Main

Despite representing only 2% of all tumours, central nervous system (CNS) cancers account for the highest average years of life lost in oncology1. The early, accurate, non-invasive detection of brain tumours and other neurologic conditions remains a major challenge because circulating biomarkers are either lacking or have low sensitivity. Currently, the standard of care for tumour diagnosis is intracranial biopsy2,3, a highly invasive procedure. Imaging modalities lack specificity in differentiating certain tumour types, can be challenging to interpret4,5,6 and often require substantial instrument time, cost and contrast agents or radioimaging probes. Liquid biopsies typically measure mutations in circulating tumour DNA or DNA methylation markers7,8 in accessible bodily fluids such as blood9, urine10, saliva11 or cerebrospinal fluid12. At present, there are limited liquid biopsy tests for intracranial tumour detection9, and development of intracranial tumour liquid biopsies based on tumour-derived DNA has been challenging due to the relatively low amount of circulating tumour DNA that enters the bloodstream, compared to other solid tumours13.

In this study, we developed a machine perception liquid biopsy (MPLB) approach to distinguish brain tumours in patients using peripheral blood samples and to identify the biomarkers responsible for detection. An array of quantum well nanosensors (QWNs), consisting of quantum well defect-modified single-walled carbon nanotubes (QWNTs)14,15, was assembled to acquire diverse optical responses to molecular binding events in blood plasma. Machine-learning algorithms were trained and tested on sensor responses to 739 unique patient plasma samples and detected tumour presence with 98% accuracy and identified tumour type with up to 71% accuracy. Quantitative proteomic analysis of the protein corona, eluted from the most predictive QWN, revealed that tumour-associated proteins were highly enriched on the QWN surface, and furthermore, tumour microenvironment (TME) and inflammatory response proteins were also identified. Spectral analysis of QWN responses was consistent with the hypothesis that the disease fingerprint detected by the MPLB was derived from these biomarkers. Independent patient samples recapitulated enrichment trends, validating the proteins as potential biomarkers of intracranial tumours. Our results show that the MPLB can detect and classify CNS tumours, including those that currently lack biomarkers, and enable the discovery of signal-driving biomarkers of early-stage tumours via enrichment of tumour ecosystem-secreted factors. This work appears to represent the first disease-agnostic perception-based liquid biopsy approach that concomitantly enables cancer detection, stratification and biomarker discovery.

Results

Acquisition of nanosensor responses to intracranial tumour patient plasma

To build our sample dataset for model development, we collected 690 peripheral blood plasma samples from patients diagnosed at the Department of Pathology, NYU Langone Health. These included four intracranial tumour cohorts representing glioma (n = 205), meningioma (n = 135), pituitary adenoma (n = 88) and schwannoma (n = 82), and 180 non-tumour samples (Fig. 1a and Supplementary Table 1).

Fig. 1: Nanosensor responses facilitate sensitive detection of tumour presence from plasma.

