Introduction (skip this if you know the basics): From the chemistry perspective, the world we live in is a fairly complex affair. When we consider analyzing chemicals, we realize that most analytes comprise multiple chemicals of interest (multi-analytes or complex analytes) that need to be analyzed. Unfortunately, there are no chemical sensors that can respond to all analytes. This suggests, that multiple sensors are frequently required to interrogate complex analytes. This led to the idea of integrating various chemical sensors into one sensor device called sensor array.
In the past, the leading paradigm for the design of artificial receptors and sensors was the lock & key concept proposed originally Emil Fischer in 1894, was based on the idea that chemical sensors should be shape-complementary to the target analyte to achieve high selectivity and specificity.
This, in supramolecular and analytical chemistry is usually achieved by synthesizing a preorganized receptor (attached to a signaling device such as dye or fluorophore or a microelectrode, etc., to make a sensor), which has a shape and molecular properties of binding cavity highly complementary to the target analyte to achieve the perfect recognition and binding. This is pretty hard to do! And, within the context of multianalytes, one has to have several (perhaps many) of these hard-to-get sensors to address the composition of the multianalyte of interest.
Nature mastered this over millions of years and now we have various natural protein receptors, enzymes, antibodies, etc. These are expensive, require gentle conditions (no bases, acid, metal ions, varying temperature, etc.). And then, their shelf-life is relatively low. Naturally, we would prefer to have artificial receptors made of as few atoms as possible (to be inexpensive and stable) or polymers. But now we are in the realm of organic synthesis and materials chemistry!
Using few atoms (compared to thousands of atoms in a few kDalton of protein) makes artificial receptors and sensors (when you attach a signaling device to the receptor) cheaper and sturdier. Also, it is harder to achieve the perfect recognition and binding with such limited resources!
Therefore, driven largely by the interest in complex analytes, the focus in molecular sensing is shifting from selective sensors toward cross-reactive sensors, that are not selective (i.e., are cross-reactive). One such cross-reactive sensor is prone to responding to several analytes within the multianalyte. This is what is happening, for example in our nose or tongue, which has only five types of taste buds and can distinguish only among five distinct qualities of taste. But the key part of the process is that several of the taste buds send a combined signal to the brain, and the brain remembers that this combination of biases corresponds to a specific taste. This means that what is being recognized is the composite signal, which we call response pattern. In chemistry, we have something similar: the so-called chemical noses and tongues, etc. These are arrays of sensors that utilize cross-reactive recognition elements to identify an analyte based on the response pattern that originates from a number of non-selective sensors.
Sensor arrays & Optical Sensor Arrays: One of the important outcomes of the supramolecular analytical chemistry is the understanding of molecular aspects of molecule or ion - receptor recognition and binding and knowledge how to integrate receptors and signaling devices (molecular devices) to obtain an unambiguous signal. We call this modestly sensor design. The design of a sensor array starts with considering the multianalyte under study (liquid, gas), enumerating the components, and a carefully considering whether all the analytes are known to us of whether there might be an unexpected or unknown component (such as unknown toxin). In our group, we are analyzing mostly liquieds (although we have worked on analyses of gasses - air comprising gasses of explosives from deteriorating landmines, Chem. Eur. J. 2012, 18, 12712-12718; Org. Lett., 2008, 10, 3681-3684).
Another important decision relates to the nature of the signal that will be observed. This will inform our choice of signalling molecular device. If the observed signal is current, a conductive polymer or a short-range spacer comprising sulfur atoms that bind to gold electrodes as it is, for example, in some of our OTFT-based sensors may of advantage, Anal. Chem. 2016, 88, 1092-1095; Chem. Com. 2015, 51, 17666-17668. For fluorescence-based read-out signal, a bright fluorophore will be desired. We prefer to work with fluorescence as it is sensitive and simple to read (spectra or images) compared to, for example colorimetry.
The most important aspect for design of sensor array are the nature of the individual sensors, the receptor-fluorophore chemicals to be used. Fortunately, we can get a lof of that information from binding studies (NMR, fluorescence, UV-Vis, NIR spectroscopy and MS spectrometry). Then we have a good idea about the binding affinity and stoichiometry, magnitude of the response, etc. Do we need to do that? No. But it will inform our selection of the sensor chemicals. Still, we don't have to do that. But we will need to employ some literature precedents and perhaps also chemical intuition, and most importantly a larger number of individual sensors to build in some robustness and redundancy which will help to cover varying affinity, various stoichiometries and perhaps low magnitude of response (Chem. Soc. Rev. 2010, 39, 3954-3979). An array with a large number of sensors will quickly generate a large amounts of data we will have to deal with during the statistical analysis. After all, the statistical analyses, analysis of variance, and artificial intelligence embedded in various chemometric tools are quite likely to be able sort the important and junk data. One may even start with the brute force (many sensors) and use the chemometrics tools to weed out the sensors that do not provide important data and optimize the array for maximum analytical performance with lowest number of sensors (J. Am. Chem. Soc., 2013, 135, 7705-7712; J. Am. Chem. Soc. 2008, 130, 10307-10314).
