Introduction

Tricarboxylic acid (TCA) cycle, as the main energy metabolism pathway of the cell, is prevalent in almost all living aerobic organisms [1]. Occurring completely in mitochondria, the series of chemical reactions generate energy via the oxidation of acetylcoenzyme A (CoA) derived from carbohydrates, proteins and fatty acids [2]. Metabolites in the TCA cycle showed different degrees of changes in metabolism characteristics of multitudinous physiological diseases, especially in cancer [3]. Even in conditions of excessive nutrition, such metabolism pathways are specifically rewired during process of carcinogenesis for meeting uncontrolled cancer cell proliferation [4]. In TCA cycle of cancer cell, over-expressed malate dehydrogenase and under-expressed aconitase cause the deficiency of the intermediate metabolite α-ketoglutarate acid, which is a critical co-substrate factor can inhibit HIF-1α producing [5]. Additionally, over-expressed succinate has a negative feedback effect and efficiently inhibits prolyl hydroxylases because a dioxygenase can catalyze the hydroxylation process of proline residues in Hif-1α, leading to its degradation. This helps to stabilize Hif-1α. What more, over-expressed succinate has other consequences: protein succinylation and activation of the pro-inflammatory receptor GPR91, both of which are cancer-related targets [6]. Histopathological examination is employed as the main method for clinical cancer diagnosis, but limits by low sensitivity and invasiveness [7]. One TCA cycle metabolism targeted detection approach may achieve the noninvasive, real time, rapid cancer cell discrimination.

Traditional sensor employs the “lock-key” strategy for highly selective and specific receptor analysis, showing obvious advantages in the controlled interference and background [8]. However, in complicated analyte detection or unknown structure analysis, traditional sensors often suffer from some interference to limit practical applications [9]. As complement to traditional chemical sensor, sensor array is inspired by enormous cross-reactive receptors inside the mammalian olfactory system, exhibiting great potential in discriminating feature characteristic information of similar complex mixtures [10]. Receptors of sensor array, with low specifical and selective characteristics, exhibit different affinities to combine with analytes. This “weak lock-key” recognition mechanism allows all the response signals extraction for further data mining [11, 12]. In recent year, sensor array, employing standard data analysis methods and excellent chemical discrimination capability, is highlighted as one mainstream study for complicated multilevel system analysis, especially in bioanalysis [13,14,15,16]. Ma et al., built a luminescence sensor array including 5 double positive charges-based emission luminogens, which show good aggregation-induced emission in the mitochondria, achieving identification of normal cells and cancer cells [17]. Mohammad et al., use nanoparticles, organic dyes and metalloporphyrins to develop a paper-based colorimetric sensor array for detecting COVID-19 disease by tracking metabolites of urine samples with good accuracy and sensitivity [18]. Si et al., structured one target miRNA-responsive DNA hydrogel-based SERS sensor array and successfully achieved the quantification and recognition of multiple cancer related miRNAs with the detection limit of 0.11 nM [19]. Currently, the signal sources of sensor array are provided by the vast majority of multiple array points. Achieving more signals with fewer sensor array points can simplify detection operations and increase the value of practical applications.

Metal–organic framework (MOF) with ordered chemical and porous crystalline structure is formed by coordination bonds between organic-ligands and metal-ions [20]. Interesting properties, such as diverse 3D geometrical skeleton and high surface areas, endow MOFs widely applications in sensors [21]. Profiting from manifold coordination modes and high coordination of lanthanide ions, lanthanide-based metal–organic framework (Ln-MOF) exhibits special luminescence capability such as visible emission light, large stokes displacement, and pure color [22]. Peng et al., composed a Tb-functionalized Zr-MOF for highly selective and sensitive for thiabendazole detection in oranges with rapid response, low detection limit and good anti-interference [23]. Zhou et al., developed a 3 lanthanide (Ce, Tb, Eu) based sensor array for rapid discrimination of multiple metal ions according to different binding affinities toward metal ions [24]. Present Ln-MOFs were reported as single probe or the member of sensor array points. Little of existing studies reported multiple lanthanide ions-doped MOFs in building multiple luminescence signals sensor array.

