Urinary detection of early responses to checkpoint blockade and of resistance to it via protease-cleaved antibody-conjugated sensors

Abstract

Immune checkpoint blockade (ICB) therapy does not benefit the majority of treated patients, and those who respond to the therapy can become resistant to it. Here we report the design and performance of systemically administered activity sensors conjugated to anti-programmed cell death protein 1 (αPD1) antibodies for the monitoring of antitumour responses to ICB therapy. The sensors consist of a library of mass-barcoded protease substrates that, when cleaved by tumour and immune proteases, are released into urine, where they can be detected by mass spectrometry. By using syngeneic mouse models of colorectal cancer, we show that random-forest classification trained on mass-spectrometry signatures from a library of αPD1-conjugated mass-barcoded activity sensors for differentially expressed tumour and immune proteases can be used to detect early antitumour responses and to discriminate resistance to ICB therapy driven by loss-of-function mutations in either the B2m or Jak1 genes. Our data supports the use of activity-based biomarkers for early on-treatment response assessment and classification of refractory tumours based on resistance mechanisms.

One-sentence editorial summary:

A library of systemically administered protease-cleavable sensors conjugated to anti-programmed cell death protein 1 antibodies allows for the early urinary detection and monitoring of antitumour responses to immune checkpoint blockade therapy in mice.

Immune checkpoint blockade (ICB) therapy has transformed the treatment of cancer for patients across a broad range of malignancies^1,2. ICB involves the administration of antibodies that block inhibitory checkpoint molecules, such as the cytotoxic T lymphocyte-associated protein 4 (CTLA4) or the programmed cell death protein 1 (PD1), to reinvigorate an antitumour T cell response. Despite the potential for ICB to produce durable clinical outcomes, a large fraction of patients do not derive clinical benefit^1,3. Objective response rates remain below ~25% in many cancer types, largely due to immunosuppressive factors in the tumour microenvironment (e.g., regulatory T cells or myeloid-derived suppressor cells) and primary tumour-intrinsic mutations¹. In addition, responsive tumours can acquire resistance during therapy such as in metastatic melanoma where up to one-third of patients with initial responses to ICB therapy eventually relapse³. Both primary and acquired resistance are driven by mechanisms that enable tumour cells to evade antitumour immune responses, including defects in antigen presentation or in the interferon gamma (IFNγ) response pathway^3,4. Therefore, developing noninvasive biomarkers of immune response and resistance to ICB has emerged as a clinical priority⁵.

Patient responses to ICB therapy are currently assessed using a combination of radiographic, tumour, and serum biomarkers⁵. Radiographic evaluation by Response Evaluation Criteria in Solid Tumours (RECIST) is the standard assessment method and occurs after the first cycle of ICB therapy, which consists of 3–4 doses administered within an 8–12-week window^6–8. The observation of atypical patterns of response to ICB has motivated continual refinement to the timing and frequency of radiographic assessment such as the development of immune-related response criteria (e.g., irRC, irRECIST) to account for phenomenon like pseudoprogression^5,9. Tumour biomarkers such as programmed death-ligand 1 (PD-L1) expression have been shown to enrich for populations with clinical benefit, but have limitations as predictive biomarkers as at least ~40–50% patient tumours with PD-L1 positivity do not experience objective responses^5,10. Other tumour biomarker strategies, such as assessing on-treatment changes in tumour mutational burden by whole exome sequencing¹¹, are promising and have been found to correlate with αPD1 response. However, these approaches require serial biopsies, which in practice are not typically collected over the course of therapy with attendant patient risks. Therefore, considerable interest is focused on identifying noninvasive biomarkers to allow longitudinal and quantitative assessment. These include quantifying changes in T cell clonality or circulating tumour DNA levels, which have been shown to be detectable within 3–4 weeks of treatment and correlate with objective response and overall survival^12–14. These studies highlight the considerable interest and need for noninvasive and longitudinal assessment strategies to track response and resistance to ICB therapy early on-treatment.

Proteases play fundamental roles in cancer biology, immunity, and antitumour responses and therefore may provide a mechanism to evaluate ICB therapy. Tumour-dysregulated proteases (e.g., matrix metalloproteases, cathepsins) are involved in proteolytic cascades that modify the tumour microenvironment during angiogenesis, growth, and metastasis^15,16. In addition, T cell-mediated tumour control is primarily carried out by granzymes, which are serine proteases, released by cytotoxic T cells¹⁷. The ubiquity of protease dysregulation has motivated the development of molecular imaging probes for visualizing tumour or T cell proteases^18–21, as well as synthetic biomarkers for multiplexed quantification of protease activity from urine^22–29. Building on these studies, we developed activity sensors to detect tumour and immune proteases during treatment as activity-based biomarkers of response and resistance. These activity sensors consist of mass-barcoded protease substrates conjugated to ICB antibodies that during the course of treatment are cleaved by proteases, triggering the release of reporters that filter into urine. After urine collection, cleaved reporters are quantified by mass spectrometry according to their mass barcode. In preclinical animal models, we show that binary classifiers trained on protease signatures by machine learning indicate on-treatment responses as early as the start of the second dose and differentiate B2m and Jak1 resistance with high sensitivity and specificity.

Results

Antibody-peptide therapeutic sensors retain target binding and in vivo therapeutic efficacy.

We first characterized target binding and therapeutic efficacy of ICB antibody-peptide conjugates. As a representative formulation, we coupled a fluorescently labelled peptide substrate selective for murine granzyme B (GzmB; substrate: IEFDSG²⁷) to αPD1 (clone 8H3) by cross-linking surface amines on the antibody with a terminal cysteine on the peptide to form an αPD1-GzmB sensor conjugate (αPD1-GS) (Fig. 1a, Supplementary Fig. 1a). This substrate has well-characterized cleavage kinetics and specificity^30–32 and has been used in molecular imaging probes^21,33,34 and activity biomarkers²⁷ to detect GzmB in vivo. To determine whether peptide conjugation would interfere with PD1 binding, we varied reaction stoichiometry to synthesize conjugates with different peptide:antibody ratios (0, 1, 3, 5, 7) (Supplementary Fig. 1b) and quantified binding to recombinant PD1 by ELISA. We observed negligible differences in EC₅₀ at a 1:1 ratio compared to unmodified αPD1 (3.6 vs. 2.1 nM respectively) (Fig. 1b) but at higher ratios, a gradual reduction in binding (up to 24 nM at a 7:1 ratio) (Supplementary Fig. 1c). To confirm that these results were not clone dependent, we coupled GzmB peptides to another αPD1 clone (29F.1A12) at a 1:1 ratio and found that target binding was likewise preserved between αPD1-GS and unconjugated antibody (EC₅₀ = 0.15 nM vs. 0.18 nM) (Fig. 1c). We did not observe significant differences in peptide-to-antibody ratios that would indicate batch-to-batch variation (Supplementary Fig. 1d). Based on these results, we used a 1:1 conjugation ratio for all subsequent studies.

Fig. 1 |. Antibody binding and therapeutic efficacy are unaffected by peptide conjugation.

a, αPD1-GzmB sensor conjugates (αPD1-GS) consist of αPD1 therapeutic antibody decorated with reporter-labelled GzmB peptide substrates (AA sequence: IEFDSG). b, ELISA assays comparing binding affinity of αPD1-GS and unconjugated αPD1 with recombinant PD1 (rPD1) using the mouse αPD1 clone 8H3 (log(agonist) vs. normalized response fitting function, n = 3 technically independent wells, error bars depict standard error of mean (s.e.m.)). c, ELISA assays comparing binding affinity of αPD1-GS with unconjugated αPD1 using the rat αPD1 clone 29F.1A12 (log(agonist) vs. normalized response fitting function, n = 3 technically independent wells, error bars depict s.e.m.). d, Representative flow cytometry histogram showing PD1 expression of CD8+ tumour-infiltrating lymphocytes (TILs) isolated from MC38 tumours. The same sample was divided and stained with either αPD1-GS, αPD1, or IgG1 isotype control. e, Quantified plot of PD1 expression showing the median fluorescence intensity (MFI) of samples stained with either αPD1-GS, αPD1, or IgG1 isotype control (one-way ANOVA with Tukey’s post-test and correction for multiple comparisons, ns = not significant, n = 10 biologically independent wells). f, Tumour growth curves of MC38 tumours treated with αPD1-GS, αPD1, or IgG1 isotype control immune checkpoint blockade treatment (ICB Tx) (two-way ANOVA with Tukey’s post-test and correction for multiple comparisons, ****P < 0.0001, n = 6 biological replicates, error bars depict s.e.m.).

We next evaluated target binding of αPD1-GS to tumour-infiltrating lymphocytes (TILs) isolated from MC38 tumours since ligand presentation of plate-bound recombinant PD1 may differ from endogenous PD1 expressed by T cells. We used the MC38 colon adenocarcinoma syngeneic tumour model because these cancer cells have a high mutation burden, which has been shown to lead to an endogenous T cell infiltrate following αPD1 monotherapy³⁵. Flow cytometry analysis of CD8+ TILs stained with either αPD1-GS or unmodified αPD1 showed statistically equivalent PD1 expression by median fluorescence intensity (MFI), indicating that peptide conjugation did not significantly affect target binding to endogenous PD1 expressed on cell surfaces (n = 10, Fig. 1d, e). We further confirmed that peptide conjugation did not affect therapeutic efficacy by comparing antitumour responses. Following a treatment schedule that involved four doses of antibody to C57BL/6 mice bearing MC38 tumours, we observed no statistical difference in tumour burden in mice given αPD1-GS or unmodified αPD1. Both formulations resulted in smaller tumours that were statistically significant compared to animals given IgG1 isotype control (P < 0.0001, n = 6, Fig. 1f, Supplementary Fig. 2). Taken together, these data demonstrate that coupling peptides at a low molar ratio to αPD1 does not affect target binding nor in vivo therapeutic efficacy.

