A Rapid and Robust Diagnostic for Liver Fibrosis Using a Multichannel Polymer Sensor Array

Abstract

Liver disease is the fifth most common cause of premature death in the Western world, with the irreversible damage caused by fibrosis, and ultimately cirrhosis, a primary driver of mortality. Early detection of fibrosis would facilitate treatment of the underlying liver disease to limit progression. Unfortunately, most cases of liver disease are diagnosed late, with current strategies reliant on invasive biopsy or fragile lab-based antibody technologies. We report a robust, fully synthetic fluorescent polymer sensor array that rapidly (in 45 minutes) detects liver fibrosis from low-volume serum samples with clinically relevant specificity and accuracy, using an easily readable diagnostic output. The simplicity, rapidity and robustness of this method make it a promising platform for point-of-care diagnostics for detection and monitoring of liver disease.

Point of Care (PoC) diagnostics based on blood serum allow for rapid and accurate diagnosis of more disease states than any other body fluid, and can be administered at the hospital bedside, at local clinics, or in the patient’s home.^[1–3] PoC diagnostics can greatly improve care in both urban and rural communities by enabling more frequent monitoring of patient health, yielding both lower costs and shorter analysis times.^[4] In all areas of biomedicine, continuous longitudinal collection of health-data can lower patient mortality rates by facilitating earlier intervention.^[5] Serum is however a challenging medium for sensors, containing thousands of different proteins, at concentrations ranging over ten orders of magnitude, as well as salts, carbohydrates and lipids.^[6] Serum-based PoC diagnostics must be quick, robust, low-cost, and use small sample volumes of serum. Additionally, serum-based diagnostics must be designed from materials that are resistant to fouling by protein adsorption and degradation by enzymes, yet can still operate with clinically relevant specificity and sensitivity.

Liver disease is a particularly significant yet underexplored target for PoC blood testing, despite its prevalence and socio-economic costs. In contrast to cancer and heart disease, mortality from liver disease has increased over the last 30 years and is now the fifth most common cause of middle-aged death in Western society. Liver disease costs health services tens of billions of dollars each year,^[7] and could affect an estimated 30% of people in the US.^[8]

Disease severity, prognosis, and response to treatment for liver disease are largely determined by the stage of liver fibrosis (scarring).^[9] Liver health is strongly manifested in blood composition,^[10,11] with immonosensing platforms such as the ‘Enhanced Liver Fibrosis’ (ELF) used to assess and monitor fibrosis progression from serum without invasive standard of biopsies currently in use.^[11,12] These blood tests quantify multiple serum biomarkers to provide a measure of liver fibrosis, shortening time to treatment and improving assessment of patient prognosis.^[12,13] The instability of the bioconjugates used for biomarker detection, however, requires sending of samples to centralized pathology laboratories for analysis, increasing cost and complexity of tests, and delaying diagnosis and treatment for patients.^[14,15]

Cross-reactive ‘chemical nose/tongue’ sensing arrays have emerged as a strategy to rapidly profile complex chemical and biological systems using robust synthetic receptors.^[16,17] These array-based sensors generate patterns from the sample that are subsequently classified to generate algorithms for identifying analytes. Synthetic solution-based arrays have been successful in ‘fingerprinting’ and distinguishing proteins spiked in serum^[18] and in cell and cell lysate sensing.^[19–21] Pattern-based serum sensing, however, has not been widely demonstrated, with neither examples based on robust multiplexed (single well) sensing, nor examples that can stage a disease.^[22–24] Such a “hypothesis-free” approach would allow for disease detection using multiple known and unknown biomarkers in a single assay.

We present here a robust, multiplexed fluorescent polymer-based sensor platform that detects liver fibrosis from a small-volume serum sample, with clinically relevant accuracy. The sensor elements have been engineered to act both as cross-reactive recognition and transduction elements, with modulated fluorescence provided by simple chemical moieties. This chemical approach generates a modular and reproducible array design, simplifying implementation relative to most or multi-part sensor systems.^[18,25] By mixing three of the chemically stable polymers in a single, multiplexed array, an information-rich output (4 fluorescent channels) is generated from a single sample measurement (Figure 1). This array can accurately distinguish non-fibrotic patients from those with early stage liver fibrosis, in our cohort of 65 benchmarked patient samples. Significantly, the polymer sensor does not degrade in ambient conditions, dramatically increasing the viability of this platform for PoC diagnostics relative to the biologicals used in current methods.

