Proteolysis Activity Mapping and Substrate Discovery Platform for Identifying Tumor-Activated Biosensors

Abstract

Dysregulated extracellular proteolytic activity is a prominent hallmark of cancer and can thus be exploited for tumor detection and therapeutic development. However, the discovery of tumor-responsive probes has been hindered by the lack of methods to directly screen proteolytic events in specific tissue samples. Here, we report PSurf, a platform that enables the identification of tissue-specific protease sensors with tissue specimens. Through differential selection of tumor-specific sequences over healthy tissue, PSurf identifies context-specific tumor-activated probes that precisely distinguish metastatic lesions in lung tissue slices. Using these substrates, we engineered nanobody-targeted biosensors that release urinary reporters upon tumor-specific cleavage in vivo, enabling precise noninvasive tumor detection in a murine lung metastasis model. PSurf provides a foundation for developing conditionally activated agents through tissue-specific activity mapping and probe discovery.

Introduction

Dysregulation of extracellular tumor microenvironment (TME) proteases is a well-established hallmark of cancer, driving fundamental processes such as angiogenesis, tumor growth, and metastasis. These enzymes play a pivotal role in modulating the TME by mediating matrix degradation, receptor activation, immune signaling, thereby influencing tumor progression and therapeutic response within a specific disease context. Harnessing protease activities to sense and respond to specific TMEs has therefore emerged as a promising strategy for detecting and treating disease with precision¹. Local proteolytic activation has been utilized to develop imaging agents^{2, 3} and activity-based diagnostics⁴ for early cancer detection and treatment monitoring. In addition, conditional activation of therapeutics can aid the balance between potency and toxicity^5-7. This design principle has been implemented to create prodrugs⁸, pro-antibodies, pro-cytokines⁹ and antibody-drug conjugates (ADCs)^{10, 11}, several of which have advanced to regulatory approval or clinical trials^{12, 13}. Tailored proteolytic profiling holds promise to guide clinical intervention for tumor metastasis, the leading cause of cancer mortality, through the development of activatable tools to precisely detect, monitor, and treat malignant lesions.

Although significant progress has been made over the past decade in characterizing tumor-associated proteolytic activities, most studies have focused on identifying substrates of individual recombinant proteases selected based on transcriptomic data^14-17. However, with more than 500 proteases encoded in the human proteome¹⁸ and only a fraction available in active recombinant forms, these approaches interrogate a limited fraction of the protease family¹⁹. Moreover, studies based on purified enzymes do not fully capture the integrated protease activities occurring within the TME, where multiple proteases function simultaneously and are tightly regulated by cofactors, prodomains, and natural inhibitors such as tissue inhibitors of metalloproteinases (TIMPs)^20-22. Synthetic peptide libraries have contributed substantially to defining protease specificity, yet their scale is often constrained by cost and size, with most containing only hundreds of peptides^23-29. Mass spectrometry profiling approaches have further expanded our ability to study protease biology^{17, 25, 30-34}. In particular, N-terminomics strategies, including terminal amine isotopic labeling of substrates (TAILS), High-Efficiency Undecanal-Based N-Termini Enrichment (HUNTER), and subtiligase-mediated N-terminal labeling methods, enable mapping of physiological proteolytic processes^35-38. In addition, multiplex substrate profiling by mass spectrometry (MSP-MS) integrates synthetic peptide libraries with MS to define protease cleavage motifs in complex protease mixtures^{39, 40}. While powerful, these technologies are primarily designed to map endogenous proteolysis events or extract specificity rules rather than directly generate synthetic substrate sequences optimized for translational applications. Thus, there remains a need for platforms that can directly discover and screen for synthetic protease substrates within tissue environments, bridging protease activity profiling with substrate engineering.

Here, we report a tissue proteolytic mapping and substrate discovery platform termed PSurf (Proteolysis Signature discovery via yeast surface display). Distinct from acellular platforms, PSurf harnesses genetically encoded yeast surface display coupled with a differential selection protocol to uniquely enable identification of context-specific proteolytic events and synthetic optimized proteolysis substrates in tissue environments. The PSurf library, comprising over 10⁶ randomized 7-amino-acid sequences, offers 100- to 10,000-fold greater sequence coverage than existing synthetic substrate libraries. We first demonstrated PSurf’s capabilities using recombinant cathepsin B, identifying both canonical and novel substrates with distinct pH sensitivities. We subsequently applied PSurf to murine models of lung metastasis and identified tumor-activated substrates that spared healthy tissues. From these selections, we developed fluorogenic probes that outperform conventional metalloproteinase-activatable substrates. One substrate was further engineered into a membrane-inserting tissue imaging probe, enabling localization of disseminated lung tumor nodules both on tissue and in vivo. Furthermore, DNA-barcoded tumor-trafficking nanobodies fusions to the PSurf substrates enabled systemic, multiplexed proteolytic activity assessment from biofluids, and, when combined with CRISPR-Cas-mediated urinary readout, allowed rapid disease classification distinguishing tumor-bearing from healthy animals (Fig. 1a).

Figure 1. Design, construction, and validation of the yeast library.

(a) Workflow schematic for protease profiling using tissue samples. Organs and tumors are harvested from animal models, and extracellular fluid samples are subjected to substrate profiling using the PSurf yeast display library. Differential selection eliminates non-specific cleavage sequences and enriches tumor-specific substrates. Selected substrates are validated through in vitro, ex vivo, and in vivo assays, resulting in novel tissue-activated substrate sequences. (b) Diagram of the yeast display strategy for tissue protease analysis. Peptides flanked by epitope tags are displayed on the yeast surface. Protease activity removes one tag, and yeast are stained with detection antibodies for analysis and sorting. Customized fluorescence-based sorting-based, multi-round selection strategies generate an enriched yeast pool, which is amplified and analyzed by next-generation sequencing (NGS) to identify tissue-sensitive substrates. (c) Tag pair selection using a TEV substrate. Schematics and flow cytometry results show three tag pairs flanking a TEV substrate peptide treated with TEV protease over time. (d) Validation of the peptide display strategy with TEV substrates. V5- and HA-tag-flanked TEV substrates were incubated with wild-type TEV protease for 24 hours. Yeast display detected sequence-dependent cleavage, with the non-canonical sequence ENLYFQV showing reduced cleavage compared to the canonical ENLYFQG. Data represent the mean of n = 2 replicates. (e) A TEV mutant (L56V, S135G, S219V) displayed similar activity toward both canonical and non-canonical substrates. Data represent the mean of n = 2 replicates. (f) Design and validation of the randomized PSurf library flanked by V5 and HA tags. The DNA logo shows base frequencies per position from NGS, consistent with the expected NNK codon distribution. Figs. 1a, 1b included graphics created with BioRender.com.

Collectively, our findings establish PSurf as a powerful chemical biology platform for customizable, high-resolution profiling of proteolytic activity and substrate discovery in biological samples. By complementing existing methods that focus on native protease substrates and recombinant protease studies, PSurf holds the potential to advance understanding of extracellular proteolytic networks and support the design of protease-activation strategies for precision diagnostics and therapy.

Results

Library design and construction

PSurf employs a double-tagged peptide display approach, where each peptide displayed on the yeast surface is flanked by two epitope tags recognized by fluorescently-labeled antibodies (Fig. 1b). When proteolysis occurs, one tag is cleaved off, allowing tracking of cleavage events by measuring fluorescence change through cell sorting. Yeast populations displaying peptides with specific proteolytic activities are sorted for next-generation sequencing (NGS) analysis or cultured for further rounds of selection.

To optimize the tagging strategy, we used the Tobacco Etch Virus (TEV) protease substrate, ENLYFQG. The substrate was fused to the C-terminus of Aga2p for yeast surface display and three tag pairs were compared: HA/cMyc, HA/V5, and V5/HA (N-/C-terminal orientations) (Fig. 1c). Yeast cells displaying the peptides were incubated with TEV protease and analyzed using flow cytometry. Among the designs, the V5/HA tag combination exhibited homogeneous display and dose- and time-dependent reduction in HA signal upon proteolysis (Fig. 1c, d), and was selected as the optimal design for library construction.

To assess whether the display strategy can capture protease substrate preferences, we compared the canonical TEV protease substrate (ENLYFQG) with a non-canonical substrate (ENLYFQV), observing significantly reduced cleavage of the latter, as expected⁴¹ (Fig. 1d). In contrast, cleavage was comparable between substrates when using a TEV protease mutant (TEV L56V, S135G, S219V), which is known to have broader substrate tolerance⁴² (Fig. 1e). We next used NNK degenerate codons (N: A/T/C/G, K: G/T, Extended Data Fig. 1a, b) to construct the PSurf library displaying randomized seven-amino-acid peptides, allowing exploration of a large sequence space, including sequences not found in native proteins⁴³. The peptide length of seven amino acids was chosen to balance coverage of central recognition motifs and flanking residues while maintaining sufficient sequence diversity within the practical size constraints of a yeast library⁴⁴. A yeast surface display plasmid was transformed into Saccharomyces cerevisiae, along with the amplified DNA library containing homologous flanking sequences to enable recombination, generating a library of 3x10⁸ transformants. 58% of the population expressed both epitope tags as detected by flow cytometry (Extended Data Fig 1c). NGS analysis of 5x10⁶ transformants showed nucleotide and amino acid frequencies closely aligned with expectations (Fig. 1f, Extended Data Fig. 1d-e, Supplementary Table 1).

To understand the population of yeasts lacking the C-terminal HA tag, we examined this population and identified the two primary contributions as stop codons introduced by NNK usage (Extended Data Fig. 1e) and substrates of yeast proteases (Extended Data Fig. 3a, b). 34% of this population encoded full-length peptides with strong enrichment of LxxR and RR motifs, (Extended Data Fig. 3b). These motifs match the known cleavage sequences of the yeast protease Kex2, indicating the peptide cleavage occurred during secretion to the cell surface and demonstrated that PSurf can unbiasedly reveal endogenous protease substrate preferences (Extended Data Fig. 3a)³². The library was further sorted to enrich for the uncleaved peptide population, generating the final PSurf library used for subsequent proteolysis profiling studies (Extended Data Fig. 1d, 1e right panel).