αPD1-GS is sensitive to GzmB activity during T cell killing and stable in circulation.

We next tested the ability of αPD1-GS to monitor GzmB activity in a T cell killing assay. To quantify cleavage activity by fluorimetry, we coupled GzmB peptides containing a fluorophore-quencher pair (5FAM-AIEFDSG-CPQ2) to αPD1. (Fig. 2a). We assessed substrate specificity by incubating αPD1-GS with fresh mouse serum, tumour-associated proteases (e.g., cathepsin B, MMP9), or coagulation and complement proteases (e.g., C1s, thrombin). While incubation with recombinant GzmB led to a rapid increase in sample fluorescence, incubation with mouse serum or recombinant proteases did not result in detectable increases in fluorescence that would indicate cross-cutting of our sensors (Fig. 2b). We quantified GzmB cleavage kinetics of αPD1-GS by Michaelis-Menten kinetics and found a k_cat/K_m of 1.54 × 10⁴ M⁻¹ s⁻¹ (Fig. 2c) compared to 3.43 × 10⁴ M⁻¹ s⁻¹ for the free substrate. These values were consistent with previously studies that reported values ranging from ~10²-10⁵ M⁻¹ s⁻¹ for mouse GzmB cleaving the same or similar substrate sequences (IETDSG, IEFD)^27,30,33,34. In addition, cleavage rates were statistically equivalent across batches (Supplementary Fig. 3).

Fig. 2 |. Sensing T cell killing of tumour cells by antibody-GzmB sensor conjugates.

a, Schematic showing αPD1 antibody conjugated to fluorescently quenched peptide substrates for GzmB. Upon incubating these conjugates with transgenic Pmel T cells and B16 tumour cells, secreted GzmB cleaves peptide substrates, separating the fluorescent reporter from the internal quencher and resulting in an increase in sample fluorescence. b, In vitro protease cleavage assays showing normalized fluorescence (relative fluorescence units, RFU) of αPD1-GS after incubation with recombinant GzmB (blue), mouse serum (red), and other bystander proteases (n = 3 technically independent wells). c, Michaelis-Menten curve of initial cleavage velocity (V₀) at varying substrate concentrations for GzmB substrate (GS) when unconjugated (in presence of unmodified αPD1) and when conjugated to αPD1. k_cat and K_m were determined by fit to Michaelis-Menten equation (n = 3 technically independent wells, error bars depict s.e.m.). d, ELISA quantification of GzmB from T cell killing assays in which Pmel T cells were incubated with B16 target cells at different T cell to target cell ratios (one-way ANOVA with Dunnett’s post-test and correction for multiple comparisons, ****P < 0.0001, n = 4 biologically independent wells, error bars depict s.e.m.). e, Bar plot quantifying percent of cell cytotoxicity as measured by lactate dehydrogenase (LDH) assay from cocultures of Pmel T cells with B16 target cells (one-way ANOVA with Dunnett’s post-test and correction for multiple comparisons, **P = 0.004, ***P < 0.001, n = 3 biologically independent wells, error bars depict s.e.m.). f, Activity assays showing sample fluorescence after incubating αPD1-GS, αPD1, and an αPD1 conjugate with control substrates (αPD1-CtrlSub) with cocultures of Pmel T cells with B16 target cells (two-way ANOVA with Tukey’s post-test and correction for multiple comparisons, **P = 0.0075, ****P < 0.0001, n = 3 biologically independent wells, error bars depict s.e.m.). g, Half-life measurements of intact αPD1-GS and unconjugated αPD1 antibody (one phase decay fitting function, n = 3 biological replicates, error bars depict s.e.m.).

To evaluate αPD1-GS activation in the context of a T cell killing assay, we cocultured Pmel T cells with gp100-expressing B16 melanoma cells at increasing effector to target cell ratios (0, 1, 5, 10) and verified statistically significant increases in both supernatant GzmB by ELISA and target cell death by lactose dehydrogenase (LDH) release (n = 3, Fig. 2d, e). Under these co-culture conditions, we observed significant increases in fluorescence only in cocultures incubated with αPD1-GS, but not in control wells containing unmodified αPD1 antibody or αPD1 conjugated with a control peptide substrate (5FAM-ALQRIYK-CPQ2) (n = 3, Fig. 2f). We also did not observe αPD1-GS activation in cocultures of OT1 T cells and B16 cancer cells, which do not express the OVA antigen (P < 0.0001, n = 4, Supplementary Fig. 4).

We next considered the in vivo stability of αPD1-GS. Free peptides are rapidly cleared from circulation due to degradation by circulating endo- and exoproteases and/or renal clearance (Supplementary Fig. 5)^26,36–39 but can have improved pharmacokinetic profiles when conjugated to an antibody or protein scaffold^40,41. Therefore, we quantified the plasma concentration of uncleaved αPD1-GS following intravenous administration to determine peptide stability. We developed an indirect ELISA that uses plate-bound PD1 to capture αPD1-GS and a detection antibody specific for the FAM reporter at the terminus of the peptide substrate to differentiate between cleaved and uncleaved conjugates (Supplementary Fig. 6a). In validation assays, we compared ELISA signals from samples that contained αPD1-GS with or without preincubation with recombinant GzmB. Whereas αPD1-GS was readily detected compared to unmodified αPD1, we observed dose dependent reduction in signals for αPD1-GS samples treated with GzmB (n = 3, Supplementary Fig. 6b, c), validating the ability to discriminate between cleaved and uncleaved conjugates. Using this assay, we determined that the circulation half-life of uncleaved αPD1-GS was several hours and statistically equivalent to unmodified αPD1 antibody (3.9 ± 1.3 h vs. 6.5 ± 4.2 h, n = 3) (Fig. 2g), indicating peptide stability in circulation.

Noninvasive detection of early on-treatment response to ICB therapy.

We evaluated αPD1-GS to noninvasively detect response in C57BL/6 mice bearing syngeneic MC38 tumours. We confirmed significant increases to the number and percentage of GzmB+CD3+CD8+ TILs (P < 0.001, n = 9–10) but not in CD3+CD8− T cell subsets in response to two doses of αPD1-GS compared to control mice that received an isotype antibody conjugated with the same peptide (Iso-GS) (Fig. 3a–c), indicating that CD8+ TILs are the cells that respond to ICB therapy in this model. Using αPD1-GS and Iso-GS conjugates labelled with a near-infrared (NIR) fluorophore (VT750) (Extended Data Fig. 1a), we found that probe biodistribution in the tumour and major organs (liver, spleen, tumour draining lymph nodes (dLNs), lungs, kidneys, heart, brain) were statistically identical (Extended Data Fig. 2a–c). To determine whether αPD1-GS was cleaved in the tumour, we designed an activatable probe in which the GzmB substrate was flanked by a NIR fluorophore (800CW) and a quencher (QC1) (Extended Data Fig. 1b). We confirmed that incubation of activatable αPD1-GS and Iso-GS with recombinant GzmB led to increased sample fluorescence (by ~1.5 and ~1.6 fold respectively) (Extended Data Fig. 2d). In mice bearing MC38 tumours that previously received two doses of unmodified αPD1 or isotype control antibodies before given activatable αPD1-GS or Iso-GS, we found that NIR fluorescence was statistically equivalent in all non-tumour organs (liver, spleen, dLNs, lungs, kidneys, heart, brain) whereas NIR fluorescence significantly increased by ~1.4 fold in the tumour (P = 0.040, n = 4–5; Extended Data Fig. 2e, f). Together, these results indicate that αPD1-GS accumulates in MC38 tumours and is cleaved in response to treatment.

Fig. 3 |. Urinary detection of ICB therapeutic response by administration of antibody-GzmB sensor conjugates.

a, Representative flow cytometry plots, b, count, and c, frequency of cells expressing intracellular GzmB within CD3+CD8+ and CD3+CD8− TIL subsets that were isolated from MC38 tumours treated with either αPD1-GS or IgG1 isotype antibody conjugated with the GzmB peptide substrates (Iso-GS) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.011, ****P < 0.0001, n = 9–10 biological replicates, error bars depict s.e.m.). d, Tumour growth curves of MC38 tumour-bearing mice treated with either αPD1-GS or Iso-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.022, ****P < 0.0001, n = 6–7 biological replicates, error bars depict s.e.m.). Black arrows denote the treatment time points. e, Normalized urine fluorescence of mice with MC38 tumours after each administration of αPD1-GS or Iso-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, **P = 0.0093, ****P < 0.0001, n = 6–7 biological replicates, error bars depict s.e.m.). f, Receiver-operating-characteristic (ROC) analysis showing the diagnostic specificity and sensitivity in differentiating between mice treated with aPD1-GS vs. Iso-GS using urine signals on the second (AUC = 0.857, 95% CI = 0.643–1.00) or the third dose (AUC = 1.00, 95% CI = 1.00–1.00). g, Representative flow cytometry plots, h, count, and i, frequency of cells expressing intracellular GzmB within CD3+CD8+ and CD3+CD8− TIL subsets that were isolated from CT26 tumours treated with either αPD1-GS + αCTLA4 or matched isotype antibody controls (Iso-GS + Iso) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, **P = 0.0013 (h) and 0.0016 (i), n = 7 biological replicates, error bars depict s.e.m.). j, Tumour growth curves of CT26 tumour-bearing mice treated with combination therapy of αPD1-GS and αCTLA4 or combination of matched isotype controls (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.033, ****P < 0.0001, n = 7–14 biological replicates, error bars depict s.e.m.). Black arrows denote the treatment time points. k, Normalized urine fluorescence of mice with CT26 tumours after each administration of αPD1-GS and αCTLA4 or matched isotype controls (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ****P < 0.0001, n = 7–14 biological replicates, error bars depict s.e.m.). l, ROC analysis showing the diagnostic specificity and sensitivity of αPD1-GS in differentiating between responders to ICB combination therapy from off-treatment controls using urine signals on the second (AUC = 0.949, 95% CI = 0.856–1.00) or the third dose (AUC = 0.92, 95% CI = 0.795–1.00).

