolomic model and the Gardner system, respectively, with results compared to the training set. The predictive power for clinical pregnancy was systematically compared between morphological, metabolomic, and combined models using ROC curve analysis on both datasets, providing a comprehensive evaluation of model accuracy and generalizability.

Reporting summary

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