Key regulatory elements of the TGFβ–LRRC15 axis predict disease progression and immunotherapy resistance across cancer types

Significance

Leucine-rich repeat-containing protein 15 (LRRC15) is a clinically tractable readout of an immunotherapy-resistance-defining transforming growth factor-beta (TGFβ) subprogram in mesenchymal-derived cancer cells and cancer-associated fibroblasts, yet the regulatory architecture by which TGFβ exposure is converted into LRRC15 expression has remained unresolved, constraining selective interventions distinct from the dose-limited pan-TGFβ inhibitors that have failed clinically. Integrating functional compound screening, scRNA-seq, and gene regulatory network inference, we identify four upstream regulators (MMP2, SPARC, WNT5B, and TGFβR2), each individually necessary for TGFβ-induced LRRC15 expression in three biologically distinct cancer cell models. The resulting five-gene signature outperforms LRRC15 alone in predicting progression and anti-PD-1 resistance across three independent immune-excluded tumor cohorts, nominating a precision target distinct from global TGFβ inhibition.

Abstract

Transforming growth factor-beta (TGFβ) has dual roles in cancer, initially suppressing tumors but later promoting metastasis and immune evasion. Efforts to inhibit TGFβ have been largely unsuccessful due to significant toxicity and indiscriminate immunosuppression. Leucine-rich repeat-containing protein 15 (LRRC15) is a TGFβ-regulated antigen expressed by cancer cells of mesenchymal origin and cancer-associated fibroblasts (CAFs). In preclinical studies, ablation of TGFβ-driven LRRC15+ CAFs enhances effector functions of CD8+ T cells. However, the pathobiological mechanisms associated with TGFβ’s upregulation of LRRC15 expression in cancer cells remain unclear. Using an integrated approach combining functional compound screening with scRNA-seq, we reveal key genomic features regulating TGFβ’s ability to increase LRRC15 expression on cancer cells. Construction of gene regulatory networks converged our analyses on four key genes (MMP2, SPARC, TGFβR2, and WNT5B) central to TGFβ-induced LRRC15 pathobiology in cancer cells. Validation of these genes in cell models and their use in predicting immunotherapy responses highlight their potential in refining immunotherapy strategies and personalizing cotreatment options.

TGFβ signaling pathway, recognized for its dual role in cancer biology, acts as a tumor suppressor in early-stage malignancies but promotes progression, metastasis, and immune evasion in aggressive cancers and advanced stages of disease (1–3). This paradoxical behavior complicates targeting of TGFβ in oncology, as its functions vary widely from tumor suppression to the facilitation of epithelial–mesenchymal transition (EMT) and therapy resistance (4–6). TGFβ also plays a role in both primary and acquired resistance to treatment; signaling initially contributes to an immunosuppressive microenvironment that shields emerging tumors from immune surveillance. During acquired resistance, TGFβ signaling intensifies, enabling cancer cells to evade therapies by promoting invasiveness, metastasis, and maintaining an immunosuppressive microenvironment (7–9).

LRRC15 has been recognized for its role as surrogate marker for TGFβ signaling and its association with immunosuppression and therapy resistance in aggressive TGFβ-mediated malignancies, notably in osteosarcoma (OS) (10). Its predominant association with CAFs in the tumor microenvironment (TME) of immune-suppressed, metastatic, and aggressive primary tumors further underscores its therapeutic relevance (11). This study primarily focuses on the regulation of LRRC15 in cancer cells of mesenchymal origin to uncover cancer cell-intrinsic mechanisms underlying TGFβ-driven pathobiology and immunotherapy resistance. Investigations into the TGFβ-driven TME have identified LRRC15 as a part of an 11-gene signature that correlates with immune checkpoint therapy (ICT) resistance in patients (12–14). Furthermore, recent data from engineered mouse models demonstrated that targeted depletion of LRRC15+ CAFs markedly diminishes tumor fibroblast content and augments CD8+ T cell efficacy (13). While LRRC15-positive CAFs are predominantly associated with the immunosuppressive functions and contribute to T cell exhaustion in solid tumors such as Pancreatic ductal adenocarcinoma (PDAC) (13), this study focuses on its regulation in mesenchymal-derived cancer cells to uncover intrinsic mechanisms that may drive LRRC15 biology, which may also parallel LRRC15’s role in fibroblasts (15–17). This approach complements CAF-centric studies and reveals regulators extendable to stromal contexts (14). Our previous independent verification using a novel radiotheranostic antibody-based approach confirmed that targeting of LRRC15+ cells significantly decreased tumor burden and disease progression, along with suppression of genes linked to TGFβ-driven immunotherapy resistance (18).