To evaluate the potential for serial on-treatment response assessment, we quantified the concentration of cleaved fluorescent reporters in urine samples that were collected within 3 hours after each dose was administered (day 7, 10, 14) (Fig. 3d, Supplementary Fig. 7a). At the start of the first dose on day 7, urine signals from both cohorts of mice were statistically identical as expected. By contrast, urine signals were significantly elevated in mice treated with αPD1-GS at the start of the second dose on day 10 (P = 0.0093, n = 6–7) when tumours were statistically equivalent in volume compared to control mice that received Iso-GS (255 mm³ vs. 441 mm³, P = 0.68, n = 6–7). This difference in urine signals was further accentuated by the start of the third dose on day 14 (P < 0.0001, n = 6–7) (Fig. 3e). We also observed that changes in the urine signals were significantly anticorrelated to tumour volume in individual mice at both dose 2 and 3 (R = −0.65, −0.73; Supplementary Fig. 8a, b), which was further supported by the significant anticorrelation between tumour fluorescence for activatable αPD1-GS (Extended Data Fig. 2f) and tumour growth (R = −0.91; Supplementary Fig. 9). Receiver operator characteristic (ROC) analysis of reporter levels in urine samples revealed an area under curve (AUC) of 0.86 and 1.00 for dose 2 and 3 respectively (Fig. 3f), indicating the ability to differentiate ICB response with high sensitivity and specificity.

We further sought to confirm urinary detection in a different preclinical model using BALB/c mice bearing syngeneic CT26 tumours that respond to combination therapy (αPD1 and αCTLA4) but minimally to monotherapy (αPD1 or αCTLA4)^42,43. Compared to matched isotype control conjugates, monotherapy with either αPD1-GS or αCTLA4-GS did not result in statistical differences in tumour burden and urine signals across all doses (Extended Data Fig. 3). By contrast, combination treatment with αPD1-GS and αCTLA4 resulted in higher GzmB+ expression in CD8+ TILs (P = 0.0013, n = 7) but not in CD3+CD8− TILs (Fig. 3g–i); significantly lower tumour burden (P < 0.0001, n = 7–14, Fig. 3j, Supplementary Fig. 7b); and significant increases in urine signals at the start of the second or third dose (AUROC = 0.95 and 0.92 respectively, Fig. 3k, l). Similar to results observed in the MC38 study, urine analysis indicated response to treatment several days before tumour volumes were statistically different compared to control mice (day 14 vs. 17), and urine signals were significantly anticorrelated with tumour volume for individual mice (Supplementary Fig. 8c, d). Collectively, these results showed that αPD1-GS indicated response to ICB treatment as early as the start of the second dose with high sensitivity and specificity.

Tumour resistance via loss of function of B2m or Jak1 leads to reduced GzmB activity.

Two mechanisms of tumour resistance to ICB include loss-of-function (LOF) mutations in B2M, a protein subunit of major histocompatibility complex I (MHC I), and JAK1, an essential signalling protein of the IFNγ response pathway^3,4. To model resistance, we knocked out (KO) B2m or Jak1 from wildtype (WT) MC38 tumour cells with CRISPR/Cas9. We validated KO cells by TIDE (Tracking of Indels by Decomposition) analysis⁴⁴ (Supplementary Fig. 10a), loss of surface expression of MHC I (H2-Kb) in B2m^−/− cells by flow cytometry (Supplementary Fig. 10b), reduction in GzmB and IFNγ expression by OT1 T cells after co-culture with OVA-pulsed B2m^−/− MC38 target cells (P = 0.031, n = 3, Supplementary Fig. 10c), and lack of upregulation of H2-Kb and PD-L1 following IFNγ stimulation of Jak1^−/− cells (Supplementary Fig. 10d). To confirm resistance to ICB therapy, we treated mice bearing B2m^−/− or Jak1^−/− MC38 tumours with either αPD1 or IgG1 isotype control alongside WT tumours as a positive control. Whereas αPD1 treatment of WT tumours resulted in significantly smaller tumours and improved survival (median survival time (MST) = 30) relative to isotype control (MST = 21) (P < 0.0001, n = 25, Fig. 4a, Supplementary Fig. 11), no statistical differences in tumour burden and overall survival were observed in mice with B2m^−/− or Jak1^−/− tumours. Together, our data confirmed that LOF mutations in B2m and Jak1 render MC38 tumours resistant to αPD1 therapy.

Fig. 4 |. ICB resistance by loss of function mutations in B2m and Jak1 leads to reduced GzmB activity.

a, Tumour growth curves of mice bearing WT (left), B2m^−/− (middle), or Jak1^−/− (right) MC38 tumours treated with αPD1 or matched IgG1 control (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ****P < 0.0001, n = 15–25 biological replicates, error bars depict s.e.m.). b, Representative flow cytometry plots, c, count, and d, frequency of GzmB+ cells within CD8+ TILs isolated from MC38 WT or knockout tumours treated with αPD1 or IgG1 (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.012, **P = 0.0044, ***P < 0.001, n = 5 biological replicates, error bars depict s.e.m.). e, Counts of GzmB+ stem-like cells (PD1+TIM3-CD39-) and f, counts of GzmB+ terminally differentiated effectors (PD1+TIM3+CD39+) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.022, ***P < 0.001, n = 5 biological replicates, error bars depict s.e.m.). g, Normalized urine fluorescence of mice with MC38 WT or resistant tumours after administration of αPD1-GS or Iso-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ***P < 0.001, n = 6–7 biological replicates, error bars depict s.e.m.).

We next analysed CD8+ TILs to characterize GzmB expression and T cell phenotype in resistant tumours. As expected, PD1 blockade resulted in significantly higher numbers of GzmB+CD8+ TILs in WT tumours (P = 0.0044) and not in B2m^−/− or Jak1^−/− tumours; however, both WT and B2m^−/− tumours had significantly elevated percentages of GzmB+ cells (P < 0.001 for WT, P = 0.012 for B2m^−/−, Fig. 4b–d). To understand why this increased frequency of GzmB-expressing cells did not result in response to αPD1 therapy for B2m^−/− tumours (Fig. 4a, Supplementary Fig. 11), we examined two functionally distinct subsets of CD8+ T cells that have been recently reported to mediate responses to ICB therapy: stem-like T cells (PD1+TIM3−CD39−) and terminally differentiated effector cells (PD1+TIM3+CD39+; Supplementary Fig. 12)^45–52. Stem-like CD8+ T cells retain polyfunctionality, undergo self-renewal, persist long term, and differentiate into terminally differentiated effectors but are poorly cytotoxic. By contrast, terminally differentiated effector CD8+ T cells have superior cytotoxicity but are short lived. In WT tumours, αPD1 treatment resulted in a significant increase in the population of GzmB+ terminally differentiated T cells (16.3-fold, P < 0.001), consistent with previous studies^45,52, with no significant increase in GzmB+ stem-like cells (Fig. 4e, f). Neither GzmB+ stem-like cells nor GzmB+ terminally differentiated effectors increased after αPD1 treatment of Jak1^−/− tumours. By contrast, we found that PD1 blockade of B2m^−/− tumours significantly increased the number of GzmB+ stem-like CD8+ T cells (18.6-fold, P = 0.022) but not the number of GzmB+ terminally differentiated effectors (Fig. 4e, f). These results support that ICB treatment of both B2m^−/− and Jak1^−/− tumours do not increase the number of terminally differentiated effector T cells; therefore, we postulated that αPD1-GS would not be able to discriminate resistance. To test this, we analysed urinary cleavage products and found that both B2m^−/− and Jak1^−/− tumours did not result in elevated urine signals after αPD1 treatment compared to WT tumours (Fig. 4g). Taken together, these data indicate that αPD1-GS lacks the ability to differentiate between ICB-resistant B2m^−/− and Jak1^−/− tumours.

Protease dysregulation in tumour resistance to ICB therapy.