Herein, terbium (Tb) and europium (Eu) were employed to constructure the double lanthanide-based MOF (Tb/Eu MOF) with multiple luminescence emission signals. Through pattern recognition and labeling experiments, the quantitative and qualitative detection capabilities of the sensor array have been proved. This novel single-particle sensor array could realize rapid accurate detection of TCA cycle metabolites and discrimination of cancer cells via significant host–guest interaction, showing a great potential for clinical diagnosis of cancer.

Experimental section

Materials, regents and cells

1,3,5-tris(4-carboxyphenyl) benzene, terbium (III) nitrate pentahydrate, europium (III) nitrate hexahydrate, oxalacetic acid, glycerol phosphate disodium salt, α-ketoglutaric acid, L-ribose, phosphoenolpyruvic acid monopotassiummalic acid salt, adenosine 5-triphosphate disodium salt, fumaric acid, L-tyrosine, L-methionine, L-cysteine, L-threonine, L-aspartic acid, L-valine, L-Isoleucine, L-histidine were purchased from Adamas Inc (Shanghai, China). N,N-Dimethylformamide, L-arginine, L-glutamic acid and citric acid were bought from Aladdin Industrial Inc (Shanghai, China). Human normal cells (HAMSC, EC), breast cancer cells (MCF-7), lung cancer cells (A549), cervical cancer cells (Hela), leukemia cells (K562) and liver cancer cells (HCCLM3) were obtained from ATCC cell bank.

Transmission electron microscopy from Transmission electron microscopy (TEM), JEOL Ltd., Japan was employed in morphology characterization. FL970 spectrophotometer (Techcomp Ltd, China) was used for signal measurements. One 3K15 High frequency centrifuge (Sigma, Germany) was applied for centrifugation step. All aqueous solutions were prepared with deionized water (resistivity, 18.2 MΩ cm-1) from Millipore Direct-Q Water system (Mosheim, France).

Synthesis of Tb/Eu MOF

3 kinds of Tb/Eu MOF were synthesized depending on a previous report via adjusting molar ratios of terbium (Tb) and europium (Eu) (99.5:0.5, 99:1, 98:2) [25]. Specifically, terbium (III) nitrate pentahydrate and europium (III) nitrate hexahydrate were mixed with ratios of 99.5:0.5. Then, the mixture (0.15 mmol totally) and 1,3,5-tris(4-carboxyphenyl) benzene (0.15 mmol) were dissolved in 30 mL N,N-Dimethylformamide. The obtained solution was transferred into 50 mL Teflon equipped stainless steel autoclave and heated under 150 ℃ for 3 h to produce Tb/Eu MOF 1. The Tb/Eu MOF was washed and centrifuged 3 times with N,N-dimethylformamide and ethanol, and the volume was made constant to 10 mL by passing through N,N-dimethylformamide. Tb/Eu MOF 2 and Tb/Eu MOF 3 were synthesized by similarly pathway.

Fabrication of sensor array and data analysis

The units of the excitation and emission wavelengths are nanometers (nm). Under a 300 nm light excitation wavelength, Tb/Eu MOF shown 7 characteristic peaks, which were 360 nm, 490 nm, 546 nm, 596 nm, 617 nm, 652 nm and 700 nm, respectively. The maximum excitation peak at 652 nm is discarded for subsequent analysis because of the small value. In the presence of guest molecules, a significant host–guest interaction happens broke and reconstruct the energy transfer pathway between 1,3,5-tris(4-carboxyphenyl) benzene, terbium and europium in Tb/Eu MOF showing significant signal intensity change in each emission. Maximum emission peaks were extracted as feature signals. In the equation, the initial intensity (Ix) and final intensity (Fx) are the luminescence intensities after and before the addition of the analyte. Data before reaction were recorded the as I360, I490, I546, I596, I617, I700 while data after reaction were recorded as F360, F490, F546, F596, F617, F700. In the data analysis, such as pattern recognition, Eq. 1 was applied to normalize the data to reduce the relative error. In optimal experience and quantification detection, changes of Euclidean distance were defined with Eq. 2.