Figure 1.

a) Generation of a fluorescent fingerprint through serum protein-polymer interactions, giving b) a fluorescent fingerprint. c) Exemplar outputs for healthy and fibrotic patients used for d) discriminant analysis for fibrosis detection. e) The molecular structures of the fluorescent polymers - m:n ~ 9:1. f) The interaction of the dye and their environment leads to modulation of their fluorescence through changes in physical arrangement, solvation and charge, with pyrene providing two fluorescence channels, one main emission and one from an excimer

Our sensor array is based on a poly(oxanorborneneimide) (PONI) random co-polymer scaffold,^[26] chosen for its ease of modification and good compatibility with biological media.^[27] The polymer featured benzoate (Bz) monomers to provide protein recognition and reactive sites for dye attachment using NHS-ester chemistry, with the overall fluorophore loading be controlled by proportionate mixing of the two monomer units in the PONI backbone (Figure 1e). The number of repeat units (ca. 40) was kept low to enhance stability in serum. This scaffold was decorated with environmentally responsive fluorescent dyes that act both as cross-reactive recognition and transduction elements, providing a straightforward array design.

Three PONI-polymer sensor elements were synthesized bearing pendant pyrene (Py), dapoxyl (Dap) and PyMPO dyes (Figure 1e). Overall, the concise 3-polymer sensor generates 4 channels from a single sample measurement. Each polymer displayed a change in fluorescence intensity upon the addition of specific proteins to the polymer solution, due to changes in the ionic strength, pH and supramolecular interactions of the dyes (Figure 1f).^[28,29] The Py polymer displays a principle emission at 380 nm and an excimer emission at 480 nm.^[30] The former band is ratiometrically sensitive to the polarity of the pyrene microenvironment, and the latter to the physical separation of multiple pyrenes.^[31] Dapoxyl and PyMPO gave emission in the yellow/orange region of the spectrum at 580 and 570 nm respectively, but had well separated excitation bands (330 and 416 nm) providing spectral resolution in the mixed system (Figure S1).

Initial experiments were performed by testing the array against 40 μL human serum samples, the amount available from a single drop finger-stick sample, and hence suitable for PoC applications.^[32] Increases in fluorescence intensity were observed for all polymers in differing ratios on mixing with serum. Whilst some red or blue shifting of the peaks was also seen, but the intensity changes were the major factor (Figure S2). In the first tests, the ability of the array to measure perturbations in protein levels in human serum was tested by spiking analyte proteins (Human serum albumin (HSA), immunoglobulin G (IgG), transferrin (Trf), fibrinogen (Fib) and alpha-1-antitrypsin (a1AT)) into diluted or full human serum (Figure S3 and S4, Table S1). Full details are given in the Supporting Information.

Therefore, the array was tested to determine whether it could ‘fingerprint’ liver fibrosis in a serum sample, using the hypothesis-free approach to provide a potentially clinically-relevant assay. The Enhanced Liver Fibrosis (ELF) test, based on three serum biomarkers hypothetically linked to liver fibrosis, was used as our benchmark, due to its use as a gold-standard for fibrosis detection in a wide range of liver diseases.^[11,33] Sixty-five human serum samples were previously quantified for hyaluronic acid (HA), PIIINP (N- terminal propeptide of Type III collagen), and TIMP-1 (a tissue inhibitor of metalloproteinase) with the commercial ELF test, to generate an ELF score for each sample (Table 1).

Table 1.

Values of HA, PIIINP and TIMP-1 used to calculate the ELF score for each sample. The range and mean are given for the three fibrosis groups (healthy, mild-moderate and severe) as determined on the basis of the ELF score

Fibrosis Group	n		Range of value (mean)
		HA [ng/mL]	PIIINP [ng/mL]	TIMP-1 [ng/mL]	ELF Score
Healthy	16	4.72–17.62 (9.63)	2.32–9.51 (6.66)	147.0–235.3 (197.9)	7.03–7.94 (7.64)
Mild-mod	17	14.45–118.68 (56.82)	6.76–17.07 (10.35)	159.5–279.0 (230.9)	8.23–10.34 (9.39)
Severe	17	92.44–811.86 (367.83)	11.33–57.27 (22.46)	193.2–693.1 (347.5)	10.50–13.38 (11.69)