To discriminate resistance, we postulated that a multiplexed set of probes against T cell and tumour proteases could provide the ability to distinguish resistance based on signature analysis. We therefore quantified differentially expressed proteases by sequencing the transcriptomes of WT, B2m^−/−, and Jak1^−/− MC38 tumours after two doses of either αPD1 or IgG1 (n = 5). By t-Distributed Stochastic Neighbour Embedding (t-SNE) analysis, we observed three distinct gene clusters corresponding to WT, B2m^−/−, and Jak1^−/− tumours (Fig. 5a). Gene set enrichment analyses (GSEA)⁵³ confirmed enrichment of immune pathways (e.g., IFNγ response, IL2-STAT5 signalling, inflammatory response, complement) in WT tumours in response to PD1 therapy, with minimal enrichment or downregulation in B2m^−/− and Jak1^−/− tumours, respectively (P < 0.05, Fig. 5b, Supplementary Fig. 13a). To compare with patient ICB responses, we performed GSEA on bulk tumour RNA-Seq data from advanced melanoma patients treated with αPD1 monotherapy¹¹ that were classified into complete or partial responders (CR + PR), progressive disease (PD), or stable disease (SD) based on RECIST criteria⁵⁴. We observed enrichment in immune pathways that were similar to murine tumours (e.g., IFNγ response, IL2-STAT5 signalling, complement) in CR + PR relative to PD (P < 0.05, Fig. 5b, Supplementary Fig. 13b).

Fig. 5 |. Proteases are dysregulated in ICB response and resistance.

a, t-SNE plot showing global transcriptional profiles of WT, B2m^−/−, and Jak1^−/− MC38 tumours treated with αPD1 or IgG1 isotype control (n = 5). b, Left: Gene set enrichment analyses (GSEA) comparing gene set signatures of all mouse tumours and treatment groups relative to WT tumours receiving isotype control treatment (n = 5 biological replicates). 6 gene sets were shown from the canonical Hallmark gene sets³⁴, with 4 immune- and 2 tumour-associated gene sets. Only the gene sets that are significantly different (false discovery rate < 0.05) between the two groups being compared were shown. Red colour indicates upregulation in the first group, and blue indicates downregulation. The size of the circle represents the nominal enrichment score (NES). Right: similar GSEA analyses using human data from melanoma patients treated with αPD1 monotherapy¹¹. Gene set signatures of the two patient groups (Complete Response (CR) + Partial Response (PR), and Stable Disease (SD)) were compared to patients with Progressive Disease (PD). c, Top: Volcano plots summarizing the extracellular and transmembrane proteases differentially expressed between WT MC38 tumours treated with αPD1 or IgG1 (n = 5 biological replicates). The threshold for differentially expressed genes (opaque dots) was defined as P value < 0.05 and |log2(fold change)| ≥ 1. Bottom: waterfall plot showing the fold changes in transcript levels of proteases that are differentially expressed between these two groups. The proteases are grouped into the families of interest while the remaining are greyed out. d, Waterfall plot showing the fold changes in transcript levels of proteases that are differentially expressed between αPD1 treated B2m^−/− and Jak1^−/− tumours (n = 5 biological replicates). e, Waterfall plot showing the fold changes in transcript levels of proteases that are differentially expressed between human tumours from responders (CR + PR) and non-responders (PD).

To identify proteases dysregulated in ICB response and resistance, we compared RNA transcripts levels of WT tumours on αPD1 or IgG1 treatment and observed that the top differentially expressed proteases, as selected by a log2 fold change threshold greater than 1, were from the granzyme, metalloproteinase, and cathepsin families of enzymes (P < 0.05, Fig. 5c, Extended Data Fig. 4a). By comparison, B2m^−/− tumours on αPD1 treatment showed broader dysregulation that included proteases from the complement, coagulation, and caspase families compared to Jak1^−/− tumours (log2 fold change > 1, P < 0.05, Fig. 5d, Extended Data Fig. 4b). Similar to our mouse models, human melanoma tumours in patients¹¹ that had a complete or partial response to ICB were characterized by significant upregulation of ~20 proteases across the same protease families relative to progressive disease (log2 fold change > 1, P < 0.01, Fig. 5e). By unsupervised hierarchical clustering, protease expression profiles were primarily grouped into CR+PR compared to PD (Extended Data Fig. 4c). Taken together, these data indicate that proteases are differentially regulated during response and resistance to ICB therapies.

Multiplexed detection of protease activity by mass spectrometry.

We next designed substrates for our multiplexed library to detect the proteases differentially expressed in ICB response and resistance (Fig. 6a). We compiled published substrate sequences for five target protease families – granzymes, metalloproteinases, coagulation and complement proteases, caspases, and cathepsins – and synthesized a candidate library of 66 fluorogenic substrates, which consisted of 6–11 amino acids flanked by a fluorophore (FAM) and a quencher (Dabcyl). We tested each substrate against 17 recombinant proteases (2+ per family) and quantified cleavage efficiency based on the fold change in fluorescence at 60 minutes (Fig. 6b, Supplementary Fig. 14). To facilitate downselection, we applied t-SNE analysis and observed 4 major substrate clusters: cluster 1 contained substrates preferentially cleaved by metalloproteases, cluster 2 by metalloproteases and cathepsins, cluster 3 by coagulation and complement proteases, and cluster 4 by granzymes and caspases (Fig. 6c). From each cluster, we selected 3 or more representative substrates to form a final library of 14 substrates. Each substrate in this set was characterized by a 2–22 fold increase in fluorescence in the presence of target proteases (Fig. 6d), and the majority of substrate pairs (76%) had a Spearman’s correlation coefficient (Rs) less than 0.5, indicating low redundancy of the library (Supplementary Fig. 15).

Fig. 6 |. Mass-barcoded peptide sensors for multiplexed detection of protease activity.

a, Schematic of the peptide substrate screen to identify candidate substrates for multiplexed activity sensor library. b, Fluorescence cleavage assays of representative substrates against recombinant proteases of interest. Each cleavage trace represents the average of 3 independent replicates. c, t-SNE plot showing unsupervised clustering of 66 candidate substrates into major clusters. d, Heat map summarizing the log2 fold change in fluorescence of 14 selected substrates at 60 min after addition of the respective recombinant protease (n = 3 technically independent wells). Signals were row-normalized before plotting. e, Calibration curves of mass barcodes as quantified by LC-MS/MS. MS2 peak area from each mass barcode used to label representative substrates is normalized by peak area of an internal standard to obtain peak-area-ratio (PAR) (n = 3 technical replicates, error bars depict s.e.m.). f, Bar plot showing corresponding mass reporter signals (PAR) from mixtures of αPD1- or IgG1-peptide conjugates (two-way ANOVA with Tukey’s post-test and correction for multiple comparisons, n = 3 technical replicates, error bars depict s.e.m.).

To enable multiplexed detection by mass spectrometry, we designed 14 mass barcodes by enriching the peptide reporter glutamate-fibrinopeptide B (GluFib) (EGVNDNEEGFFSAR) with different distributions of stable isotopes. As described previously²³, this approach allows multiple reporters that share the same MS1 parent mass to be differentiated by unique quantifier MS2 fragments by tandem mass spectrometry (MS/MS) (Supplementary Table 1). For validation, we derivatized our 14-plex substrate library with mass barcodes and confirmed that MS2 signals were linearly correlated with substrate concentrations (R² ≥ 0.96, Fig. 6e) and that the mass barcoded substrates conjugated to αPD1 or IgG1 antibody were quantifiable after cleavage (n = 3, Fig. 6f). Our results showed that our substrates are sensitive to cleavage by dysregulated proteases in the context of ICB response and resistance, and mass-barcoding allows multiplexed quantification of substrates.

Binary classification of response and resistance by 14-plex library.

To assess the potential of our 14-plex activity sensor library to detect early on-treatment response to ICB therapy, we administered 14-plex αPD1 or IgG1 conjugates to mice bearing WT MC38 tumours at days 7, 10, and 13 (Fig. 7a). At each timepoint, urine samples were collected within three hours after intravenous administration and cleavage fragments were quantified by mass spectrometry. Urinary signals from doses 2 and 3 were normalized to dose 1 to account for pre-treatment baseline activity. We applied random forest classification to the data split into training and test sets by 5-fold cross validation and repeated this procedure 100 times to obtain the average area under the ROC curve (AUC)⁵⁵. Under these conditions, our library of activity sensors discriminated αPD1-treated mice (n = 25) from isotype controls (n = 15) with high accuracy (AUC = 0.92 [95% CI = 0.88–0.95], sensitivity (Se) = 87%, specificity (Sp) = 86%) as early as the start of the second dose, with statistically identical classification performance at dose 3 (AUC = 0.93 [0.90–0.95], P = 0.650, paired Student’s t-test) (Fig. 7b). To assess the relative weight of each probe, we quantified the feature importance score and observed that probes L2-8, L3-7 and L2-1 had the largest contribution to classification accuracy with aggregate scores for dose 2 and 3 above 0.6 compared to scores of 0.3 and below for all other probes (Fig. 7c). These three probes were selective for granzymes, metalloproteinases, and cathepsins, including substrate L2-1 which was the same sequence previously used in αPD1-GS (Fig. 3). Based on the marked difference in feature importance scores, we further tested whether L2-8, L3-7, and L2-1 alone were sufficient to classify ICB responses, and found that the 3 probe set classified response with AUCs greater than 0.9 for both doses (dose 2 AUC = 0.95 [0.93–0.97]; dose 3 AUC = 0.91 [0.87–0.93]) with no statistical reduction in accuracy compared to the 14-plex panel (P = 0.147 on dose 2, P = 0.317 on dose 3, Fig. 7d). These data indicated that our activity sensors discriminated ICB responders as early as the second dose with 3 probes out of the 14-plex set.