Identifying cathepsin substrates and pH responses

We first applied the library to recombinant cathepsin B (CTSB), a prognostic biomarker associated with multiple solid tumors⁴⁵, that has widely been developed as a protease trigger for payload release in ADCs and as a therapeutic target for inhibitor development^{46, 47}. An input library size of 2x10⁷ was selected to preserve high sequence coverage while remaining compatible with the experimental throughput imposed by cell sorting speed. The library was initially treated with the enzyme at pH 4.4, a condition expected to support maximal CTSB activity. The selected population was subsequently subjected to more stringent conditions with reduced enzyme concentration and incubation time (Fig. 2a, Extended Data Fig. 2a, b, Supplementary Fig. 1). In alignment with previous reports, NGS analysis revealed gradual enrichment of the leucine-valine-glycine (LVG) motif across successive rounds of selection ^{48, 49}.

Figure 2. Identifying cathepsin substrates.

(a) Selection schematic for CTSB substrates identification. (b) Flow cytometry plots and sequence logos from PSurf selection with hCTSB at pH 4.4. A strongly enriched LVG cleavage motif was observed under stringent selection conditions (10 nM hCTSB). Sorting gates are indicated by dashed lines. (c) Flow cytometry plots and sequence logos showing substrate cleavage motifs at pH 6.4 and 7.4. Unlike the pH 4.4 results, motifs such as RR and VR were strongly enriched. Sorting gates are indicated by dashed lines. (d) Z-score plots showing dipeptide enrichment profiles for hCTSB at different pH levels. The y-axis represents the first amino acid; the x-axis the second. Distinct dipeptide motifs were enriched under each pH condition. (e) On-yeast cleavage activity of hCTSB with newly identified substrates. RRLYAFL and RMLVGFV were highly sensitive to hCTSB at both pH 4.4 and 6.4; LVGVLLF showed significantly selective activity at pH 4.4, and VRFLAAA at pH 6.4. Yeast displaying these peptides were incubated with 70 nM hCTSB for 72 hours. Cleavage activity is shown as the negative log of HA-AF488 signal normalized to peptide expression (V5-AF647 signal) and to untreated controls. Data represents the mean of n = 3 replicates; error bars indicate standard deviation. STSSTSS served as a no-activity control and GRLVGFD⁴⁹ served as a positive control. Unpaired two-tailed t-test was used for pH-dependent activity comparisons ****P < 0.0001 for LVGVLLF. *P=0.0456 for VRFLAAA. (f) Schematic of the probe and cleavage assay comparing RRLYAFL to the conventional substrate KSGGGK. A synthetic peptide labeled with FAM/CPQ2 was used; cleavage releases the fluorophore, generating a fluorescence signal. RRLYAFL showed enhanced cleavage by hCTSB. Data represents the mean of n = 3 replicates.

PH sensitivity is an important characteristic of cathepsins, enabling them to modulate activity across cellular compartments. We next investigated CTSB’s pH-dependent substrate preferences by performing similar selection cycles at pH 5.4, 6.4, and 7.4 (Fig. 2c, Extended Data Fig. 2c). Compared to the pH 4.4 selections, smaller library population was cleaved at higher pH values, reflecting the expected decrease in CTSB activity ⁵⁰. NGS results indicated that as pH increased, enrichment of the LVG motif decreased, while RR and VR motifs became more prominent (Fig. 2d, Extended Data Fig. 2d, e). RR and VR has previously been identified as a sensitive hCTSB substrate at pH 7.0 and 6.4, in line with our findings⁵¹. Notably, a VR-derivative peptide, Valine-Citrulline (Val-Cit), is one of the most widely used ADC linkers⁵². Citrulline is a deaminated form of arginine. Val-Cit is preferred over VR in ADCs because it is more stable in human plasma while retaining robust cleavage activity by cathepsins.⁵³ These results highlight the ability of PSurf to identify context-specific enzyme cleavage features.

Since cathepsins primarily promote tumor progression in the extracellular space, where pH markedly above the lysosomal pH but often lower than normal tissue, there is a strong interest in identifying novel cathepsin substrates or peptide-based inhibitors that function at tumor microenvironmental pH to better target tumor-associated cathepsins. We nominated four peptides from the pH 6.4 screen to be displayed individually on yeast and tested their CTSB sensitivity (Fig. 2e, Supplementary Table 2). Consistent with the motif enrichment results, the VR-containing peptide VRFLAAA showed selectivity at pH 6.4, while the LVG-containing peptide LVGVLLF showed selectivity at pH 4.4 (Fig. 2e). Peptides containing motifs enriched at both pH 4.4 and 6.4 maintained activity across pHs, as seen with peptide RMLVGFV (motif RM) and peptide RRLYAFL (motif RLY) (Fig. 2b, c). From our analysis, the novel substrate RRLYAFL emerged as the most enriched sequence in the pH 6.4 NGS dataset, representing 12% of all sequences. To assess its activity relative to known cathepsin substrates, we designed synthetic fluorogenic substrates, consisting of 7 amino acids flanked by a fluorophore and quencher, where cleavage leads to time-dependent fluorescence increase (Fig. 2f, left). Testing this probe alongside a CTSB-responsive probe, the newly identified substrate exhibited faster cleavage kinetics than the known CTSB substrate at an acidic extracellular pH characteristic of the tumor microenvironment (6.4) (Fig. 2f). These results demonstrate that our PSurf library can effectively identify both known and novel cathepsin substrates.

Samples derived from living organisms exhibit greater complexity than purified proteases due to the presence of diverse protease families, endogenous inhibitors, competing substrates, metal ions, and other components that modulate protease activity. To evaluate the system’s ability to characterize biologically complex samples, we applied PSurf to a mouse upper intestinal washout sample, which resulted in cleavage of 24% of displayed peptides in 24 hours and 10% of peptides in a subsequent 4-hour selection cycle (Extended Data Fig. 3c-e, Supplementary Fig. 2). In agreement with the expected high concentrations of trypsin in this sample, NGS analysis revealed clear arginine enrichment in the cleaved pool and further arginine enrichment across all seven amino acid positions in the more stringent selection. These results demonstrate that PSurf can effectively capture specific proteolytic activities in biological samples and that increased selection stringency enhances enrichment for tissue-sensitive substrates, consistent with observations in the CTSB experiments.

Identify tumor-selective sequences through differential selection

We next applied PSurf to tumor samples. We harvested tumor bearing lungs and healthy organs from two lung metastasis mouse models derived from CT26 cells (colorectal cancer) and B16-F10 cells (melanoma). Both models developed pulmonary metastatic nodules within 10-14 days after intravenous injection. We extracted the soluble fraction of the ECM (sECM) from healthy tissues (liver, lung, kidney, and spleen) and tumor lesions in the lung based on a filter centrifugation method (Fig. 3a, Extended Data Fig. 4a)⁵⁴. In some contexts, we prepared tissue extracellular samples (ECS) by gentle mechanical dissociation and collected extracellular proteins⁵⁵. Both tissue processing methods were carefully performed to minimize altering compartmentalization and regulation or cell lysis, to represent tissue extracellular environments.

Figure 3. Profiling of tissue extracellular fluids using PSurf.

(a) Schematic and protocol overview for isolating and profiling tissue-derived extracellular fluids. (b) Z-score plots showing single amino acid enrichment and depletion across tissues and tumors (liver, lung, kidney, spleen, CT26 CRC lung metastases, and B16 lung metastases). Residues with a significant increase compared to pooled mean of Z-scores are framed in red, indicating enrichment. Residues with a significant decrease compared to the pooled mean are framed in blue, indicating depletion. Unpaired two-tailed independent t-tests were performed and p-values are provided in Supplementary Table 5. (c) Frequencies of enriched amino acids in the input library and each sample type. (d) Schematic of the differential selection strategy. The tumor-enriched population was sequentially screened against liver, lung, kidney, and spleen samples. After each round, the non-cleaved pool was carried forward to remove substrates cleaved by healthy tissues. The final non-cleaved pool was screened against tumor extracellular fluid to further enrich for tumor-sensitive substrates. (e) IceLogos showing per-position amino acid enrichment and depletion before and after differential selection, normalized to the initial library. The pool selected from global profiling was enriched in leucine and phenylalanine, while the final pool following differential selection displayed greater sequence diversity. IceLogos were generated using the IceLogo web server. Figs. 3a, 3d included graphics created with BioRender.com.

The PSurf library was incubated with the extracted sECM samples for 24 hours, resulting in substrate cleavage levels of 11% and 10% in CT26 and B16 tumor-bearing lung, respectively, and 7% in healthy lung (Fig. 3a, Extended Data Fig. 4b, Supplementary Fig. 3). Healthy spleen, kidney, and liver exhibited cleavage of 8%, 9%, and 10%, respectively. The buffer control led to 6% cleavage, reflecting residual pre-cleaved peptides by yeast, though the mean fluorescence intensity indicated a reduced degree of cleavage compared to tissue-derived samples. The cleaved fractions were harvested, pre-cleaved peptides were removed by sorting, and the resulting yeast pool was then subjected to a second round of selection with tissue sECM samples and analyzed by NGS sequencing.

Interestingly, across both healthy tissue and tumor samples, NGS revealed strong enrichment of sequences containing phenylalanine (Phe), leucine (Leu), and other hydrophobic residues (Fig. 3b, c; Extended Data Fig. 4c, d), characteristic of substrates for several serine proteases⁵⁶. For example, GZMH and CTSG possess chymotrypsin-like activity and preferentially cleave after large hydrophobic residues like Phe, as evidenced by their efficient cleavage of the Phe-Leu-Phe synthetic substrate^{57, 58, 59}. Other secreted serine proteases, such as KLK7⁶⁰, NE, and PR3⁶¹, also exhibit substrate preferences for Leu or Phe. In addition, some MMPs (MMP1, 2, 3, 7, 8, 9, 12, 13, 14) have been reported to prefer leucine at the P1′ position¹⁶. Thus, the observed enrichment of Phe/Leu-rich sequences may reflect the integrated activity of proteases acting within the sECM of healthy and tumor tissues. These findings suggest that tissue samples may share common cleavage signatures driven by conserved protease substrate preferences.

To distinguish tumor-specific substrates from those commonly cleaved in both tumor and healthy tissues, we leveraged yeast display’s capacity for repeated positive and negative selection, a strategy used in nanobody and antibody discovery to isolate binders targeting specific epitopes or conformational states of antigens^62-64 (Fig. 3d). We initiated this differential selection process using the cleaved population obtained from the initial tumor sECM screens (Extended Data Fig. 5a). To deplete substrates also cleaved by healthy sECM, this pool was sequentially incubated with sECM from healthy liver, lung, kidney, and spleen, collecting the non-cleaved fraction at each step and culturing for the next selection step. Finally, the remaining non-cleaved pool was exposed to tumor-derived sECM, and the resulting cleaved fraction was collected, enriching for substrates highly sensitive and specific to tumor-associated proteolytic activity (Supplementary Fig. 4).

