Development of LRRC15-binding disulfide-constrained peptides for PET imaging of cancer-associated fibroblasts Authors: Benjamin J. Tombling, Fei Cai, Timothy J. Wendorff, Annie Ogasawara, Herman S. Gill, Jeff N. Tinianow, Andrew Chang, Aaron T. Balana, Lingling Peng, Stephen E. Miller, Benjamin T. Walters, Aaron Lictao, Qinying Yu, David C. DeWitt, Yuehua Wei, Sunny Z. Wu, Jawahar Sudhamsu, Akshay T. Krishnamurty, Simon P. Williams, Jan Marik, Yingnan Zhang, Henry R. Maun, Daniel Kirchhofer Published: 2026 ABSTRACT Leucine-rich-repeat-containing protein 15 (LRRC15) is selectively expressed on cancer-associated fibroblasts (CAFs) and constitutes a promising biomarker for imaging the tumor microenvironment. Using a combinatorial library approach, assisted by machine learning, we developed disulfide-constrained peptides (DCPs), notably ML-YSD-07 and ML-PD-03, that demonstrate subnanomolar affinities for murine LRRC15 (muLRRC15) and specifically localize onto muLRRC15-expressing fibroblasts. PET imaging with 18F-radiolabeled ML-YSD-07 exhibits specific tumor accumulation in a murine pancreatic cancer model highly enriched with LRRC15-expressing CAFs. Crystal structures of apo-muLRRC15 and of ML-YSD-07-bound muLRRC15 show that the DCPs evolved to adopt a distinct binding conformation that efficiently interacts with a flat epitope on muLRRC15. Collectively, this work identifies potent, molecularly engineered LRRC15-binding peptides and further highlights LRRC15 as a valuable CAF biomarker for cancer imaging applications. Introduction Tumor-targeted therapies have traditionally focused on agents that act directly on cancer cells. However, the tumor microenvironment surrounding cancer cells is a highly complex and heterogeneous environment that can account for up to 90% of the tumor mass, and the composition of the tumor microenvironment can govern therapeutic response to treatment1,2. Cancer-associated fibroblasts (CAFs) represent a major component of the tumor microenvironment and, as a result, receptors expressed on the surface of CAFs have attracted interest as biomarkers for therapeutic and diagnostic purposes3–5. An extensively studied CAF surface biomarker is fibroblast activation protein (FAP), with several FAP-targeted therapeutic and imaging agents currently under clinical evaluation6–17. Recognizing the potential of FAP-targeted agents, there is a growing interest in exploring additional CAF biomarkers to broaden therapeutic and diagnostic applications. Leucine-rich-repeat-containing protein 15 (LRRC15) is a cell surface receptor present on a population of myofibroblastic CAFs that become the dominant fibroblast type in multiple advanced solid tumor types, including pancreatic ductal adenocarcinoma (PDAC)18,19. LRRC15’s association with PDAC is particularly intriguing given the challenging disease prognosis and poor response to current standard of care therapy. Fibroblasts show minimal to no LRRC15 expression in non-tumor tissue, and become programmed to express LRRC15 by elevated levels of TGFß in the tumor microenvironment18,19. In mouse tumor models of PDAC, LRRC15-expressing CAFs promote tumor growth while the selective depletion of LRRC15-expressing CAFs significantly reduced tumor growth, improved tumor infiltrating CD8⁺ T cell effector function, and enhanced the response to immune checkpoint blockade therapy19,20. In addition, clinical analysis of human tumors from patients with bladder cancer, which shares molecular features with pancreatic cancer, has shown that elevated levels of an LRRC15⁺ CAF gene signature in the tumor environment correlated with poor patient response to anti-PD-L1 immunotherapy19. Collectively, these studies show that LRRC15-expressing CAFs play an active role in tumor progression and immunosuppression. Notably, single-cell and spatial profiling studies demonstrate partial overlap but distinct differences between FAP⁺ and LRRC15⁺ fibroblast compartments21,22, underscoring the value of LRRC15-targeted approaches as complementary to FAP. While these findings have sparked interest in LRRC15 as a potential cancer-associated biomarker within the tumor microenvironment, the biological function of LRRC15 and its interaction partners are still subject to investigation23–27. To date, LRRC15-targeting moieties are restricted to large biologics, such as monoclonal antibodies and single chain variable fragments18,28,29. For example, antibody-drug conjugates and radiolabeled antibodies are being developed and explored in clinical and preclinical studies for osteosarcoma and glioblastoma18,28–30. Peptides are a clinically validated modality and have desirable properties as diagnostic agents, such as rapid renal clearance, ability to bind target proteins with high affinity and selectivity, and convenience of site-selective functionalization31–33. Most importantly, their smaller molecular size compared to biologics provides enhanced tissue penetration and rapid accumulation in tumors34. Disulfide-constrained peptides (DCPs), such as cystine-knot peptides, are naturally occurring and highly diverse peptide scaffolds found in plants and in animal toxins and are commonly attributed with high thermal, chemical, and enzymatic stability35. In addition, they are amenable to protein engineering and directed evolution techniques to achieve binding to a protein target of interest, which has led to the exploration of DCPs as imaging agents for several tumor-related biomarkers36–40. In this study, we introduce an approach that integrates combinatorial library affinity selections with machine learning to engineer DCPs with high affinity for murine LRRC15 (muLRRC15). Enrichment data from combinatorial library binding selections were used to train a sequence-based machine-learning model to generate improved DCPs with subnanomolar binding affinities. One of these identified high affinity binders, ML-YSD-07, was radiolabeled with 18F for non-invasive positron emission tomography (PET) imaging and demonstrated significant and specific accumulation in murine pancreatic tumors enriched with muLRRC15-expressing fibroblasts. These findings highlight the value of DCPs as drug design scaffolds for targeted tumor imaging, and the peptides identified are valuable tools to further explore LRRC15 as a selective biomarker for fibroblast-enriched tumors. Results Identification of 1st generation DCPs that bind to murine LRRC15 We recently established a DCP-phage display platform for discovering peptide leads against protein targets41–43, and we performed binding selections using four different cystine-knot peptide scaffolds with amino acid randomizations in surface-exposed loops (Fig. 1a; see Supplementary Table S1 for library designs). The libraries were panned against the extracellular leucine-rich-repeat domains of muLRRC15 (amino acids Y22–S475, see Supplementary Fig. S1), which is predicted to be a well-structured region according to AlphaFold244. We focused our library panning efforts on muLRRC15 to enable in vivo characterization of the DCP hits, since transgenic human LRRC15 mouse models are not yet established. Nine DCPs from the four scaffolds were enriched during the affinity selection and these hits were chemically synthesized. Their binding affinities to muLRRC15 ranged from KD 0.45 to 34 µM, as determined by surface plasmon resonance (SPR) (Supplementary Table S2). Hits derived from the Ecballium elaterium trypsin inhibitor II (EETI-II) scaffold had the strongest binding affinities, but binding kinetics for all hits were characterized by fast dissociation rates based on the SPR sensorgrams (Fig. 1b, Supplementary Fig. S2, S3). Notably, the parent EETI-II scaffold showed no binding to muLRRC15 (Supplementary Fig. S2). EETI-based hits had mutations across loops 1 and 5, and four of the five EETI-based hits had an extended amino acid composition of the loop 1 region (eight vs six residues for the parent scaffold, respectively). Interestingly, all EETI-based hits competed with the strongest DCP-phage binder (EE-D1) for muLRRC15 binding, suggesting a conserved muLRRC15 binding epitope (Fig. 1c). The DCP EE-D1 had the strongest affinity for muLRRC15 determined by SPR (KD = 0.45 µM), which was confirmed by isothermal calorimetry (Supplementary Fig. S4a) and, therefore, we focused our efforts on using EE-D1 as a scaffold to engineer improved binding affinity.Fig. 1Discovery of muLRRC15-binding DCPs using phage-displayed libraries based on naturally occurring cystine-knot scaffolds.a 1st generation DCP affinity selections were performed using phage display and four DCP scaffolds with loop randomizations were displayed on the g8 M13 bacteriophage major coat protein. ChbTx Charybdotoxin (PDB 2CRD), EETI-II Ecballium elaterium trypsin inhibitor II (PDB 2IT7), Ast-A Asteropsin A (PDB 2LQA), Mch-1 Momordica charantia 1 (PDB = 2M2Q). Portions of this figure were created with BioRender.com. b Representative SPR sensorgram showing concentration-dependent binding of EE-D1 to immobilized muLRRC15 (data are representative of three independent experiments). c Competition-based ELISA assay showing binding of EE-D1-phage to immobilized muLRRC15 in the presence of synthesized EETI-based 1st generation DCPs (data shown are from a single experiment). Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu. Discovery of muLRRC15-binding DCPs using phage-displayed libraries based on naturally occurring cystine-knot scaffolds. a 1st generation DCP affinity selections were performed using phage display and four DCP scaffolds with loop randomizations were displayed on the g8 M13 bacteriophage major coat protein. ChbTx Charybdotoxin (PDB 2CRD), EETI-II Ecballium elaterium trypsin inhibitor II (PDB 2IT7), Ast-A Asteropsin A (PDB 2LQA), Mch-1 Momordica charantia 1 (PDB = 2M2Q). Portions of this figure were created with BioRender.com. b Representative SPR sensorgram showing concentration-dependent binding of EE-D1 to immobilized muLRRC15 (data are representative of three independent experiments). c Competition-based ELISA assay showing binding of EE-D1-phage to immobilized muLRRC15 in the presence of synthesized EETI-based 1st generation DCPs (data shown are from a single experiment). Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu. Improving the affinity of EE-D1 by use of a machine-learning-guided platform In order to design 2nd generation phage libraries, we first identified the EE-D1 residues important for muLRRC15 binding by use of a phage-based alanine scanning approach, in which the non-cysteine amino acids were preferentially mutated between the wild type residue or alanine45. All residues in loop 1 of EE-D1 were found to contribute to binding based on preferential enrichment of the wild type residue (Supplementary Fig. S4b), which was unsurprising since this loop was fully randomized in the 1st generation libraries. However, loop 5 residues, which were also randomized in the 1st generation libraries, had a negligible contribution to binding based on preferential enrichment of the non-wild type residue at these positions, suggesting they could be further optimized. Several residues within loops that were not mutated in the 1st generation libraries (loops 2, 3, and 4), did not contribute to binding, suggesting site-directed mutagenesis could provide optimized residues. Therefore, several 2nd generation DCP-phage libraries were designed with mutations across most of the peptide sequence using a combination of hard or soft randomization to allow for maximal library diversity (Supplementary Fig. S4c). We also designed libraries that explored the effects of extending the loop 5 length from five up to seven residues. Since EE-D1 phage binding to immobilized muLRRC15 could be detected when fused to the gene 3 (g3) phage coat protein (Supplementary Fig. S4d), all subsequent libraries were displayed as g3 fusions, which is preferable for identifying high affinity binders during selections due to its much lower copy number per phage particle compared to the gene 8 (g8) phage coat protein. Following phage panning of the 2nd generation libraries against muLRRC15, enriched phage were analyzed using next generation sequencing (NGS). The peptide sequences and enrichment data from the 2nd