a, Left: schematic of diagnostic platform workflow, including summary of sample classes and sample sizes used in the study for initial model training, testing and external institution validation. Right: covalently modified SWCNTs, comprised of eight functional groups (4-sulfonatebenzene-, 4-methoxybenzene-, 4-nitrobenzene-, N,N-diethylaminobenzene-, 3,4,5-trifluorobenzene-, 4-carboxybenzene-, 3,5-dichlorobenzene-, 4-bromobenzene-) were wrapped with five polymers ((GT)15, (TAT)4, (GCCCCG)3, (AT)15, DMG-PEG2000) to generate an array of 21 physicochemically distinct QWNs. Centre: these QWNs were then combined with patient plasma samples to probe biomolecular binding interactions using high-throughput NIR spectroscopy. Bottom: spectral features were extracted and used to train classification models to predict presence and class of disease. b, Representative fluorescence emission spectra of F3-DMG-PEG2000 QWN from tumour and non-tumour samples. Spectra are the average of three technical replicates. c, PCA of QWN spectral features. Each point corresponds to an individual patient sample, coloured by sample class. d, Model performance as a function of increased number of sensor features. The highlighted model, comprising 12 spectral features, was selected for subsequent studies. e, ROC curve of cross-validation and test set for tumour identification. External institution validation samples were tested on the trained model. Dashed unity line denotes the line of no discrimination, representing the theoretical performance of an informationless classifier. f, ROC curve of test set performances for binary classification of different tumour classes. Dashed unity line denotes the line of no discrimination. g, Analysis of feature importance between tumour and non-tumour cohorts of QWN features for (1) polymer excipient (2) spectral feature and (3) quantum well covalent modification. MSKCC, Memorial Sloan Kettering Cancer Center; PPV, positive predictive value; CV, cross-validation; glio., glioma; men., meningioma; pit., pituitary adenoma; sch, schwannoma; G, guanine; C, cytosine; A, adenine; T, thymine; ({\lambda }_{E11}), E11 wavelength; ({\lambda }_{E11-}), E11− wavelength; ({I}_{E11}), E11 peak intensity; ({I}_{E11-}), E11− peak intensity; SO3, 4-sulfonatebenzene-; MeO, 4-methoxybenzene-; NO2, 4-nitrobenzene-; NEt2, N,N-diethylaminobenzene-; F3, 3,4,5-trifluorobenzene-; CO2H, 4-carboxybenzene-; Cl2, 3,5-dichlorobenzene-; Br, 4-bromobenzene-. Panel a created with BioRender.com.

Source data

To build a sensor array with diverse protein coronas and optical responses to proteins, we constructed 21 QWNs comprising quantum well defect-modified, polymer-wrapped carbon nanotubes (Fig. 1a, Supplementary Fig. 1 and Supplementary Table 2)14,15. We introduced covalent quantum defects onto the nanotube surface using a panel of aryl diazonium salts bearing substituents spanning a large range of Hammett σ values (–0.72 to 0.78). The substituents, 4-sulfonatebenzene-, 4-methoxybenzene-, 4-nitrobenzene-, N,N-diethylaminobenzene-, 3,4,5-trifluorobenzene-, 4-carboxybenzene-, 3,5-dichlorobenzene- and 4-bromobenzene-, were chosen to maximize diversity of chemical responsivities of the quantum well defect emission sites in the array. Each QWN generated a unique spectral profile, emitting narrow fluorescence bands at 1,000 nm (intrinsic bandgap fluorescence, E11) and at ∼1,140 nm (quantum-well defect-induced emission, E11−)16 (Supplementary Fig. 1).

To further diversify the sensitivity of QWNs in the nanosensor array, we selected four single-stranded DNA sequences ((GT)15, (TAT)4, (GCCCCG)3, (AT)15) and one amphiphilic polymer (1,2-dimyristoyl-rac-glycero-3-methoxypolyethylene glycol-2000 (DMG-PEG2000)) (Fig. 1a) that varied in sequence composition, structure and surface coverage. The sequences (GT)15 and (TAT)4, are known to form ordered surface patterns on carbon nanotubes based on previous experimental and computational studies17,18, while (GCCCCG)3 has been predicted to adopt similarly ordered conformations19, and (AT)15 is characterized by high surface coverage20 and a disordered surface patterning. DMG-PEG2000, introduced a non-DNA-based corona with distinct physicochemical properties (for example, charge). This set was chosen to facilitate a range of interactions with adsorbed biomolecules and to maximize sensor array diversity (the full list of defect/polymer combinations is given in Supplementary Table 2).

We acquired near-infrared (NIR) fluorescence spectra by introducing each patient sample to wells holding each of the 21 QWNs and exciting samples with 577-nm light (representative spectra from two samples are shown in Fig. 1b). We derived six spectral features from the fluorescent emission peaks of each QWN and observed that the QWN spectral features demonstrated diverse responses to the patient samples (Supplementary Fig. 2). Principal component analysis (PCA) of the QWN spectral responses identified two clear clusters corresponding to the tumour/non-tumour sample cohorts, with some overlap between clusters (Fig. 1c).