In this study, we evaluated the complex interplay of TGFβ signaling and LRRC15 expression within MSC-derived tumors. We show a bifurcated cellular response to TGFβ activity; certain cancer cells undergo a rapid induction of LRRC15, while expression is unaltered in others. To further explore the transcriptional dynamics detected in TGFβ-responsive LRRC15+ cells, we comprehensively investigated how this observation is indicative of TGFβ-mediated tumor pathobiology. Integration of high-throughput screening (HTS) to identify small molecules that modulate TGFβ-mediated LRRC15 induction, along with scRNAseq to identify the gene networks and key regulators of LRRC15 inducibility, identified four genes (TGFβR2, SPARC, MMP2, and WNT5B) that activate TGFβ-mediated LRRC15 expression. Further evaluation in patient tumor cohorts revealed that these activating genes, alongside LRRC15, are a prognostic determinator for patient response to immunotherapy and the progression of aggressive malignancies.

Results

TGFβ1 Induces LRRC15 Expression in Cancer Cells.

Previous studies have shown that TGFβ1 induced LRRC15 expression in MSCs under supraphysiological conditions over extended periods of time (19). In this study, we investigated the response of LRRC15 expression in cancer cells to physiologically relevant TGFβ1 concentrations, simulating conditions within the TME (20). Nine LRRC15+ cancer cell lines (CALU1, KASUMI2, SAOS2, U2OS, HUO9, NCI-H196, U118, U87, RPMI7951) (18) underwent 24-h incubation in reduced-serum media prior to reintroduction of TGFβ1 at concentrations from 0 to 10 ng/mL (Fig. 1A). Subsequently, plasma membrane-associated LRRC15 was detected via flow cytometry using the AlexaFluor-647 labeled anti-LRRC15 IgG1 antibody (DUNP19). The specificity of DUNP19 has been validated in engineered LRRC15+/− cell lines, other orthogonal assays, and is further confirmed in this study via siRNA-mediated LRRC15 knockdown (Fig. 5A).

Fig. 1.

Dissection and modulation of the TGFβ–LRRC15 axis in select cancer cell lines. (A) Flow cytometry (n = 3 to 4 per line) reveals two response groups: TGFβ noninducible (red) and TGFβ inducible (blue, P < 0.0001). LRRC15 staining shown as mean ± SD. (B) Confocal microscopy of TGFβ-inducible RPMI7951 cells ± TGFβ (nuclei = cyan, LRRC15 = magenta). (C) TGFβR1 inhibitor galunisertib dose-dependently suppressed LRRC15 in RPMI7951 cells (P < 0.0001), confirming pathway specificity. LRRC15 data are an average of the staining intensity. (D) HTS workflow: (i) LRRC15+ cells were pretreated with library compounds, followed by TGFβ1 induction and LRRC15 staining; (ii) compounds were classified by mechanism of action; (iii) of 1,280 compounds, 282 were nontoxic (Z-score > −3), and 72 hits (LRRC15 > 3 SD). (E) Heatmap of LRRC15 Z-scores (red = increase, blue = decrease) by compound class (Left) and hit compounds (Right).

Fig. 5.

High expression of LRRC15 activating genes predicts immunotherapy resistance and tumor progression. (A) LRRC15-targeted siRNA reduced LRRC15 in TGFβ-responsive cell lines, overcoming TGFβ-induced upregulation (P < 0.0005). (B) siRNA k/d of target genes identified by in silico and compound screening (n = 3 per condition) produced inconsistent responses across cell lines, underscoring the influence of genetic background on the TGFβ–LRRC15 pathway. (C) Heatmap of LRRC15 staining intensity after siRNA k/d across increasing TGFβ concentrations. K/d of TGFβR2, SPARC, WNT5B, and MMP2 reduced LRRC15 in OS (HUO9), whereas effects in melanoma (RPMI7951) and glioblastoma (U118) were less consistent. (D) In TCGA LUSC tumors, LRRC15 correlated with MMP2 (r = 0.77) and SPARC (r = 0.79) but weakly with WNT5B and TGFβR2; the four-gene signature correlated with LRRC15 (r = 0.81). (E–G) Kaplan–Meier curves showing the predictive power of the five-gene signature for (E) survival in LUSC patients on anti-PD1 (P = 0.044), (F) 2-y survival in metastatic clear cell renal cell carcinoma patients on anti-PD1 (P = 0.035), and (G) prognosis in mUC patients on atezolizumab. Patients were stratified by high (red) versus low (blue) expression.