generation libraries (Fig. 2a) were used to train a sequence-based machine-learning (ML) language model to generate a respective true enrichment score for each sequence. The NGS data was first subjected to quality control measures to retain sequences with high confidence, leaving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~160,000 unique sequences that would be used to train the model. Potential bias was mitigated by normalizing the enrichment scores for each library based on round-over-round abundance changes. Using the respective true enrichment scores, the model generated a predicted binding score for a given peptide sequence (Fig. 2b). A non-specific binding score for a given sequence was also determined based on NGS enrichment data from negative selections against streptavidin/neutravidin during the 2nd generation phage panning. It is important to highlight that no information regarding muLRRC15 sequence or structure, or DCP structure was provided to the model. The accuracy of the model was tested by comparing the predicted binding scores of four 1st generation EETI-based DCP sequences that had the same amino acid length, for which SPR binding data were available (Fig. 2c). The good correlation, albeit limited to very few data points, provided confidence in the performance of the ML model.Fig. 2An overview of the machine-learning (ML)-guided combinatorial library platform used in this study for the discovery of high-affinity DCP binders to muLRRC15.a, b, 2nd generation phage libraries were individually screened against muLRRC15 and NGS enrichment scores for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~160,000 DCP sequences were used to train a ML model to predict target binding for each sequence. c Analysis of the predicted binding score compared to the experimentally determined binding affinities of 1st generation EETI-based DCPs. KD values shown are mean ± s.d. (N = three independent experiments). The strong correlation provided confidence in the performance of the ML model. d The model then generated 10 million new sequences within the DCP framework, which were subsequently filtered to 50,000 top sequences, considering both predicted binding scores and length diversity, while ensuring a non-specific binding score <−0.1 based on prior empirical observations. e The 50,000 sequences, designated as the ML-library, were screened against muLRRC15 using both yeast surface display and g3-phage display. Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu. An overview of the machine-learning (ML)-guided combinatorial library platform used in this study for the discovery of high-affinity DCP binders to muLRRC15. a, b, 2nd generation phage libraries were individually screened against muLRRC15 and NGS enrichment scores for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~160,000 DCP sequences were used to train a ML model to predict target binding for each sequence. c Analysis of the predicted binding score compared to the experimentally determined binding affinities of 1st generation EETI-based DCPs. KD values shown are mean ± s.d. (N = three independent experiments). The strong correlation provided confidence in the performance of the ML model. d The model then generated 10 million new sequences within the DCP framework, which were subsequently filtered to 50,000 top sequences, considering both predicted binding scores and length diversity, while ensuring a non-specific binding score <−0.1 based on prior empirical observations. e The 50,000 sequences, designated as the ML-library, were screened against muLRRC15 using both yeast surface display and g3-phage display. Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu. Next, the model computationally generated 10 million DCP sequences (Fig. 2d), following the DCP framework present in the 2nd generation libraries, that were predicted to have a high target binding score. It is important to highlight that these sequences were not present in the training (second-generation) or test (first-generation) data sets. To minimize non-specific binding, sequences were filtered using a non-specific binding score threshold of <−0.1 (lower values indicate reduced non-specific interactions). This threshold was chosen based on prior empirical observations, where sequences below this cutoff exhibited desirable specificity while maintaining sufficient sequence diversity. To select the final 50,000 sequences for library construction, we aimed to balance predicted binding scores with sequence length diversity rather than simply selecting the highest-scoring sequences. Subsequently, we selected the top 40,000 30-mers, 7500 31-mers, and 2500 32-mers, which together are referred to as the ML-library (see Fig. 2e and Supplementary Fig. S5a for the predicted binding scores and non-specific binding scores for the ML-library sequences, respectively). A 2D-UMAP analysis plot, where the distance between any two points approximates their embedding-based similarity in a high-dimensional space, showed that the ML-library contained DCP sequences with different properties than the enriched DCPs from the 2nd generation libraries (Supplementary Fig. S5b). This highlights the ability of the model to combine enrichment data from the different libraries to generate distinct sequences. Interestingly, the model identified that the Y5WWE/DWE10 motif in loop 1, which was very similar to the EE-D1 sequence (Y5WWDWE10), was highly conserved for the predicted high-scoring binders, suggesting that this motif is important for binding. In summary, the ML-library incorporated target-focused mutations across 12–13 residues (depending on the overall sequence length, see Supplementary Fig. S5c) and was subject to combinatorial display. ML-derived DCPs have strong affinity for muLRRC15 Binding selections of the ML-library were performed using two different display platforms, g3-phage display and yeast surface display. The latter method offers several advantages, such as normalized target-binding and peptide display levels, and endoplasmic reticulum chaperones-assisted peptide folding46. Following iterative rounds of selection, 13 DCPs from the ML-yeast surface display (ML-YSD) library and 10 DCPs from the ML-phage display (ML-PD) library were chosen for synthesis and binding affinity measurements (see Table 1, Supplementary Figs. S6, S7). DCPs were predominately from the 30-residue scaffold, with three 31-residue DCPs also selected, however, no 32-residue DCPs were selected due to weak enrichment during library screening. Two DCPs, ML-YSD-07 and ML-PD-03, had subnanomolar binding affinities for muLRRC15 (KD 0.48 nM and 0.36 nM, respectively), a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~1000-fold improvement over the top 1st generation hit EE-D1. Binding kinetics indicated that the improved affinity of ML-YSD-07 and ML-PD-03 was predominantly driven by a slower dissociation rate (koff) compared to other high-affinity DCPs identified (Fig. 3a and Supplementary Table S3). Based on the highly conserved Y5WWEWE10 motif in loop 1 between the ML-DCPs and the 1st generation hit EE-D1, the affinity improvement was driven by residues from other engineered loops and the mutated residues at positions 3 and 4 in loop 1. Notably, the seven ML-DCPs with binding affinities <10 nM were all from the 30-residue scaffold and had a conserved W25GV27 motif in loop 5. Altogether, we demonstrated that the ML-derived DCPs achieved stronger binding affinities compared to the 1st generation hits (Table 1 and Supplementary Fig. S8). This supports the model’s ability to generate DCPs, not present in the training data, with improved affinity.Fig. 3ML-DCPs have high binding affinity and selectivity for muLRRC15.a Representative SPR sensorgrams showing concentration-dependent binding of ML-YSD-07 (top) and ML-PD-03 (bottom) to immobilized muLRRC15. Experimental curves are shown in black. Kinetic binding fits, based on a 1:1 model, are shown in red. The kinetic constants kon and koff, and the calculated equilibrium dissociation constant KD are shown as mean ± s.d. (N = three independent experiments). b Representative histograms of flow cytometry plots showing that fluorescently labeled DCPs ML-YSD-07-N488 and ML-PD-03-N488, with their N-termini functionalized with Alexa488 (N488), only bind to murine 3T3 cells transfected with muLRRC15 (3T3-muLRRC15), but not to the parental 3T3 cells or KPR3070 murine PDAC cancer cells (N = three independent experiments). c Representative fluorescence images show co-localization of ML-PD-03-N488 with the outer cell membranes (CellBrite) of 3T3-muLRRC15 cells, not with the parental 3T3 cells. Cell nuclei were imaged using NucBlue. In the overlay images, CellBrite = red, NucBlue = blue, and ML-PD-03-N488 = green. Images were captured 1 h after the addition of 500 nM ML-PD-03-N488 to cells. Images are representative of two independent experiments. Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu.Table 1Sequences and binding affinities of the EETI-based DCPsDCP12345678910111213141516171819202122232425262728293031muLRRC15KD (nM)Loop 1Loop 2Loop 3L4Loop 5EETI-IIGCPRILMRCKQDSDCLAGCVCGPNGFCGN.B.1st generationEE-H2GCWWEDPICKQDSDCLAGCVCQWGYDCG4060 ± 510EE-D1GCAQYWWDWECKQDSDCLAGCVCREFENCG450 ± 50EE-E12GCDQYWWAWECKQDSDCLAGCVCPDGGWCG1300 ± 200EE-E1GCYQHWWPWECKQDSDCLADCVCWNQVWCG1600 ± 200EE-E3GCTQYFYEWECKQDSDCLAGCVCDWGSWCG756 ± 61ML-YSDML-YSD-01GCAQYWWEWECELDTDCLAGCVCRWGENCG190 ± 4ML-YSD-02GCDQYWWEWECPLDADCLAGCVCREFENGCG330 ± 60ML-YSD-03GCDQYWWEWECRYDPDCLAGCVCREFENGCG24 ± 1ML-YSD-04GCTQYWWEWECQYDTDCLAGCVCRELENGCG110 ± 30ML-YSD-05GCDQYWWEWECRLDPDCLAGCVCRDFETCG20.0 ± 0.5ML-YSD-06GCTQYWWEWECPHDSDCLAGCVCRDFESCG17.00 ± 0.06ML-YSD-07GCDRYWWEWECRYDTDCLAGCVCTWGVTCG0.48 ± 0.03ML-YSD-08GCDRYWWEWECEHDADCLAGCVCTWGVNCG140 ± 6ML-YSD-09GCDEYWWEWECSHDADCLAGCVCRWLENCG130 ± 9ML-YSD-10GCDRYWWEWECRLDPDCLAGCVCEWGLTCG13.0 ± 0.2ML-YSD-11GCDRYWWEWECPYDPDCLAGCVCRDFENCG20.5 ± 0.5ML-YSD-12GCTRYWWEWECEYDTDCLAGCVCRWGENCG150 ± 10ML-YSD-13GCDRYWWEWECPYDADCLAGCVCRWGENCG62 ± 2ML-PDML-PD-02GCTRYWWEWECPYDTDCLAGCVCEWGVNCG16.8 ± 0.3ML-PD-03GCTQYWWEWECPHDTDCLAGCVCEWGVTCG0.36 ± 0.01ML-PD-04GCEQYWWEWECPHDTDCLAGCVCEWGVNCG5.5 ± 1.0ML-PD-05GCERYWWEWECRYDTDCLAGCVCEWGVNCG5.08 ± 0.16ML-PD-07GCDRYWWEWECRHDADCLAGCVCEWGVNCG10.6 ± 0.2ML-PD-08GCEQYWWEWECTHDADCLAGCVCEWGVTCG1.60 ± 0.05ML-PD-09GCEQYWWEWECRHDPDCLAGCVCVWGVNCG2.22 ± 0.26ML-PD-12GCERYWWEWECPLDTDCLAGCVCEWGVNCG16.9 ± 0.8ML-PD-13GCDRYWWEWECPHDADCLAGCVCVWGLNCG33.7 ± 2.0ML-PD-14GCTQYWWEWECEHDTDCLAGCVCEWGVTCG2.64 ± 0.10The parent EETI-II sequence is shown for reference. Cys residues are shown in bold. Binding affinities for muLRRC15 were measured by SPR with either steady state or kinetic fitting models (see methods). KD are shown as mean ± s.d. (N = 3–4 independent experiments). N.B. = no binding detected at 100 µM.DCPs are grouped into 1st generation DCPs, ML-derived DCPs identified by yeast surface display (ML-YSD) and ML-derived DCPs identified by phage display (ML-PD). ML-DCPs have high binding affinity and selectivity for muLRRC15. a Representative SPR sensorgrams showing concentration-dependent binding of ML-YSD-07 (top) and ML-PD-03 (bottom) to immobilized muLRRC15. Experimental curves are shown in black. Kinetic binding fits, based on a 1:1 model, are shown in red. The kinetic constants kon and koff, and the calculated equilibrium dissociation constant KD are shown as mean ± s.d. (N = three independent experiments). b Representative histograms of flow cytometry plots showing that fluorescently labeled DCPs ML-YSD-07-N488 and ML-PD-03-N488, with their N-termini functionalized with Alexa488 (N488), only bind to murine 3T3 cells transfected with muLRRC15 (3T3-muLRRC15), but not to the parental 3T3 cells or KPR3070 murine PDAC cancer cells (N = three independent experiments). c Representative fluorescence images show co-localization of ML-PD-03-N488 with the outer cell membranes (CellBrite) of 3T3-muLRRC15 cells, not with the parental 3T3 cells. Cell nuclei were imaged using NucBlue. In the overlay images, CellBrite = red, NucBlue = blue, and ML-PD-03-N488 = green. Images were captured 1 h after the addition of 500 nM ML-PD-03-N488 to cells. Images are representative of two independent experiments. Portions of this figure were created