IceLogos⁶⁵ generated from cleaved and non-cleaved populations revealed distinct tissue-specific cleavage features (Extended Data Fig. 5b) and clear differences before and after differential selection (Fig. 3e). While Leu and Phe were highly enriched in the input sequences from both tumor models, these residues were depleted in the output population, where enriched residues included arginine, proline, asparagine, glycine, alanine, and aspartate (Extended Data Fig. 6). We selected 10 substrate sequences (Supplementary Table 3) based on several selection criteria: ‘Top hit’ and ‘2^nd hit’ are sequences that had the highest NGS reads count for each treatment; ‘Per-position enrichment’ describes sequences comprising the most frequent amino acid in each position; ‘Top unique hit’ is a sequence with the highest NGS reads count not identified in other tissues, and ‘Shared sequence’ was present in all samples. We then tested their tissue specificity on yeast and nominated four peptides preferred by tumor tissue for further biochemical and in vivo analyses (Extended Data Fig. 5c).

Biochemical characterization of PSurf-identified substrates

To further characterize the PSurf-identified sequences, we synthesized these peptides as fluorogenic probes. Lung ECS from mice bearing CT26 metastases and healthy animals were incubated with the probes to assess differences in substrate cleavage rates (Fig. 4a, b). Lung samples from lipopolysaccharide (LPS)-treated mice were also included to distinguish tumor-specific from inflammation-associated proteolytic activity. Four PSurf-identified substrates were benchmarked against the previously reported MMP substrates PVPLSLVM and PLGLRSW^{66, 67}, which contain the conserved MMP cleavage motif PXXL⁶⁸, as well as AIEFSD, a known substrate of granzyme B⁶⁹. The kinetic fluorescence assays revealed that all four PSurf-identified probes were cleaved more rapidly in tumor samples than in healthy lung samples and exhibited minimal cleavage in inflamed lung tissue (Fig. 4c). When applied to blood samples from tumor-bearing and healthy mice, we did not detect differences in fluorogenic probe activation (Extended Data Fig. 7a), suggesting these probes are sensing tissue-level proteolytic changes rather than systemic changes in circulation.

Figure 4. PSurf-derived substrates show tumor-specific cleavage distinct from canonical MMP substrates.

(a) Schematic of fluorogenic probes report proteolysis by tissue ECS. (b) Kinetic fluorescence traces; DNLGRAF shows higher cleavage in CT26 tumor-bearing versus healthy lung ECS (mean of n = 3). (c) Cleavage rates for four PSurf probes, MMP probes (PVPLSLVM and PLGLRSW), and GZMB probe (AIEFDS) in CT26 tumor-bearing, inflamed, and healthy lung ECS. Rates are slopes of linear trace regions (mean ± s.e.m.; healthy and CT26, n = 7; inflamed, n = 3). Two-way ANOVA with Dunnett’s test. p-values are provided in Supplementary Table 5. (d) DNLGRAF cleavage with protease-class inhibitors (150 μM; mean, n = 2). (e) Comparison of DNRGRAF versus DNLGRAF (same DNLGRAF data as in c; mean ± s.e.m., n = 7). p-values are provided in Supplementary Table 5. (f) Alanine scan of DNLGRAF on-yeast after 24 h CT26 ECS treatment (normalized means ± s.d., n = 3; one-way ANOVA, P < 0.05; Supplementary Table 5). (g) Mass spectrometry of CT26 ECS, intact DNLGRAF, and DNLGRAF after CT26 ECS treatment showing three cleavage fragments with corresponding mass peaks. For all panels, significance is indicated with asterisk corresponding to ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05. Fig. 4f included graphics created with BioRender.com.

We next sought to identify which tumor-associated protease families contribute to the substrate cleavage. Transcriptomic analysis of patient samples⁷⁰ revealed upregulation of proteases from multiple families in colorectal adenocarcinoma and skin cutaneous melanoma compared to respective healthy tissues (Supplementary Fig. 5a). Comparison of dysregulation in these two cancer type highlights both shared upregulated proteases, including GZMB, MMP1, MMP9, and unique signatures, including upregulation of MMP3, MMP7 in CRC. We assessed the cleavage specificity of the substrates using recombinant catalytic domains of MMP1, 2, 3, 7, 8, 9, 10, 11, 12, 13, 14 (Extended Data Fig. 7b). As expected, PVPLSLVM was cleaved by various MMPs. In contrast, three of the PSurf-identified substrates, DNLGRAF, PFLYLFG, and AYGRLYR, were not cleaved by any of the tested MMPs nor serine proteases upregulated in CRC, including FAP¹⁴, GZMB⁶⁹, and PCSK9⁷¹, nor by the cysteine protease CTSB, enzymes that have frequently been targeted in recombinant enzyme-based substrate discovery. Among the four new substrates, PARMLHI was the only sequence partially cleaved by certain MMPs (Extended Data Fig. 7b). The discovery of these substrates thus expanded the repertoire of tumor-associated protease sensing beyond the conventional set of enzymes.

To determine which protease families contribute to substrate cleavage, we profiled probe sensitivity using a panel of protease inhibitors, each tested at 150 μM. This analysis did not indicate selective inhibition by a single protease or protease class. Instead, cleavage was most strongly reduced by inhibitors targeting serine and metalloprotease families, particularly when applied in combination (Fig. 4d, Extended Data Fig. 7c, Supplementary Fig. 5b). Specifically, cleavage of DNLGRAF and AYGRLYR was strongly inhibited by the broad-spectrum serine and metalloprotease inhibitor epigallocatechin gallate (EGCG) and by the serine protease inhibitor nafamostat, while showing minimal response to cysteine protease inhibitors. Dose-response analysis across multiple inhibitor concentrations yielded IC₅₀ values consistent with previously reported ranges (19 μM for nafamostat and 57 μM for EGCG) (Supplementary Fig. 5b)^{72, 73}. PARMLHI, consistent with its partial cleavage by MMPs (Extended Data Fig. 7b) was inhibited by MMP inhibitors marimastat and phosphoramidon. In contrast, PFLYLFG was more strongly inhibited by cysteine protease inhibitors, indicating involvement of proprotein convertases and suggesting distinct protease sensitivity compared to the other substrates. These findings indicate that tissue selection identified substrates are responsive to multiple tumor-associated proteases and represent distinct classes of protease-reactive probes compared to conventional MMP substrates.

To investigate how amino acid composition affects substrate sensitivity, we compared substrate DNLGRAF to a substrate derived from the most enriched amino acid at each position - DNRGRAF (Extended Data Fig. 5b). Notably, DNRGRAF differs from the DNLGRAF sequence by only a single residue at the position 3 (R to L), yet showed a substantially reduced ability to differentiate disease states (Fig. 4e). To further investigate the contribution of individual positions, we performed an alanine scan on the DNLGRAF substrate. Each of the six non-alanine residues was individually substituted with alanine, and cleavage activity by CT26 ECS was assessed using yeast display (Fig. 4f). All alanine variants exhibited reduced cleavage compared to the original sequence. Particularly, substitutions at positions 3, 4, and 5 (LGR) resulted in the most pronounced decreases. These findings indicate that efficient substrate activity depends on the precise sequence context rather than on only motif enrichment.

Lastly, we used matrix-assisted laser desorption ionization–time of flight mass spectrometry (MALDI-TOF MS) to identify specific cleavage sites. The untreated probe showed peaks corresponding to the full-length probe and a species lacking its fluorophore (Fig. 4g, Extended Data Fig. 7d). Following incubation with CT26 ECS, three additional peaks emerged, corresponding to cleavage between residues L/G, G/R and A/F (Supplementary Table 4). To confirm that these fragments resulted from tissue-derived activity, we monitored the accumulation of higher-mass fragments (>1000 Da) over time. The progressive increase in fragmented signal supported time-dependent proteolysis (Extended Data Fig. 7d, e). These results indicated that the DNLGRAF substrate undergoes cleavage at multiple specific sites.

These biochemical characterizations demonstrate that PSurf-identified substrates undergo preferential activation in tumor tissue, that protease recognition and cleavage depend on the precise sequence context of the substrate, and that multiple protease families contribute to substrate activation and may cleave at multiple sites.

Tumor-activated probe identifies metastases with spatial resolution

To spatially detect the proteolytic activity in tissues, we designed a probe using the DNLGRAF substrate (MBvivo-DNLGRAF) that enables tissue retention of the probe upon cleavage through electrostatic interactions with a nearby cell membrane only after proteolytic activation (Fig. 5a) ⁷⁴. We first examined the tumor-specific activation of this membrane-binding probe in vivo by comparing retention of the probe in the lungs following intravenous administration. We administered this probe to mice with and without CT26 lung metastases and imaged the lungs ex vivo. The probe demonstrates increased activation in tumor-bearing lungs (Fig. 5 b-c) and probe signal correlates with the tumor burden as quantified by tumor autofluorescence (Fig. 5d).

Figure 5. New substrate-based spatial probe delineates CT26 metastases in whole tissue sections.

(a) Schematic of MB-DNLGRAF, a membrane-binding fluorogenic probe that emits fluorescence following proteolytic cleavage of the DNLGRAF substrate. (b) Schematic of administration of MBvivo-DNLGRAF in mice with and without CT26 lung metastases and fluorescence images of the lungs ex vivo. (c) Bar graph showing mean fluorescence intensity of MBvivo-DNLGRAF in the lungs with and without CT26 metastases. Data represents the mean ± s.e.m. of n = 8 animals for CT26 and n = 7 for healthy. Unpaired two-tailed t-test was used. **P = 0.0026. (d) Correlation analysis of tumor autofluorescence (left panel) or lung weight (right panel) with MBvivo-DNLGRAF fluorescence showing a positive correlation in healthy lungs (gray) and CT26 lungs (orange). (e) Representative images from 2 experiments of mouse lung tissue with CT26 metastases stained with hematoxylin and eosin, anti-CA9 or anti-CD31 antibodies, or the MB-DNLGRAF probe with or without a protease inhibitor cocktail. Hematoxylin (blue), eosin (red), CA9 (red), CD31 (vasculature, cyan), MB-DNLGRAF (orange). L: healthy lung; T: tumor nodule.