However, a good knowledge of analyte and carefully conducted binding studies will allow us to use few of the semi-finalist sensors selected for large response and binding affinity to match the analyte profile. Even then, some redundancy may be desirable. Other aspects that need to be considered is the volume of the analyte, solvent (water or organic), approximate concentration, presence of solids (field drainage, river mud), etc. This is simply because if one uses organic solvents, the plastic array substrate might be compromised (multi-well plates), glass bottom unglued, etc. In such a case, glass or quartz multi-well plate may need to be used. On the other hand, "muddy" analytes may still be used, but because nobody want to filter the analyte, a bottom reading will be used (assuming that the dirt stays on top of the sensor).
However, a good knowledge of the analyte and carefully conducted binding studies will allow us to use few of the semi-finalist sensors selected for large response and binding affinity to match the analyte profile. Even then, some redundancy may be desirable. Other aspects that need to be considered is the volume of the analyte, solvent (water or organic), approximate concentration, presence of solids (field drainage, river mud), etc. This is simply because if one uses organic solvents, the plastic array substrate might be compromised (multi-well plates), glass bottom unglued, etc. In such a case, glass or quartz multi-well plate may need to be used. On the other hand, "muddy" analytes may still be used, but because nobody want to filter the analyte, a bottom reading will be used (assuming that the dirt stays on top of the sensor).
Examples of sensor arrays from our laboratory: As noted above, a sensor array could be a commercially available multi-well plate (see below), a bundle of optical fibers (Science 2000, 287,451-452) or many proprietary microchip-like formats (Illumina Microarray Technology comes to mind), or even a collection of data obtained from different sensors during different measurements at different times (by different people). This is not entirely desirable as it could lead to potentially deleterious artefacts and introduce inaccuracies.
We work with two types of arrays. The first is a glass-based microscope slide with ultrasonically drilled micro-wells (diameter: 950 um, depth 250 um, analyte volume 200 nL) (A). Similarly, a quartz plate holds 1536 micro-wells in a spatial arrangement that matches the distribution/arrangement of wells in a standard 1536-well plate. This means that this plate can be read by multi-well plate readers used in high-throughput screening (HTS) (A). A large plate with several fluorescence-based arrays is being inspected under UV-light by a student (B). Small microscope slide-based array with deposited sensor materials (C). Confocal image of a single well shows the individual slices stacked up to form a 3D image of the micro-well (D). The last slice shows the diameter of the well at the opening (D). This allows calculating the well volume: 200 nL.
What is important on the images (B), (C), and (D) is that the sensor compounds are actually dissolved in a hydrophilic polymer (Chem. Eur. J., 2013, 19, 8497-8506; J. Am. Chem. Soc. 2007, 129, 7538-7544). The sensor molecules are colored dust and would fall out from the wells upon handling. The sensor therefore doped into a hydrophilic polymer which the forms a 10 um thick film (D) that sticks to the glass, absorbs water with the analyte components, which then interact with the actual sensor molecules. This way, one can use muddy water as analyte, whole urine (Chem. Com. 2016, 52, 1827-1830; J. Am. Chem. Soc. 2013, 135, 7705-7712) (no filtration of the cells or proteins is necessary), or even whole blood serum (Angew. Chem. Int. Ed. 2007, 46, 7849-7852; Org. Biomol. Chem., 2016, 14, 7459-7462). Once again, the lipids, proteins, etc., are stuck to the top of the sensor-polymer film while the analytes (in our case phosphates such as ATP) diffuse through the polymer with the water. The change in fluorescence is then measured from the bottom using a UV scanner.
When the analyte with the sensor molecules embedded in the polymer film, the optical properties of the sensor molecules change. Such changes can be captured by scanner. It looks like this: The image of the array chip is captured by a scanner (A). The images of the actual responses (A), i.e., the change in color from the wells for the sensors 1-8 are cut out from the image (B), the colors of the image are inverted (to obtain black background) and the resulting color image is deconvoluted to individual color channels (Red/Green/Blue). That means that one color image is split into three images, each with different intensity of the lighter spots (non-zero pixels; the zero is black=the background). Then the non-zero pixels are integrated, i.e., converted into a number of pixels that are not black. The result is a group of numbers for each well for each of the RGB channels (C). These numbers from all analytes or concentrations for each repetition go to a spreadsheet for statistical program for evaluation. Note that the array (A) shows only one trial. These arrays must be run in many copies (8-12) so that one has the repetitions to be able to judge how robust is the analysis.