in BioRender. Maun, H. (2026) https://BioRender.com/jxu0mbu. Table 1Sequences and binding affinities of the EETI-based DCPsDCP12345678910111213141516171819202122232425262728293031muLRRC15KD (nM)Loop 1Loop 2Loop 3L4Loop 5EETI-IIGCPRILMRCKQDSDCLAGCVCGPNGFCGN.B.1st generationEE-H2GCWWEDPICKQDSDCLAGCVCQWGYDCG4060 ± 510EE-D1GCAQYWWDWECKQDSDCLAGCVCREFENCG450 ± 50EE-E12GCDQYWWAWECKQDSDCLAGCVCPDGGWCG1300 ± 200EE-E1GCYQHWWPWECKQDSDCLADCVCWNQVWCG1600 ± 200EE-E3GCTQYFYEWECKQDSDCLAGCVCDWGSWCG756 ± 61ML-YSDML-YSD-01GCAQYWWEWECELDTDCLAGCVCRWGENCG190 ± 4ML-YSD-02GCDQYWWEWECPLDADCLAGCVCREFENGCG330 ± 60ML-YSD-03GCDQYWWEWECRYDPDCLAGCVCREFENGCG24 ± 1ML-YSD-04GCTQYWWEWECQYDTDCLAGCVCRELENGCG110 ± 30ML-YSD-05GCDQYWWEWECRLDPDCLAGCVCRDFETCG20.0 ± 0.5ML-YSD-06GCTQYWWEWECPHDSDCLAGCVCRDFESCG17.00 ± 0.06ML-YSD-07GCDRYWWEWECRYDTDCLAGCVCTWGVTCG0.48 ± 0.03ML-YSD-08GCDRYWWEWECEHDADCLAGCVCTWGVNCG140 ± 6ML-YSD-09GCDEYWWEWECSHDADCLAGCVCRWLENCG130 ± 9ML-YSD-10GCDRYWWEWECRLDPDCLAGCVCEWGLTCG13.0 ± 0.2ML-YSD-11GCDRYWWEWECPYDPDCLAGCVCRDFENCG20.5 ± 0.5ML-YSD-12GCTRYWWEWECEYDTDCLAGCVCRWGENCG150 ± 10ML-YSD-13GCDRYWWEWECPYDADCLAGCVCRWGENCG62 ± 2ML-PDML-PD-02GCTRYWWEWECPYDTDCLAGCVCEWGVNCG16.8 ± 0.3ML-PD-03GCTQYWWEWECPHDTDCLAGCVCEWGVTCG0.36 ± 0.01ML-PD-04GCEQYWWEWECPHDTDCLAGCVCEWGVNCG5.5 ± 1.0ML-PD-05GCERYWWEWECRYDTDCLAGCVCEWGVNCG5.08 ± 0.16ML-PD-07GCDRYWWEWECRHDADCLAGCVCEWGVNCG10.6 ± 0.2ML-PD-08GCEQYWWEWECTHDADCLAGCVCEWGVTCG1.60 ± 0.05ML-PD-09GCEQYWWEWECRHDPDCLAGCVCVWGVNCG2.22 ± 0.26ML-PD-12GCERYWWEWECPLDTDCLAGCVCEWGVNCG16.9 ± 0.8ML-PD-13GCDRYWWEWECPHDADCLAGCVCVWGLNCG33.7 ± 2.0ML-PD-14GCTQYWWEWECEHDTDCLAGCVCEWGVTCG2.64 ± 0.10The parent EETI-II sequence is shown for reference. Cys residues are shown in bold. Binding affinities for muLRRC15 were measured by SPR with either steady state or kinetic fitting models (see methods). KD are shown as mean ± s.d. (N = 3–4 independent experiments). N.B. = no binding detected at 100 µM.DCPs are grouped into 1st generation DCPs, ML-derived DCPs identified by yeast surface display (ML-YSD) and ML-derived DCPs identified by phage display (ML-PD). Sequences and binding affinities of the EETI-based DCPs The parent EETI-II sequence is shown for reference. Cys residues are shown in bold. Binding affinities for muLRRC15 were measured by SPR with either steady state or kinetic fitting models (see methods). KD are shown as mean ± s.d. (N = 3–4 independent experiments). N.B. = no binding detected at 100 µM. DCPs are grouped into 1st generation DCPs, ML-derived DCPs identified by yeast surface display (ML-YSD) and ML-derived DCPs identified by phage display (ML-PD). To confirm that these peptides can bind selectively to cell-surface expressed muLRRC15, ML-YSD-07 and ML-PD-03 were functionalized with the fluorescent dye A488 at their respective N-termini for cell-binding studies. Both labeled DCPs specifically bound to mouse fibroblasts engineered to express muLRRC15 (3T3-muLRRC15) but showed no binding to wildtype 3T3 cells which have no detectable surface expression of muLRRC15 (Fig. 3b). Importantly, 3T3 cells endogenously express LRR-containing proteins, such as Islr (Meflin)47, suggesting that the DCPs may exhibit selectivity for LRRC15 over other LRR-containing proteins. In addition, no DCP binding was observed to KPR3070 (KPR) murine PDAC tumor cells (KPR; Kras LSL.G12D/wt;p16/p19 fl/wt;p53LSL.R270H/wt;Pdx1.Cre) that do not express muLRRC15 (Fig. 3b)48. Confocal microscopy images showed that ML-PD-03 co-localizes with the outer cell membrane of 3T3-muLRRC15 cells, demonstrating that cell binding is due to muLRRC15 at the surface (Fig. 3c). In summary, the high-affinity DCPs derived from the ML-library selectively recognized cell surface muLRCC15 on fibroblasts. ML-derived DCPs adopt a rearranged disulfide connectivity To assess whether the ML-derived DCPs retained the cystine knot disulfide conformation present in the EETI-II scaffold49, we carried out a solution-based disulfide mapping study on ML-PD-05 as a representative of the predominant 30-residue DCP hits. Stepwise disulfide bond reduction using tris(2-carboxyethyl)phosphine (TCEP) followed by selective cysteine alkylation at low pH using N-ethylmaleimide enabled detection of the paired cysteine residues. Interestingly, the disulfide-connectivity of ML-PD-05 was distinct from the parent EETI-II scaffold. ML-PD-05 adopts a C1–C4, C2–C3, C5–C6 disulfide-connectivity (cysteine residues are numbered from the peptide N-terminus; see Supplementary Fig. S9), and therefore, no longer adopts a cystine knot (C1–C4, C2–C5, C3–C6). Given the similar sequence motifs in the engineered loops of all 30-residue ML-DCPs, they likely adopt the same non-cystine knot disulfide connectivity. Our results, in addition to previous studies43,50, demonstrate a propensity of the EETI-II scaffold to form different disulfide connectivities during engineering campaigns. Identification of the DCP binding region by hydrogen deuterium exchange-mass spectrometry (HDX-MS) To obtain information of the DCP binding site on muLRRC15 we performed hydrogen deuterium exchange-mass spectrometry (HDX-MS) experiments with muLRRC15 in the absence and presence of a DCP. As a representative DCP we used EE-2-1, which was identified from the 2nd generation phage panning prior to machine learning. EE-2-1 shares many of the key features of the high-affinity ML-DCP binding hits, such as the extended loop 1 with its Y5WWEWE10 motif, and the W25G26 motif in loop5 (see Supplementary Fig. S10a and b for sequence and binding data for EE-2-1, respectively). Upon binding, muLRRC15 dynamics will be most affected at or proximal to the binding site and therefore will be expected to have the largest change in exchange rates. High sequence coverage (87%) of muLRRC15 residues was detected, and sets of amide protons displayed slower exchange rates in the presence of EE-2-1 (Supplementary Fig. S10c). Using the AlphaFold2 predicted structure of muLRRC1544, the slower exchanging residues are clustered together on the concave face, mainly spanning across leucine-rich-repeats (LRRs) 1-7. This suggests that this region contains the DCP binding site (Supplementary Fig. S10d), most likely residing in the center of the perturbed region. Crystal structure reveals molecular details of ML-YSD-07 interaction with muLRRC15 To investigate the DCP binding interaction with muLRRC15 at the atomic level, we aimed to solve a muLRRC15:DCP complex structure using X-ray crystallography. Initial efforts to generate crystals of a muLRRC15:DCP complex were unsuccessful. Therefore, we developed LRRC15-binding Fabs to act as structural chaperones, since Fab antibody fragments have been extensively used to facilitate the structure determination of challenging macromolecular complexes51. We performed a Fab phage display campaign against muLRRC15 using the Fab YSGX library, where amino acid diversity at the complementary determining regions (CDRs) is restricted to Tyr, Ser, Gly and X (X represents equimolar mixture of all amino acids except for Cys, Tyr, Ser and Gly)52. In order to select for Fabs that do not compete with DCP, the library screening against muLRRC15 was carried out in the presence of excess ML-YSD-07. The identified Fab-E1 (see Supplementary Fig. S11a for Fab-E1 sequence) was expressed in mammalian cells and bound to muLRRC15 with a KD of 48.2 nM, as determined by SPR (Supplementary Fig. S11b). We confirmed that Fab-E1 did not compete with DCP for binding to muLRRC15 (Supplementary Fig. S11c), indicating that it is suitable for structural studies. We solved the X-ray crystal structure of ML-YSD-07 bound to muLRRC15 in complex with Fab-E1 at 2.6 Å resolution (Fig. 4a). For this, we used the extracellular domain of muLRRC15 (Y22–S475) with a C416S mutation to avoid potential disulfide formation of the unpaired cysteine at high concentrations. The structure showed that the extracellular domain of muLRRC15 adopts a typical LRR fold, with the 15 LRRs displaying parallel ß-strands on the concave face and diverse secondary structural features, including helices and loops, in the inter-strand convex face of the LRRs. The LRRs are capped at both the N- and C-terminal ends by cysteine-rich domains (LRRNT and LRRCT). muLRRC15 contains a glycosylation site at N369, however, this residue was not near the ML-YSD-07 binding site (there was insufficient electron density to model the sugar at the second glycosylation site, N75). ML-YSD-07 binds to the concave, ß-strand face of muLRRC15 and occupies a flat interface of 803 Å2 (Fig. 4b). Notably, muLRRC15 contact residues determined from the crystal structure agreed with the buried binding region determined by HDX-MS (Supplementary Fig. S11d), with the crystal structure confirming that ML-YSD-07 predominately forms binding contacts with residues from LRRs 2–5. The disulfide connectivity of ML-YSD-07 in the crystal structure (C1–C4, C2–C3, C5–C6) was in agreement with the solution-based disulfide mapping data for ML-PD-05 (Fig. 4c and Supplementary Fig. S9), confirming that the DCPs no longer display a cystine knot fold. Fab-E1 binds to residues in the LRRCT domain of muLRRC15, with the Tyr-rich heavy chain CDRs providing nearly all the binding interactions (Supplementary Fig. S11e). Interestingly, Fab-E1 shows no binding to human LRRC15 (huLRRC15) up to 300 nM (Supplementary Fig. S11b), which is likely due to amino acid differences in the Fab-E1 binding epitope.Fig. 4Crystal structure of the ML-YSD-07:muLRRC15:Fab-E1 complex.a Overview of the complex. Cartoon representation of muLRRC15 (gray) with ML-YSD-07 (green) bound to the concave face and Fab-E1 (orange for heavy chain and yellow for light chain) bound to the muLRR15 C-terminal region of the extracellular domain. b ML-YSD-07 (green) binding interface with muLRRC15 (gray), with muLRRC15 contact residues within 4 Å of ML-YSD-07 shown in yellow and labeled in black italics. The ML-YSD-07 contact residues are shown as sticks and labeled in green. c The tertiary structure and 2-D representation of the loops and disulfide-bond connectivity of ML-YSD-07. Disulfide bonds are represented as yellow sticks. Cysteines are sequentially numbered starting at the N-terminus. d–g Close up of the key binding interactions between ML-YSD-07 (green) and muLRRC15 (gray), with electrostatic (red dotted lines), H-bonds (blue dotted lines), and π-stacking (orange dotted lines) interactions highlighted. d E10 of ML-YSD-07 interacts with K177, N179, and R203 of muLRRC15. e W7, W9, W25 account for half of the binding interface and bind to shallow pockets on the muLRRC15 surface. In addition, the indole nitrogen of W7 and W9 interact with the side chains of E153 and Q155, respectively. muLRRC15 residues within 4 Å of these three ML-YSD-07 Trp residues are shown as yellow sticks. f Y5 and W6 of ML-YSD-07 pack into the hydrophobic core of the DCP and stabilize the DCP fold and binding conformation. The C-terminus of ML-YSD-07 is near the muLRRC15 surface, whereas the N-terminus is solvent-exposed and ideally suited for functionalization. g R12 of ML-YSD-07 forms electrostatic interactions with E206 and E230 of muLRRC15, and Y13 of ML-YSD-07 interacts with Y157 (H-bond) and Y205 (π -stacking) of muLRRC15. Crystal structure of the ML-YSD-07:muLRRC15:Fab-E1 complex. a Overview of the complex. Cartoon representation of muLRRC15 (gray) with ML-YSD-07 (green) bound to the concave face and Fab-E1 (orange for heavy chain and yellow for light chain) bound to the muLRR15 C-terminal region of the extracellular domain. b ML-YSD-07 (green) binding interface with muLRRC15 (gray), with muLRRC15 contact residues within 4 Å of ML-YSD-07 shown in yellow and labeled in black italics. The ML-YSD-07 contact residues are shown as sticks and labeled in green. c The tertiary structure and 2-D representation of the loops and disulfide-bond connectivity of ML-YSD-07. Disulfide bonds are represented as yellow sticks. Cysteines are sequentially numbered starting at the N-terminus. d–g Close up of the key binding interactions between ML-YSD-07 (green) and muLRRC15 (gray), with electrostatic (red dotted lines), H-bonds (blue dotted lines), and π-stacking (orange dotted lines) interactions highlighted. d E10 of ML-YSD-07 interacts with K177, N179, and R203 of muLRRC15. e