$$Mx=(Fx-Ix)/Ix$$
(1)
$$Euclidean\;distance=\sqrt{M_{360}^2+M_{490}^2+\dots+M_{700}^2}$$
(2)

SPSS 22, a multivariate analysis software, was used to deep analysis. In this work principal components analysis (PCA), hierarchical clustering analysis (HCA) and linear discriminant analysis (LDA) were employed as the main pattern recognition methods. PCA is usually used to dig deeper into multidimensional data structures for reducing dimensionality via data characteristics [26]. Combining with factor analysis, PCA could calculate the information contribution of each sensor array point [27, 28]. LDA measures the difference of samples by means of supervised learning, which can accurately classify samples [29]. The “leave one-out” cross validation (LOO), as a utilitarian function, brings the individual samples back to the LDA model one by one for verification, so as to obtain the accuracy of LDA model. Radial basis function neural network (RBFN) is a special feed forward neural-network, can predict the trend of the model based on the existing data.

Detection of tricarboxylic acid cycle metabolite

18 metabolic markers, which basic physical and chemical information were shown in Table 1, were dissolved in 50% ethanol and water solution. In the specific process, 50 μL provided Tb/Eu MOF scatters in 150 μL 50% ethanol solution mix with metabolic marker and finally determined to 250 μL. 6 channel emission peaks (360 nm, 490 nm, 546 nm, 596 nm, 617 nm and 700 nm) were captured to performed normalized analysis. Mx was used for further pattern recognition and calculation. Euclidean distance changes were employed for optimizing the reaction time of Kea and Tb/Eu MOF. It also was used for quantitative detection of Val.

Table 1 Information of 18 TCA cycle metabolites
Full size table

Discrimination of cancer cell

2 human normal cells (HAMSC, EC) and 5 cancer cells (MCF-7, A549, Hela, K562, HCCLM3) were tested via Tb/Eu MOF. The medium was washed with normal saline to obtain a cell suspension, and then the cells were diluted with purified water to 10,000 cells per milliliter to further break up and extract the supernatant.50 μL provided cell lysing reagent reacted with 50 μL Tb/Eu MOF in 150 μL 50% ethanol solution. The luminescence intensity was measurement when the reaction comes to stabile.

Result and discussion

Fabrication of sensor array

Herein, the Tb/Eu MOF as the sources of optical signal synthesized by a simple and green way. In the presence of new molecules, the solvent molecules in the MOF framework are easily removed and replaced due to the large coordination numbers of lanthanides (Tb and Eu), showing significant host–guest interaction [30]. Specifically, the original energy-connecting loop path between 1,3,5-tris(4-carboxyphenyl) benzene, Tb and Eu is broken. A new energy transfer pathway between guest molecule, 1,3,5-tris(4-carboxyphenyl) benzene, Tb and Eu is reconstructed, showing significant luminescence intensity change in each emission. Scheme 1 shows the entire material synthesis steps and the general reaction mechanism.

Scheme 1
scheme 1

Synthesis of Tb/Eu MOF and reaction mechanism

Full size image

The luminescence spectrums of 3 Tb/Eu MOF with different formulations are shown in Fig. 1A. Under excitation of 300 nm, 3 Tb/Eu MOF shows characteristic peaks with different heights placed respectively at 360 nm, 490 nm, 546 nm, 596 nm, 617 nm and 700 nm. Characteristic peaks are generated because of energy transfer between 1,3,5-tris(4-carboxyphenyl) benzene, Tb and Eu. After being excited, 1,3,5-tris(4-carboxyphenyl) benzene emits a signal at 360 nm. Notably, in the 3 MOF, the signal at 360 nm was lower than other peaks, indicating that 1,3,5-tris(4-carboxyphenyl) benzene acts as an energy donor to transfer energy to other elements [31]. This evidence also proves the successful synthesis of Tb/Eu MOF. Significantly, due to the lowest energy level of 5D0, Eu is allowed to accept the energy from both 1,3,5-tris(4-carboxyphenyl) benzene and Tb in priority. The emission peaks at 490 nm, 617 nm and 700 nm were produced by the 5D4 → 7F6, 5D4 → 7F3 and 5D4 → 7F0 electronic transition of Tb, respectively [32]. The emission peak at 596 nm was produced by the5D0 → 7F1 electronic transition of Eu [33]. 5D4 → 7F5 electronic transition of Tb and 5D4 → 7F5 electronic transition of Eu were mutually influenced to product final comprehensive emission peak at 546 nm. The luminescence spectra of Tb/Eu MOF 2 are consistent with those reported in the literature, exhibiting a more balanced luminescence profile. This indicates that the uniform contribution of each sensor array point does not interfere with the array identification ability by overweighting individual array points. Tb/Eu MOF 2, selecting as the sensor array for subsequent experiments, will not cause the excessive weight of individual point to interfere with the recognition ability.