Classifier differentiates between intracranial tumour and non-tumour blood samples

To determine if the spectral fingerprint of QWN responses could unambiguously differentiate intracranial tumour patients from non-tumour plasma, we developed machine-learning models to discriminate between cohorts. We first used analysis of variance (ANOVA)-based feature selection to determine the most significant spectral features from the QWN responses21. We then used the CatBoost algorithm22 to assess the performance of different feature combinations, ranging from 2 to 58 features (Supplementary Fig. 3a). To optimize the training model, we tuned hyperparameters via Bayesian hyperparameter optimization to maximize the F1-score of 5-fold cross-validation training data averaged across folds (Supplementary Table 3). The average cross-validation accuracy varied between 0.87 and 0.98 (Fig. 1d). The best-performing model exhibited a test F1-score of 0.98, an accuracy of 98%, an 88% sensitivity at 95% specificity, and an area under the curve (AUC) of the receiver operating characteristic (ROC) curve of 0.98 (Fig. 1e). Strikingly, as evidenced by the high overall prediction accuracy, our models were able to detect tumours at a range of WHO grades, and over 20% of our glioma cohort was comprised of WHO grade 1–2 tumours (Supplementary Fig. 4a). The statistical assessment of age, sex and class imbalance on model performance confirmed negligible bias (Supplementary Fig. 4b–d and Supplementary Table 4).

To determine the generalizability of this method, we assembled a validation cohort of samples from two external institutions. For this analysis, we synthesized a new set of QWNs. Using the new batch of nanosensors, we collected spectral responses from 20 primary tumour plasma samples and 29 non-tumour plasma samples from Northwestern University and the Memorial Sloan Kettering Cancer Center (Fig. 1a and Supplementary Table 5). The QWN responses clustered consistently with our initial cohort of 690 patient samples, confirming robust sensor responses (Supplementary Fig. 5). The performance of this external validation dataset of the previously trained model was comparable to the internal test set with an F1-score of 0.897, an accuracy of 89.8%, a specificity of 93.1% at a sensitivity of 98%, and an AUC of 0.95 (Fig. 1e), suggesting reproducible sensor synthesis and robustness of spectral responses to patient samples

We investigated the potential of the QWN array to discriminate between intracranial tumour types. The QWN optical responses to plasma from patients harbouring the four intracranial tumour types (glioma, meningioma, pituitary adenoma and schwannoma) did not separate into individual clusters via PCA (Supplementary Fig. 6a). We trained CatBoost classification algorithms to differentiate between each pair of tumours (Fig. 1f and Supplementary Fig. 6b–g). The classifier trained to differentiate between glioma and meningioma scored the best among the differentiation tasks, with a weighted F1-score of 0.71, an accuracy of 71%, a precision of 0.74 and an AUC of 0.73 (Supplementary Fig. 6b). The other models performed with accuracies ranging between 65% and 71% (Supplementary Fig. 6b–g; detailed model statistics are given in Supplementary Table 6). The detection models were not able to robustly identify all clinically relevant differences, however. For example, detection of the isocitrate dehydrogenase 1 (IDH-1) mutation status of glioblastoma patients gave a test set AUC of 0.51, performing no better than chance (Supplementary Fig. 7).

3,4,5-F3-DMG-PEG2000 QWN drives machine prediction performance

We aimed to identify which QWN features were responsible for the disease-specific sensor responses. The DMG-PEG2000-wrapped QWNs were statistically overrepresented in the significant feature set (Fig. 1g(1), ({\chi }^{{2}=27.34;,P < 1.7\times {10}}{-7})), suggesting that the DMG-PEG2000 wrapping may facilitate the binding of disease-relevant analytes to the QWN surface. The single most important sensor for the classification task was the 3,4,5-trifluoro aryl functionalized nanotube with a DMG-PEG2000 wrapping, termed the 3,4,5-trifluoro-DMG-PEG2000 nanosensor, which alone detected tumours with 93% accuracy and differentiated between glioma and meningioma with 64% accuracy (Supplementary Fig. 3). The response was further improved by the addition of other sensors. The E11− wavelength spectral feature was generally the most important for model development as assessed by the scikit-learn feature selection toolkit (Fig. 1g(2)), consistent with previous work suggesting that the quantum well defects can elicit enhanced environmental sensitivity23. Correlation analysis between spectral responses to patient plasma samples (Supplementary Figs. 8 and 9) indicated that many spectral features transduced a substantial degree of mutual information, highlighting the need for feature selection.