W7, W9, W25 account for half of the binding interface and bind to shallow pockets on the muLRRC15 surface. In addition, the indole nitrogen of W7 and W9 interact with the side chains of E153 and Q155, respectively. muLRRC15 residues within 4 Å of these three ML-YSD-07 Trp residues are shown as yellow sticks. f Y5 and W6 of ML-YSD-07 pack into the hydrophobic core of the DCP and stabilize the DCP fold and binding conformation. The C-terminus of ML-YSD-07 is near the muLRRC15 surface, whereas the N-terminus is solvent-exposed and ideally suited for functionalization. g R12 of ML-YSD-07 forms electrostatic interactions with E206 and E230 of muLRRC15, and Y13 of ML-YSD-07 interacts with Y157 (H-bond) and Y205 (π -stacking) of muLRRC15. Ten ML-YSD-07 residues from loops 1, 2, and 5 are engaged in binding contacts with muLRRC15 (Supplementary Fig. S12a). The identified Y5WWEWE10 motif in loop 1, which was highly conserved in all binding hits derived from the ML-library, provides several key binding contributions. Residue E10, which is part of the short helix from residues E8–Y13, is inserted into a deep pocket and is positioned to form electrostatic interactions with K177 or R203, and an H-bond with N179 of muLRRC15 (Fig. 4d). Residues W7 and W9 of ML-YSD-07 occupy hydrophobic patches on muLRRC15’s surface and the indole groups of W7 and W9 form H-bond interactions with E153 and Q155, respectively (Fig. 4e). Although E8 was highly enriched during library screening, it is solvent exposed and may therefore be important for folding. Residue R4 packs closely with R249 of muLRRC15 and may form an Arg-Arg π-stacking interaction (Supplementary Fig. S12b). Interestingly, Y5 and W6 do not contact muLRRC15 residues but instead stabilize the DCP’s hydrophobic core through intramolecular packing (Fig. 4f), potentially compensating for the loss of the stabilizing cystine knot motif in ML-YSD-07 and related DCPs. Loop 5 residue W25 was highly enriched for all high affinity DCPs, and it occupies a hydrophobic pocket covering LRRs 2–4 (Fig. 4e). The three binding residues W7, W9, and W25 account for half (397 Å2) of the total buried surface area. The importance of W9 and W25 for binding was confirmed by the complete lack of muLRRC15 binding by ML-YSD-07 in which these two residues were mutated to Ala (Supplementary Fig. S12c). T24, V27, and T28 form hydrophobic packing away from the binding interface and appear to stabilize the turn in loop 5 that positions W25 to optimally interact with muLRRC15 (Supplementary Fig. S12d). The effect of non-interacting DCP residues on affinity is further supported by the 10-fold reduced binding affinity of ML-PD-05, which has three mutations compared to ML-YSD-07, all of which are non-contact residues, including two of the loop 5-stabilizing residues (T24E and T28N). Loop 2 residues of ML-YSD-07 also form key interactions with muLRRC15, as R12 forms salt bridge interactions with E206 and E230, and Y13 forms an H-bond with Y157 in addition to π-stacking with Y205 (Fig. 4g). In addition, D14, T15, and D16 all form side-chain electrostatic interactions with K273 (Supplementary Fig. S12e), despite both D14 and D16 not being subject to randomization during library designs. Overall, the binding affinity data of the ML-DCPs were in agreement with the experimental structure. We also solved the crystal structure of muLRRC15 alone and found it to be very similar to the complexed state, with an overall all-atom RMSD of 1.28 Å. The ML-YSD-07 binding interface was also structurally identical, indicating that the DCP did not induce a structural change upon binding (Supplementary Fig. S13a). The same was true for the Fab-E1 binding epitope (Supplementary Fig. S13b). Interestingly, a comparison of the experimental muLRRC15 structure with the predicted AlphaFold2 structure (all-atom RMSD is 2.52 Å) showed that, although the conformations of the individual leucine-rich-repeats and the ML-YSD-07 binding region were similar, the curvature of the extracellular domain was slightly more pronounced for the AlphaFold2 model (Supplementary Fig. S13c). Functionalizing the DCP ML-YSD-07 for PET imaging Due to their subnanomolar affinities for muLRRC15, ML-YSD-07 and ML-PD-03 emerged as prime candidates for in vivo imaging. Proteolytic stability experiments in mouse plasma showed that ML-YSD-07 was slightly more stable than ML-PD-03, with 86% and 62% intact peptide remaining at 1 h and 4 h, respectively (Supplementary Fig. S14). This stability was considered sufficient for imaging purposes given the short imaging duration and desired rapid clearance from circulation. Because ML-YSD-07 showed affinity similar to ML-PD-03 yet offered modestly better plasma stability, we prioritized ML-YSD-07 for in vivo evaluation. Positron emission tomography (PET), combined with Computed Tomography (PET/CT), enables non-invasive imaging of solid tumors and requires target-specific probes to be labeled with a positron-emitting radioisotope33. We selected the positron-emitting radioisotope 18F for labeling ML-YSD-07, as this is a clinically validated imaging radionuclide and has nuclear decay characteristics favorable for high resolution PET images53. The ligand RESCA chelates aluminium mono[18F]fluoride ([18F][AlF]2+) (Fig. 5a) at ambient temperature and mild conditions, enabling efficient radiolabeling of peptides54. We focused on direct conjugation of (+)-RESCA-TFP (2,3,5,6-tetrafluorophenol ester) to the free N-terminal amine of ML-YSD-07 (Supplementary Fig. S15a) since it is orientated away from the muLRRC15 binding site (see Fig. 4f). The RESCA-functionalized DCP, RESCA-ML-YSD-07, was successfully synthesized (Supplementary Fig. S15b) and retained subnanomolar affinity for muLRRC15 (Fig. 5a and Supplementary Fig. S15c, d) and was subject to 18F radiolabeling.Fig. 5PET/CT imaging of KPR murine PDAC tumors with [18F]AlF -RESCA-ML-YSD-07.a The DCP ML-YSD-07 was functionalized at the N-terminal amide with the chelating ligand RESCA, which captures ([18F]AlF)2+. The kinetic constants kon and koff, and the calculated equilibrium dissociation constant KD for RESCA-ML-YSD-07 (without [18F]AlF)2+) binding to muLRRC15 are shown as mean ± s.d. (N = three independent experiments). b Representative PET/CT images from two KPR tumor-bearing mice at 45–60 min after [18F]AlF-RESCA-ML-YSD-07 administration. The tumors are highlighted with a white dashed box. B = bladder; K = kidneys. c Radioactivity concentrations in the tumor and muscle were measured as the percentage of the injected dose per gram (% ID/g) during the 60 min dynamic PET scans. Data represents the mean ± s.d. (n = five animals); solid line, mean; shaded area, s.d. PET/CT imaging of KPR murine PDAC tumors with [18F]AlF -RESCA-ML-YSD-07. a The DCP ML-YSD-07 was functionalized at the N-terminal amide with the chelating ligand RESCA, which captures ([18F]AlF)2+. The kinetic constants kon and koff, and the calculated equilibrium dissociation constant KD for RESCA-ML-YSD-07 (without [18F]AlF)2+) binding to muLRRC15 are shown as mean ± s.d. (N = three independent experiments). b Representative PET/CT images from two KPR tumor-bearing mice at 45–60 min after [18F]AlF-RESCA-ML-YSD-07 administration. The tumors are highlighted with a white dashed box. B = bladder; K = kidneys. c Radioactivity concentrations in the tumor and muscle were measured as the percentage of the injected dose per gram (% ID/g) during the 60 min dynamic PET scans. Data represents the mean ± s.d. (n = five animals); solid line, mean; shaded area, s.d. RESCA-ML-YSD-07 was successfully radiolabeled with ([18F]AlF)2+, affording the conjugate [18F]AlF-RESCA-ML-YSD-07 at a molar activity of 151.7 GBq/µmol and with a radiochemical purity of 96% (Supplementary Fig. S15e). Therefore, we proceeded to perform PET/CT with [18F]AlF-RESCA-ML-YSD-07 using an established subcutaneous PDAC cancer model, in which female C57BL-6 mice were implanted with KPR murine PDAC tumor cells which develop a muLRRC15-expressing enriched fibroblast tumor microenvironment20. Once the tumors reached ~400 mm3 in size they were analyzed by FACS, confirming that the tumors contained a high level of LRRC15-expressing CAFs (PDPN+LRRC15+) amongst the total fibroblast (PDPN+CD31−) population (see Supplementary Fig. S16), in line with previous reports20. In representative images from dynamic PET scans, the KPR tumors could be clearly visualized in the right flank of the animals (Fig. 5b and Supplementary Fig. S17). At the 45–60-min post-injection time frame, average tissue uptake in tumor of [18F]AlF-RESCA-ML-YSD-07 was 3.50 ± 1.00 % ID/g compared to 0.89 ± 0.21 % ID/g in healthy muscle (Fig. 5c). Uptake in tumor was higher than in blood, bone, and lung, which were 1.71 ± 0.44, 2.50 ± 0.55, and 2.26 ± 0.91 % ID/g, respectively (Supplementary Fig. S18a, b). Moderate levels of tracer were detected in the bone, likely due to ([18F]AlF)2+ dissociation from the RESCA ligand, and quantification of tracer in the lungs was challenging due to their large air volume and respiratory motion. The tracer was predominately excreted through the renal pathway, as high levels of [18F]AlF-RESCA-ML-YSD-07 were present in the bladder and kidneys; although, a fraction of radioactivity was found in the gall bladder and intestine indicating a partial hepatobiliary clearance of the tracer or its radio-metabolites (Supplementary Fig. S18a, b). Biodistribution over the course of the dynamic PET scans showed that the tumor uptake of the tracer was stable 60 min post injection (Fig. 5c), and kinetic profiles of tumor-to-blood and tumor-to-muscle ratios suggest specific tracer binding with a slow washout (Supplementary Fig. S18c). Altogether, the PET imaging shows that the strong affinity of the DCP to muLRRC15, driven by a slow dissociation rate, translated to excellent tumor uptake and retention in vivo. Consistent with these observations, Patlak graphical analysis of the dynamic PET data demonstrated a linear uptake phase in tumors indicative of slowly reversible tracer accumulation (parametric images of Patlak Ki are shown in Supplementary Fig. S19 as projections alongside orthogonal slices through the tumor), further supporting specific and stable binding of [18F]AlF-RESCA-ML-YSD-07 to muLRRC15 in vivo. DCPs show weak binding to human LRRC15 Following the successful imaging of muLRRC15-expressing CAFs using the PDAC tumor model, we examined the binding affinity of the DCPs to human LRRC15 (huLRRC15) to assess their cross-species activity. SPR binding analysis showed that the three strongest muLRRC15-binding DCPs, ML-YSD-07, ML-PD-03, and ML-PD-08, had relatively lower affinities to huLRRC15 (KD of 123 nM, 85 nM, and 232 nM, respectively; Supplementary Fig. S20a). The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~150-250-fold reduced binding affinities compared to muLRRC15 were due to both slower association and faster dissociation binding kinetics (compare Supplementary Fig. S20b with Fig. 3a). Despite lower binding affinity to huLRRC15, ML-YSD-07 demonstrated clearly detectable binding to surface-expressed huLRRC15 of engineered 3T3 cells (Supplementary Fig. S20c). Although mouse and human LRRC15 share 84% sequence identity in their LRR domains, seven amino acid substitutions within the DCP binding epitope (derived from the ML-YSD-07:muLRRC15:Fab-E1 crystal structure; see Supplementary Fig. S20d) may potentially explain the reduced binding affinity to the human ortholog. Binding selectivity of the DCPs to LRRC15 over other LRR-containing proteins We explored the binding selectivity of the DCPs towards other LRR-containing proteins that share sequence homology with LRRC15. A protein BLAST search analysis using the extracellular LRR regions of muLRRC15 and huLRRC15, against their respective murine and homo sapiens databases in UniProt, identified proteins that share moderate sequence identity scores to LRRC15 (Supplementary Tables S7, S8). We were particularly interested in the homology of the LRR-containing proteins in relation to the major DCP binding site (LRR2–5 region of LRRC15) and sequence alignment analysis of the top BLAST hits indicated poor conservation of binding residues (Supplementary Fig. S21). To validate the binding