These assessments revealed variable responses to TGFβ, with minimal changes in LRRC15 levels in CALU1 (percent change in LRRC15 staining = −21.94%, P = 0.2805), KASUMI2 (−8.16%, P < 0.0005), SAOS2 (+4.89%, P = 0.0517), and U2OS (+18.98%, P = 0.0008), while pronounced increases were noted in HUO9 (+210.26%, P = 0.0220), NCI-H196 (+70.82%, P < 0.0005), U118 (+173.06%, P < 0.0005), U87 (+53.65%, P < 0.0005), and RPMI7951 (+76.45%, P = 0.0038) (Fig. 1A). As our study focuses on TGFβ’s induction of LRRC15 as opposed to its other downstream effectors, the latter cell lines were described as “TGFβ-responsive” in all subsequent analyses for their ability to upregulate LRRC15 after treatment with exogenous TGFβ1. Among the TGFβ-responsive cell lines, we focused on three models representing aggressive malignant tissues from different anatomical origins: HUO9 (osteosarcoma), RPMI7951 (malignant melanoma), and U118 (glioblastoma multiforme). Using a confocal microscopy-based approach with an AlexaFluor-647-DUNP19, we confirmed a significant increase in LRRC15 expression in cells treated with TGFβ compared to controls as shown for the RPMI7951 cell line (Fig. 1B). Furthermore, pretreatment with the TGFβ receptor inhibitor galunisertib before addition of TGFβ1 effectively suppressed LRRC15 induction in HUO9, RPMI7951, and U118 cells, demonstrating the role of TGFβ in modulating expression (Fig. 1C).

Next, we sought to annotate the molecular mechanisms impacting TGFβ’s regulation of LRRC15 (Fig. 1 D, i–iii). We conducted HTS using the LOPAC1280 library, which comprises 1,280 well-annotated small molecules spanning various mechanisms of action and drug classes (Fig. 1 D, ii and iii). LRRC15+ cells were incubated in reduced-serum media to deplete endogenous cytokines before treatment with the compound library and TGFβ (Fig. 1 D, i). Following compound incubation, cells were treated with TGFβ and assessed for LRRC15 expression via confocal microscopy (Fig. 1 D, i). All screened compounds were categorized by class and mechanism, and the normalized results were presented as their average effect on LRRC15 expression (Fig. 1 E, i). Compounds that significantly inhibited TGFβ-driven LRRC15 expression, defined as those reducing the integrated intensity of LRRC15 by more than three SD from the plate mean, were identified as hits (Fig. 1 E, ii). In addition, compounds were assessed for toxicity; those reducing cell viability >3 SD from the mean were excluded from subsequent analyses. Ultimately, we identified 26, 22, and 24 hits in the HUO9, RPMI7951, and U118 cell lines, respectively (SI Appendix, Table S1).

Established TGFβR2 antagonists such as SB-525334 and RepSox further validated the specificity of the assay and the mechanism by which LRRC15 is upregulated (SI Appendix, Table S1). Notably, various histaminergics, antivirals, and hormone modulators, for example chloroquine, Tenidap, and hydrocortisone, also reduced LRRC15, underscoring the diverse roles of TGFβ in inflammation and immune modulation via diverse biological pathways (Fig. 1E). Other compound classes, including potassium channel inhibitors and angiotensin inhibitors, increased LRRC15 (Fig. 1E). These compound hits were particularly intriguing given LRRC15’s previously reported role in SARS-CoV-2 infection and structural similarity to the angiotensin-converting enzyme ACE2 (21), suggesting some utility in preventing viral infection. All gene targets from the effective compounds within our screen were compiled for further analysis.

Single-Cell RNA Sequencing of TGFβ-Responsive Cell Lines Reveals Distinct LRRC15-Related Signatures.