MPLB platform reveals potential biomarkers driving disease-specific QWN responses

To investigate the molecular species driving the MPLB disease response, we conducted quantitative proteomic analysis on non-tumour plasma samples and on samples from glioblastoma and meningioma patients (Fig. 2a) to identify the protein corona of the single most predictive QWN, 3,4,5-trifluoro-DMG-PEG2000. After washing away weakly bound proteins from the QWN corona, we eluted the remaining proteins (Fig. 2a and Supplementary Fig. 10a,b)24 and conducted quantitative mass spectrometry analyses on three independent protein corona extractions for each cohort, labelled using 18-plexed TMTpro tandem mass tags to enable sample multiplexing25 (Fig. 2a and Supplementary Fig. 10a,b). PCA revealed that non-tumour and tumour sample clusters separated, while glioblastoma and meningioma clusters overlapped (Supplementary Fig. 10d), mirroring the QWN optical responses (Fig. 1c and Supplementary Fig. 6a). The protein compositions of the glioblastoma and meningioma samples were substantially correlated, with a Pearson coefficient of 0.86. This, in contrast to 0.75 for glioblastoma and non-tumour samples and 0.78 for meningioma and non-tumour samples (Supplementary Fig. 10e), suggests greater similarity between the protein corona compositions of the tumour samples, and more differences between non-tumour and tumour samples, further recapitulating the trends in QWN model performance.

Fig. 2: QWN corona proteomics reveals differential protein abundance and functional enrichment in non-tumour and tumour samples.

a, Schematic of QWN corona proteomics workflow. b, Volcano plot of log2(FC) versus −log10(P) between tumour (n = 12) and non-tumour (n = 6) patient plasma protein corona extracts; protein IDs labelled with |FC| > 4 and P< 5 × 10−4 are highlighted. Samples were analysed using two-sided Welch’s t-test, adjusted for multiple comparisons at a 1% FDR. c, log2(FC) of tumour (n = 12) compared to non-tumour (n = 6) samples of selected protein classes, indicating elevated and depleted expression across many classes of proteins. d, TMTpro relative abundance of specific proteins associated with tumour samples (n = 12) compared to non-tumour (n = 6). Data are reported as mean ± s.e.m. e, Gene Ontology enrichment analysis of proteins with |log2(FC)| > 1. Circle size indicates number of significantly enriched proteins with functional annotation and circle colour indicates statistical significance. Statistical test was conducted via one-sided Fisher’s exact test, with Benjamini–Hochberg correction at 5% FDR. f, Gene Ontology depletion analysis of proteins with |log2(FC)| > 1. Circle size indicates number of significantly depleted proteins with functional annotation and circle colour indicates statistical significance. Statistical test was conducted via one-sided Fisher’s exact test, with Benjamini–Hochberg correction at 5% FDR. g, Upper right: Venn diagram of proteins detected via LC–MS/MS in matched protein corona extraction (n = 18) and whole-plasma input (n = 18). Lower left: average percentage abundance of the 10 most prevalent plasma proteins from mass spectrometry analysis in whole plasma (n = 18) and their abundance in protein corona extractions (n = 18). TMT, tandem mass tag; RP, reverse phase; STAGE, stop and go extraction; RTS, real-time search; CID, collision-induced dissociation; HCD, higher-energy collisional dissociation; ADD, adaptor protein; CCT, chaperonin-containing TCP-1; CD, cluster of differentiation; CHMP, charged multivesicular body protein; COP, coatomer protein; CST, cystatin; CUL, cullin; HNRN, heterogeneous nuclear ribonucleoprotein; PSM, proteasome subunit; RP, ribosomal subunit; SERPIN, serine protease inhibitor; TUB, tubulin; Alb, albumin; TF, transferrin; APOB, apolipoprotein B; C3, complement C3; APOA1, apolipoprotein A1; IGHM, immunoglobulin heavy constant μ; A2M, α-2-macroglobulin; SERPINA3, serine protease inhibitor A3; SERPINC1, serine protease inhibitor C1. Panel a created with BioRender.com.