selectivity of the DCPs, we selected two of the most homologous proteins to LRRC15, based on sequence and structure: insulin-like growth factor-binding protein complex acid labile subunit (IGFALS) and leucine-rich-repeat-containing G-protein coupled receptor 4 (LGR4) (Supplementary Fig. S22a, b). Remarkably, neither ML-YSD-07 nor ML-PD-03 showed any detectable binding to human IGFALS, human LGR4, or mouse LGR4 at 10 µM (Supplementary Fig. S22c), demonstrating >20,000-fold selectivity for muLRRC15. Although more LRR-containing proteins need to be investigated to fully characterize the binding specificity, the low preservation of key binding residues combined with our binding analysis provides evidence that ML-YSD-07 and ML-PD-03 specifically target the LRRC15 epitope rather than broadly recognizing the LRR fold. Discussion In this study, we implement a platform using combinatorial library display and ML to develop peptides that bind to muLRRC15, with ML-YSD-07 and ML-PD-03 displaying subnanomolar affinity and excellent selectivity for muLRRC15-expressing fibroblasts. As shown by HDX-MS epitope mapping studies and X-ray crystallography, the DCPs interact with a flat binding region on the concave muLRRC15 surface and utilize a complementary planar contact region comprised of residues located on three surface loops. One of these peptides, ML-YSD-07, was radiolabeled with 18F for PET imaging studies and it successfully visualized murine PDAC KPR tumors, validating LRRC15-expressing fibroblasts as a tumor imaging target. The herein identified DCPs were derived from the cystine-knot peptide EETI-II, which was previously used to develop target-specific DCP binders41,50,55–57. Our muLRRC15-binding DCPs did not display a cysteine-knot motif and had evolved to form a rearranged cysteine connectivity and altered overall conformation. The fold of the herein derived DCPs, exemplified by ML-YSD-07, is apparently unique, as no structurally similar folds are found in the PDB database (using directory 240101), according to Foldseek58. Although the cystine-knot motif is typically attributed with desirable pharmaceutical properties, its ability to evolve to form new thermodynamically stable folds with moderate plasma stability substantially expands the structural diversity of DCP-based libraries to enhance target-specific lead peptide identification. The flat binding surface of ML-YSD-07, essential for recognizing the shallow muLRRC15 interface, arose from the rearranged disulfide connectivity, which induced a complete reorganization of the primary binding loops. This structural adaptation allowed the DCP to engage LRRC15 through an extensive interface comprising 10 contact residues—a striking example of this scaffold’s capacity to evolve binding surfaces suited for challenging, low-curvature protein targets. The ability of the DCPs to feature a different fold following library screening and ML analysis highlights the advantage of training the ML model based on primary amino acid sequences alone, with no structural or target protein information provided. Although recent advancements in protein structure-based prediction tools have demonstrated remarkable accuracy44,59,60, such models are often guided by priors derived from known folds, which may favor canonical disulfide connectivities61. In contrast, our sequence-based approach was able to identify peptides with a rearranged connectivity that achieve high target affinity. The power of the ML model is exemplified by its accurate prediction of DCP motifs that are involved in muLRRC15 binding, such as the Y5WWEWE10 loop 1 motif of ML-YSD-07, which was largely preserved from the 1st generation hit EE-D1 (Y5WWDWE10) and was subject to very limited diversity in the ML-library. The crystal structure confirmed that this motif was indeed important for both DCP structure (Y5, W6) and binding to muLRRC15 (W7, W9, E10). The structure further revealed that additional residues from loop 1, and residues from loops 2 and 5 formed key binding interactions with muLRRC15, driving affinity improvements compared to EE-D1. The conservation of key binding residues in loops 1, 2 and 5 suggests that all identified high-affinity DCPs engage the same binding site on LRRC15 in a similar manner. Overall, the ML model was able to successfully generate unique sequences with strong target binding affinity after analyzing enrichment scores data from the second-generation libraries. In this project, fixed thresholds were applied as a pragmatic strategy to enrich for the top predicted binders while filtering out non-specific sequences; future work with larger affinity datasets will enable robust calibration of model scores against experimental KD values. Moreover, while we trained separate models for target-specific and non-specific binding to simplify training and maintain interpretability, future extensions could explore multi-task or adversarial architectures with a shared encoder, which may provide a more principled trade-off between specificity and affinity. Following the discovery of DCPs with high-affinity for muLRRC15, we developed [18F]AlF-RESCA-ML-YSD-07, a peptide tracer that enables imaging of LRRC15-expressing tumors in vivo. Upon injection to KPR tumor-bearing mice, the tracer was able to clearly image the tumor, even though the cancer cells themselves, which are of epithelial origin, are LRRC15-negative. Therefore, in agreement with immunohistochemical and immunofluorescence analysis18,19, the PET signal originates from the LRRC15-expressing fibroblasts that populate the tumor stroma. Because receptor levels on specific CAF populations can vary due to their heterogeneous origins, tracers require efficient tumor localization and strong target binding affinity. The small size of [18F]AlF-RESCA-ML-YSD-07 favoring tissue penetration and its strong subnanomolar binding affinity to muLRRC15 translated into strong tumor accumulation within 20 min of tracer administration and good tumor retention after 1 h. The favorable tumor-to-blood and tumor-to-muscle uptake ratios for [18F]AlF-RESCA-ML-YSD-07, in addition to the linear uptake phase in the Patlak analysis, further support target-specific tumor localization and retention. We observed that relatively high levels of [18F]AlF-RESCA-ML-YSD-07 were also retained in the kidneys, which however was not LRRC15 mediated as negligible LRRC15 expression levels are present in healthy mouse tissues20. Kidney accumulation is a common limitation of peptide-based PET tracers, which is largely attributed to reabsorption and metabolism of the tracer in the proximal convoluted tubule. Strategies such as the use of alternative radioprosthetic groups and protease-cleavable linkers have been proposed to reduce non-specific uptake in healthy kidney tissue62–64, though further work is required to determine whether these approaches can be broadly applied. [18F]AlF-RESCA-ML-YSD-07 exhibited rapid systemic clearance, despite slow target dissociation, limiting the risk of immunogenicity which is typically minimal for peptide-based PET tracers65. In general, the overall low accumulation of the DCP tracer in healthy tissues emphasizes the attractiveness of LRRC15 as a CAF biomarker. In addition to PDAC, a LRRC15-expressing CAF signature associated with elevated TGFß levels is also present in breast, lung, ovarian, colon, renal, esophageal, and bladder cancers, as well as stomach adenocarcinoma19, highlighting the broad utility of LRRC15-targeting peptide imaging agents, exemplified by [18F]AlF-RESCA-ML-YSD-07. LRRC15 has also been identified as a potential marker for platinum-based resistance in patients with ovarian cancer66, suggesting that LRRC15-targeting imaging tracers can be used to guide patient treatment for various types of cancer. Complementary to characterizing LRRC15 expression in tumors, DCP tracers can be used to monitor target engagement of LRRC15-directed therapies, such as the antibody-drug conjugate ABBV-085 that showed promising results in a phase I trial for osteosarcoma18,30. LRRC15-targeted PET enables highly sensitive and quantitative detection of LRRC15-expressing stromal compartments, providing complementary biological information to existing fibroblast-targeted tracers such as FAPI. Importantly, because LRRC15 marks a biologically distinct subset of cancer-associated fibroblasts involved in immunosuppression and therapy resistance, LRRC15-targeted PET imaging has the potential to offer orthogonal insights beyond those accessible with current fibroblast-directed imaging. However, enhancing the binding affinity of the identified DCPs for huLRRC15 will be essential for their further development. The ability of DCPs to achieve subnanomolar affinity in binding to the flat muLRRC15 surface demonstrates that LRR domains are amenable to peptide-based targeting, opening opportunities to explore the biology of LRR-containing proteins involved in diverse signaling pathways67. The concave surfaces of LRR domains are common interactions site for ligands, such as glycoprotein hormones68, and this study suggests that engineered DCPs offer an attractive moiety to study LRR-containing protein function and therapeutic pathway modulation. Vital to this, we demonstrated that the high affinity DCPs identified in this study selectively bind to LRRC15 over other LRR-containing proteins (LGR4 and IGFALS) that share sequence homology with LRRC15. Although the LRR framework is structurally conserved, sequence variation within the repeats is highly prominent and suggests that target selectivity is achievable. In summary, this study demonstrates the power of ML-enhanced peptide combinatorial library screening to discover DCPs with altered disulfide connectivity, which gave rise to a divergent peptide conformation that was perfectly adapted to interact with muLRRC15, a challenging protein target featuring an inconspicuous surface topology. The ability of the ML model to accurately predict binding fitness, based on peptide sequence enrichment from phage display alone, suggests that this platform can be widely adopted across peptide discovery efforts. The successful translation of high-affinity DCPs into a functional PET imaging tracer of muLRRC15-expressing CAFs in a mouse PDAC model provides a possible foundation for the further development of a huLRRC15-specific PET imaging agent for cancer-imaging purposes. LRRC15 is an attractive biomarker across multiple solid tumor indications, and the peptides identified in this study could be valuable lead candidates to monitor LRRC15 dynamics during tumor development. Methods Ethics statement All animal studies were conducted with the approval of and in accordance with the ethical guidelines of Genentech’s AAALAC accredited Institutional Animal Care and Use Committee. For the tumor xenograft studies, mice were monitored daily for health and tumor progression. The tumor burden never exceeded the maximum permitted limit of 2000 cubic mm. Recombinant protein expression and purification For library affinity selections and binding studies, the extracellular regions of muLRRC15 (Uniprot ID Q80X72; residues Y22–S475) and huLRRC15 (Uniprot ID Q8TF66; residues Y22–S475) with C-terminal AviTag and His6 tag were expressed in Trichoplusia ni (T. ni) cells. T. ni cells (10 L) were maintained in ESF921 (Expression Systems) at 27 °C, and cells were infected at 2 ×106 cells/mL with a multiplicity of infection of 2. After 48 h, the fermentation was treated with 500 mL of 1 M Tris pH 8.0, 50 mL of 1 M CaCl2, and 10 mL of 1 M NiCl2, followed by centrifugation and filtration. Harvested supernatant containing crude protein was purified by nickel affinity chromatography (Ni-NTA agarose, Qiagen) by eluting with 300 mM imidazole in 50 mM Tris pH 7.5 0.3 M NaCl. Proteins were further purified by size exclusion chromatography (SEC) using a Sephacryl S-200 column (GE Healthcare) equilibrated with 20 mM HEPES pH 7.2, 100 mM NaCl. Protein-containing fractions were combined, concentrated, and stored at −80 °C. MuLRRC15 and huLRRC15 were site-specifically biotinylated at the AviTag using the BirA500 biotin-protein ligase kit (Avidity LLC). Briefly, 2 mL of 13 µM (0.7 mg/mL) LRRC15 was added to 200 µL BiomixA (0.5 M bicine buffer, pH 8.3) and 240 µL BiomixB (100 mM ATP, 100 mM MgOAc, 500 μM d-biotin), followed by the addition of BirA (1 mg/mL, 8 µL). Reactions were carried out for 2 h at room temperature, and biotinylation of the LRRC15 proteins was confirmed by electrospray ionization mass spectrometry (ESI-MS). The biotinylated proteins were then re-purified by SEC as described above. For structural studies, the extracellular region of muLRRC15 (residues 22–475 with a C416S point mutation, muLRRC15.22–475.C416S) with N-terminal His6 and Flag tags was expressed and purified as described above. Flag-tag cleavage was performed by adding 30 µL of EKMaxTM enterokinase (Invitrogen) to 1 mL of 8 mg/mL muLRRC15.22-475.C416S in 20 mM HEPES pH 7.2, 150 mM NaCl and incubating reactions overnight at room temperature. Flag-tag cleavage was confirmed by ESI-MS and proteins were purified by SEC as described above. Protein deglycosylation was performed by adding 3 µL of 1 mg/mL endoF3 (New England Biolabs) to 1 mL of 2.1 mg/mL muLRRC15.22–475.C416SΔFlag and incubating reactions for 2 h at room temperature. Deglycosylation was confirmed by ESI-MS and the final protein was re-purified by SEC equilibrated with 20 mM HEPES pH 7.2, 100 mM NaCl as described above. 