We further applied this technique to stain CT26 lung metastasis tissue sections using a modified probe format based on the DNLGRAF substrate that enables both fluorescence activation and local membrane insertion upon cleavage (MB-DNLGRAF). DNLGRAF was fused to a membrane-binding peptide⁷⁵, with a fluorophore at the N-terminus and a quencher at the C-terminus, enabling cleavage-dependent fluorescence activation. Cleavage by tissue proteases separates the fluorophore from the quencher, leading to dequenching and restoration of fluorescence (Fig. 5e). Hematoxylin and eosin staining identified metastatic lesions, and regions of interest (ROIs) were used to correlate hematoxylin signal with probe fluorescence. MB-DNLGRAF fluorescence showed a strong positive linear correlation with hematoxylin staining (R² = 0.87) (Extended Data Fig. 8a, Supplementary Fig. 6). Mean fluorescence intensity was 2.2-fold higher in tumor lesions compared to healthy lung regions (Extended Data Fig. 8b). To confirm protease-dependent probe activation, we applied a cocktail of inhibitors identified from the protease inhibitors screen (Extended Data Fig. 7c) to tissue sections, which reduced tumor-associated fluorescence by 67.3% (Extended Data Fig. 8b). This result demonstrates that probe activation in the tumor is proteolysis-dependent.

This probe format was also validated in solution using ECS from healthy lungs and CT26 metastasis-bearing lungs. Consistently, cleavage rates were significantly higher in tumor ECS compared to healthy tissue, and activity was markedly reduced in the presence of inhibitors (Extended Data Fig. 8c-d).

These results demonstrate that the PSurf-identified substrate DNLGRAF undergoes tumor-specific protease cleavage in vivo, in solution and in ex vivo tissue slices, enabling spatial detection of proteolysis activities.

In vivo protease–activated biosensors enable disease classification via urinary readout

We next leveraged the tumor-activated protease substrates to engineer in vivo biosensors for biofluidic tumor detection. In a modular sensor panel, each of the four PSurf-identified substrates (cleavable linkers) is individually attached to an in vivo trafficking scaffold (e.g., nanobody, nanoparticle) and encoded with a chemically stabilized, single-stranded oligonucleotide that enables multiplexing and converts local TME protease activity into a measurable and amplifiable urinary signal (Fig. 6a). Following systemic administration, the biosensor accumulates in the TME where tumor-associated proteases cleave the substrate and liberate the DNA barcode. This barcode re-enters circulation, is size-filtered by the kidneys, and concentrates in the urine, where it can then be noninvasively detected through activating CRISPR-Cas12a as an indicator of tumor presence (Fig. 6b)⁶⁶.

Figure 6. In vivo activation of tumor-targeted PSurf substrates reports disease state via urinary readout.

(a–c) Schematics of the nanosensor design and experimental workflow. (a) The nanosensor consists of a CA9-targeting nanobody fused to a protease-cleavable substrate conjugated to a Maleimide-DBCO linker and a DNA barcode for detection. (b) Following intravenous injection, nanosensors accumulate in tumor nodules via nanobody targeting. Tumor-associated proteases cleave the substrate linker, releasing the DNA barcode into circulation. Free barcodes are filtered by the kidneys and concentrated in the urine. (c) Mice with and without CT26 lung metastases were administered a cocktail of four nanosensors incorporating the DNLGRAF, PARMLHI, PFLYLFG, and AYGRLYR substrates. Representative images of nanobody lung accumulation. Urine was collected over 1 hour post-injection and analyzed for barcode concentration using Cas12a activation and a fluorogenic reporter assay. (d) Box plot comparing Cas12a activation, measured by reporter DNA trans-cleavage rate, for each corresponding nanosensor substrate. Signal from Cas12a activation was normalized to the average healthy urine signal for each barcode. n = 12 mice per tumor group; n = 8 mice per healthy group; one-way ANOVA with Sidak’s multiple comparisons test: DNLGRAF, ***P < 0.001; PARMLHI, **P = 0.0082; PFLYLFG, **P = 0.0025; AYGRLYR, *P = 0.0109. Box plots display the interquartile range from the first quartile (Q1, 25th percentile) to the third quartile (Q3, 75th percentile), with the median (50th percentile) represented by a line within the box. Whiskers extend to the minimum and maximum values. (e) ROC analysis demonstrating the nanosensor panel’s ability to classify disease status using a 4-plex multiple logistic regression model compared to DNLGRAF alone. The dashed red line indicates a random classifier with an AUC of 0.5. Table shows AUC of ROC curve for each sensor with 95% confidence interval (C.I.). Fig 6b included graphics created with BioRender.com.

To improve sensor retention in the tumor, we engineered a nanobody against carbonic anhydrase IX (CA9), a target antigen selected due to its overexpression in CRC and other solid tumors⁷⁶(Extended Data Fig. 9a) and its mechanistic link to tumor acidosis-driven protease alteration and local invasion^77-79. In transcriptomic analysis of cancer types prone to metastasize to the lung, CA9 was upregulated 53-fold in primary colon adenocarcinoma compared to healthy tissue and across disease stages (Extended Data Fig. 9b). In CT26 murine lung metastasis model, histological staining with an anti-CA9 antibody confirmed specific CA9 expression in tumor lesions and minimal expression in adjacent normal tissue (Extended Data Fig. 9c).

To create the in vivo sensor, we fused protease-activable peptides or uncleavable controls to the CA9 nanobody through a rigid peptide linker designed to minimize steric hindrance to protease access. A cysteine was incorporated to the nanobody for site-specific labelling via maleimide-functionalized linker⁶⁶. We reengineered a CA9 nanobody⁸⁰ with optimal recombinant yield through stabilizing mutations⁸¹ of five residues identified in the protein stability prediction tool INPS-MD⁸² (Extended Data Fig. 9d). Incorporation of these mutations significantly improved nanobody expression while preserving target binding, as confirmed by staining of CT26 lung metastasis tissue with a dye-labeled CA9 nanobody (Extended Data Fig. 9e, Supplementary Fig. 7a). Biodistribution studies showed increased uptake of the CA9 nanobody in tumor-bearing lungs (Extended Data Fig. 9f-h, Supplementary Fig. 8).

To evaluate the biosensors in the CRC lung metastasis model, mice with and without CT26 lung metastases were administered a cocktail of four biosensors, each incorporating a distinct PSurf-identified substrate paired with a unique DNA barcode (Supplementary Fig. 7b, c). One hour after administration, when DNA barcodes reach peak bladder accumulation⁶⁶, urine was collected from each mouse (Fig. 6c, Extended Data Fig. 10a). Urinary DNA barcode abundance was quantified by incubating urine sample with Cas12a and barcode-specific guide RNAs. Cas12a activation, indicating the presence of each DNA barcode, was measured via collateral cleavage of a fluorescent reporter DNA (Extended Data Fig. 10b). All biosensors tested successfully distinguished tumor-bearing and healthy mice, as evidenced by elevated Cas12a activation in urine from tumor-bearing mice (Fig. 6d). Receiver operating characteristic (ROC) analysis revealed strong diagnostic performance, with area under the curve (AUC) values ranging from 0.82 to 0.89 for individual sensors, where an AUC of 1.0 represents perfect classification (Fig. 6e and Extended Data Fig. 10d-f). A control sensor with a 7xGlycine linker rather than a cleavable substrate showed no difference in urinary readout, corroborating that the reporter signal reflects protease activation (Extended Data Fig. 10c).

We further applied multiple logistic regression to integrate signals from all four biosensors into a unified predictive model. Previous studies have shown that using 3 to 5 biosensors offers sufficient predictive power in isogenic mouse models, with marginal gains from adding additional sensors^{66, 83}. This multiplexed approach further improved classification performance, yielding an AUC of 0.93 (Fig. 6e). These results demonstrate the potential of PSurf-discovered substrate biosensors for noninvasive tumor detection and highlight the value of combining biosensors into a multiplexed panel for robust disease classification.

Discussion

Dysregulated proteolytic activity in specific disease microenvironments presents compelling opportunities for the development of cancer diagnostics, fluorescence-guided surgery probes, and conditionally activated therapeutics. Progress, however, is limited by the small pool of disease-selective substrate sequences. Most substrates are identified using recombinant enzyme assays guided by transcriptomic data, yet transcript levels often do not reflect actual enzymatic activity in tissues. Additionally, substrates identified using individual enzymes do not capture the collective activity of the many protease families present in native tissue environments. Consequently, there remains an unmet need for methods that map proteolytic activity and discover optimal, disease-relevant substrates directly within tissue samples.

In this study, we report PSurf, a platform that, to our knowledge, is the first to enable synthetic substrate screening directly with tissue samples for translational applications. PSurf builds on key strengths of yeast surface display, including compatibility with high-resolution cell sorting and the ability to perform iterative selection to enrich rare cleavage events. We further introduce platform innovations such as double-tagging strategies and tailored differential selection workflows using ECM-derived samples, which together enable sensitive profiling of proteolytic activity in tissue-derived samples and identification of optimal tissue-specific substrates (Supplementary Fig. 9). PSurf expands the substrate discovery space compared with conventional synthetic peptide libraries by allowing exploration of large and diverse sequence repertoires, while reducing synthesis costs through library renewability. Although methods such as TAILS and HUNTER effectively identify natural proteolysis targets in tissues, they are not designed to discover optimal substrates for protease-activated modalities. By enabling direct screening of synthetic, non-proteome derived peptides that may offer improved stability and specificity compared with sequences derived from natural proteins, PSurf extends yeast display applications beyond their traditional uses^{32, 43, 62, 84-87}.

We demonstrated the utility of PSurf through several applications. First, the tailorable nature of the substrate discovery platform was demonstrated through identification of novel CTSB substrates under variable pH conditions, which can be applied for designing new ADC linkers or pH-dependent cathepsin inhibitors. Additionally, PSurf-discovered substrates showed improved specificity to lung metastases than widely used MMP probes and low reactivity in healthy and inflamed lung tissue. Importantly, we developed nanobody-based multiplexed biosensors integrating these sequences, and validated their in vivo utility for distinguishing tumor-bearing from healthy animals. This overall consistency across multiscale validation underscores the robustness of PSurf technology for optimal substrate identification. The identified substrates are promising for translation into tumor-detection sensors and conditionally activated therapeutics to minimize off-tumor effects. Further, the platform presents on opportunity for tumor-specific substrate discovery across tumor types, expanding the toolkit for detecting and responding to the TME with greater specificity.