Fig. 7 |. Urinary classification of ICB response and resistance.

a, Schematic of our pipeline to develop urinary classifiers of ICB response and resistance. b, Area under the ROC curve (AUC) analysis showing the diagnostic specificity and sensitivity of random forest classifiers based on 14-plex activity sensor library in differentiating between αPD1-treated WT tumours (n = 25 biological replicates) and IgG1-treated controls (n = 15 biological replicates) using urine signals on dose 2 (AUC = 0.92, 95% CI = 0.88–0.95) or dose 3 (AUC = 0.93, 95% CI = 0.90–0.95). c, Feature importance analysis revealing the probes that are important for response monitoring. Probes with higher important scores, produced by random forest, contribute more to the diagnostic performance. The pie charts above individual probes show the protease families that are monitored by each probe. d, AUC analysis of random forest classifiers based on the top 3 probes (L2-8, L3-7, L2-1) for response monitoring (dose 2: AUC = 0.95, 95% CI = 0.93–0.97, dose 3: AUC = 0.91, 95% CI = 0.87–0.93). e, AUC analysis of random forest classifiers based on 14-plex library in differentiating between αPD1-treated B2m^−/− (n = 15 biological replicates) from Jak1^−/− MC38 (n = 15 biological replicates) tumours using urine signals on dose 2 (AUC = 0.77, 95% CI = 0.71–0.82) or dose 3 (AUC = 0.91, 95% CI = 0.86–0.94). f, Feature importance analysis revealing the probes that are important for resistance stratification. g, AUC analysis of random forest classifiers based on the top 5 probes (L2-11, L2-20, L2-19, L3-16, and L2-9) for resistance stratification (dose 2: AUC = 0.80, 95% CI = 0.74–0.84, dose 3: AUC = 0.91, 95% CI = 0.86–0.94). h, Scatter plot showing feature importance scores of all 14 probes in the multiplexed panel for response monitoring and resistance stratification. The highlighted probes belong to the downselected probe sets that achieve comparable diagnostic performance in these classification tasks as compared to using the entire multiplexed panel. i, Schematic of prospective study design in which the original cohort of αPD1-treated WT tumours (n = 25) and IgG1-treated controls (n = 15) was used to train a response monitoring classifier based on the top 3 probes (L2-8, L3-7, L2-1), and classifier accuracy was tested in an independent prospective cohort (n = 15 biological replicates each, αPD1 and IgG1). j, AUC analysis of prospective response classification on dose 2 (AUC = 0.73) and dose 3 (AUC = 0.87).

We conducted similar longitudinal experiments to assess the ability of our multiplexed library to stratify refractory tumours based on B2m^−/− (n = 15) or Jak1^−/− (n = 15) LOF mutations (Fig. 7a). Following urine quantification by mass spectrometry, random forest classification resulted in an AUC of 0.77 (95% CI = 0.71–0.82, Se = 84%, Sp = 65%) on dose 2, which significantly increased to 0.91 (95% CI = 0.86–0.94, Se = 87%, Sp = 81%; P < 0.0001) on dose 3 (Fig. 7e). By feature importance analysis, we observed that a larger number of probes contributed to resistance classification where the top 5 probes had aggregate scores above 0.45 while the previous top ICB response probes, L2-8, L3-7 and L2-1, were in the bottom half by rank order (Fig. 7f). We further asked whether a minimal probe set could stratify resistance and by iterative analysis, we found that the top 5 probes (L2-11, L2-20, L2-19, L3-16, and L2-9) classified B2m^−/− from Jak1^−/− resistance with statistically equivalent performance to the full 14-plex library (dose 2 AUC = 0.80 [0.74–0.84], P = 0.430; dose 3 AUC = 0.91 [0.86–0.94], P > 0.999; Fig. 7g). Given that this subset of 5 probes did not contribute to the response monitoring classifier, we compared the importance score for all 14 probes for both classification tasks and found a strong negative correlation (R = −0.90) between the top probes for response monitoring (L2-1, L3-7, and L2-8) and stratifying resistance (L2-11, L2-20, L2-19, L3-16, and L2-9) (Fig. 7h). Last, to determine whether our classifiers can identify responders when applied prospectively, we trained a classifier on the 3-plex response monitoring probes (L2-1, L3-7, and L2-8) in our original cohort of mice (n = 40) and applied the classifier to an independent cohort of mice (n = 30) that was not involved in classifier training (Fig. 7i). We found that prospective classification discriminated responders from non-responders with high accuracy (dose 2 AUC = 0.73, Se = 100%, Sp = 73%; dose 3 AUC = 0.87, Se = 91%, Sp = 93%; Fig. 7j). Collectively, our data indicated that binary classifiers trained on urine samples discriminate response and resistance to ICB therapies in mouse models.

Discussion

In light of the central role proteases play in T cell cytotoxicity and tumour biology, our study focused on demonstrating protease sensors as an activity-based approach to track early response and resistance to ICB therapies. We showed that αPD1-peptide conjugates act as therapeutic sensors that carry out the dual roles of reinvigorating T cell function and reporting on treatment response by the release of protease-cleaved reporters into urine for noninvasive detection. Our results with a single αPD1-GS probe to quantify GzmB activity in vivo showed that urinalysis of cleavage fragments anticipated response as early as the start of the second dose before tumour volumes began to diverge between treated and untreated animals. By transcriptomic analysis, we identified proteases across five families that were broadly dysregulated in tumours harbouring B2m^−/− or Jak1^−/− loss-of-function (LOF) mutations. This list of proteases formed the basis of a bespoke 14-plex library of activity sensors that allowed binary classifiers trained on urine samples by machine learning to stratify the mechanism of resistance with high diagnostic accuracy. Our results support the development of activity sensors for noninvasive and longitudinal assessment of response and resistance to ICB therapies.

GzmB is the most potent pro-apoptotic granzyme and its release from granules accompanied by perforin is a primary mechanism by which CD8+ T cells exert tumoricidal activity. Compared to other tumour biomarkers (e.g., PD-L1⁵⁶, tumour mutational burden⁵⁷, T cell-inflamed gene expression profile⁵⁸, microsatellite instability⁵⁹) and serum biomarkers (e.g., circulating tumour DNA^14,60, T cell receptor clonality^12,13, memory phenotypes^12,13,61) under investigation, GzmB is a direct biomarker of T cell cytotoxicity, and its expression has been shown to be significantly upregulated in patient tumours responsive to αPD1 and αCTLA4 therapies^62–66. GzmB expression, therefore, has potential as an early biomarker of ICB response. Recent work on a peptide-based positron emission tomography probe that irreversibly binds to GzmB^21,67 and activatable imaging probes that either fluoresce^33,34,68 or radiolabel cell membranes⁶⁹ upon GzmB cleavage demonstrated that high GzmB signals predicted early response to checkpoint therapy before changes in tumour volumes were apparent in animal models. Similarly, we observed that tumour treatment with αPD1-GS therapeutic sensors led to quantifiable levels of cleaved peptides in urine that anticipated responders from isotype controls as early as the second dose and before tumour volumes significantly diverged. Furthermore, changes in urine biomarker levels were significantly anticorrelated with tumour volume at an individual mouse level. Looking ahead, these results support a potential strategy for clinical testing in which urine samples are collected after the first dose to establish baseline activity for individual patients, followed by longitudinal monitoring to assess changes in urine biomarker levels at subsequent doses. This strategy is analogous to measuring biomarker “velocity,” which is used with the prostate specific antigen test^70,71 and would aid in reducing variability due to factors like differences in patient baseline and sampling error from a single measurement^72–74.

GzmB expression by itself, however, is not a specific biomarker of ICB response but rather a general biomarker of T and NK cell cytotoxicity that could be elevated under confounding conditions such as immune-related adverse events^75–80 and reactivation of latent viral infections^81–84. As we showed, our GzmB sensor lacked the ability to differentiate mechanisms of resistance that similarly result in loss of T cell cytotoxicity. A similar challenge exists for imaging probes that have thus far targeted GzmB to track ICB response. Therefore, we investigated whether a mass-barcoded set of probes could provide the ability to assess response and resistance to ICB therapy by signature analysis. By transcriptomic analysis, we found that proteases are broadly dysregulated across multiple enzyme families both in tumours that respond to therapy and in tumours that harbour LOF mutations in B2m or Jak1 genes that underpin resistance to checkpoint inhibitors^3,4. These proteases informed the design and selection of a 14-plex library of probes that broadly covered protease cleavage space to provide the ability to generate high-dimensional data by mass spectrometry for classifier training. We observed that although the same library was used in our animal studies, separate subsets of 3 to 5 probes were ranked highest in importance depending on whether the use case was response monitoring (L2-1, L3-7, and L2-8) or stratifying resistance mechanisms (L2-11, L2-20, L2-19, L3-16, and L2-9). These probes were strongly anticorrelated (R = −0.90), and binary classifiers that were trained only on these minimal probe sets recapitulated the diagnostic performance of the entire 14-plex library without reductions in classification accuracy in both retrospective and prospective studies. These observations lend support for a potential future strategy for human testing that involves using the same superset of probes to train separate classifiers for each intended use case. Following classifier validation, a downselection process could then be employed to reduce the number of probes to a minimal set. This strategy may ensure the ability to generate high-dimensional data while reducing regulatory burden associated with the need to test the safety and immunogenicity of separate probe compositions.