1st generation DCP libraries construction and panning All phage libraries were constructed as previously described69 using the phagemid pComb3 template. The Kunkel mutagenesis approach was used to build the 1st generation DCP libraries (see Supplementary Table S1 for library designs) as previously described41. Residues were randomized with the degenerate codon encoding 19 natural amino acids except cysteine. Oligonucleotides for mutagenesis were synthesized using 28 custom mixes of trimer phosphoramidites encoding for 19 amino acids at equimolar concentration (Glen Research). The libraries were fused via their C-terminus to the N-terminus of gene 8 (g8) M13 bacteriophage major coat protein, with a gD (HSV-2 envelope glycoprotein D)-tag fused to the N-terminus of the DCP sequences. All phagemids were transformed into E. coli and grown at 37 °C overnight to display fusion proteins. The propagated phage were purified and resuspended in 1 mL PBT buffer (PBS, 0.5% (w/v) BSA and 0.1% Tween20). For three rounds of selections, biotinylated muLRRC15 (14 µg/mL) was immobilized to Dynabeads™ MyOne™ Streptavidin T1 (Invitrogen) followed by adding DCP-phage in PBT buffer for 1 h at room temperature. The fourth round of selection was performed by immobilizing muLRRC15 (5 µg/mL) on neutravidin-coated Nunc Maxisorb immunoplates (Invitrogen), followed by DCP-phage incubation. For all selections, captured phage were eluted with 400 µL of 100 mM HCl for 5 min and eluates were quenched with 50 µL of 1 M Tris pH 11.0. Single phage clones from the third and fourth round of selections were amplified and analyzed by phage ELISA against immobilized biotinylated muLRRC15 using an anti-M13 horseradish-peroxidase (HRP)-conjugated antibody (expressed in-house) to detect phage binding. An anti-gD HRP-conjugated antibody (expressed in-house) was used to detect peptide display levels on the phage to normalize the muLRRC15 binding ELISA signals. Phage titration ELISAs against immobilized biotinylated muLRRC15 were performed as previously described70. Phage clones with positive ELISA signal were subjected to Sanger DNA sequencing69. Phage competition binding assay Oligonucleotides for the DCP EE-D1 were ordered from Genenlink and phage displaying EE-D1 fused via its C-terminus to the N-terminus of the gene 3 (g3) M13 bacteriophage major coat protein (EE-D1-g3) were constructed using the phagemid pComb3 template and purified using standard protocols69. Biotinylated muLRRC15 (2 µg/mL) was immobilized on neutravidin-coated Nunc Maxisorb immunoplates (Invitrogen). EE-D1-g3 (final OD268 of 0.1) was pre-mixed with chemically synthesized DCPs titrated at 5-fold dilutions from 100 µM in PBT buffer. EE-D1-g3 phage that remained bound to immobilized muLRRC15 in the presence of the DCPs was detected using an anti-M13 HRP-conjugated antibody (expressed in-house). Ala-phage library construction and panning A shotgun-scanning Ala library (EE-D1-Ala) was constructed as previously described into the phagemid pComb3 template45. Briefly, non-cysteine and non-alanine residues were preferentially mutated between the wild-type residue of EE-D1 and alanine. Oligomers were ordered from Integrated DNA Technologies (IDT) and DNA degeneracies are represented by IUB code: K = G/T, M = A/C, N = A/C/G/T, R = A/G, S = G/C, W = A/T, Y = C/T. The respective codon for each mutated amino acid and prevalence of amino acids are shown in Supplementary Fig. S4b. The EE-D1-Ala library was fused to g8 as described above. Selections were performed against pre-immobilized biotinylated muLRRC15 (14 µg/mL) using Dynabeads™ MyOne™ Streptavidin T1 (Invitrogen). Captured phage were eluted as described above. Phage pools after each round were subjected to amplicon sequencing targeting the randomized peptide region and subject to next generation sequencing (NGS) on the Illumina MiSeq platform. Results were analyzed using a heatmap plot comparing the enrichment levels of the parent EE-D1 residue to the input library. 2nd generation DCP libraries construction and panning Oligonucleotides for the 2nd generation phage libraries were from Genelink and IDT, and libraries were constructed using the phagemid pComb3 template using standard protocols69. Library designs were based on the EE-D1 sequence, and randomized peptide loops either carried 100% of randomized amino acids (hard randomization) or roughly 50% of the wildtype sequences and 50% of randomized amino acids (soft randomization) (Supplementary Fig. S4c). The libraries were fused via their C-terminus to the N-terminus of the g3 M13 bacteriophage major coat protein, with a gD-tag fused to the N-terminus. Selections were performed against pre-immobilized biotinylated muLRRC15 (14 µg/mL), and captured phage were eluted as described above. Negative selections were performed in parallel against streptavidin-coated magnetic beads to determine non-specific binding, and phage were eluted as described above. After four rounds of selections and enrichment (see Supplementary Fig. S23), phage pools after each round were subjected to NGS. Deep learning model We employed a proprietary transformer-based language model developed and maintained by DeepSeq.AI (San Francisco, CA) to learn the relationship between peptide sequences and their enrichment behavior during phage display selection71. The core architecture is based on RoBERTa, a masked language model initially pretrained on curated protein sequence datasets, including UniProt and MGnify72,73. To adapt the model for our regression task, we appended a regression head consisting of a pooling layer followed by a fully connected linear layer, which outputs continuous binding scores. Notably, the model requires only primary amino acid sequences as input and does not utilize any structural information or binding affinity measurements during training. Data processing NGS data were collected from 2nd generation DCP-phage display selections targeting muLRRC15. A negative control selection against streptavidin-coated magnetic beads was included to quantify non-specific binding. Raw reads were demultiplexed, quality-filtered based on Phred scores, and translated into peptide sequences. Sequences with read counts below a predefined threshold were removed to ensure statistical robustness. After filtering, 160,545 unique peptide sequences remained for downstream analysis. For each peptide, an enrichment score was calculated as the fold change in normalized read counts across selection rounds, providing a proxy for target binding strength. To correct for batch effects and ensure comparability across samples and selection rounds, enrichment scores were further processed using quantile normalization. This normalization ensured that the distribution of scores remained consistent between datasets while preserving rank order. Model training and evaluation strategy Two separate machine-learning models were fine-tuned: one to predict muLRRC15-specific binding and the other to identify non-specific binders. The training dataset for target-specific prediction was derived from the muLRRC15 selection screen, while the non-specific model was trained using the control screen against streptavidin/neutravidin lacking antigen. To evaluate model generalization, we employed a five-fold cross-validation strategy in which sequences were clustered based on pairwise similarity and split into disjoint folds. This ensured that training and validation folds represented distinct sequence classes, reducing the risk of overfitting and simulating prediction on unseen peptide scaffolds. For each fold, 80% of the data were used for training and 20% held out for evaluation. This hold-out strategy was scaffold-aware and designed to assess generalization across peptide families rather than individual variants. Training was conducted using the Adam optimizer with a learning rate of 1e−5 and a linear warm-up schedule over the first 300 steps, followed by exponential decay (rate = 0.01). The loss function was mean squared error (MSE), and early stopping was applied based on validation set performance to prevent overfitting. Model training was performed on NVIDIA A100 GPUs using mixed-precision (FP16) training to accelerate convergence and reduce memory usage. True hold-out test To evaluate whether a model trained exclusively on sequence-derived enrichment data, without access to affinity measurements or structural information, can predict actual binding affinity, we conducted a true hold-out test using peptide sequences with experimentally determined affinities (KD) measured by surface plasmon resonance (SPR). These sequences were entirely excluded from both training and validation sets and originated from scaffolds not seen during model development, ensuring a stringent test of model performance. For assessment, predicted binding scores were compared against the corresponding SPR-derived KD values for several 1st generation binders. Correlation was quantified using the coefficient of determination (R²) from a linear regression analysis. This comparison, while limited to four sequences, was used as a check to confirm that the model captured meaningful binding information (Fig. 2c). More importantly, model-designed DCPs tested experimentally exhibited significantly stronger binding affinities than the 1st generation hits (Supplementary Fig. S8; p < 0.0001, Mann–Whitney U test), providing direct evidence of the model’s ability to generate high-affinity binders. Generation of the machine-learning library (ML-library) Using the trained model, 10 million DCP sequences, not present in the training or test data set, were computationally generated within the DCP framework used for the 2nd generation libraries. A multi-step filtering and selection process was applied to construct the final sequence library. First, sequences with a non-specific binding score below −0.1 were retained, a threshold derived from prior empirical observations to minimize off-target interactions while maintaining sequence diversity. Next, sequences with high predicted binding scores were prioritized. For reference, the 1st generation DCP binder EE-D1 (KD 0.45 µM) had a predicted binding score of 1.4. Based on this threshold, sequences with scores above 2.1 were expected to exhibit significantly stronger muLRRC15 binding affinities than EE-D1 while maintaining a sufficient number of binders for further selection. Within this high-scoring population, sequence length distribution was skewed, with 91% consisting of 30-mers, 8% having 31 amino acids, and 1% having 32 amino acids. Given this bias, the final selection was adjusted to ensure a broader sequence length representation. The final library composition included 80% of the highest predicted enrichment score sequences as 30-mers, 15% as 31-mers, and 5% as 32-mers, ensuring a balance between high predicted affinity and sequence diversity. The resulting 50,000 sequences (40,000 30-mers, 7500 31-mers, and 2500 32-mers) constituted the ML-library. Phage-displayed