The result of all that data transformation (A) is a dataset (spreadsheet), which is then analyzed by chemometric methods. We use principal component analysis (PCA) (B) which is a non-supervised method that allows you to explore the data (variance, clustering, et.), or linear discriminant analysis (LDA), which is slightly different in the sense (because it is a supervised method) and allows for determination of classification accuracy. We also use hierarchical clustering analysis (HCA) (C) which allows for classification but most importantly provides into similarities of the properties observed. The large the distance on y-axis (Distance) the more dissimilar the behavior (closer=similar). For nice explanation see this site.
One thing you might have not noticed: This simple array of cross-reactive sensors comprised 8 sensors. And yet, it was able to identify 10 analytes! Somewhere along the way, we have beaten the selective "lock-and-key" sensors at this game. Because with the great selective and expensive sensors, you would need 10 of them to analyze 10 analytes. And that is not all! Below we show how the very same array also recognizes different concentrations of analytes. So, 8 sensors recognizes 10 analytes and also their concentrations! That is the power of the weak. Yes, the individual sensors are not selective (are cross-reactive) and are less difficult/expensive to synthesize. But they deliver an analysis superior analysis to the best sensors. This is the power of the weak!
The above showed a qualitative determination of analytes, i.e., helped to classify (identify) the analytes. But one can also perform a quantitative analyses using arrays. If you look really carefully at the wells, you see that the color changes from top down (as the analyte concentration increases) (A). For example, sensor 1 is orange without analyte and becomes read. Sensor 8 started as pink, but upon addition of the analyte turned green. These changes can be quantified (colors deconvoluted and non-zero pixels integrated) and graphs that are very similar to binding isotherms are obtained for each sensor and each color channel and each concentration (in many copies). Resulting (large dataset) is then analyzed by machine-learning analysis (that is the Artificial Intelligence) and your computer gives you the results in ca 30 sec. All this is very nicely explained in J. Am. Chem. Soc. 2007, 129, 7538-7544 and Chem. Soc. Rev. 2010, 39, 3954-3979.
The quantitative analysis, which showed isotherm-like dependence of color on the analyte concentration matches closely the isotherms obtained during the binding studies. This confirms that the array behaves reasonably, i.e. follows the supramolecular behavior of the sensors and confirms that our array design was correct and that we understand the chemical behavior correctly. It also means that it is unlikely that we measure some unknown artifacts.
But as good as the above graphs are, at the end of the day we have 8 sensors with three channels (RGB) and many analytes provide a large response datasets. As mentioned above, these can be easily analyzed by machine learning (and embedded Artificial Intelligence) and other methods of chemometrics provide us with graphical outputs we can easily understand.
In drug analysis, we also developed arrays of sensors that are not only colorful and change their color in the presence of analytes, but the sensors are at the same time also fluorescent. In the figure below, you see the colorimetric response (and there is an observable pattern of changes) (A). But there is the same array under UV light, and entirely different pattern of changes emerges! This array can determine non-steroidal drugs as well as a metabolite of drug Ritalin (ritalinic acid, major metabolite of the psychostimulant drugs methylphenidate, brand name Ritalin). But the important point is that we can double, quadruple the information obtained from a single array! This makes the array mote robust and more sensitive. After all, the fluorescence is more sensitive than colorimetric response: You can see the difference in the colors, but they are weaker at low analyte concentration. The response pattern is clearer under the UV light, isn't it?
Speaking of fluorescence, we will make a small detour. The colorimetric sensor arrays in glass microchips are slightly bit difficult to read (scanning, REG channel deconvoluting, etc.). A large number of arrays are performed in multi-well plates, which have standard dimensions, you can purchase a simple inexpensive plate reader to obtain the read-out data, etc. For array purposes, the 96 (working vol 100-300 uL, 384 (30-100 uL), and 1536-well (2-5 uL) plates are the most practical for high-throughput screening (HTS) (Figure below, A). Most plate readers are compatible with 96-1536 well formats. For colorimetric readings you need the clear-bottom plates (B), for fluorescence-based assays the plastic-only black ones (C). Then you just pipette in your solvent, sensor, and an analyte and use a plate reader to read the absorbance or fluorescence at wavelengths that correspond to your sensors, the plate reader computer writes the data and statistical software then analyzes the data for you.
The sensor arrays are not useful only for identification and determination of analytes such as anions (Chem. Eur. J. 2018, 24, 4879-4884; Chem. Sci. 2013, 135, 7705-7712), metals (Angew. Chem. Int. Ed. 2012, 51, 2345-2348; J. Am. Chem. Soc., 2008, 130, 10307-10314; Anal. Chem. 2008, 80, 7451-7459), drugs (J. Am. Chem. Soc. 2017, 139, 14954-14960; J. Am. Chem. Soc. 2013, 135, 15238-15243), toxins (J. Am. Chem. Soc. 2012, 134, 20021-20024), etc. The arrays could be set up also to elucidate and predict, for example, yields and enantiomeric excess in organic chemistry transformations.