Several key areas warrant future study. Our bioconjugation strategy crosslinks peptides to surface amines on lysines, which is the basis for at least four FDA-approved antibody-drug conjugates (Kadcyla®, Mylotarg®, Besponsa®, Lumoxiti®) that are not purified to a particular drug-antibody ratio or isoform. Future formulations could be developed with site-specific chemistries like enzymatic ligation^85,86 or proximity-induced labeling⁸⁷. Transcriptomic analysis of a large set of resistant tumours (i.e., primary, adaptive and acquired) with different mechanisms of action (e.g., absence of antigen presentation, insensitivity to T cells, genetic T cell exclusion³) would further serve to nominate differentially expressed proteases and determine the extent of conservation across cancer types and ICB therapies (e.g., αPD1 versus αCTLA4). Moreover, our observation in B2m^−/− tumours that stem-like cells are GzmB+ suggests that they can be activated by antigen-presenting cells, yet the number of GzmB+ terminally differentiated effector cells did not increase in response to therapy. Additional studies are warranted to elucidate the mechanisms by which stem-like cells differentiate into terminally differentiated effectors in the context of resistant tumours (e.g., WT versus B2m^−/−). Last, differences between preclinical models and human response to ICB therapy should be noted. Although ~500 proteases are considered true orthologues of human proteases⁸⁸, the murine genome encodes more proteases (~630 versus ~550) and granzymes⁸⁹ (~10 versus 5) compared to the human genome. Moreover, given that proteases that are closely related cleave similar substrates such as the matrix metalloproteinases⁹⁰, cathepsins⁹¹ and caspases⁹², our peptide selection process did not exclude substrates with broad selectivity for proteases within a family nor did it involve a negative selection step in vivo to account for background noise from circulation or organ-associated proteases. Therefore, assigning protease specificity to our urine signals will be challenging without developing probes with exquisite selectivity for target proteases, which may be possible with non-natural amino acids^93,94, or mathematical algorithms to deconvolve complex protease signatures^95,96. Looking forward, phase 1 studies are necessary to establish the safety of αPD1-peptide conjugates, which we anticipate to be well tolerated in humans given their composition is similar to protease-activatable masked antibodies⁴⁰ and T cell engagers⁹⁷ that are undergoing clinical efficacy studies. Overall, our results support ICB antibody-conjugated protease sensors as an activity-based biomarker approach to noninvasively track early response and resistance to ICB therapies from urine.

Methods

Animals.

6- to 8-week-old female mice were used at the outsets of all experiments. Pmel (B6.Cg-Thy1a/Cy Tg(TcraTcrb)8Rest/J) and OT1 (C57BL/6-Tg(TcraTcrb)1100Mjb/J) transgenic mice were bred in house using breeding pairs purchased from Jackson Lab. C57BL/6 and BALB/c mice for tumour studies were purchased from Jackson Lab. All animal procedures were approved by Georgia Tech IACUC (protocol #KWONG-A100193).

Peptide substrate synthesis.

FITC-labelled GzmB substrate peptides ((FITC)AIEFDSGc; lower case letters = d-form amino acids) were synthesized by Tufts University Core Facility and used for in vivo formulations. FITC-labelled GzmB substrate peptides with internal quencher ((5-FAM)aIEFDSG(K-CPQ2)kkc) were synthesized by CPC Scientific and used for all in vitro activity assays. Quenched fluorescent peptides for screening candidate substrates ((5-FAM)-{substrate}-(K-DABCYL)-amide)) were synthesized by Genscript or in-house, as described below. Mass-barcoded peptides for multiplexed urinary monitoring (eGVndneeGFFsAr(ANP)GG-{substrate}-GGC, ANP = 3-amino-3-(2-nitrophenyl)propionic acid) were synthesized in-house using the Liberty Blue Peptide Synthesizer (CEM). The peptide synthesis scale used was 0.025 mmol, and low-loading rink amide resin (CEM) was used. Amino acids (Chem-Impex) were resuspended in DMF (0.2 M), as were all synthesis buffers. Activator buffer used was diisopropylcarbodiimide (DIC; Sigma) (0.25 M) and the activator base buffer was Oxyma (0.25 M; CEM) while the deprotection buffer was piperidine (20% v/v; Sigma) supplemented with Oxyma (0.1 M). Crude peptides were purified on 1260 Infinity II HPLC system (Agilent) until a purity of 80% was achieved. Peptide mass and purity were validated by LC-MS (Agilent) and Autoflex MALDI-TOF mass spectrometer (Bruker).

Antibody-peptide conjugation.

Conjugation was performed by cross-linking antibodies and peptides using the heterobifunctional crosslinker succinimidyl iodoacetate (SIA; Thermo), which contains amine-reactive NHS ester to label antibody lysines and sulfhydryl-reactive iodoacetate to couple peptide cysteines. Free αPD1 (kind gift of Dr. Gordon Freeman, Dana-Farber, with the help of Dr. Rafi Ahmed, Emory University, clone 8H3; BioXCell, clone 29F.1A12), αCTLA4 (BioXCell; clone 9H10), and matched isotype control (BioXCell) antibodies were first reacted to SIA for 2 hours at room temperature (RT) in the dark, and excess SIA were removed by buffer exchange using Amicon spin filter (30 kDa, Millipore). Cysteine-terminated peptides were mixed with mAb-SIA and reacted overnight at RT in the dark to obtain mAb-peptide conjugate. The conjugates were purified on a Superdex 200 Increase 10–300 GL column using AKTA Pure FPLC System (GE Healthcare). Endotoxin was removed from the samples by phase separation with Triton X-114 (Sigma) at 2% final volume ratio⁹⁸. Final endotoxin concentrations were quantified by Pierce LAL Chromogenic Endotoxin Assay Kit (Thermo). Protein concentrations were determined by Pierce Protein Assay Kit (Thermo). Conjugates were buffered exchanged into PBS and sterile filtered before in vivo usage. Conjugation ratios (i.e., peptide:antibody ratio) of fluorescently labelled peptides were determined by UV/Vis spectroscopy (NanoDrop; Thermo) using ε (FITC, 490 nm) = 68,000 M⁻¹ cm⁻¹ and ε (Ab, 280 nm) = 210,000 M⁻¹ cm⁻¹ with subtraction of FITC absorbance at 280 nm using correction factor of 0.3. PAR was controlled by varying reaction stoichiometry of SIA (molar ratio between 5 and 50) and peptide (10 and 100). SIA and peptide molar ratios of 5 and 10 were selected to generate conjugates with 1–2 peptides per mAb. Conjugation of mass-encoded peptides was validated by MALDI using Autoflex mass spectrometer (Bruker).

PD1 binding.

Binding of αPD1 conjugates to recombinant PD1 ligand was quantified using an ELISA assay developed in house, in which a high protein binding plate was coated with 1 ug/mL of recombinant Mouse PD-1 Protein (R&D, 9047-PD-100). Binding of intact αPD1-GS conjugates was quantified in a sandwich ELISA using the same PD1-coated plate. After sample incubation, αFITC mAb (Thermo, 13-7691-82; 1:800 dilution staining concentration) was used for secondary staining. ELISA development was performed according to well-established protocol⁹⁹.

Recombinant protease cleavage assays.

For in vitro assays, αPD1 was conjugated with GzmB peptide substrates carrying an internal CPQ2 quencher to allow cleavage detection by fluorescent measurements. αPD1-GS (1.3 μM by peptide) was incubated in PBS at 37 °C with fresh mouse serum, murine granzyme B (0.17 μM; Peprotech), human thrombin (13.5 μM; HaemTech), mouse thrombin (12.5 μM; HaemTech), cathepsin B (1.5 μM, R&D), C1r (1.43 μM; Sigma), C1s (1.80 μM; Sigma), MMP9 (0.1 μM, R&D). Sample fluorescence (Ex/Em = 485 nm/528 nm) was measured for 60 minutes using Cytation 5 plate reader (Biotek). Batch-to-batch cleavage velocity was quantified by incubation of three independently synthesized batches of quenched αPD1-GS (0.31 μM) with murine granzyme B (0.35 μM), with interpolation of fluorescence to concentration using a standard curve of 5-FAM. Cleavage activity during T cell killing (k_cat/K_m) was determined by incubation of quenched αPD1-GS at varying substrate concentrations with murine granzyme B (65 pM, quantified by ELISA, Thermo, BMS6029) contained in supernatant from co-culture of activated OT1 T cells and EG7-OVA target cells (5 effectors per target cell), with interpolation of fluorescence to concentration using a standard curve of 5-FAM.