Fig. 1
figure 1

The luminescence spectra of Tb/Eu MOF1, 2 and 3 (A), The luminescence physical map under excited at 365 nm light (B). TEM image (C), graph of particle size distribution (D) and EDS spectrum of Tb/Eu MOF 2 (E) were exhibited. Fluorescence attenuation fitting delay at 544 nm (F) and 612 nm (G)

Full size image

Figure 1B exhibit the luminescence physical map under excited at 365 nm light. The structure and morphology of as-prepared Tb/Eu MOF 2 were characterized by TEM, shown as Fig. 1C. Tb/Eu MOF 2 are spherical in shape and well dispersed with an average diameter of 0.18 μm, which similar to previous reports. Under visible light, the solution shows milk white color. Figure 1D showed the size distribution frequency diagram of the prepared material. Result of energy dispersive spectroscopy (EDS) further demonstrated the successful synthesis of Tb/Eu MOF (Fig. 1E). Figure 1F and G showed the luminescence lifetime curves of the two emission peaks of MOFs after being excited by 300 nm pulsed laser. All curves showed second-order exponential decay, and the calculated lifetime values of the two emission peaks are 33.74 μs (544 nm) and 816.54 μs (612 nm), respectively. The above luminescence lifetime at the μs level corresponds to Zhang et al. 's finding that "The luminescence lifetimes (τ) of lanthanide and transition metal complexes are at micro/millisecond (μs/ms) and hundreds/thousands nanosecond (ns) levels, respectively [34]". For the TCAs in our study (e.g., ATP), their fluorescence lifetimes are extremely short, typically at the ns level. With the extension of time after excitation, the fluorescence and scattered light of targets in the detection system are weakened one by one until the end, while the long-luminous life of MOFs still maintains the optical signal. The MOFs do have the potential for background-free detection of the analyte, which provides the theoretical basis for the successful detection and discrimination of TCA by this sensor array.

Feasibility analysis

Gpd, Glu, Phc and Met were employed to evaluate the feasibility of detecting metabolic markers via Tb/Eu MOF 2 (Fig. 2). The non-radiative energy from 1,3,5-tris(4-carboxyphenyl) benzene could be transferred to the lanthanide metal via the "antenna" effect, showing significant lanthanide element emission. When Met added, the enhancement of π − π stacking could suppress the electronic transition and the energy transfer from 1,3,5-tris(4-carboxyphenyl) benzene to Tb and Eu, exhibiting that all emission peaks were reduced. When Gpd, Glu and Phc added, the original energy transfers were broken. Ligand energy cannot be transferred to lanthanide elements, showing the emission of 1,3,5-tris(4-carboxyphenyl) benzene enhanced significantly while the characteristic peaks of Tb and Eu decreased obviously. At the same time, we found that there is no exact correspondence between the increase of the peak at 360 nm and the decrease of other peaks, this may be related to the energy conversion ratio and efficiency. Moreover, in Glu and Gpd test, the sharp luminescence intensity increases at 360 nm are obviously different while the decrease of other peaks is very similar. That may be due to that the guest molecules are interspersed in the particle holes, filling the space gap and promoting the emission of the ligand. The disparate abilities of guest molecules to intercalate and bind with elements of MOF, showing diverse effects on characteristic peaks at different positions, are the detection basis of Tb/Eu MOF sensor array.