After identifying compounds that could disrupt the TGFβ–LRRC15 pathway, we performed scRNAseq on the cell lines utilized in the compound screen to analyze molecular signaling pathway differences that contributed to TGFβ responsiveness. In addition to the screened cell lines, three TGFβ nonresponsive cell lines (CALU1, KASUMI2, SAOS2) were also sequenced and used in subsequent differential gene and pathway analyses (Fig. 2A and SI Appendix, Fig. S3). Agreeing with the microscopy data in Fig. 1C, LRRC15 was not expressed in all cells in our scRNAseq data and showed intercellular variability within each cell line (SI Appendix, Fig. S3). To reduce gene expression sparsity, we used a high-dimensional weighted gene coexpression network analysis (hdWGCNA) metacell aggregation approach (22), which aggregates neighboring cells based on gene expression similarity and averages their expression to create metacells that are robust to scRNAseq dropouts (Fig. 2A). Using the metacell approach for DEG analysis has been shown to mitigate sparsity and technical variability in scRNAseq data while still capturing the biological heterogeneity within broader cell types (22, 23). This enabled us to identify DEGs and transcriptomic pathways that were exclusive to TGFβ-responsive cells. While the lower dimension UMAP representation of the gene expression data did not separate the six cell lines by TGFβ inducibility (Fig. 2B), we identified sets of DEGs with consistent effects in TGFβ-responsive cells (Fig. 2C). Key cell proliferation and noncanonical TGFβ signaling genes, including TGFβ activator LTBP1 (latent-transforming growth factor beta-binding protein 1) and Wnt ligand WNT5B, were upregulated in TGFβ-responsive cells. Interestingly, T cell modulators and inflammatory driver genes such as IL-7R (interleukin-7 receptor) (24) were downregulated in TGFβ-inducible cell lines, providing insights into the potential immunoregulatory mechanisms within the TGFβ–LRRC15 pathway (Fig. 2C). Pathway enrichment analysis of upregulated genes in the TGFβ-responsive cells showed enrichment of apoptotic processes and cytokine signaling within immune pathways, while downregulated DEGs were associated with translation and ribosomal function (Fig. 2D). Furthermore, the DEG effect, calculated using metacell expression, followed a similar pattern in single cell expression UMAP plots, indicating the robustness of the gene expression effects detected. Several genes such as SPARC, WNT5B, MMP2, and EID1, showed a stark upregulation of expression within TGFβ-responsive cells (Fig. 2E). We also observed distinct subclusters in some cell lines in the UMAP plots. Since >90% TGFβ-inducible DEGs were consistently expressed between the subclusters across cell lines, the subclustering does not confound the transcriptional differences observed between TGFβ-inducible and noninducible cell lines.

Fig. 2.

ScRNAseq reveals distinct transcriptional signatures in TGFβ-inducible cancer cell lines. (A) Schematic of the scRNAseq workflow. (B) UMAP of TGFβ-inducible (blue) and noninducible (red) cell lines; cells did not cluster by inducibility. (C) Metacell-level heatmap of large-effect DEGs (|log2FC| > 0.5) across both groups, shown as average log2FC (red = high, blue = low). Asterisks denote FDR (*<0.05, **<0.01, ***<0.001). (D) Pathway enrichment of DEGs between groups: upregulated genes were enriched in apoptotic processes and cytokine signaling, downregulated genes in translation and ribosomal function. (E) UMAPs from (B), split by inducibility, showing single-cell expression of key genes upregulated in TGFβ-inducible cells.

Network Analysis of TGFβ-Responsive Transcriptional Signatures.

Beyond the individual associations of genes to TGFβ-inducibility of LRRC15, understanding gene network activity can highlight the molecular programs and key regulators driving the induction. Gene networks can further elucidate the molecular mechanisms involved in the TGFβ–LRRC15 axis. Here, we employed two complementary scRNAseq-based gene network modeling approaches to capture both GRNs and gene coexpression signatures that distinguished the TGFβ-responsive cell lines (Fig. 3A). GRNs consist of nodes and edges representing genes and regulatory interactions, respectively. To construct a GRN for the TGFβ–LRRC15 axis, we used SCING (25), a bagging gradient-boosting machine learning approach to predict regulatory relationships between genes. This method was previously shown to effectively predict gene perturbation effects and uncover key driver genes for disease (25). We incorporated all samples to build an aggregate GRN and identified 41 highly connected SCING GRN modules through Leiden clustering (26). Gene coexpression networks, on the other hand, identify coregulated gene modules based on correlations. We constructed a gene coexpression network across samples using hdWCGNA based on gene correlation (22), identifying 25 coexpression modules. The expression levels of individual gene regulatory or coexpression modules from both SCING and hdWCGNA networks were computed for each cell and metacell using hdWGCNA’s module eigengene calculation (Fig. 3A) (22). Comparison between the networks from SCING and hdWGCNA showed high agreement among the network modules (SI Appendix, Fig. S2C).