To screen for candidate substrates for the mass-barcoded library, fluorescently quenched peptide substrates (10 μM) were incubated in manufacturer-recommended buffers at 37°C with recombinant proteases (25 nM). Our set of human recombinant proteases included granzyme A, granzyme B, MMP1, MMP3, MMP7, MMP9, MMP13, caspase 1, caspase 3, cathepsin G, cathepsin S (Enzo), human thrombin, human factor Xia (HaemTech), C1r, fibroblast activation protein alpha (FAP), tissue plasminogen activator (tPA), and urokinase-type plasminogen activator (uPA) (R&D systems). Assay buffers were formulated in Tris or HEPES buffers with salts (NaCl, CaCl₂), pH (7.4–8.0 for most, 4.5 for cathepsin S), blocking reagent and detergent (bovine serum albumin, Brij-35, Triton X-100, CHAPS), reducing agent (DTT, EDTA), and stabilizer (glycerol) selected for each protease based on manufacturer recommendations or published conditions. Sample fluorescence (Ex/Em = 485 nm/528 nm) was measured for 180 minutes using Cytation 5 plate reader (Biotek). Enzyme cleavage rates were quantified as relative fluorescence increase over time normalized to fluorescence before addition of protease. Hierarchical clustering was performed in python, using log2 fluorescence fold change at 60 minutes. A positive cleavage event was defined as having fluorescence signal more than 2-fold above background. Correlation analysis with Spearman coefficient was done on the cleavage patterns of all peptide substrates for selection of 14 substrates for library construction. These peptide substrates were paired with isobaric mass reporters based on the GluFib peptide (eGVndneeGFFsAr) and synthesized using Liberty Blue peptide synthesizer (CEM).

Sensing protease activity during T cell killing.

B16-F10 cells (ATCC) were cultured in DMEM supplemented with 10% FBS and 1% penicillin-streptomycin (Thermo). CD8+ T cells were isolated from either OT1 or Pmel (Jackson Labs) splenocytes by MACS using CD8a Microbeads (Miltenyi). Cells were activated by seeding in 96-well plates pre-coated with anti-mouse CD3e (1 μg/ml working concentration, Clone: 145-2C11, BD) and anti-mouse CD28 (2 μg/ml working concentration, Clone: 37.51, BD) at 2×10⁶ cells/ml in RPMI 1640 supplemented with 10% FBS, 100 U/ml penicillin-streptomycin, 1X non-essential amino acids (Gibco), 1 mM sodium pyruvate, 0.05 mM 2-mercaptoethanol, and 30 U/ml hIL-2 (Roche). After 2 days, cells were washed and transferred to untreated culture flasks for expansion. Between day 4 to 6 after activation, activated T cells were washed before being coincubated with 3×10⁴ B16 target cells at various T cell to effector cell ratios. After 48 hours, coculture supernatants were collected for LDH and GzmB measurements by the Pierce LDH Cytotoxicity Assay Kit (Thermo) and GzmB Mouse ELISA Kit (Thermo, BMS6029) respectively. To assess sensor activation during T cell killing, cocultured of T cells and target cells were spiked in with either αPD1-GS, αPD1 conjugated with control peptide (LQRIYK), and unconjugated αPD1. After 48 hours, fluorescence of coculture supernatant were measured using Cytation 5 plate reader (Biotek).

Circulation half-life.

For half-life characterization, unconjugated αPD1 or αPD1-GS (100 μg) was administered i.v. to naïve C57BL/6 mice (Jackson Labs). At several time points following administration, blood was collected into Capillary Tubes (VWR), and serum was isolated by centrifugation. Serum concentrations of unconjugated αPD1 and αPD1-GS were determined by the PD1 binding and intact PD1 ELISA respectively. Half-life of unbound peptide was determined by administering 1 nmol of granzyme B substrate labelled with amine-reactive VivoTag S-750 (VT750; PerkinElmer; full peptide sequence: (k-VT750)GGsIEFDSGGGs(Pra)c; lower case letters = d-form amino acids, Pra = propargylglycine) and collecting serum as described above. Serum concentrations were determined by imaging using Odyssey CLx Imaging System (LI-COR).

Tumour models.

CT26 (ATCC), MC38 (kind gift of the NCI and Dr. Dario Vignali, University of Pittsburgh), and B2m^−/− vs. Jak1^−/− MC38 tumour cells were cultured in DMEM supplemented with 10% FBS and 1% penicillin-streptomycin (Thermo). Cells were grown to ~70% confluence before being trypsinized for tumour inoculation. On the day of inoculation, C57BL/6 and BALB/c mice were shaved and injected s.c. into the left flank with either 1×10⁶ MC38 or CT26 cells respectively. Tumour burden was monitored until average tumour volume, quantified as 0.52 × length × width × depth, was approximately 100 mm³ before initiating treatment. Mice were administered with αPD1 and/or αCTLA4 antibody-sensor conjugates or matched isotype control (100–150 μg/injection) every 3 or 4 days.

Biodistribution of antibody and probe cleavage.

For biodistribution of antibody, αPD1 or matched isotype control was labelled with VT750 (PerkinElmer) by NHS ester labelling to antibody lysines at non-saturating conditions (molar ratio 3:1), then conjugated to GzmB substrate using SIA and administered to MC38 tumour-bearing mice (day 11, 150 μg). After 24 hours, mice were transcardially perfused with PBS, and organs were excised and imaged using Odyssey CLx Imaging System (LI-COR).

For organ localization of probe cleavage, we synthesized a functionalized GzmB substrate (Ac-rrkGGsIEFDSGGGs(Pra)C-CONH2; lower-case: d-amino acid, Ac: acetylated N-terminus, Pra: propargylglycine) in order to introduce a free amine (on lysine) and alkyne (on propargylglycine) to couple quencher near N-terminus and fluorophore near C-terminus, respectively, and a C-terminal reduced cysteine to crosslink to antibody. We included arginines to increase solubility for aqueous bioconjugation reactions and GGS linkers flanking the GzmB substrate (IEFDSG) to reduce effects on GzmB activity by the nearby fluorophore and quencher, using D-amino acids to avoid introducing sites for proteolytic degradation. The N-terminus was acetylated to improve stability to exoproteases. We then coupled a near-infrared quencher (QC1-NHS ester (LI-COR) reacts with amine) to the substrate, cross-linked the substrate to αPD1 or matched isotype control (using succinimidyl iodoacetate to crosslink amines on the antibody with thiol on the substrate), and labelled the substrate with a near-infrared fluorophore (800CW-azide (LI-COR), reacts with alkyne by copper(I)-catalysed alkyne-azide cycloaddition) to produce the activatable GzmB sensor. Specifically, 800CW was coupled to peptide in presence of copper(II) sulphate, L-ascorbic acid, tris(3-hydroxypropyltriazolylmethyl) amine, and aminoguanidine hydrochloride (all reagents from Sigma). After, probe was purified by FPLC as described above. MC38 tumour-bearing mice were first treated with 2 doses of unmodified αPD1 or matched isotype (0.15 mg), then with activatable αPD1-GS or isotype-GS on the third dose. After 24 hours, mice were transcardially perfused with PBS, and organs were excised and imaged using Odyssey CLx Imaging System (LI-COR).

Flow cytometry analysis of intratumoural T cells.

Tumour dissociation and staining for flow cytometry. Less than 1g of murine tumours were enzymatically and mechanically dissociated using Mouse Tumour Dissociation Kit (Miltenyi) and gentleMACS Dissociator (Miltenyi). TILs were then isolated from the single cell suspension using a density gradient with Percoll Centrifugation Media (GE Life Sciences) and DMEM Media (10% FBS, 1% penicillin-streptomycin) at 44:56 volume ratio. TILs were counted with Trypan Blue (Thermo), and approximately 1×10⁶ viable cells per sample were stained for flow cytometry analysis. Cells were first stained for surface markers in FACS Buffer (1x DPBS, 2% FBS, 1 mM EDTA, 25 mM HEPES). Intracellular staining was performed using eBioscience Intracellular Fixation & Permeabilization Buffer Set (Thermo). All antibodies were used for staining at 1:100 dilution from stock concentrations. Stained cells were analysed by LSRFortessa Flow Cytometer (BD).

Antibody clones: CD45 (30-F11), CD8 (53–6.7), CD44 (IM7), PD-1 (29F.1A12), TIM3 (RMT3-23), CD39 (Duha59), CD4 (RM4-5), NK1.1 (PK136), CD19 (6D5), GZMB (GB12). Viability was accessed by staining with LIVE/DEAD Fixable Dye (Thermo).

Urinary detection of therapeutic response and resistance to ICB therapy.

At 3 hours after administration of ICB antibody-sensor conjugates, urine was collected and analysed for noninvasive detection of therapeutic response and resistance. FITC reporters were isolated from urine samples by immunoprecipitation using Dynabeads (Thermo) decorated with αFITC antibody (GeneTex) with elution in acetic acid (5% v/v) and neutralization in Tris buffer (2 M, pH 12). Sample fluorescence was measured by Cytation 5 plate reader (Biotek), and reporter concentrations were determined by using a known FITC ladder. To compare αPD1 and isotype across doses, urine signals were normalized to the average isotype signal on the respective dose. To compare WT and resistant tumours, urine signals on dose 3 were normalized to the baseline signal (i.e., dose 1) for each individual mouse. Concentrations of isobaric mass reporters were quantified by Syneous Health (Morrisville, NC) using UV cleavage of ANP residue and LC-MS/MS. Peak area ratios were quantified by area of reporter to area of known internal standard to normalize between sample runs, then interpolated to concentration using calibration ladders for each reporter.

Cas9 knockout of B2m and Jak1.