This sample library represented an ‘averaged’ disease landscape, reflecting the spectrum of liver diseases encountered in hospital practise, across age and gender, with equal representation of healthy patients and patients with moderate and severe fibrosis. Whilst age and gender can be used as a proxy for “risk factor”, they were not used here, nor in the ELF scoring of the samples.^[33] Fifty samples were categorized into three groups on the basis of their ELF score: healthy (ELF < 8.0), mild-moderate fibrosis (8.0 ≤ ELF > 10.5) or severe fibrosis (ELF ≥ 10.5), set as per National Institute for Health and Care Excellence (NICE) guidance for liver fibrosis.^[34] The second set of 15 samples was set aside as an independent validation set (Table S2).

Serum was added to the polymer sensor solution in a standard microplate for fibrosis detection studies. Samples were measured in replicate and the ratiometric change in each fluorescent readout used to generate the fluorescent fingerprint of each sample (Figure S5 and Table S3). Sensor response was generated in 30–45 minutes, much faster than current methods requiring multiple hours.

The fluorescent patterns generated from mixing the polymer and serum samples were processed with a simple LDA (Linear Discriminant Analysis) model. Each polymer displays a change in fluorescence intensity upon the addition serum containing various proteins to the polymer solution, due to changes in the ionic strength, pH and supramolecular interactions of the dyes (Figure 1f). The relative change in the emission intensity (I/Io) of each polymer is recorded for each sample in the ‘training’ dataset (Figure 1c). The data processing with LDA takes the four recorded polymer emissions for each sample and creates a linear combination of the input data – a ‘score’. This is done in such a way as to minimise the variance between samples of the same group (e.g. all ‘healthy’ samples have a similar score) whilst maximising the difference between samples of different groups (scores for ‘healthy’ and ‘fibrotic’ samples are as different as possible) (Figure 1d). Alternative, non-linear models such as Quadratic Discriminant Analysis, or Support Vector Machines were also tested, but had issues of overfitting the data in this case.

For the samples of unknown liver health in the ‘test’ dataset, the four polymer emissions are recorded for each as before. These data were compared quantitatively to the training set through their Mahalanobis distance^[35] to the previously defined groups (e.g. healthy or fibrotic), a technique that provides effective classification of new samples.^[20,36]

A diagnostic test for healthy, mild-moderate or severe fibrosis was developed by training this model against the first 50 patient samples. This classification model can distinguish between the three individual groups with 60% accuracy (Figure 2a, Figure S6), with the most misclassification occurring between mild-moderate and severe fibrosis. An independent reference sample set was analyzed with the same model (n=15, across all classes), with LDA giving 66.7% accuracy using the same 3-group model (Table S6).

Figure 2.

a) LDA models built on the training set for a 3-group model provides 60% accuracy as echoed in the test data. b) ROC analysis for this model showed that most misclassification arose between mild-moderate and severe fibrosis. c) LDA performed using 2 groups only – healthy vs. all fibrosis. The box plot gives data max/min (x) and the tails are set at 1.5 times the interquartile range. The box gives the upper and lower quartiles, the median, and the mean (•). The histogram is marked with normal distributions fitted to the full data. d) ROC analysis of the two-group diagnostic study, with accuracy improved to 80% and an AUROC of 0.89.

Notably, the array could discriminate between healthy samples and those from patients with fibrosis, a critical distinction of interest to clinicians. Thus, further analysis was undertaken using the Healthy group vs a total Fibrotic group combing Mild/Moderate and Severe into one class.

Multiple common liver biomarkers were measured in the samples and correlated to the classification accuracy (Figure S7) Some small correlations between TIMP-1 levels and misclassification of the fibrotic samples were evident, but a lack of overall correlation between total protein concentrations or key proteins and the misclassified results indicate that it is indeed multiple biomarkers being analysed to generate the result. Ultimately it is this signature that is determined and can be linked back to the disease; and this is an area we currently investigating further for future biomarker discovery and improvements to fibrosis detection.