This work provides one of the first global profiles of tissue cleavage, revealing a proteolytic preference for phenylalanine and leucine residues in both healthy tissues and tumors. Such finding is consistent with a prior report capturing Phe and Leu enrichment³⁹ and highlights the need for tissue-based differential selection to uncover context-specific probes. Biochemical analysis of a nominated tumor-reactive probe identified multiple cleavage sites within the DNLGRAF substrate and showed that multiple residues across the 7-amino acid-sequence are essential for cleavage. By contrast, synthesized peptide libraries often reveal only short motifs and may miss important flanking regions that contribute to proteolytic specificity; thus a 7-mer library, rather than shorter sequences used in previous reports can be advantageous for discovering tumor-sensitive probes. The likely cleavage of identified substrates by multiple enzymes, as suggested by inhibitor and MS studies, highlights a critical distinction of our selection technology compared to single enzyme-based screens. Notably, three of the nominated probes were not cleaved by recombinant MMPs, implying involvement of other protease families. One limitation is that many inhibitors tested are promiscuous and were applied at a high concentration to capture maximal activity across all inhibitors, reflecting overall sensitivity to the broad protease families rather than pinpointing individual proteases. Targeted follow-up using related recombinant enzymes, selective inhibitors, or genetic knockouts will be required to identify the specific proteases involved. These findings demonstrate PSurf’s capacity to identify tissue-responsive substrates without prior knowledge of protease expression or individual cleavage preferences. For clinical translation, we will need to test these substrates in additional lung metastasis models, validate in patient samples, optimize the probe designs, and thoroughly evaluate pharmacokinetics, biodistribution, and safety/toxicity⁸⁸.

Overall, PSurf is a broadly applicable platform for tissue-level extracellular proteolysis profiling, substrate discovery, and the development of protease-responsive diagnostic and therapeutic agents across a broad range of clinical applications in oncology, inflammation, aging and other fields that require substrates tailored to high-fidelity tissue proteolytic activity. This technology addresses a critical gap in the field of tumor-activated substrate discovery and provides a foundation for novel strategies that harness proteolytic activity as a hallmark of disease. To promote further research and facilitate broad application, we will make the PSurf library and protocols available for non-profit scientific use.

Methods

Ethical regulations.

All animal studies are approved by the Institutional Animal Care and Use Committee (IACUC) at the researcher’s institution, under protocol number IPROTO202200000081. Standard animal care facilities including housing chambers, sterile hoods, anesthetization, and ethical endpoint euthanization are required to properly carry out these experiments. All experiments were conducted in compliance with institutional and national guidelines and supervised by the veterinarian staff at one’s institution. Female BALB/c or C57BL/6 mice are used for experiments. Mice were obtained from Jackson Laboratories or Envigo Life Sciences and ranged in age from 6 to 8 weeks at the start of the study.

Animal models.

Tumor samples were harvested from two lung metastasis mouse models, one for metastatic colorectal cancer and the other for metastatic melanoma. The models were established by injecting 10⁵ cells CT26 colon adenocarcinoma cells into BALB/c mice or B16-F10 melanoma cells into C57BL/6 mice in 100 μL phosphate buffered saline (PBS) into the tail vein of female, 6-8 week-old female BALB/c or C57BL/6 mice. Tumor cells circulated through the bloodstream and seed in the lungs, forming visible metastatic nodules after 10-14 days. Metastatic nodules were harvested from the lungs of these mice for analysis and blood samples were collected prior to euthanasia. Simultaneously, healthy tissues, including lung, liver, kidney, and spleen, were collected from age-matched mice of the same strains to serve as controls. To model inflamed lungs, C57BL/6J mice were intraperitoneally injected with lipopolysaccharide (LPS, purified from E. coli 0111:B4) at 10 mg/kg. Three hours after LPS dosing, mice were euthanized and major tissues were collected. All tissue samples were flash-frozen immediately upon collection to preserve protease activities at maximum levels.

Upper intestinal washout sample preparation.

Mice were euthanized following institutional guidelines, and the small intestine was rapidly excised and placed in ice-cold PBS. The intestinal lumen was gently flushed with cold PBS using a 1 mL syringe, with approximately 1 mL of PBS per flush, to collect the luminal washout contents.

sECM sample preparation.

A 100 mg tissue sample was placed onto a 35 μm nylon mesh cell-strainer cap positioned on a 1.5 mL tube and centrifuged at 800×g for 10 minutes at 4°C, separating the extracellular matrix soluble fraction (sECM) from the tissue, while not lysing the cells. The extracted sECM fraction was adjusted to a total volume of 500 μL using a buffer containing 300mM sodium chloride, 50 mM Tris pH=7.5, 10 mM calcium chloride and 20 μM zinc chloride which serve as cofactors for MMPs and other metal ion-dependent proteases, and 0.02% Brij-35 to stabilize the proteases activity.

Tissue processing.

Tissues were gently dissociated with a GentleMACS tissue dissociator in gentleMACS^™ C Tubes in PBS. Extracellular proteins were collected by centrifuging tissue dissociates at 4°C for 10 minutes at 4000 x g. The supernatant was reserved and the total protein concentration was quantified by BCA assay. The tissue ECS was diluted to 0.5 mg/mL with PBS.

Plasmid construction.

Plasmids were constructed using standard molecular cloning techniques. DNA fragments encoding peptides displayed on the yeast surface (tagA-TEVs-tagB) with various tested epitope tags were synthesized by Integrated DNA Technologies and cloned into the pCL2 plasmid, which encodes N-terminal Aga2p, downstream of the AGA2 gene, using the Gibson assembly protocol. Non-canonical TEVs and substrate sequences selected from library screening were cloned using overlap extension PCR. All plasmid constructs were sequence-verified by Sanger or whole plasmid sequencing. All plasmids used in this study are summarized in Supplementary Table 11, and the full sequences are available as a source data file.

Yeast culture and peptide display.

Strain EBY100 was used for all yeast experiments. Competent yeast preparation and DNA transformation were performed using the Frozen-EZ Yeast Transformation II Kit (Zymo Research, catalog no. T2001). Following plasmid transformation, yeast cells were plated on Trp-selective SDCAA agar plates (5.4 g sodium phosphate dibasic anhydrous, 7.44 g sodium phosphate monobasic anhydrous, 15 g agar, 182 g sorbitol, in 1 L double-distilled water (DDW), pH 6.0) and incubated at 30°C for 2–3 days until visible colonies formed. Single colonies were picked and cultured overnight in SDCAA liquid media (20 g glucose, 6.7 g yeast nitrogen base without amino acids, 5 g casamino acids, 10.4 g sodium citrate, in 1 L DDW, pH 4.5) at 30°C and 220 RPM. Overnight cultures were refreshed by diluting to an optical density at 600 nm (OD₆₀₀) of 1 (where one OD₆₀₀ unit corresponds to approximately 1×10⁷ cells/mL) and incubated at 30°C until reaching an OD₆₀₀ of 2–5. When the desired OD was reached, the cultures were centrifuged at 2000xg for 5 minutes and resuspended in SGCAA induction media (20 g galactose, 2 g glucose, 6.7 g yeast nitrogen base without amino acids, 5 g casamino acids, 5.4 g sodium phosphate dibasic anhydrous, 7.44 g sodium phosphate monobasic anhydrous, in 1 L DDW, pH 6.0). The cultures were diluted back to an OD₆₀₀ of 1 and incubated for 24-48 hours at 20°C with shaking at 220 RPM. All media were sterilized by autoclaving or filtration prior to use.

Yeast library construction.

The substrate region of the gene encoding the peptide was randomized using overlap PCR extension. This was performed using the V5/HA tagged TEV substrate construct as the template and a primer containing a 7×NNK degenerate codon sequence (Extended Data Fig. 1a). Two sets of primers were used. The first set amplified the upstream region of the substrate, including the AGA2 sequence at the 5′ end and the DNA encoding the linker located upstream of the substrate (forward primer 1: ATTGCTTCAGTTTTAGCACAGGAAC; reverse primer 1: GCCTCCGCTTCCACTAGTG). The second set amplified the downstream region, incorporating the randomized substrate. The forward primer included a 19-bp overlap with the first PCR fragment, followed by a 7×NNK degenerate codon region as the substrate, with the 3′ end homologous to the downstream linker (forward primer 2: CACTAGTGGAAGCGGAGGCNNKNNKNNKNNKNNKNNKNNKGGCGGTAGTGGGGCTAGCTA; reverse primer 2: TTGTAATACGACTCACTATAGGGCG). Flanking primers were positioned 60 bp upstream and downstream of the XhoI and BamHI restriction sites in the template plasmid for yeast homologous recombination. The PCR fragments were gel-purified using the Zymoclean Gel DNA Recovery Kit (Zymo Research) and eluted in DDW. A secondary PCR reaction was performed to assemble the purified fragments into a single fragment, using forward primer 1 and reverse primer 2. The final product was purified and concentrated using the DNA Clean & Concentrator-25 kit (Zymo Research). The plasmid backbone was digested with XhoI and BamHI restriction enzymes (Thermo Scientific, catalog no. FD0694, FD0054) and gel-purified. Both the backbone and the randomized PCR fragment underwent ethanol precipitation to remove salts that could interfere with electroporation efficiency. Fresh yeast culture (EBY100) grown in YPD media was used for electroporation. 300 mL Yeast cells were grown to mid-log phase (OD₆₀₀ = 1.5–3), split into 50 mL tubes, resuspended in 100 mM lithium acetate with 10 mM DTT, and incubated for 10 minutes at 30°C with the cap loosened. The cells were washed twice with ice-cold DDW, resuspended in 250 μL DDW (resulting in 500 μL including the cell volume) and split into two 250 μL tubes. Cells in each tube were mixed with 4 μg PCR product and 1 digested backbone, and kept on ice for electroporation. Electroporation was conducted in a pre-chilled 2 mm cuvette with a 1:4 backbone-to-insert ratio using an ECM 830 square-wave electroporator (500 V, 15 ms, 1 pulse). After pulsing, 1 mL of YPD was immediately added for recovery. A total of 11 electroporations were performed, and the transformed yeast cells were pooled into a 50 mL flask and incubated at 30°C for 1 hour without shaking. Following incubation, a 1:100 dilution in SDCAA media was used to create serial 1:10 dilutions for library diversity estimation. The remaining cells were centrifuged at 900×g for 5 minutes, washed with 40 mL SDCAA media, and centrifuged again. The SDCAA-washed pellet was resuspended in 800 mL of SDCAA media (2x400 mL in 2L flasks) and incubated at 30°C for 24-48 hours until saturation (OD₆₀₀ of 8–10). The yeast cells were centrifuged at 2000×g for 5 minutes, resuspended in 10 mL SDCAA media supplemented with 30% glycerol, aliquoted, and stored at −80°C in a styrofoam container until use.