Source data

After elution, we observed 2,017 QWN-enriched proteins, 1,155 of which were enriched in tumour patient sample proteins, as compared to 862 proteins enriched in non-tumour individuals (Fig. 2b). A full list of significantly enriched and depleted proteins is provided in Extended Data Table 1 The five most significantly enriched proteins in the corona of tumour patients were ENPP2, LILRB3, FCAR, BCL2L15 and CAPG (Extended Data Table 1). These proteins indicate a protein corona composition derived from tumour samples influenced by proteins involved in inflammation26, immune modulation27,28 and lipid signalling29, released from the TME, and by systemic inflammatory responses. The five most significantly depleted proteins in the corona of tumour patients were TPP2, FASN, LPL, DDI2 and PDE5A, (Extended Data Table 1). These proteins are involved in various roles including proteostasis30, lipid metabolism31 and RNA regulatory mechanisms32.

Gene Ontology analysis33 revealed differentially enriched proteins in the tumour samples (|log2(FC)| > 1; FC, fold change). The proteasome 19S and 20S subunits (PSMs) were significantly depleted on the surface of the carbon nanotube in cancer patients (Fig. 2c). In contrast, the ribosomal subunits (RPs) were consistently more abundant in tumour samples (Fig. 2c), consistent with previous studies showing enrichment of RPs in glioblastomas34.

We found that multiple S100A proteins were notably upregulated in tumour samples, especially S100A8, S100A9, S100A11 and S100A12 (Fig. 2d). These S100A proteins have been recognized as potential biomarkers in various neurological tumours35,36,37. Additionally, our analysis indicated that ANXA1, 3, 6 and 11 were significantly enriched in the tumour samples (Fig. 2d), aligning with previous research38. Furthermore, TME and inflammatory response proteins were identified, including FCAR, LILRB2/B3, SIRPA/B3 and CLEC5A/12A. The vast majority of the enriched proteins (shown in Fig. 2b,c) have not previously been reported as tumour biomarkers. Figure 3e,f reports highlighted GO biological processes of differentially enriched proteins (| log2(FC)| > 1), indicating that enriched proteins are associated with immune responses, leukocyte activation and chemotaxis (Fig. 2e), while the depleted proteins are linked to proteolysis and catabolysis (Fig. 2f). The data suggest that a combination of tumour-specific, tumour microenvironment, and systemic inflammatory and immunological proteins from tumour samples were enriched on the QWNs (Extended Data Table 1).

Fig. 3: Identification of circulating molecular markers of cancer and potential tumour-specific biomarkers.

a, Plot of glioblastoma/non-tumour log2(FC) versus meningioma/non-tumour log2(FC), highlighting correlation between tumour enrichment and depletion profiles. Highlighted proteins in light green are the top 10 tumour non-specifically enriched (upper-right quadrant) or depleted (lower-left quadrant), representing potential molecular markers of cancer, agnostic to cancer type. Dashed yellow box with arrow indicates highlighted region of plot that is expanded in b. b, Magnified view at x intercept and y intercept reveals potential cancer type-specific differentially enriched protein species. Red, enriched (x > 0, y = 0) or depleted (x < 0, y = 0) in glioblastoma samples compared with non-tumour samples; blue, enriched (x = 0, y > 0) or depleted (x = 0, y < 0) in meningioma samples compared with non-tumour samples.

Source data

Protein corona analysis facilitates the detection of low-abundance biomarkers

To evaluate the extent to which our QWN protein-enrichment approach supports the detection of low-abundance proteins, we analysed the differences in the proteomic data provided by QWN-eluted plasma compared to whole-plasma samples. We conducted an 18-plex TMTpro experiment using matched patient plasma samples (non-tumour, glioblastoma and meningioma) and compared it to the protein corona extracts detected from the QWN. In contrast with the 2,017 proteins identified from the QWN, we could reliably quantify only 259 proteins in the undepleted plasma extracts (Fig. 2g, inset). Statistical analysis revealed that several of the proteins enriched in tumours, including S100A8, S100A9 and ANXA1, were detectable using bulk plasma input (Supplementary Fig. 11), but the vast majority of tumour-enriched proteins identified from QWN elution were not quantified in the bulk plasma input. In matched samples, albumin was over 10-fold depleted, and transferrin was 17-fold depleted in the protein corona compared with whole plasma (Fig. 2g). Of the top 10 most abundant proteins in the undepleted plasma samples, only APOA1 showed a relative enrichment (1.4-fold) in the protein corona compared with the whole plasma. This finding is consistent with the greater repertoire of protein types detected in protein corona (n = 2,017) when compared with undepleted plasma (n = 259).