Fig. 2
figure 2

Feasibility analysis

Full size image

Condition optimization

Reaction time of Kea and Tb/Eu MOF were optimized according to the change of Euclidean distance, which is a common distance definition between points in multi-dimensional space (Figure S1). The error bars are described by the 95% confidence interval. Finally, 10 min was chosen as the reaction time.

Detection of tricarboxylic acid cycle metabolites

TCA metabolites in cancer cells show marked difference with that in normal cells to meet uncontrolled cancer cell proliferation. Herein, 18 small molecule organics were chosen as representative TCA cycle metabolites. The Ix and Fx of 18 representative metabolites under 4 concentrations (50 μM, 100 μM, 200 μM, 300 μM) were captured at 6 emission peaks of MOF. Via data normalization processing (Eq. 1), obtained Mx was utilized into pattern recognition. Significantly, these 4 concentrations include the clinical detection criteria for most TCA metabolites. To distinguish and identify analytes qualitatively is meaningful. For discriminating 18 representative metabolites, LDA was applicated as a linear supervised learning method aims to construct and optimize orthogonal dimensions to mine the differences (Fig. 3). In the discriminant scatterplot formed by the first 4 discriminant functions, 18 metabolites can be accurately classified by the array at 4 concentrations with high LOO accuracy.

Fig. 3
figure 3

LDA results of 18 metabolites at 4 concentrations (A, B, C, D). weight analysis for each sensor array points in 4 concentrations (E)

Full size image

In Fig. 3E, although 18 metabolites can be distinguished successfully at low concentrations (50 μM), weight coefficients of characteristic peak 1 and 6 are very uneven shows that the impact of peak 1 on the sensor array is much higher than that of peak 2, which is unfavorable to the construction of truth. As the concentration increases, the weight of the characteristic peak tends to be stable, which proves that the sensor array can adapt to the detection of multiple concentrations.

Take the data set when the concentration is 300 μM as an example to analyze LDA. Figure 3C shows first 4-dimensional scatter plot for discrimination, in which discrimination function 1 and 2 could provide 76.6% and 22.2% valid information respectively. Though the first 2 discrimination functions could prove 98.8% valid information for classification, there is still 1 sample of Iso is misclassified into Thr. Via intra-group correlation analysis, discrimination function 1 is more sensitive to luminescence change at 360 nm, and less correlated to other characteristic peaks. Function 2 is more sensitive to luminescence change at 490 nm, 546 nm, 596 nm, 617 nm, and less sensitive to peak changes at 360 nm, 700 nm. The reason for the misclassification may be due to the information neglect at 700 nm. In this case, discrimination function 3 and 4 act as supplementary information of the first 2 functions can provide key support for the correct discrimination. The specific expressions of the first 4 discriminant functions are as following.

$$\begin{array}{l}{\mathrm Y}_1=1.485{\mathrm M}_{360}+0.211{\mathrm M}_{490}-0.336{\mathrm M}_{546}-0.260{\mathrm M}_{596}+1.209{\mathrm M}_{617}-0.476{\mathrm M}_{700}\;\left(\mathrm P<0.001\right)\\{\mathrm Y}_2=0.477{\mathrm M}_{360}-0.360{\mathrm M}_{490}+0.219{\mathrm M}_{546}+0.363{\mathrm M}_{596}+1.175{\mathrm M}_{617}-0.227{\mathrm M}_{700}\;(\mathrm P<0.001)\\\begin{array}{l}{\mathrm Y}_3=-0.010{\mathrm M}_{360}-0.364{\mathrm M}_{490}+1.066{\mathrm M}_{546}+1.063{\mathrm M}_{596}-1.759{\mathrm M}_{617}+0.325{\mathrm M}_{700}\;(\mathrm P<0.001)\\{\mathrm Y}_4=0.247{\mathrm M}_{360}+0.477{\mathrm M}_{490}-2.473{\mathrm M}_{546}+0.642{\mathrm M}_{596}+1.596{\mathrm M}_{617}-0.358{\mathrm M}_{700}\;(\mathrm P<0.001)\end{array}\end{array}$$