ML-library construction and panning Oligomers for the 50,000 sequences that represent the ML-library were ordered from TWIST Bioscience. The libraries were fused via their C-terminus to the N-terminus of the g3 M13 bacteriophage major coat protein, with a gD-tag fused to the N-terminus. Two rounds of selections were performed against pre-immobilized biotinylated muLRRC15 (5 µg/mL) to Dynabeads™ MyOne™ Streptavidin T1 (Invitrogen). The third round of selection was performed by capturing biotinylated muLRRC15 (2 µg/mL) to neutravidin-coated Nunc Maxisorb immunoplates (Invitrogen). Captured phage were eluted as described above. Single phage clones were amplified and analyzed in phage titration ELISAs against biotinylated muLRRC15. Phage pools after each round were subjected to NGS. Yeast surface displayed ML-library construction and panning Oligomers for the ML-library were PCR amplified and the yeast surface display library was prepared and purified as previously described71. Library selections and FACS sorting were performed using standard protocols74. Four rounds of panning were performed in solution using biotinylated muLRRC15 (500 nM, 250 nM, 100 nM, and 100 nM, for rounds one to four, respectively). FACS sorting was performed with anti-Myc labeling (primary - chicken IgY anti-c-Myc, Invitrogen; secondary - Alexa Fluor 488-goat anti-chicken IgG, Invitrogen) to normalize peptide display, and streptavidin-allophycocyanin (APC) (Invitrogen) to detect muLRRC15 binding using a FACS sorter Aria II Fusion with purity mode (BD Biosciences). Individual yeast clones were selected, analyzed by FACS and subjected to Sanger DNA sequencing. Sorted libraries were subjected to NGS. Peptide synthesis and functionalization Peptides were synthesized using standard 9-fluorenylmethoxycarbonyl (Fmoc)-based solid-phase peptide synthesis. The linear peptides (0.5 mg/mL) were oxidized in optimal folding conditions using either 0.1 M ammonium bicarbonate pH 9.0, 2 mM reduced glutathione, 0.5 mM oxidized glutathione, 4% (v/v) DMSO, 0.3 M arginine; or 0.1 M ammonium bicarbonate pH 8.0, 1 mM reduced glutathione, 50% (v/v) DMSO for 24 h at room temperature with shaking conditions. Oxidized peptides were purified by C18 Sep-Pak (Waters) using preparative reversed-phase (RP)-HPLC column using a linear gradient of 1–100% (v/v) acetonitrile containing 0.05% (v/v) trifluoroacetic acid (TFA) and analyzed by ESI-MS. All non-functionalized DCPs used in this study were synthesized at CS Bio or WuXi AppTec. All peptides were purified to >90% as determined by analytical RP-HPLC. The DCPs ML-YSD-07-N488 and ML-PD-03-N488 were fluorescently labeled with AF-Dye-488 at the free amino group of the N-terminal Gly1 residue to generate ML-YSD-07-N488 and ML-PD-03-N488, respectively. DCPs (100 µL, 3 mM in DMF) were diluted to 300 µL in 100 mM sodium bicarbonate pH 8.0, followed by addition of powdered AF-Dye-488 NHS ester (Lumiprobe) (5 molar eq.). Reactions were mixed for 1 h at room temperature and reaction progress was monitored by analytical RP-HPLC and ESI-MS. The fluorophore-labeled DCPs were purified by RP-HPLC with a 5 µm C18 100 Å pore size, 21.2 ×100 mm column (Phenomenex) using a linear gradient of 20–70% (v/v) acetonitrile containing 0.05% (v/v) TFA over 20 min. Surface plasmon resonance Surface plasmon resonance (SPR) experiments were performed at 25 °C using a Biacore T200 instrument (Cytiva). Biotinylated muLRRC15 was captured to a level of 250 RU and 1000 RU for kinetic affinity and steady state affinity, respectively, on a streptavidin sensor chip (Cytiva). All peptides were prepared in HBS-EP buffer (10 mM HEPES pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.05% v/v Surfactant P20). A reference channel and blank injection were subtracted (double referencing) from the muLRRC15-captured channel. Six serial three-fold dilutions of peptide concentrations were injected for 120 s at 30 µL/min. Peptides where the affinities were derived from steady state affinity were tested using both a single cycle and multi cycle injection method (affinities derived from the different methods were in agreement with each other). Peptide affinities derived from the kinetic model analysis were solely tested using the multi cycle injection method. Information regarding peptide maximum concentrations and the model used to determine peptide affinities for muLRRC15 are shown in Supplementary Table S4. Biotinylated huLRRC15 was captured to a level of 500 RU on a streptavidin sensor chip (GE healthcare). Peptide maximum concentrations for determining ML-YSD-07, ML-PD-03, and ML-PD-08 binding affinities for huLRRC15 were 1000 nM and affinities were derived using the steady state affinity model. Fab-E1 was tested with a maximum concentration of 300 nM and muLRRC15 binding affinity was determined from the kinetic model analysis using the multi cycle injection method. Data were analyzed using BIAevaluation software (Cytiva, v.1.1.1) using a 1:1 Langmuir binding model. The KD values were the mean ± s.d. of 3–4 independent experiments. Rmax and Chi2 values were the mean ± s.d. of 3–4 independent experiments (Supplementary Table S5). SPR competition binding experiments with Fab-E1 and ML-PD-03 were carried out using the A-B-A injection method. ML-PD-03 (5 nM) was injected over the muLRRC15 immobilized surface for 120 s to occupy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}~80% of all available binding sites, before a solution of ML-PD-03 (5 nM) and Fab-E1 (500 nM) were immediately injected over the muLRRC15:ML-PD-03 surface for 120 s. The experiment was also performed in the absence of ML-PD-03 to determine if Fab-E1 had reduced binding response when muLRRC15 was pre-complexed with ML-PD-03. Isothermal titration calorimetry All measurements were performed on a MicroCal 200 iTC instrument (Malvern Panalytical) in 50 mM Tris pH 7.5, 150 mM NaCl, 3% (v/v) DMSO. MuLRRC15 was loaded in the sample cell at 20 μM and EE-D1 was loaded into the syringe at 200 μM. Each titration experiment consisted of an initial 0.4 μL injection, followed by 13 injections of 3 μL. The stirring speed was set to 750 rpm. The cell temperature was kept constant at 25 °C. Results represent the data from a single experiment. Data were fitted to a one-site binding model. Peptide stability assays in mouse plasma Mouse plasma was thawed and centrifuged at 3750 rpm for 15 min at 4 °C to remove blood cells. Plasma supernatant was adjusted to pH 7.25–7.35. The three DCPs ML-YSD-07, ML-PD-03 and EETI-II were added to 50 µL plasma to a final DCP concentration of 1 µM and incubated for 0 min, 15 min, 1 h, and 4 h. At the respective time points, reactions were quenched with 350 µL acetonitrile with 5% (v/v) formic acid (FA) and vortexed for 25 min. Samples were centrifuged at 3750 rpm for 15 min at 4 °C and supernatants were analyzed by LC-MS/MS (Thermo Fisher Scientific Inc.). The percentage of peptide remaining was quantified by measuring the peak area from the LC-MS/MS chromatograms at 214 nm and normalized to the levels detected at the 0 h time point. Results represent the data from two independent experiments. Cell culture The NIH/3T3 cell line (referred to as 3T3) was purchased from ATCC. The 3T3-muLRRC15 and 3T3-huLRRC15 cell lines were generated by transfecting 3T3 cells with a PiggyBac Transposon Vector System carrying the muLRRC15 gene (pb_EF1a_muLRRC15) and the huLRRC15 gene (pb_EF1a_huLRRC15) by lipid-based transfection using Lipofectamine 2000 (Invitrogen). Cells were then selected based on high muLRRC15 or huLRRC15 expression by FACS using the anti-LRRC15 antibody M25 (expressed in-house). The KPR3070 mouse PDAC cell line (KPR; Kras LSL.G12D/wt;p16/p19 fl/wt;p53 LSL.R270H/wt;Pdx1.Cre), that has no LRRC15 expression, was previously described48. All cell lines were cultured in DMEM with 10% heat-inactivated FBS and 2 mM L-glutamine. Flow cytometry 3T3-muLRRC15, 3T3, KPR3070, and 3T3-huLRRC15 cells were washed in FACS buffer (PBS supplemented with 0.5% (w/v) BSA and 3 mM EDTA) and resuspended to 2 ×105 cells per sample. Cells were incubated with fluorophore-labeled DCPs ML-YSD-07-N488 or ML-PD-03-N488 (1 µM) for 10 min at 4 °C and subsequently washed twice in FACS buffer. Dead cells were stained with SYTOXTM Blue (1:10,000, Invitrogen). Cells were analyzed using a FACSymphony (BD Biosciences) and data were visualized using FlowJo (v.10.10.0). Samples were analyzed from three independent experiments. Fluorescence microscopy 3T3 and 3T3-muLRRC15 cells were plated in a black µ-Plate 96 well plate (ibidi) at 20,000 cells/well and were grown in DMEM with 10% FBS and 2 mM L-glutamine for 48 h at 37 °C, 5% CO2. Cells were treated with 500 nM of ML-PD-03-N488 in DMEM with 10% FBS and 2 mM L-glutamine for 1 h. Cells were washed with Hanks’ Balanced Salt Solution (+CaCl2 +MgCl2) (HBSS). Cells were then treated simultaneously with CellBrite® Steady 650 Membrane Staining reagent (Biotium) (1:10000 dilution in HBSS) and NucBlue™ Live ReadyProbes™ Reagent (Hoechst 33342, Invitrogen) (5 drops added to 10 mL HBSS) for 10 min to stain outer cell membranes and nuclei, respectively. Fluorescent images were captured using a Leica TCS SP8 confocal microscope with a HC PL APO CS2 63×/1.40 Oil objective. Hoechst 33342 was imaged with excitation at 405 nm and emission collected from 413 to 486 nm. ML-PD-03-N488 was imaged with excitation at 488 nm and emission collected from 501 to 602 nm. CellBrite® Steady 650 was imaged with excitation at 633 nm and emission collected from 646 to 703 nm. Cells were incubated at 37 °C with 5% CO2 during imaging. Partial reduction for disulfide mapping ML-PD-05 (1 mg) was dissolved in 500 µL of 10 mM glycine buffer pH 3.0 containing 20% (v/v) acetonitrile. Tris(2-carboxyethyl)phosphine (TCEP) (2 eq.) was added and the reaction was incubated at 37 °C for 30 min. LC-MS detected the formation of major [M + 2] and [M + 4] peaks. N-Ethylmaleimide (NEM) was then added to a final concentration of 50 mM, to cap the free thiols, and left at room temperature for 1 h. The crude material was purified by preparative RP-HPLC (C18 column, 5 μm, 10 × 250 mm, 100 Å, Phenomenex) using a linear gradient of 20–40% (v/v) acetonitrile containing 0.05% (v/v) TFA over 20 min. The eluted sample which contained both the [M + 2xNEM] and [M + 4xNEM] was lyophilized. The sample was dissolved in 100 µL 50% (v/v) acetonitrile and fully reduced with 20 mM dithiothreitol (DTT) at 37 °C for 1 h. Then iodoacetamide (IAA, 40 mM) was added to alkylate all remaining free thiols at pH 8.5, and the reaction was incubated at room temperature for 1 h. The sample was immediately analyzed by LC-MS/MS (Agilent 6545XT AdvianceBio LC Q/TOF), and the two species [M + 2xNEM + 4IAA] and [M + 4xNEM + 2IAA] were separated for targeted MS/MS analysis by a Poroshell 120 EC-C18 column (1.9 µm, 2.1 ×60 mm) using a linear gradient of 5–60% (v/v) acetonitrile containing 0.05% (v/v) FA over 5 min. The MS2 fragment ions were analyzed to identify the chemical modified cysteines (see Supplementary Fig. S9c, d). Hydrogen deuterium exchange - mass spectrometry (HDX-MS) muLRRC15 (30 µM) was incubated with the DCP EE-2-1 (33 µM) (1:1.1 molar ratio) for 1 h in 20 mM Tris-HCl pH 7.5, 100 mM NaCl. A sample was also prepared with no DCP present. Samples (3 μL each) were diluted into 60 μL of deuterated labeling buffer (10 mM HEPES pD 7.1, 150 mM NaCl) to initiate isotopic labeling. The exchange process was slowed after variable labeling times by quenching the reactions with equal volume of 4 M GdmCl, 1 M glycine pH 2.5. The samples were injected into a temperature-controlled chamber at 1 °C for online proteolytic digestion and chromatographic separation. Samples were passed through a pepsin and fungal protease XIII column (w/w, 1:1) (NovaBioassays), followed by loading onto a BEH C8 Vanguard trap column (Waters) where they were washed for 3 minutes. Following this, digested muLRRC15 fragments were analyzed using a BEH C18 HPLC column (Waters Acquity UPLC) using a linear gradient of 1–100% acetonitrile and analyzed by ESI-MS (ThermoScientific Exploris 480) to measure deuterium labeling. Experimental manipulations were performed using a custom-built Leap HDX Pal DHR platform (Leap Technologies). Samples