Yeast selection.

The PSurf library was cultured in SDCAA media, diluted to an OD₆₀₀ of 1, and incubated at 30°C until OD₆₀₀ = 2–5. The culture was then diluted back to OD₆₀₀=1 in SGCAA media for induction at 20°C over 24–48 hours. Once saturated (OD₆₀₀=8–10), 2×10⁷ cells were collected in a 1.5 mL tube, washed with buffer, and resuspended in a protease-containing sample. This number of cells was chosen to allow high sequence diversity while balancing throughput of experiments. A control tube with buffer only was prepared for comparison, and both were incubated at 20°C for the desired duration. After incubation, the library samples were washed twice with cold TBST containing 1% BSA (TBSTB) and stained with anti-V5-647 and anti-HA-488 antibodies (Supplementary Table 6) on ice for 30 minutes, protected from light. The stained cells were washed twice with TBSTB, resuspended in 1 mL TBSTB, and analyzed using FACS. During FACS sorting, the untreated control sample was first analyzed to define gates for the non-cleaved and cleaved populations. The non-cleaved gate included cells showing both V5 and HA signals (diagonal population), while the cleaved gate included cells with comparable V5 signals but reduced HA signals. The same gates were applied to the treated samples, and cells were sorted while processing >1×10⁷ events to preserve sequence diversity. Sorted cells were centrifuged at 2000×g for 5 minutes, resuspended in SDCAA media, and grown to saturation. Saturated cultures were re-induced, stained, and sorted to isolate the non-cleaved population. This step excluded substrates containing stop codons or those cleaved by endogenous yeast proteases. The non-cleaved population was then treated again with the protease-containing sample while including a buffer-only control for gate setting. In subsequent rounds, both non-cleaved and cleaved populations of the treated samples were sorted, cultured, and prepared for NGS. To ensure sequence diversity was preserved, each sorting step processed number of yeast cells equal to or exceeding 10 times the number of cells used as the input population.

NGS sample preparation.

Samples for amplicon-EZ NGS analysis (Azenta Life Sciences) were prepared using PCR amplification with barcoded primers (Supplementary Table 1). A total of 5×10⁶ cells were collected in a 200 μL PCR tube and resuspended in 20 μL of 20 mM sodium hydroxide, followed by incubation at 95°C for 45 minutes to release plasmid DNA from the yeast cells. The heated sample was centrifuged in a microcentrifuge for 2 minutes, and 4 μL of the supernatant was used as the template for PCR amplification. The resulting PCR fragments were separated on a 1.5% agarose gel, gel-purified, and their concentrations measured using a NanoDrop spectrophotometer. Equal masses of each fragment were combined to prepare a multiplexed amplicon-EZ NGS sample.

NGS data processing.

NGS data were processed using a custom Python script (https://gitfront.io/r/ialgov/xujkmkKDfAEW/PYSurf/). The analysis began with de-multiplexing sequencing reads into individual samples based on their DNA barcodes. Next, the script filtered the DNA reads to retain only those containing the linker sequence downstream of the substrate region, extracting the 21 bp substrate sequences located upstream of the linker. The DNA sequences were translated into amino acid sequences, and any substrate exhibiting 70% or greater similarity to the TEVs template sequence used for library construction was excluded. Identical sequences were then collapsed and counted, with the resulting data saved in a FASTA file. Sequence names indicated the number of times they appear in the data Seq1_249 representing the most abundant sequence (Seq1), found 249 times (N) in the dataset. Sequences that appear only once in the data were removed as well, keeping only sequences that appear two times or more (N>2).

NGS data analysis and visualization.

NGS data analysis was conducted using custom Python scripts tailored to the specific analytical needs of the study. Sequence logos were generated from the weighted sequences (considering ‘N’ in the frequency calculations) using a custom R script based on the ‘ggseqlogo’ package. IceLogos were generated using the IceLogo web server (https://compomics.github.io/projects/iceLogo), using all N>2 sequences from the relevant NGS results, with the input populations as the reference sequences. Motif analyses such as single amino acid, dipeptide, and per-position enrichment analyses were performed using a Python script (https://gitfront.io/r/ialgov/xujkmkKDfAEW/PYSurf/). The script first calculates the frequencies of a motif of interest in both the reference data (input library data) and the sample (output sample data). To include N in these calculations, the script extracts the number of times each sequence was found from the sequence name and multiplies the sequence by that number. For results that include raw amino acids frequencies such as the amino acids distribution analyses, this is the data used. For enrichment calculations of Z-scores, a log2 transformation is performed on the counts of amino acids to compress the data before performing the analyses. The script utilizes the log2-transformed frequencies calculated from the reference file and the sample file to calculate z-scores, indicating motif enrichment (positive values) or depletion (negative values) across the dataset. Z-scores indicate the distance of a sequence motif in the sample file from the mean of the same motif in the reference file, and expressed as the number of standard deviations. Z-scores were calculated by subtracting the frequency of a specific motif in the reference data from the frequency of that motif in the sample data, while dividing the result by the standard deviation of the motif frequency in the reference data. Graphs and heatmaps were created in GraphPad Prism.

Characterization of individual peptide substrates on yeast.

1x10⁶ yeast cells per well were plated in a 96-well assay plate and treated with either a protease-containing sample or buffer control sample at 20°C. The incubation temperature of 20°C was selected to maintain optimal conditions for yeast viability and protease activity. At designated time points, treated cells were collected, washed with TBSTB, and stained with anti-V5-647 and anti-HA-488 antibodies (Supplementary Table 6) for 30 minutes on ice, protected from light. Following staining, the samples were washed three times with TBSTB to remove unbound antibodies and analyzed by flow cytometry using a CytoFLEX instrument (Beckman Coulter, CytExpert Version 2.4.0.28). Cleavage activity was quantified by calculating the ratio of the mean fluorescence intensities (MFI) of the anti-HA-488 and anti-V5-647 signals. Negative log (-log) transformation was performed on the calculated ratios to present activity as increased values for the ease of results interpretation. For positional alanine scanning of DNLGRAF substrate, the same procedure was performed on yeast displaying single-alanine mutants at the different positions of the peptide. The measured signal was subjected to reciprocal transformation (1/read value) followed by normalization to the DNLGRAF mean signal.

Ex vivo Proteolytic Cleavage Assays.

Fluorogenic peptide substrates were synthesized by CPC Scientific with fluorophore and quencher FAM and CPQ2 or Cy7 and QC1. Lyophilized peptide probes were reconstituted in DMF to 2mM, then diluted to 100 μM in 20% (vol/vol) DMSO in water, before a final dilution to 6 μM in water. 5 μL of each 6 μM fluorogenic probe were distributed into a flat, clear-bottom black untreated polystyrene 384-well plate, which was then briefly centrifuged. 25 μL tissue ECS (0.5 mg/mL in PBS) or recombinant enzyme (0.36 ng/μL in assay buffer) were added to each well, and the plate was briefly centrifuged before immediately transferring to a spectrophotometer for reading. The fluorescence at 37°C was measured every 2 minutes for 1 hour (ex/em 485nm/535nm for FAM dye or 756nm/779nm for Cy7). For proteolytic cleavage assays with inhibitors, inhibitors (Supplementary Table 7) were combined with CT26 lung metastasis tissue ECS and fluorogenic probes. Inhibitors were pre-incubated with ECS for 30 minutes prior to addition of the fluorogenic probe. Inhibitors, ECS, and probe were combined at a final concentration of 150 μM, 0.27 mg/mL, and 1 μM, respectively.

MALDI-TOF MS analysis.

Samples for MALDI-TOF MS analysis were prepared as previously described⁸⁹. CT26 lung mets ECS diluted in an assay buffer to 1 mg/mL was incubated with DNLGRAF probe (5FAM-GDNLGRAFG-K(CPQ2)-GC) at a final concentration of 50 μM. In the desired time points (0, 5, 15 and 60 minutes), 15 μL of the reaction was mixed with 15 μL of 8M guanidinium hydrochloride to stop the reaction. Control tubes containing CT26 lung mets ECS alone, and probe alone, were incubated in the same conditions, while sampling in time points 0 and 60 minutes. Samples were desalted and peptides were enriched using a 10 μL ZipTips C18 pipette tips (Millipore) according to the manufacturer instructions, and eluted in 5 μL of 70% acetonitrile containing 0.1% trifluoroacetic acid (TFA) in water. For MALDI-TOF MS analysis, 1 μl of α-Cyano-4-hydroxycinnamic acid (α-CHCA) matrix was mixed with 1 μl sample solution, then 1 μl of the combined mixture was spotted and allowed to dry. Mass spectrometry data were acquired using microflex LRF (Bruker). The mass spectrum data were analyzed using Bruker flexAnalysis software.

Transcriptomic Analysis.

Human transcriptomic data was accessed using OncoDB, a database which compiles patient data from The Cancer Genome Atlas (TCGA) and the Genotype-Tissue Expression (GTEx) Portal. From a list of dysregulated genes in COAD, genes were filtered by protease identity and sorted by fold change in expression over corresponding healthy tissue.

Tissue sectioning and staining.