Identification of cancer type-specific protein enrichment

Next, we focused on identifying QWN-eluted proteins that exhibited specific differential expression between either glioblastoma and non-tumour samples or between meningioma and non-tumour samples. There was substantial collinearity between the protein enrichment profiles between non-tumour/glioblastoma and non-tumour/meningioma samples (Supplementary Fig. 10e), suggesting that a large component of the variation was influenced by protein-binding interactions not specific to tumour class (Fig. 3a). By evaluating along the x and y intercepts, we identified cancer type-specific differences in the composition of the QWN protein corona (Fig. 3b).

Along the x intercept, we identified multiple glioblastoma-specific protein biomarkers in the composition of the QWN protein corona, in comparison to non-tumour samples (Fig. 3b and Extended Data Table 2). Known biomarkers, including matrix metalloproteinase 3 (MMP3)39, apolipoprotein A4 (APOA4)40 and several RPs (RPL4, RPL7, RPL18, RPS9, RPL6)34, were enriched 2- to 3.5-fold on the protein corona of the glioblastoma samples, whereas they were not differentially enriched in the meningioma protein corona samples. Other differentially enriched proteins were not previously recognized as implicated in glioblastoma, but are variously involved in extracellular matrix (ECM) remodelling, supporting growth, enhancing stress resistance and response, or are related to immune function pathways (Extended Data Table 2).

Our analysis along the y intercept revealed meningioma-enriched proteins compared to non-tumour samples, all unreported as meningioma biomarkers (Fig. 3b and Extended Data Table 3). Enrichment of immunological proteins (HLA-B, MARCO, CD5L)41,42,43 and ECM components (AEBP1)44 are consistent with the immune and structural composition of meningiomas. Additionally, the enrichment of lipid metabolism proteins (APOC1, C1QTNF1) supports the existing literature on changes in meningioma lipid metabolism45. Interestingly, many of the proteins identified throughout our proteomic analyses were not normally secreted proteins (Extended Data Table 3), which suggests that the observed proteins result from cellular fragments released from the tumour into the bloodstream.

Identification of differentially enriched proteins between glioblastoma and meningioma patients driving the prediction model

We assessed the protein-binding profiles potentially undergirding the model performance of our glioblastoma and meningioma classifier (AUC, 0.73; Fig. 1f). We integrated findings from three TMTpro protein corona extraction experiments to enhance the statistical signal and identify potential proteins contributing to model performance (Fig. 1f and Supplementary Fig. 6b). After combining the protein quantification results from the three experiments, we observed a total of 817 unique QWN-enriched unique proteins across all three experiments. Of these, 645 were enriched in glioblastoma patient samples, while 172 proteins were enriched in meningioma patient samples (Fig. 4a). We found that various protein classes were differentially bound to the sensors across groups, including many members of the septin family, heat-shock proteins and the protein kinase C-related family. These proteins serve diverse roles in ECM and structural remodelling, and in immunological functions (Fig. 4b). Our results showed that several serpins, including SERPINA1, SERPINA3 and AGT, were relatively enriched in the protein corona of glioblastoma patients (Fig. 4a,c). Notably, SERPINA3 has previously been identified as a diagnostic biomarker of glioblastoma40,46. Additionally, MMP339 and APOA440 have been reported as upregulated proteins in glioblastoma patients, which aligns with our findings (Fig. 4c)40,46. We found that ENPP2, CA6, OIT3 and IGHD were relatively enriched in the protein corona blood plasma of meningioma patients, playing roles in lipid processing and immune responses (Fig. 4a,c). Importantly, the degree of overlap with the literature further suggests that both known and unknown protein disease biomarkers influenced the sensor responses.