were analyzed using the method of fragment separation75, and processed using the ExMS2 program76. Protection factors were determined using a previously reported empirical method77. Fab phage display and identification of Fab-E1 The Fab YSGX phage library, based on the humanized Fab-4D5 scaffold, was generated as previously described52. Three rounds of selections were performed against pre-immobilized biotinylated muLRRC15 (14 µg/mL) to Dynabeads™ MyOne™ Streptavidin T1 (Invitrogen). Selection rounds four, five, and six were performed by capturing biotinylated muLRRC15 (5 µg/mL) to neutravidin-coated Nunc Maxisorb immunoplates (Invitrogen). Captured phage were eluted using 100 mM HCl, followed by quenching eluates with 1 M Tris pH 11.0, as described above. For round six, selections were performed in the presence of 30 µM ML-YSD-07. Single phage clones identified after round six selections were amplified and analyzed in phage ELISA against biotinylated muLRRC15 in the presence of ML-YSD-07 (1 µM). Phage clones with positive ELISA signal in the presence of ML-YSD-07 were subjected to Sanger DNA sequencing. All phage clones sequenced corresponded to a single clone named Fab-E1. Fab-E1 was expressed in Chinese hamster ovary (CHO) cells and purified by protein G resin-based affinity chromatography followed by SEC purification using a Sephacryl S-200 column (GE Healthcare) equilibrated with 20 mM histidine acetate pH 5.5, 150 mM NaCl. Apo-muLRRC15 structure determination by X-ray crystallography The extracellular domain of muLRRC15 (residues 22–475 with C416S point mutation) was concentrated to 11 mg/mL in 20 mM HEPES pH 7.2, 100 mM NaCl, and added to powdered EE-D1 in excess. Due to its poor aqueous solubility, non-solubilized EE-D1 was removed (the final EE-D1 concentration was not quantified), and crystallization screening was performed in sitting drops. Single crystals were obtained in 0.1 M sodium malonate pH 6.0, 12% (w/v) PEG3350 from hanging drop optimization screens. Crystals were cryo-protected in precipitant supplemented with 25% (v/v) glycerol. Diffraction data were collected at the Advanced Light Source (ALS) beamline 5.02 at 1.00004 Å beam wavelength at a temperature of 100 K. Data were processed to a 2.26 Å resolution cutoff using XDS78. Initial attempts to phase by molecular replacement using an AlphaFold2 model of muLRRC15 yielded a poor solution judged by statistics (LLG = 108.47, TFZ = 10.4) and visual inspection of the resulting density map. To determine a solution by molecular replacement, the AlphaFold2 model of mLRRC15 was split into four fragments (residues 22–170, 171–281, 282–320 and 321–470) and used as search models, which yielded a clear solution (LLG = 1421.54, TFZ = 22.1). Structure comparison revealed a less pronounced curvature of the extracellular LRR region for the crystal structure compared to the AlphaFold2 prediction. The structural model was improved with iterative rounds of building and refinement in COOT79 and PHENX80 respectively to produce the final model with an R-work/R-free of 0.2071/0.2448 (Supplementary Table S6). Of modeled residues, 95.03% exhibit Ramachandran favored angles, 4.97% allowed and 0% outliers. No density was observed for the EE-D1 peptide, suggesting that it was not bound to muLRRC15 under the crystallization conditions and, therefore, the structure represents the apo-muLRRC15 form. ML-YSD-07:muLRRC15:Fab-E1 structure determination by X-ray crystallography The extracellular domain of muLRRC15 (residues 22–475 with C416S point mutation) was added to Fab-E1 (1:1.5 molar ratio, respectively) and mixed for 1 h at 4 °C. The complex was purified by SEC using a Sephacryl S-200 column (GE Healthcare) equilibrated with 20 mM HEPES pH 7.2, 100 mM NaCl. The purified muLRRC15:Fab-E1 complex was concentrated to 16 mg/mL and added to 500 µM (final concentration) of ML-YSD-07 and mixed for 2 h at 4 °C. Sitting drop crystallization screens were conducted with 1:1 sample:precipitant ratio. Single crystals were obtained in 0.1 M Tris pH 8.5, 10% (w/v) PEG 8000 from hanging drop optimization screens. Harvested crystals were cryo-preserved in precipitant supplemented with 25% (v/v) glycerol. Diffraction data were collected at the Stanford Synchrotron Radiation Lightsource (SSRL) beamline 12-1 at 0.97737 Å beam wavelength at a temperature of 100 K. Data were processed to a 2.57 Å resolution cutoff using XDS78. Phases and an initial structure solution were obtained by molecular replacement using the previously solved X-ray crystal structure of apo-muLRRC15, and a model of Fab-E1 generated using MOE Biologics Software, with the constant and variable regions of Fab-E1 split and searched for separately. ML-YSD-07 was built de novo in COOT into clear difference density in the phased electron density map. The overall model improved with iterative rounds of building and refinement in COOT48 and PHENIX80 respectively to produce the final model with an R-work/R-free of 0.1868/0.2422 (Supplementary Table S6). Of modeled residues, 95.07% exhibited Ramachandran favored angles, 4.82% allowed and 0.11% outliers. Synthesis of RESCA-functionalization DCPs The DCP ML-YSD-07 was dissolved to 3 mM in DMF. N,N-Diisopropylethylamine (DIPEA) (10 molar eq.) was added before adding (+)-H3-RESCA-TFP (3 molar eq.) (CheMatech). Reactions were stirred for 1 h at room temperature. Reaction progress was monitored by LC-MS (Agilent 6545XT AdvianceBio LC Q/TOF) on a PLRP-S column (1000 Å, 2.1 ×50 mm, 8 µm) using a linear gradient of 15–85% (v/v) acetonitrile containing 0.05% (v/v) formic acid over 4 min. The LC-MS data was analyzed using Agilent MassHunter Qualitative Analysis v.B.07.00. The RESCA-conjugated ML-YSD-07 was purified by RP-HPLC (5 µm C18 100 Å, 21.2 ×100 mm, Phenomenex) using a linear gradient of 20–70% (v/v) acetonitrile in 0.05% (v/v) TFA over 20 min. Radiolabeling of DCP for PET Imaging [18F]fluoride (29 ± 3 GBq) was inversely eluted from a QMA light cartridge (Waters, part no. WAT023525) with 0.2 mL of 0.9% (w/v) sterile NaCl into a vial containing 10 μL of 0.5 M sodium acetate pH 4.8 and 0.7 eq. (relative to RESCA-ML-YSD-07) of 2 mM AlCl3 (Sigma-Aldrich, part no. 563919). After a 3–5 min incubation, a mixture of RESCA-ML-YSD-07 (2 μL of 10 mM in DMSO), 10 μL of 0.3 M sodium ascorbate pH 4.8, and 30 μL of 75 mM histidine, 240 mM sucrose pH 5.0 was added to the aluminium(III) [18F]fluoride ([18F]AlF2+) solution and incubated at room temperature for 15 min. The reaction mixture was diluted into 3 mL of sterile water, loaded onto a Strata-X SPE cartridge (30–60 mg, Phenomenex), washed with 3 mL of sterile water, eluted with 1 mL of ethanol, and formulated with 10 mM histidine, 10 mM methionine, 240 mM sucrose, pH 7.0 such that the ethanol concentration in the final product vial was less than 5%. The final product was analyzed by RP-HPLC using a Waters BEH C18 (100 Å, 50 ×2.1 mm) using a linear gradient of 1–100% (v/v) acetonitrile containing 0.1% (v/v) FA over 4 min. In vivo PET imaging The animal studies were conducted in accordance with the guidelines of Genentech’s AAALAC-accredited Institutional Animal Care and Use Committee. Female C57Bl-6 mice weighing 23–26 g were purchased from Charles River Laboratories. Mice were acclimated in Genentech’s animal housing facilities for at least 1 week, and were 3 months old at the study start date. Each mouse was inoculated subcutaneously with 100,000 KPR3070 cells in the right flank. Once tumors reached ~400 mm3 in size, the animals were grouped out for subsequent PET imaging sessions using the β-Cube PET scanner (Molecubes) followed by a CT scan on an X-Cube CT scanner (Molecubes). Mice (n = 5 per group) were lightly anesthetized for restraint with 3.5% sevoflurane and transferred to the scanner bed. Mice were catheterized at the lateral tail vein and ~2.4–2.9 MBq (65–79 µCi, 16–19 pmol) of [18F]AlF-RESCA-ML-YSD-07 was administered intravenously as a 1 min short infusion via the lateral tail vein catheter. A 60 min dynamic PET scan was acquired to capture the injection and subsequent biodistribution of the tracer. The respiration rate was monitored, and the body temperature was monitored and maintained under feedback control with an electric heating pad. Following the PET scan, a low-dose CT scan was acquired for attenuation correction and anatomical reference. PET images were reconstructed from the list-mode data by using Molecubes-provided histogram and reconstruction software. Regions of interest (ROIs) were drawn on multiple axial slices of the images by using VivoQuant software (Invicro). Radioactivity concentrations in the ROIs were measured as a percentage of the injected dose per gram (% ID/g), assuming 1 cm3 equivalency in 1 g of soft tissue and decay-corrected to the starting time of the scan. The maximum intensity PET projections shown in Fig. 5b, and alongside coronal, sagittal, and transverse slices through the tumor for better localization in Supplementary Fig. S17, are from the final 15 min frame (45–60 min post tracer administration) of the dynamic scans, uniformly scaled from 0 to 20% ID/g. Parametric images (Supplementary Fig. S19) displaying the tracer influx rate constant (Ki) were generated by voxelwise Patlak graphical analysis using mediastinal blood pool signal as an uncorrected image-derived input function81,82. Mouse tissue digestion Tumor and lung tissue samples of the mice following PET/CT imaging (n = five mice) were prepared similarly as previously described20. Tissues were collected, weighed and minced into small pieces. All tissues were subsequently enzymatically digested using a cocktail of dispase (Life Technologies), collagenase P and DNaseI (Roche) for 45 min at 37 °C to obtain a single-cell suspension. Red blood cells were removed from the lung samples using ammonium chloride-based red blood cell lysis buffer. Cells were counted using a Vi-CELL XR (Beckman Coulter). Cells were blocked with Fc block (2.4G2; 1:200, 553142, BD Bioscience) for 10 min at 4 °C. Cells were labeled with the following antibodies for 20–30 min at 4 °C: PDPN-ApcCy7 (Biolegend #127418; 1:200), EPCAM-BV650 (BD #740559; 1:100), CD31-PECy7 (BD #612802; 1:200), LRRC15 (M25, in-house; 1:1000) with secondary PE IgG2a (Biolegend #407108; 1:500), CD45-BUV395 (BD #564279; 1:200), CD24-BUV737 (E-Bioscience #48-0242-82; 1:100), 7AAD (Biolegend #420403; 1:100) and Calcein Violet (Thermo #C34858; 1:1000). Data were acquired using a BD FACSymphony flow cytometer and analyzed using FlowJo (v.10.10.0). The gating strategy used to determine the percentage of LRRC15+ fibroblasts is shown in Supplementary Fig. S24. Data were processed using GraphPad Prism (v.10.0.2). Statistical analyses were carried out by a Welch’s t test. IGFALS and LGR4 binding experiments Human leucine-rich-repeat-containing G-protein coupled receptor 4 (huLGR4; UniProt ID Q9BXB1; residues A25–G527) and mouse LGR4 (muLGR4; UniProt ID A2ARI4; residues A25–G527) with C-terminal AviTag and His6 tag were expressed in Trichoplusia ni (T. ni) cells and purified by nickel affinity chromatography and SEC as described above for LRRC15. Proteins were biotinylated using the BirA500 biotin-protein ligase kit (Avidity LLC) as described above. Human insulin-like growth factor-binding protein complex acid labile subunit (huIGFALS) was purchased from R&D Systems (9917-IA-050) and was biotinylated using EZ-LinkTM NHS-PEG4-Biotin (Thermo) using standard protocol. SPR binding experiments were carried out by immobilizing biotinylated huLGR4, muLGR4, or huIGFALS to a capture level of 300–400 RU using a streptavidin-coated sensor chip. All binding experiments were performed at 25 °C using HBS-EP running buffer. ML-YSD-07 and ML-PD-03 were tested at 10 µM. Recombinant mouse R-spondin 3 (R&D systems; 4120-RS-025/CF) was tested at 10 nM and used as a positive control for huLGR4 and muLGR4. Recombinant human IGF-1 (R&D Systems; 291-G1) and recombinant human IGFBP-3 (R&D Systems; 675-B3) were mixed at a 1:1 ratio and the mixture was tested at 1 nM and used as a positive control for huIGFALS. The results shown are representative of three independent experiments. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file Source data Source Data Source Data