Accuracy and sensitivity were determined through Receiver Operator Characteristic (ROC) curve analysis (Figure 2b).^[37] Improvement in a classifier is indicated by an increase in the overall summary metric of Area Under ROC Curve (AUROC), with the value ranging from 1 (perfect) and 0.5 (no better than chance), and the standard AUROC required of a diagnostic test for clinical relevancy is >0.80 (although other measures such as positive/negative predictive values must be considered too).^[38]

The LDA model was recalculated as described above to distinguish simply healthy patients from those with any degree of fibrosis (Figure 2c). This model classified the data with 80% accuracy and generated a single LDA score for each data point. In the 15-sample test set, the classification was also 80% for healthy samples vs. fibrotic samples (Table S4). The means of the two groups in the LDA were significantly different (t-test, p < 0.001) and the cut-off between healthy and fibrotic determined to be an LDA score of 0.304 (Figure 2c). On this basis, Sensitivity (the ratio of true positives to total positive values found) was calculated as 74% and Specificity as 94% (the ratio of true negatives to total negative values found). The AUROC was found to be 0.89 (Figure 2d), greater than the threshold for clinical relevance. Our new polymer-based test is fully comparable to other methods of diagnosing and staging fibrosis, such as elastography (AUROC = 0.84–0.89)^[39] and other serum biomarker tests (AUROC = 0.76–0.89).^[33,40] Therefore the result represents a substantial advance in both the use of array-based sensors for disease diagnosis, and in the detection of fibrosis, and a next step will be to recruit a larger cohort for better assessment of clinical utility.

In summary, we have fabricated a new multiplexed fluorescent polymer sensor array capable of detecting liver fibrosis using low volume serum. The accuracy and sensitivity of our hypothesis-free platform compares favourably against other leading biomarker-driven methods for detecting fibrosis, but does not require the specialist instrumentation of e.g elastography, whilst the robustness of the polymer platform is unprecedented for serum assay liver diagnostics, removing the need for cold-chain transport and storage. This combination of excellent accuracy, fast result time and stability provides a promising avenue for translation into a rapid, robust, point-of-care disease diagnostics for near-patient testing at home or in a primary care setting.

Experimental

Polymer synthesis:

Monomers and polymers were synthesized as described in Supporting Information and previous publication.^[41]

ELF characterization:

Serum samples were anonymous, unlinked, residual samples discarded after clinical evaluation from the liver clinics at the Royal Free Hospital, London, of volume between 0.5 and 1 mL and stored at −80 ^oC. The samples represented a range of etiologies of liver disease and were from a range of ages and genders. Serum had been previously collected and analysed using standard iQur protocols, as detailed in Supporting Information.

Array methodology:

The polymers were diluted and mixed in phosphate buffered saline (PBS), pH 7.4 150 mM to final concentrations of 4.7 μM for PONI-Bz-Py, 13.3 μM for PONI-Bz- Dap and 6.0 μM for PONI-Bx-PyMPO. For the spiked serum experiments 190 μL of polymer solution was loaded into a 96 well-plate, and 10 μL of serum was added. For the fibrosis sensing, it was determined that larger fluorescence changes could be achieved with a slightly larger volume of serum. Therefore, future experiments used 40 μL of serum to maximize I/I₀ while maintaining a reasonable dynamic range: 160 μL of the resultant polymer solution was loaded into a 96 well-plate following the injection of the specimen (40 μL of patient serum specimens). As a control experiment, PBS solution was injected instead of the serum specimens to account for dilution (I₀). The samples were incubated for 45 minutes, with measures made at 0, 15, 30 and 45 minutes. The emission spectra of the polymers were recorded at the optimal excitation/emission (Ex/Em) wavelengths: PONI-Bz-Py with Ex/Em 346/380 nm and an excimer emission 346/480 nm. PONI-Bz-Dap with Ex/Em 330/580 nm. PONI-Bz-PyMPO with Ex/Em 416/570 nm, using a fluorescence microplate reader, from each well plate (I) and is normalized against the PBS reference; I/ I₀.

Three to six replicates were obtained for each specimen, dependant on residual volume, and the 45 minute data was used and averaged. Standard deviations of the averages (the coefficient of variation) were 8% or less. Fifty patient training set samples and 4 channels from the change in the major excitation-emission of the three PONI-Bz polymers generated a 50 × 4 data matrix. Linear Discriminant Analysis (LDA) was applied using SYSTAT and JMP software packages. The canonical scores generated by the LDA model were used to classify the training samples and a separate test set of 15 samples; 5 each of healthy, mild and severe. In the case of healthy vs. fibrotic classification, a single canonical score was generated and significance of difference between the two groups tested with a two-group t-test. ROC analysis preformed in Origin generated an ROC curve and AUROC statistics.