Fig. 4: QWN elution reveals differential protein enrichment profiles between protein coronas of glioblastoma and meningioma patient plasma samples.

a, Volcano plot of log2(FC) versus −log10(P) between glioblastoma (n = 22) and meningioma (n = 23) samples. Samples were analysed using two-sided Welch’s t-test, adjusted for multiple comparisons at a 1% FDR. b, log2(FC) of glioblastoma (n = 22) compared with meningioma (n = 23) corona samples of selected protein classes, indicating elevated and depleted expression across many classes of proteins. c, Relative protein abundance between glioblastoma (n = 22) and meningioma (n = 23) samples of selected differentially enriched corona proteins. Data are reported as mean ± s.e.m. ACT, actin; ADD, adducin; LDHA, lactate dehydrogenase a; FCN, ficolin; HSP, heat-shock protein; IG, immunoglobulin; IDH, isocitrate dehydrogenase; IQGAP, iq motif-containing GTPase-activating protein; MMP, matrix metalloproteinase; PON, paraoxonase; PRK, protein kinase; PSM, proteasome subunit; SERPIN, serine protease inhibitor; YWH, 14-3-3 protein.

Source data

Differentially enriched QWN corona proteins generate quantitative sensor responses

We conducted protein titration experiments to assess the spectral responses of 3,4,5-trifluoro-DMG-PEG2000 to 36 of the identified disease- and corona-enriched proteins, and to a set of 22 proteins that were not enriched in the protein corona, at concentrations ranging from 246 to 0.015 μg ml−1 in 5-fold dilutions using interferent pooled healthy plasma (20%) or phosphate-buffered saline (PBS) (Supplementary Table 7). The QWNs exhibited sensitive and quantitative spectral responses to certain proteins (Fig. 5a(1)–(4), Supplementary Figs. 12 and 13). In the absence of plasma interferent (that is, in PBS), the magnitude of the sensor response at maximum protein concentration was increased ((\bar{{\lambda }_{{E}_{11}^{{-}}}:4.41;{\mathrm{{nm}}})) compared to the protein interferent condition ((\bar{{\lambda }_{{E}_{11}}{-}}}:2.27 ;{{\mathrm{nm}}})). However, overall response trends between PBS and interferent conditions were broadly consistent, particularly between ratiometric spectral parameters (Fig. 5a(4) and Supplementary Fig. 13(6)). We fitted four-parameter log-logistic (4PL) models to the dose–response curves and found that many of the sensor responses to titrated biomarkers were well-described (R2 = 0.87–0.99). Figure 5b highlights strong quantitative QWN responses to some of the most highly enriched proteins (Figs. 2a and 3b), indicating that binding interactions between these enriched proteins and QWNs can elicit substantial spectral changes to QWN emission.

Fig. 5: Nanosensor responses and validation of potential biomarkers discovered by QWN corona proteomics.

a, Heat maps of F3-DMG-PEG2000 spectral responses to 36 QWN-identified non-specific cancer and tumour-specific potential biomarkers: (1) change in E11 wavelength; (2) change in E11− wavelength; (3) change in E11− − E11 wavelength; (4) change in peak intensity ratio. b, Dose–response curves to select candidate biomarkers in the presence of plasma interferent with four-parameter log-logistic model fit overlaid (blue dashed line). Proteins were selected on the basis of quantitative logistic spectral responses (R2 > 0.7) and significance in QWN corona proteomics fold-enrichment: (1) ANXA6 E11− intensity; (2) CCT5 intensity ratio; (3) CEMIP ΔE; (4) CRYAB E11 wavelength; (5) FGL1 intensity ratio; (6) ENPP2 ΔE; (7) S100A8/A9 ΔE; (8) UBAP E11− intensity. c, Quantification of protein responses for 58 proteins, comprised of 36 QWN binders, identified via proteomics, and 22 non-binders, which were not detected or not enriched in protein corona analyses. Responders are proteins that elicited a response in one or more spectral features as determined by R2 of fit >0.7 and response at maximal concentration >1 nm or 20% change in intensity. d, ELISA protein quantification of 39 whole-plasma samples (n = 13 healthy donor, 13 glioblastoma, 13 meningioma). Data are reported as mean ± s.e.m. Samples were analysed using two-side