Organs were removed, washed in PBS, and immediately frozen in optimal cutting temperature (OCT) compound. Embedded tissues were cryosectioned into 10 μM sections, adhered to glass slides, and stored at −20°C. To stain, slides were warmed to room temperature and tissue sections were overlaid with a labeling buffer (300mM NaCl, 10mM CaCl2, 2mM ZnCl2, 50mM Tris, pH 7.5 or PBS) for 30 minutes. The buffer was removed and the sections were incubated with a labeling buffer containing MB-DNLGRAF probe (25 μM) for 4 hours at 37°C in a humidified box. For control sections, protease inhibitors (33 μM each of GM6001, Nafamostat, MG-132, Z-VAD(OMe)-FMK, and Gliotoxin) were included in the PBS labeling buffer at both of these steps. The buffer was then removed and the slides were washed with PBS (3 X 5 minutes), then incubated for 30 minutes at room temperature in a blocking buffer (PBS supplemented with 5% v/v fetal bovine serum). The buffer was removed and replaced with a blocking buffer containing primary antibodies (CD31, BD Biosciences 550274; CA9, Abcam ab15086; CA9 nanobody) (Supplementary Table 6) and incubated at 37°C for 1 hour. Slides were washed with PBS (3 x 5 minutes), then incubated with PBS containing secondary antibodies (AF594 anti-rat, ThermoFisher A-11081; AF647 anti-rabbit, ThermoFisher A-31573) at room temperature for 30 minutes, then washed in PBS (3 X 5 minutes) and mounted with ProLong^™ Gold Antifade Mountant with DNA Stain DAPI (Invitrogen). Hematoxylin and Eosin staining was performed according to kit manufacturer guidelines (Abcam ab245880). Slides were scanned on an Olympus VS120 slide scanner. Image analysis was performed on QuPath version 0.5.1⁹⁰.

Membrane-binding probe administration and ex vivo imaging.

An additional membrane-binding probe was synthesized by CPC Scientific for in vivo use, termed MBvivo-DNLGRAF. (Suc-DGlu-DGlu-DGlu-DGlu-DGlu-DGlu-DGlu-DGlu-C6-Asp-Asn-Leu-Gly-Arg-Ala-Phe-Gly-DArg-DArg-DArg-DArg-DArg-DArg-DArg-DArg-DArg-C6-DLys(Cy5)-NH2). 4.5 nmol of the probe reconstituted in PBS was injected intravenously and mice were euthanized 4 hours after administration for tissue imaging. Lungs were resected and weighed prior to fluorescence imaging (BioRAD ChemiDoc MP Imaging system). For dye visualization, excitation at 649 nm and a 700–730 nm filter were used. For tumor burden quantification, the tissues were also imaged using excitation at 450–490 nm and a 518–546 nm filter. Images were processed and quantified using ImageJ. To quantify the tissue area, a gaussian filter was applied to the images.

Cloning and synthesis of recombinant nanobodies.

A gene fragment encoding a CA9-directed nanobody fused to rigid linker and a protease substrate was synthesized by IDT and amplified by PCR for Gibson cloning into a pET-28 expression vector. Amino acid sequences for the anti-CA9 nanobody and rigid linker can be found in Supplementary Table 10. BamHI and SacI restriction sites flanking the protease substrate were used to exchange the substrate. Annealed oligos encoding alternative substrates were incorporated through ligation. The constructs were transformed into chemically competent BL21 E. coli, which were confirmed by Sanger sequencing to contain the correct insert. 5mL starter cultures were grown overnight at 37°C and used to inoculate 500 mL kanamycin-supplemented Terrific Broth expression cultures. Expression cultures were grown at 37°C and induced by addition of isopropyl β-d-1-thiogalactopyranoside (IPTG) to a final concentration of 0.5mM when the culture reached optical density at 600 nm (OD₆₀₀) of 0.6-0.8. The culture was incubated overnight at 18°C. The bacterial pellet was lysed with B-PER lysis buffer (Thermo Fisher Scientific) and sonication, then purified using Ni-NTA agarose (Qiagen). Purified nanobodies were confirmed by SDS-polyacrylamide gel electrophoresis and dialyzed with PBS. For tissue-section staining, the CA9 nanobody was labeled with IVISense 645 MAL fluorescent dye by room temperature incubation with the nanobody overnight at 300 rpm shaking followed by centrifugal filtration to remove unconjugated dye (Amicon Ultra Centrifugal Filter MWCO = 3.5 kDa).

Synthesis of nanobody-based DNA barcoded nanosensors.

50 mM tris(2-carboxyethyl)phosphine (10% v/v) was added to the nanobody in 0.1 M phosphate buffer to reduce the C-terminal cysteine. This mixture was reacted with sulfo DBCO-maleimide linker (10 equiv.) (Click Chemistry Tools) in 0.1 M phosphate buffer (pH 7.0) at room temperature overnight after which the excess linker was removed via dialysis (Slide-A-Lyzer^™ MINI Dialysis Devices MWCO = 3.5 kDa, ThermoFisher Scientific). DNA conjugation was performed by incubating DBCO-functionalized nanobody (1 equiv.) with azide-functionalized DNA barcode with a phosphorothioate modified (0.64 equiv.) in 0.1M phosphate buffer at room temperature overnight. Excess DNA barcode was removed via centrifugal filter at 4000 x g (Amicon Ultra Centrifugal Filter MWCO = 10 kDa). The product was confirmed via SDS–polyacrylamide gel electrophoresis analysis and ssDNA barcode was quantified using spectrophotometry. Sequences of DNA-barcoded synthetic urine biomarkers utilized in this study are listed in Supplementary Table 9.

Nanosensor administration and urine collection.

Barcoded nanobodies were pooled to create a cocktail of 0.5 nmol of each nanosensor in 100 μL sterile PBS. The sensors were intravenously injected into mice with and without CT26 lung metastases. Immediately following sensor injection, mice were injected with 100 μL PBS subcutaneously to support hydration. Urine was collected for 1 hour following nanosensor administration.

Detection of DNA barcodes in urine.

Urine samples were centrifuged at 800 x g for 5 min to remove sediment. The administered cocktail was diluted 1:100 in water to enable comparison between the signal generated from the injected dose and the urine. Cas12a was activated for 30 min at 37°C as previously described⁹¹. The Cas12a activation reaction was then used in the detection assay. Cleavage of the ssDNA reporter substrate with dye FAM and quencher QC1 was monitored by spectrophotometry every 2 minutes for 3 hours using ex/em 485 nm / 535 nm. The slope of the linear region of the resulting kinetic curve was used to quantify Cas12a activation.

Nanobody Biodistribution.

Purified, recombinant CA9 nanobody at 0.5 mg/mL was reacted with IVISsense 750 NHS dye (Perkin Elmer) at a ratio of 3 nmol dye to 1 nmol nanobody overnight with shaking at 300 rpm. Unreacted dye was removed using centrifugal filtration (Amicon Ultra Centrifugal Filter MWCO = 3.5 kDa). 0.5 nmol dye-labeled nanobody in PBS were injected intravenously into Balb/c mice with or without CT26 lung metastases, 16 days following tumor induction. Animals were euthanized 2 hours after nanobody administration and the heart, lung, spleen, kidneys, and liver were resected and weighed prior to fluorescence imaging (BioRAD ChemiDoc MP Imaging system). For dye visualization, excitation at 755–777 nm and a 813–860 nm filter were used. For tumor burden quantification, the tissues were also imaged using excitation at 450–490 nm and a 518–546 nm filter. Images were processed and quantified using ImageJ. To quantify the tissue area, a gaussian filter was applied to the images

Statistical analysis.

All graphing and statistical analyses were performed in GraphPad Prism (v10.4.1 (532)) or Microsoft Excel (Microsoft Office 2021). Unpaired one-way ANOVA tests were used for multiple comparisons assuming Gaussian distribution and equal SD unless otherwise noted. A two-tailed P-value was used to determine statistical significance for all analyses. P < 0.05 was considered statistically significant. ns: P>0.05, *: P<= 0.05, **: P<=0.01, ***: P <=0.001, ****: P<=0.0001. GraphPad Prism was used to perform multiple logistic regression for the 4-plex sensor output and generate ROC analyses.

Extended Data

Extended Data Figure 1. Library design, construction, and characterization.

(a) Overlapping PCR for constructing randomized peptide library. Two sets of primer pairs amplified two DNA fragments with an overlapping region containing randomized NNK codons during the first round of PCR (i). PCR using the first and last primers assembled the fragments (ii) into a single fragment (iii) for integration into a yeast display vector to display the randomized substrate peptide (iv). (b) DNA gel indicating two fragments (300 and 500 bp) and their assembly into an intact 800 bp fragment. Representative images of 4 experimental repeats of the 300 and 500 bp fragments and 22 repeats of the 800 bp fragment. (c) Flow cytometry plots showing yeast populations displaying intact peptides, cleaved peptides, or non-expressing populations before and after sorting, along with corresponding sequence logos. Approximately 15% of the original library exhibited partially displayed or pre-cleaved peptides before exposure to exogenous protease samples, and a second round of cell sorting reduces this population to 3%. (d) Validation of the randomized PSurf library after sorting to enrich non-cleaved population. The top panel indicates 76% of yeast displayed intact peptides with both tags. The bottom panel shows a strong correlation (R² = 0.70) between observed and expected amino acid distributions for an NNK-based library. (e) Single amino acid analysis from NGS data comparing non-cleaved and cleaved populations to the original library, revealing enrichment of stop codons in the cleaved population (left panel). Enrichment (red) or de-enrichment (blue) of individual amino acids is presented as Z-score values. NGS analysis shows that the amino acid distribution in the original constructed library positively correlates (R² = 0.66) with the expected distribution for an NNK-based library. The amino acid distribution remains highly similar in the final sorted non-cleaved library (right panel, R² = 0.99).

Extended Data Figure 2. Identification of CTSB substrates under various pH conditions.

(a) Schematic of the selection strategy for identifying CTSB substrates. (b) Flow cytometry plots (top) and sequence logos (bottom) from rounds 2 and 3 of CTSB selection at pH 4.4. Gating for sorting cleaved populations is indicated by dashed lines. Sequence logos were generated from NGS-identified sequences with N > 2. (c) Flow cytometry plots (top) and sequence logos (bottom) from round 3 of CTSB selection at pH 5.4. Gates for sorting cleaved populations are indicated by dashed lines. Sequence logos were generated from NGS-identified sequences with N > 2. (d) Correlation of dipeptide enrichment in biological replicates of yeast selections with 240 nM CTSB at pH 4.4. A strong positive correlation (R² = 0.87) indicates high reproducibility of dipeptide enrichment analysis. (e) Z-score heatmap showing the top 10 enriched dipeptides from each pH condition. Z-scores indicate enrichment relative to the untreated input library; values above 1 represent enrichment and below 1 indicate depletion.

Extended Data Figure 3. PSurf identifies yeast endogenous and upper intestin washout sample substrates.