Fig. 3.

Network analysis of TGFβ-responsive transcriptional signatures. (A) SCING and hdWGCNA were used to model gene regulatory networks and coexpression networks, respectively, in TGFβ-inducible and noninducible cell lines. Module expression was compared at the metacell level between groups to identify differential modules. (B) Heatmap of SCING and hdWGCNA module expression across cell lines. Modules upregulated in association with TGFβ responsiveness (red) included cell cycle, cytokine signaling, immune response, and apoptosis. (C) UMAPs showing single-cell expression of key module-associated biological processes in TGFβ-inducible (Top) and noninducible (Bottom) cell lines, confirming the patterns in (B). UMAP as in Fig. 2 B and E.

We identified differential module expression in SCING modules S1, S3, and S4, and hdWGCNA modules H1 and H3 (Fig. 3B) that consistently differentiated TGFβ-inducible from noninducible cell lines, and 11 other modules with consensus across five cell lines (SI Appendix, Fig. S2A). Gene Ontology and Reactome pathway analysis revealed upregulation of biological pathways related to apoptotic processes and cytokine and immune signaling, with a downregulation of translation (consistent with DEG analysis in Fig. 2D), cell motility regulation, and DNA damage response in TGFβ-inducible cell lines. The module expression patterns of individual cells in TGFβ-inducible and noninducible cell lines recapitulate the differential expression at the meta-cell level (Fig. 3C). Surprisingly, cell motility pathways were also downregulated in TGFβ-inducible cell lines (Fig. 3C), which contrasted LRRC15’s proposed role as a driver of invasiveness and cell migration in tumor cells.

Integrated High-Throughput Drug Screening and scRNAseq Data Reveal Converging Molecular Pathways Involved in LRRC15’s Regulation.

Next, we combined these results to carry out an in silico drug screen to identify the top candidate drugs and genes within the TGFβ–LRRC15 axis (Fig. 4A) and further integrated the results with those from the in vitro drug screen (Fig. 1). The L1000 database is a curated library of thousands of compounds and their corresponding affected genes, based on in vitro screening assays across different organisms, tissues, drug dosages, and timepoints (27). After filtering the database for all human drug signatures, we performed a Fisher’s exact test against hdWGCNA- and SCING-derived network modules, all significant (adjusted P-value < 0.05) DEGs, and all large-effect DEGs (|log2FC| > 0.5) to identify L1000 compounds that, upon treatment, displayed gene signatures that overlapped with our scRNAseq findings (Fig. 4A). This in silico drug screening identified 26 compounds whose gene signatures matched our scRNAseq results, and that were also present in our in vitro drug screen (Fig. 1D). We also identified 14 gene targets shared across all analyses (Fig. 4B and SI Appendix, Table S2). Visualizing these gene targets around TGFβ1 and LRRC15 in the SCING GRN revealed interconnected subnetworks with potential regulators of the TGFβ–LRRC15 axis (Fig. 4C).

Fig. 4.

Integration of high-throughput drug screening and scRNAseq reveals converging molecular pathways regulating LRRC15 in TGFβ-inducible cells. (A) In silico and in vitro drug screening approaches were combined to identify candidate compounds targeting the TGFβ–LRRC15 axis. The in silico screen utilized the L1000 drug database to map compounds against DEGs and differential hdWGCNA and SCING modules, identifying drug and corresponding genes signatures consistent across analyses. (B) Illustration of the overlap between hdWGCNA, SCING, and DEG datasets, highlighting 26 compounds hits from screening (Fig. 1) and 14 gene targets shared across analyses. (C) Visualization of gene targets in a gene regulatory network. Key driver analysis revealed potential regulators (larger nodes) of the TGFβ–LRRC15 pathway. Circle fill indicates the total number of compounds targeting each gene, and the number of analyses identifying the gene as significant is denoted by outline color intensity.

TGFβ-Induced LRRC15 Expression can be Modulated by siRNA Knockdown of Candidate Genes.