The arrays technique was, among others, used to measure enantiomeric aexcess in various chemicals including amines, amino alcohols, or diols (Nature Protocols 2020, 15, 2203-2229.). The following figure shows the construction of self-assembled sensor from a chiral diol (BINOL), formylphenyl boronic acid (FPBA) (A) and a chiral amine under study (B). Because the enantiomerically pure BINOL+FPBA for a chiral pre-complex which is chiral. Incorporation of a chiral amine leads to a mixture of diastereomers. And because each of the diastereomers (C) displays different fluorescence intensity (D), one can easily determine the enantiomeric excess (ee) in the diamine (Chem. Eur. J. 2016, 22, 10074-10080).
The same analysis of ee can be performed to determine enantiomeric excess in chiral drugs such as Atorvastatin (brand name Lipitor). Once again, each of the diastereomers display a different fluorescence (A,B), but only R,R-enantiomer is the drug (C). Chemometric analysis of the spectral data using a pattern recognition method called linear discriminant analysis (LDA) shows a clear definition and separation of the data clusters (20 repetitions) representing various enantiomeric purity from pure S,S- through the racemic mixture to the pure R,R-enantiomer (D). Inset (E) shows results of the linear regression and simultaneous analysis of two samples (red dots) of unknown ee. The error of prediction was 1.9% (plus minus 0.95%). A calibration curve for the same analysis shows simultaneous determination of unknown ee values in four different samples (the black x in the graph F). This analysis shows, how arrays can be used to evaluate ee, for example for quality control (QC) of pharmaceuticals (Chem. Eur. J. 2017, 23, 10222-10229).
Sensing arrays can be also used in optimization of reactions in a high-throughput screening (HTS) setting. The last example shows a simultaneous measurement and prediction of both product yield and enantiomeric mixture of an asymmetric reaction. For this purpose, we used the Noyori asymmetric transfer hydrogenation of benzoin (Organic Letters 1999, 1, 1119-1121.) following the Scheme below (A). Because each of the stereoisomers formed in the reaction displays a different fluorescence changes with the sensor (B), one can analyze the output data using linear discriminant analysis (LDA) (C) and determine the concentrations of individual stereoisomers.
Most interestingly, however, one can simultaneously determine the yield of the reaction as well as enantiomeric excess. In the figure below, upper left panel (A) shows the results of the determination of ee at a certain (constant) concentration. But the yield (i.e., the concentration) of the dihydrobenzoin affects the fluorescence signal and the higher the concentration of dihydrobezoins, the higher the slope of the traces in (Figure B). Naturally, because the sensor measures the concentration. At the same time, we can see that the intensity of the signal also depends on the ee. Thus, we have two independent variables (yield, ee) and fluorescence output as a dependent variable. Fortunately, when we provide these data as a training dataset to an artificial neural network (ANN) to make calibration hyperplane (not a calibration curve anymore). When we then provide the unknown sample, the artificial intelligence is able to place the data in the correct position on the hyperplane and read the answer. This is called prediction. So, what we do then is to run the reaction for different time, with different catalyst (R,R- or S,S- or a racemic catalyst R,R- : S,S- 1:1). The reading before the reaction and with the racemic catalyst are control experiments (to see if the ANN will be confused). It was not. In the Table C, we see that from each reaction we determined the ee by NMR: The data are in the pink column. Note that the yield cannot be determined so easily because NMR needs a substantial concentrations to be able to measure the ee. Then, we determined the yield and the ee fluorescence assay to figure out the yield and ee. These are the data in blue column. They match the NMR data. Then we took the fluorescence data and gave them as "unknown" samples to the ANN to see what she thinks. This is the pink column in Table C. One can clearly see that the predictions are very accurate. There is one important to point to make: The reaction samples were analyzed both as crude (including the catalyst, traces of solvents, formic acid, triethylamine) as well as recrystallized. As you can see, the assay does not care about impurities and faithfully measures only the dihydrobenzoins. Note, that with exception of the NMR, everything is done in a high-throughput screening (HTS) fashion, and to write about it takes much more time to accomplish the assay! You can read more in Chem. Eur. J. 2017, 23, 10222-10229.
As you can see, sensor arrays are very interesting and also an important topic for many different endeavors including high-throughput screening (HTS) in analytical chemistry, assays and bioassays, reaction methodology and development (to quickly screen successful substrates, conditions, catalysts) and quality control (QC). The proof that this is interesting is that you have read all the way to the end. Please, feel free to contact us if you are curious or you want to learn more by working with us.