(a) Illustration of the sample types that can be tested using the PSurf platform, including yeast endogenous protease substrates (e.g., Kex2), recombinant enzymes (e.g., human cathepsin B), animal-derived fluid samples (e.g., upper intestinal washout), and tissue or tumor sECM samples. (b) Z-score plots showing dipeptide enrichment profiles in the HA-cleaved yeast population. The y-axis represents the first amino acid; the x-axis shows either the second amino acid (left) or the fourth amino acid (right). Results reveal known cleavage motifs of the yeast protease Kex2, such as RR and LxxR. Z-scores represent enrichment relative to the input library. (c) Schematic of the selection process for identifying substrates sensitive to upper intestinal washout. (d–e) Flow cytometry and sequence logo analyses following 24- or 4-hour incubation with upper intestinal washout samples, showing enrichment of arginine-containing sequences in the cleaved population. The untreated population is shown in grey; the treated population in blue. Dashed lines indicate gates used to collect the non-cleaved (blue line) and cleaved (red line) populations. Sequence logos were generated from NGS-identified sequences with N > 2. Extended Data Fig. 3a included graphics created with BioRender.com.

Extended Data Figure 4. Tissue sECM extraction, selection, and motif analysis.

(a) Schematic and photographs illustrating the procedure for extracting sECM fluid from whole tissue samples. Approximately 100 mg of tissue was placed into a cotton mesh cap and centrifuged at 800 × g for 10 minutes at 4°C. The resulting sECM fluid was collected in the bottom tube. (b) Schematic of the selection workflow using sECM samples. (c–d) Z-score heatmaps (top) and sequence logos (bottom) showing amino acid enrichment at each position of the 7-aa substrate in buffer-treated control (c) and tissue-treated populations (d). Z-scores represent enrichment relative to the input library. Sequence logos were generated from NGS-identified sequences with N > 2. Extended Data Fig. 4a included graphics created with BioRender.com.

Extended Data Figure 5. Differential selection flow cytometry plots and selected substrates.

(a) Differentials selection flow cytometry plots showing the gating strategy used for each collected pool (dashed line), which was then subjected to the next round of treatment with the indicated healthy or tumor tissue sample. (b) IceLogos showing per-position enrichment and depletion of amino acids from the N>2 NGS sequences in the differential selection pools of yeast. The input pool contained a significant population of leucine- and phenylalanine-rich sequences, whereas the final pool demonstrated greater diversification with distinct sequence signatures. IceLogos were generated using the IceLogo web server, with the input populations as the reference sequences. (c) Bar graphs (left) and heatmap (right) comparing on-yeast cleavage of selected peptides by lung metastasis tumor samples and healthy lung tissues. Yeast was incubated with tissue samples for 6 hours. Data are shown as mean cleavage activity from n = 2 replicates. Substrates sequences selected to be synthesized as fluorogenic probes are in bold. Cleavage activity is defined as the negative log of HA-AF488 signal normalized to peptide expression (V5-AF647 signal).

Extended Data Figure 6. Motif analysis of global and differential selections results.

Z-score heatmaps illustrating dipeptide sequence enrichment and depletion relative to the input library for the N>2 NGS hits from global (a) or differential (b) selections. The y-axis represents the first amino acid, and the x-axis represents the second amino acid. For the differential selection results (b), the enrichment was plotted in blue and overlays on the global selection results (red) to highlight the differentially enriched dipeptides from these selections.

Extended Data Figure 7. Nominated substrates inhibition test and DNLGRAF cleavage site identification.

(a) Fluorogenic probe response to blood samples from mice with no tumor burden (166 mg lung) and high CT26 tumor burden (551 mg lung). Data represent mean ± s.e.m. cleavage rates normalized to healthy blood samples. (n = 3 technical replicates). (b) Heatmap showing the response of fluorogenic probes to individual recombinant enzymes. Data represent cleavage rate in RFU/min (n = 2 replicates). (c) Heatmap showing the response of fluorogenic probes to a panel of 150 μM protease inhibitors (Supplementary Table 7) in CT26 tumor ECS. Data represent percent inhibition of maximum probe cleavage with the mean of n = 2 replicates. (d) Mass-spectrometry analysis of CT26 ECS (left), DNLGRAF probe in assay buffer (middle), and DNLGRAF probe treated with CT26 ECS (right) over time. Samples were incubated in 37C for 60 minutes before quenching proteolysis using 8M guanidinium hydrochloride at selected time points (0, 60 minutes for controls, and 0, 5, 15, 60 minutes for treated probe)¹⁰⁶. CT26-treated DNLGRAF probe shows increasing intensity of the 1061.8 Da and 1978.0 peaks over time. (e) Time-dependent intensity increase of selected peaks. Intensities of the peaks identified in the high m.w. range MS scan were normalized to the 1673.8 Da peak of each time point.

Extended Data Figure 8. MB-DNLGRAF fluorescence correlates with tumor burden and protease-dependent cleavage in CT26 lung metastases.

(a) Correlation analysis of hematoxylin and MB-DNLGRAF fluorescence signals showing a strong positive correlation. Signals were quantified from ROIs drawn across lung sections (see Supplementary Fig. 6). (b) Bar graphs showing mean fluorescence intensity (MFI) of MB-DNLGRAF in tumor lesions compared to healthy tissues. Signal is significantly higher in tumor lesions compared to adjacent healthy tissue. Data represent mean ± s.e.m. across n = 10 ROIs; two-way ANOVA with Tukey’s multiple comparisons test, ****P < 0.0001 for tumor versus healthy adjacent tissue, ****P < 0.0001 for tumor versus inhibitor-treated tumor. (c) Bar graph showing MB-DNLGRAF cleavage rates in solution using healthy lung and CT26 lung ECS. Data represent mean ± s.e.m. of n = 3 lungs per group; unpaired two-tailed t-test, *P = 0.0205 for healthy versus CT26. (d) Bar graph showing MB-DNLGRAF cleavage rates in solution using CT26 lung ECS, with and without protease inhibitors. Probe activation is significantly higher in tumor ECS and is reduced by 300 μM inhibitor GM6001. Data represent mean ± s.e.m. of n = 3 lungs per group; unpaired two-tailed t-test, **P < 0.0012 for CT26 without inhibitor versus CT26 inhibitor-treated.

Extended Data Figure 9. Tumor targeting using a nanobody against pan-cancer biomarker CA9.

(a) Box and whisker plot showing human transcriptome expression of CA9 across 11 tumor types comparing tumor expression to respective normal healthy tissue expression. One way ANOVA. COAD **P=0.0016, KIRC ****P<0.0001, LUSC **P=0.0029, PAAD ****P<0.0001. Box plots display the interquartile range from the first quartile (Q1, 25th percentile) to the third quartile (Q3, 75th percentile), median (50th percentile) represented by a line within the box. Whiskers extend to the minimum and maximum values. Data represents expression from n shown in Supplementary Table 8. (b) Human colon adenocarcinoma (COAD) expression of CA9 compared to healthy colon (left) (85.0 +/− 163.0 COAD vs 1.6 +/− 1.5 Healthy, unpaired two-tailed t-test **P = 0.0012) and at each stage (right), data represent mean ± s.e.m. (c) Image showing a representative tissue section from 3 experiments of a mouse lung with CT26 lesions stained for CA9 (red) and CD31 (vasculature, blue). L: healthy lung; T: tumor nodule. (d) Predicted 3-dimensional protein structure of CA9 nanobody, with stabilizing mutations highlighted in green. Bar graph showing the predicted effect of incorporated mutations on nanobody stability. (e) Representative staining from 3 experiments of CT26 mouse lung with fluorescently-labeled, stabilized CA9 nanobody (red). (f) Normalized mean fluorescence intensity (MFI) of healthy lungs or CT26 lungs following I.V. administration of dye-labeled CA9 nanobody. Data represent mean ± s.e.m. of n = 4 lungs per group; unpaired two-tailed t-test, *P = 0.0227 for CT26 lung versus healthy lung. (g) Correlation analysis of total fluorescence of the dye-labeled CA9 nanobody and lung weight shows a positive correlation. R² = 0.57 by simple linear regression. (h) Fluorescent signal in CT26 lungs (top) or healthy lungs (bottom) from mice administered the dye-labeled CA9 nanobody or unlabeled nanobody (right).

Extended Data Figure 10. Multiplexed urinary readout reports on in vivo activation of nanosensors.

(a) Volume of urine collected from healthy and tumor-bearing groups. Data represent mean ± s.e.m. of n = 12 mice in the CT26 group and n = 8 mice in the healthy group; no significant difference (ns) by unpaired two-tailed t-test. (b) Increasing fluorescence signal from collateral cleavage of reporter DNA by activated Cas12a. Data represent Cas12a activation using urine from one healthy and one CT26 lung metastasis-bearing mouse incubated with crRNA corresponding to sensor substrate DNLGRAF. (c) Cas12a activation, measured by reporter DNA trans-cleavage rate, for a nanosensor with a control linker. Signal from Cas12a activation was normalized to the average healthy tissue urine signal. n = 8 mice per tumor group; n = 7 mice per healthy group, Unpaired two-tailed t-test: not significant (ns), P = 0.7408. Box plots display the interquartile range from the first quartile (Q1, 25th percentile) to the third quartile (Q3, 75th percentile), median (50th percentile) represented by a line within the box. Whiskers extend to the minimum and maximum values. (d) Predicted malignancy status from output of 4-plex multiple linear regression model, where 1.0 indicates predicted malignancy and 0.0 indicates predicted healthy status. Data represent mean ± s.e.m. of n = 12 mice in the CT26 group and n = 8 mice in the healthy group; unpaired two-tailed t-test, ***P = 0.0003 for healthy versus CT26. (e) Lung weight in grams. Data represent mean ± s.e.m. of n = 12 mice in the CT26 group and n = 8 mice in the healthy group; unpaired two-tailed t-test, ***P = 0.0002 for healthy versus CT26. (f) ROC analysis demonstrating each nanosensor’s ability to classify disease status independently. The dashed red line indicates a random classifier with an AUC of 0.5.

Data availability

The Cancer Genome Atlas (http://cancergenome.nih.gov) and Genotype-Tissue Expression Portal (https://www.gtexportal.org) are open access resources. The NGS datasets generated or analyzed during the current study were deposited in Zenodo and available at: https://zenodo.org/records/18063807. All data that support the findings of this study are available within the Article and Supplementary Information, Source data, or from the corresponding author upon reasonable request.

Code availability

Python and R scripts for data analysis, logos and figure plotting are available at https://github.com/ialgov/PSurf