Factor analysis of principal components analysis was employed to evaluate the contribution of each element in the sensor array. The Kaiser–Meyer–Olkin (KMO) of this data model reaches 0.672, that is, the data have a high correlation with each principal component, causing the key data information can be obtained. The KMO result indicates that PCA is feasible for data analysis. The contribution of the first six principal components was 76.619%, 22.198%, 0.981%, 0.167%, 0.024% and 0.011%, respectively Figure S2 exhibited the PCA results of 18 metabolites at 300 μM. By comprehensively analyzing the contribution ability and element matrix information of each principal component, the importance ratio of each characteristic peak in the sensor array is calculated. The weight coefficients represented by each characteristic peak are 0.183, 0.173, 0.172, 0.156, 0.172, 0.144. The weight coefficients are very close, which means that each array point contributes more evenly to the array. The array analysis result is the product of synthesizing the information of each element, and will not be greatly disturbed by a certain factor.

When the qualitative detection ability of the sensor array for metabolites is verified, Euclidean distance of the Mx (Eq. 2) as a quantitative standard is further employed to evaluate metabolite concentration (Fig. 4). The error bars are described by the 95% confidence interval. Between the concentration of Val is 50 μM to 500 μM, Euclidean distances owns a good linear relationship (R2 = 0.9755) with the Val concentration, indicating that provided sensor array has good quantitative detection ability for TCA metabolite.

Fig. 4
figure 4

Quantitative detection of Val

Full size image

Discrimination of cancer cell

Tb/Eu MOF tested human adipose-derived mesenchymal stem cells (HAMSC), human epithelial cells (EC), breast cancer cells (MCF-7), lung cancer cells (A549), cervical cancer cells (Hela), leukemia cells (K562) and hepatocellular carcinoma cells (HCCLM3). Principal component 1 and 2 could explain 89.416% and 8.720% effective information. By using PCA, the classification accuracy of 7 cells achieved to 100% (Fig. 5A). The weight coefficients of result elucidate represented by each characteristic peak are 0.179, 0.178, 0.175, 0.164, 0.136, 0.168 (Fig. 5B). Uniform contributions of each array point prove the reliability of the discrimination results. This discrimination ability also can be proved by LOO in LDA, 100% original cells group observations have been correctly classified indicating the feasibility and effectiveness of differentiation and discrimination of cells through tricarboxylic acid cycle metabolites detection. For distinguish between cancer cells and normal cells, RBFN was applied to determine the fitting relationship between the predicted sample and the actual sample. Named HAMSC and EC as NO.1 (Human Normal Cells), MCF-7, A549, Hela, HCCLM3 and K562 as NO.2 (Human Cancer Cells). Three-fourths of the samples were chosen as the experimental training group and amplified five times, and the rest one-fourth were chosen as test group. Both parts contain all 7 types of cells. As can be seen in Fig. 5D, human normal cells and human cancer cells were successfully discriminated with an average relative error of 0.0947 at a diffusion coefficient of 60 and a neuron number of 20. This proves that our sensor array has good fitting and discriminatory ability. Figure 5E shows a finger map with a 6-dimensionality vector was easily obtained through Mx with marked difference and excellent classification specificity.

Fig. 5
figure 5

PCA results for cell discrimination experiment (A). Weight coefficients of sensor array points (B). LDA results for cell discrimination experiment (C). Neural network simulation results (E). Finger map for 7 cells (E)

Full size image

Conclusion

In this work, a multi-signal Tb/Eu MOF sensor array was reported for the detection of TCA metabolites and discrimination of cells. Due to a significant host–guest interaction, provided single particle Tb/Eu MOF could discriminate the types and concentrations of TCA metabolite by luminescence intensity changes of autologous 6 characteristic peaks. 18 TCA metabolites were successfully discriminated via LDA at 4 concentrations (50 μM, 100 μM, 200 μM, 300 μM), showing the great qualitative detection ability of this sensor array. The correct detection of Val with different concentrations also proves the quantitative detection ability. In further experiments, 2 human normal cells and 5 human cancer cells were successfully classified via PCA, LDA and RBFN. This simple and accurate Tb/Eu MOF sensor array shows great potential for TCA metabolites identification and clinical application.