3.1. Construction of a Multi-Fluorescent ssDNAs/nGO Sensor
To construct a multi-fluorescent ssDNAs/nGO sensor, we designed three fluorophore-modified ssDNAs (
Figure 2A);
P1-FAM: a quadraplex-formative sequence with FAM (λ
ex max/λ
em max = 495 nm/518 nm);
P2-TAMRA: a simple repeated sequence with TAMRA (λ
ex max/λ
em max = 555 nm/575 nm);
P3-Cy5: a hairpin-structure-formative sequence with Cy5 (λ
ex max/λ
em max = 645 nm/660 nm). These ssDNAs bear different sequences, and two of these can fold into different higher-order structures, which were expected to impart the individual elements of the sensor system with differential cross-reactivity [
38,
39]. In addition, well-separated absorption and emission spectra allow the readout of independent emissions of the fluorophores (
Figure 2B).
Initially, a fluorescence titration of nGO was carried out on an equimolar mixture of the three ssDNAs (20 nM) to examine whether nGO is able to quench the fluorescence of the ssDNAs simultaneously. For instance, the fluorescence emission of
P2-TAMRA can be observed dominantly when excited at 535 nm and detected at 579 nm. As shown in
Figure 3A, the addition of nGO to a solution containing the three ssDNAs resulted in a concentration-dependent quenching of
P2-TAMRA. Although
P1-FAM and
P3-Cy5 showed a similar pronounced decrease in fluorescence emission (
Figure 3B and
Figure S1), the corresponding quenching efficacies were lower than that of
P2-TAMRA. This may be attributed to the shielding of nucleobases in
P1-FAM and
P3-Cy5, caused by the DNA folding, which could hamper
π-
π stacking interactions with nGO [
40]. Therefore, in the following sensing experiments we used a binding ratio that provides high fluorescence quenching for all ssDNAs and minimal reproducible responses with the addition of 15 µg/mL proteins (
Figure S2), i.e., 20 nM ssDNAs and 100 µg/mL nGO.
3.2. Multi-Fluorescent Signature-Based Protein Sensing
Subsequently, we tested the ability of the multi-fluorescent ssDNAs/nGO sensor to generate fluorescence signatures of proteins. For that purpose, ten proteins that vary in size and surface charges were chosen as sensing targets (
Table 1). Each protein solution (20 µL) in PBS (pH = 7.4) was mixed with solutions (180 µL) of ssDNAs/nGO conjugates in PBS (pH = 7.4) to reach a final concentration of 15 µg/mL protein in a 96-well microplate. The fluorescence signals from individual wells were recorded as (
I–
I0) at seven different channels (
Figure 2B), generating a data matrix of 7 channels × 10 proteins × 6 replicates (
Table S1). Four channels provided almost independent emissions of
P1-FAM (Ch1),
P2-TAMRA (Ch4), and
P3-Cy5 (Ch6 and Ch7). Conversely, the other three channels (Ch2, Ch3, and Ch5) were likely located between the absorption and emission spectra of two of the three fluorophores, which should allow investigating the effectiveness of using spectral crosstalk.
The thus-obtained fluorescence signatures (
Figure 4A) likely show good reproducibility for the analyte proteins. These signatures were then subjected to an LDA in order to examine whether the individual signatures differ significantly. LDA is a supervised pattern recognition algorithm that provides a graphical output that offers insight into the clustering of the data and information on the classification ability [
4]. A linear discriminant score plot revealed ten well-separated clusters corresponding to the individual proteins (
Figure 4B). In this plot, each point represents the fluorescence signature of a single analyte protein. The first discriminant score, i.e., Score (1), provided the best discrimination among the classes, which accounted for 75.6% of the total variance. We expected basic proteins such as Lys (pI = 9.2) and Cyt (pI = 9.5) to exhibit a higher binding affinity than neutral or acidic proteins, as both nGO and the ssDNAs are negatively charged at pH = 7.4. However, the first discriminant scores showed almost no correlation with the pIs of the proteins (
r = −0.17). Considering the equally low correlation with the protein size (
r = 0.18), the sum of interactions regarding various characteristics, such as electrostatic and aromatic properties, hydrophobicity, surface heterogeneity and morphology, may possibly be responsible for the output as fluorescence signatures.