CRISPR guide RNA’s were designed to target two exons in either B2m (g1: GACAAGCACCAGAAAGACCA, g2: GGATTTCAATGTGAGGCGGG) or Jak1 (g1: GTGAACTGGCATCAAGGAGT, g2: GCTTGGTGCTCTCATCGTAC) in the Mus musculus GRCm38 genome. Top and bottom guide oligonucleotides were annealed using T4 PNK (NEB) and ligated into the backbone of eSpCas9_PuroR_GFP plasmid (Sigma) using BbsI cut sites and T7 ligase (NEB). 1×10⁵ MC38 cells were transfected with gRNA-ligated eSpCas9 plasmids for 48 hours using TransIT-LT1 transfection reagent (Mirus Bio) in Opti-MEM (Thermo Fisher) and cultured for 3 passages in DMEM supplemented with 10% FBS and 1% penicillin-streptomycin (D10). Selection of transfected cells were done by supplementing culture media with 2 ug/mL puromycin (Thermo Fisher). Cells incubated with B2m-directed guides were stained with anti-mouse H-2Kb (clone AF6-88.5). H-2Kb-negative GFP-positive cells were sorted into single cells on a 96-well plate using FACSAria Fusion (BD Biosciences) and cultured for 2–3 weeks in D10. For cells incubated with Jak1-directed guides, GFP-positive cells were sorted into single cells and cultured for 2–3 weeks in D10. Clones that passed the functional assays for successful deletion of B2m or Jak1 are selected for tumour studies.

In vitro validation of knockout lines.

DNA was isolated from single-cell WT and knockout clones, and a PCR reaction was done to amplify the edited regions within B2m and Jak1 exons. The PCR products were sequenced by Sanger sequencing, and sequencing results were analysed with TIDE (Tracking of Indels by Decomposition) analysis⁴⁴ to confirm knockout efficiency. WT and knockout tumour cells were stained for H2-Kb (clone AF6-88.5) to confirm the functional loss of B2m. WT and B2m^−/− were pulsed with SIINFEKL (30 μM peptide concentration), washed, and coincubated with plate-activated OT1 T cells at 5:1 ratio of effector:target cell. After overnight incubation, cells were washed and stained for CD8 (53–6.7), IFNγ (XMG1.2), and GzmB (GB12). For IFNγ stimulation assay, WT and knockout tumour cells were incubated with recombinant murine IFNγ (Peprotech; 500 EU/mL) for 2 days and stained for surface expression of H2-Kb (AF6-88.5) and PD-L1 (10F.9G2).

Tumour RNA isolation and sequencing.

Mice bearing WT, B2m^−/−, Jak1^−/− MC38 tumours were treated with either αPD1 or IgG1 (100 ug) every 3 or 4 days. After the third administration, approximately 50 mg of tumours were dissected and rapidly frozen with dry ice and IPA. Frozen tumour samples were homogenized in MACS M Tubes (Miltenyi) using the MACS Dissociator (Miltenyi). Total RNA was isolated from the homogenate using the RNeasy Plus Mini Kit (Qiagen). Library preparation with TruSeq RNA Library Prep Kit (Illumina) and mRNA NGS sequencing (40×10⁶ paired end read) were performed by Admera Health (South Plainfield, NJ).

RNA-Seq data mapping and visualization.

Raw FASTQ reads passing quality control (FastQC v0.11.2) were aligned on the mm10 reference genome using STAR aligner (v2.5.2a) with default parameters. Aligned fragments were then counted and annotated using Rsamtools (v3.2) and Cufflinks (v.2.2.1) after a ‘dedup’ step using BamUtils (v1.0.11). t-SNE embedding results were performed in sklearn (v0.23.1) using all murine genes. Heat maps were plotted with seaborn’s (v.0.9.0) clustermap function. Rows were gaussian normalized, and the dendrograms shown for clustering come from hierarchical clustering using Euclidean distance as a metric.

Differential expression and gene set enrichment analysis.

Differential expression was performed using the edgeR package (v3.24.3) in R using the exactTest method with tagwise dispersion. For mouse data (GSE192796), TMM normalization considering mice in all treatment groups was performed to remove library size effect through the calcNormFactors function. For human data (GSE91061¹¹), TMM normalization was performed using the two groups being compared. For both datasets, differential expression was performed on Ensembl IDs before mapping to gene names. Then the identified differentially expressed genes were filtered by a list of extracellular and transmembrane endopeptidases queried from UniProt. Gene set enrichment analysis (GSEA) was performed using the fgsea package (v1.8.0) in R. To rank genes, differential expression analysis was first performed on the entire gene set. Genes are then ranked by -sign(logFC)*log(pval). Hallmark gene sets (MSigDB) were used for all GSEA analyses.

Urinary differentiation of ICB resistant mechanisms.

Random forest was used to train classifiers based on urinary mass reporter signals that differentiate therapeutic response and stratify resistant mechanisms. All urine signals were normalized on a per mouse basis by signals on the first dose. Response monitoring classifiers were trained on mean-normalized reporter concentration whereas resistance stratifying classifiers were trained on reporter concentration. For each classification task, we used five-fold cross validation by randomly leaving out one-fifth of the samples as the test set and using the remaining samples as training sets. This process was repeated 100 times, and the final performance was generated as the average area under the ROC curve (AUROC) for all train-test results. Comparisons between diagnostic performance was done by two-way paired t-test. For prospective classification, a single random forest response monitoring classifier for each dose (dose 2 and dose 3) was trained on reporter concentration. These classifiers were locked and tested against urine signals from a new cohort of mice on matching doses. Final performance was generated from a single AUROC for each dose.

Software and Statistical Analysis.

Graphs were plotted and appropriate statistical analyses were conducted using GraphPad Prism (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; central values depict the means, and error bars depict s.e.m.). Measurements were taken from independent samples, using biological replicates when possible. Whole organ images were analysed using Image Studio (LI-COR). Flow cytometry data were analysed using FlowJo X (FlowJo, LLC). Power analyses were performed using G*Power 3.1 (HHUD).

Reporting Summary.

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Extended Data

ED Fig. 1. Synthesis strategy for antibody-sensor conjugates for biodistribution of antibody and GzmB cleavage.

a, Synthetic scheme of antibody-sensor conjugate labelled with near infrared fluorophore VT750 to determine antibody biodistribution. b, Synthetic scheme of NIR activatable GzmB probe, consisting of fluorescently quenched GzmB substrate conjugated to αPD1 antibody.

ED Fig. 2. Biodistribution of antibody-sensor conjugates and GzmB cleavage.

a, Schematic of antibody-sensor conjugate labelled with near infrared fluorophore VT750 to determine antibody biodistribution. b, Representative whole organ images and c, quantification of distribution of fluorescent signal in tumour and major organs (dLNs, tumour draining lymph nodes) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ns = not significant, n = 5 biological replicates, error bars depict s.e.m.). d, Schematic and in vitro cleavage of NIR activatable GzmB probe labelled with quenched near infrared fluorophore 800CW to localize site of probe activation (FC, fold change; ****P < 0.0001, n = 3 technically independent wells, error bars depict s.e.m.). e, Representative whole organ images and f, quantification of reporter fluorescence after treatment with Iso-GS or αPD1-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, *P = 0.040, n = 4–5 biological replicates, error bars depict s.e.m.).

ED Fig. 3. Diagnostic performance of αPD1-GS in ICB nonresponsive models.

a, Average and b, individual tumour growth curves of CT26 tumour bearing mice treated with either αCTLA4-GS or matched IgG2 isotype control (Iso-GS) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ns = not significant, n = 10–11 biological replicates, error bars depict s.e.m.). Black arrows denote the treatment time points. c, Normalized urine fluorescence of mice with CT26 tumours after each administration of αCTLA4-GS or Iso-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ns = not significant, n = 10–11 biological replicates). d, Average and e, individual tumour growth curves of CT26 tumour bearing mice treated with αPD1-GS or matched IgG1 isotype control (Iso-GS) (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ns = not significant, n = 6 biological replicates, error bars depict s.e.m.). Black arrows denote the treatment time points. f, Normalized urine fluorescence of mice with CT26 tumours after each administration of αPD1-GS or Iso-GS (two-way ANOVA with Sidak’s post-test and correction for multiple comparisons, ns = not significant, n = 6 biological replicates).

ED Fig. 4. Differential expression analysis of proteases in ICB response and resistance.

a, Heatmaps showing row-normalized expression (FPKM) of proteases differentially expressed between αPD1-treated WT tumours and IgG1-treated controls (n = 5 biological replicates). b, (Left) Volcano plots summarizing differentially expressed proteases between αPD1-treated B2m^−/− and Jak1^−/− MC38 tumours (n = 5 biological replicates). The threshold for differentially expressed genes (opaque dots) was defined as P value ≤ 0.05 and |log2(fold change)| ≥ 1. (Right) Heatmaps showing row-normalized expression (FPKM) of proteases differentially expressed between B2m^−/− and Jak1^−/− MC38 tumours (n = 5 biological replicates). c, (Left) Volcano plots summarizing differentially expressed proteases between human tumours from responders (CR + PR) and non-responders (PD) (n = 5 independent patient samples). The threshold for differentially expressed genes was defined as P value ≤ 0.01 and |log2(fold change)| ≥ 1. (Right) Heatmaps showing row-normalized expression (FPKM) of proteases differentially expressed between human tumours from responders (CR + PR) and non-responders (PD) (n = 5 independent patient samples).

Data availability

The main data supporting the results in this study are available within the paper and its Supplementary Information. The sequencing datasets generated from murine tumours and analysed from human samples (Riaz, N. et al. Cell 2017) have been deposited in NCBI’s Gene Expression Omnibus and are accessible through GEO Series accession numbers GSE192796 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE192796) and GSE91061 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE91061), respectively. Other data generated and analysed during the study are available from the corresponding author on reasonable request. Source data are provided with this paper.