Then, a leave-one-out cross-validation analysis, the so-called jackknife classification procedure [
41], was performed to determine the classification potential of the multi-fluorescent ssDNAs/nGO conjugate sensor. Using a single channel afforded classification accuracies of 50%, 35%, 33%, 63%, 75%, 70%, and 33% for Ch1 to Ch7, respectively, while the accuracy increased to 97% when using all seven channels (
Table 2). Thus, it can be concluded that the sensor can acquire sufficient information to discriminate a variety of proteins from a single well. This system was able to detect ten different proteins at 15 µg/mL, ranging from 32 nM (Gal) to 1.3 µM (Cyt), which is comparable to the performance of a previously reported multi-fluorescent signature-based protein sensor [
33]. It should be noted that using merely three channels that are selective to individual fluorophores (Ch1, Ch4 and Ch6) afforded a comparable classification accuracy (98%;
Table 2), while a partial overlap between confidence ellipses was observed in the discriminant score plot (
Figure S3). The accuracy for IgG did not reach 100% in all cases shown in
Table 2, possibly due to the lower responses of ssDNAs/nGO conjugates compared to other proteins (
Figure 4A). The slight increase in accuracy for BSA upon decreasing the number of channels may be attributed to the higher levels of noise in Ch2, Ch3, Ch5, and Ch7.
Thereafter, we used 60 newly-prepared samples for a blind test, and the new cases were assigned to proteins according to their shortest Mahalanobis distances. Only four samples were misclassified when using seven channels, affording a classification accuracy of 93% (
Table S2). The accuracy only slightly decreased to 88% when using merely Ch1, Ch4, and Ch6 (
Table S2). These results suggest that it should be important to read out individual ssDNAs independently in the discrimination of proteins. It is possible that the high contributions of Ch1, Ch4, and Ch6 for protein discrimination is partly due to the higher magnitude in response compared to other channels (
Figure 4A).
3.3. Exploraion of Effective Sensing Channels for the Discrimination of Proteins
In order to gain further insight into the effective selection of channels, we investigated the relevance of individual channels on the generation of fluorescence signatures using HCA, which determines clusters on the basis of the Euclidean distances between elements of a dataset. Therein, each channel was standardized prior to the analysis based on the following equation:
z = (
x −
μ)/
σ, wherein
z is the standardized score,
x the raw response (
I–
I0),
μ the mean value of the population, and
σ the standard deviation of the population. Three clusters were observed (
Figure 5), i.e., cluster 1 includes Ch1–Ch3, cluster 2 includes Ch4 and Ch5, while cluster 3 includes Ch6 and Ch7. This result indicates a low correlation between channels included in each cluster. As estimated from
Figure 2B, Ch2 and Ch5 primarily read out the fluorescence of
P1-FAM and
P2-TAMRA, respectively. Hence, each cluster corresponds most likely to individual fluorophore-modified ssDNAs, suggesting that the use of different sequences and higher-order structures of ssDNA induce diverse cross-reactivity, which is a key feature for the generation of differential signatures. In their entirety, these results suggest that acquiring independent emissions of
P1-FAM,
P2-TAMRA, and
P3-Cy5 is critical to design accurate multi-fluorescent sensing systems, which is consistent with the results from the Jackknife classification and the blind test (
Table 2 and
Table S2).
Note that the properties of nGO should be considered to construct sensing systems with higher discrimination capability, as the interactions between nGO and proteins may play a partial role in the generation of fluorescence signatures. Given the recent progress in GO research, it has not only become possible to produce GO at lower costs and on a larger scale [
42,
43], but also to control its size, defects, and surface functionality [
44,
45,
46]. As GO with different characteristics interact differently with human cells and proteins [
44,
46], an optimization of these characteristics should improve the discrimination capability of the system.
3.4. Protein Sensing in the Presence of Human Serum
The performance of this sensing system was further evaluated for the identification of two different proteins (Cat and Myo) in the presence of interferent human serum. An estimated >10,000 proteins are present in human serum [
47], generating a challenging, complex matrix. It has been suggested that serum levels of Cat [
48] and Myo [
49,
50] could potentially be used as biochemical markers for particular diseases. Using the seven-channel system, 100% discrimination accuracy based on the jackknife classification was achieved for different concentrations of Cat (0–5 μg/mL) (
Figure 6A), and samples containing Cat and/or Myo with a total concentration of 5 μg/mL (
Figure 6B). Cat clusters moved along the x-axis with increasing concentration (
Figure 6A), while the 1:1 mixture of Cat and Myo was located between these components (
Figure 6B). These results indicate the potential of this method for the detection of proteins in solutions containing complex interferents.