To test the functional relevance of genes identified in overlapping analyses (Fig. 4) within the TGFβ–LRRC15 pathway, candidate genes were knocked down (k/d) in TGFβ-responsive cell lines via small interfering RNA (siRNA). Following a similar protocol to our compound screening approach (Fig. 1), cells were incubated in low-serum media before transfection with the siRNA construct. After transfection, cells were treated with TGFβ and assessed for LRRC15 via confocal microscopy. As a positive control, we validated siRNA transfection and k/d using three LRRC15-targeting siRNAs (Fig. 5A). Transfection with siLRRC15 reduced LRRC15 protein levels by an average of 98.61% ± 0.47% (P < 0.0005) in untreated cells, and 94.21% ± 2.02% (P < 0.0005) in TGFβ-treated cells. As an additional control in subsequent siRNA experiments, we also utilized siRNAs targeting TGFβR2, given the receptor’s established role in LRRC15 CAFs (13). We then tested siRNAs targeting various genes identified as potential nodes within the TGFβ–LRRC15 axis, including EID1, H2AFV, IQGAP1, LGALS1, MMP2, SPARC, TGFβR2, TPM4, and WNT5B (Fig. 5B). Among these, EID1 k/d significantly induced LRRC15 (P = 0.0059). As a known inhibitor of p300-mediated transcription and differentiation in fibroblast cell populations (28), k/d of EID1 may lead to upregulated p300 activity, possibly explaining the increased LRRC15. Although an intriguing avenue for further exploration, we remained focused on genes whose k/d led to reduced LRRC15 expression. In particular, we focused solely on genes coexpressed with LRRC15 to improve the feasibility of targeting the TGFβ–LRRC15 pharmacologically. Of the genes tested, MMP2, SPARC, and WNT5B, as well as TGFβR2, significantly impacted TGFβ-induced LRRC15 in two or more cell lines, with TGFβR2 and SPARC k/d significantly reducing expression across multiple concentrations of TGFβ treatment (Fig. 5 B and C). These results support the regulatory role of 4 out of the 9 tested candidate genes in activating the TGFβ–LRRC15 axis and promoting LRRC15 expression. While the three cancer cell lines tested represent distinct tumor types, each with their own transcriptional and signaling backgrounds, k/d of each of the genes disrupted TGFβ-mediated LRRC15 induction in at least one model. This pattern supports the interpretation that these genes represent bona fide modulators of the pathway, even if their functional importance is context-dependent. Notably, the genes that were experimentally validated tended to be the ones with outgoing edges to other genes in the network (Fig. 5C); outgoing network edges indicate that the gene likely functions as an upstream regulator.

While MMP2, SPARC, WNT5B, and TGFβR2 are each independently TGFβ-responsive genes in multiple biological systems, TGFβ-responsiveness alone would predict only parallel coinduction with LRRC15, not mechanistic mediation of LRRC15 induction. The siRNA loss-of-function experiments shown in Fig. 5 B and C directly distinguish these alternatives: Depletion of any of the four regulators significantly attenuated TGFβ-induced LRRC15 expression in three biologically distinct mesenchymal-derived cancer cell lines. Such necessity tests are the established standard for converting transcriptional coregulation into mechanistic dependency. Published mechanistic precedent supports each regulator independently: MMP2 proteolytically activates latent extracellular-matrix-bound TGFβ (29); SPARC modulates TGFβ bioavailability and amplifies SMAD2 signaling (30–33); and noncanonical WNT5 family ligands activate latent TGFβ through ROCK-dependent mechanisms (34). The convergence of three orthogonal discovery approaches (in vitro compound screening, scRNA-seq differential gene expression, and SCING/hdWGCNA network inference) onto the same four-node architecture, each node functionally validated by independent siRNA perturbation, establishes these regulators as bona fide mediators rather than passive cotargets of the TGFβ–LRRC15 axis. The four-node architecture is not identified in the foundational TGFβ–LRRC15 literature (13, 14, 19), because those studies employed multiday in vitro stimulation at supraphysiological TGFβ concentrations or multiweek in vivo endpoints, conditions under which a rate-limiting upstream regulator architecture would not be detectable by single-gene perturbation. The shorter timescales and physiologically relevant TGFβ concentrations of the present manuscript reveal the rate-limiting regulatory architecture that prior assay designs were not configured to identify.

Activators of LRRC15 Expression Correlate to Immunotherapy Response and Tumor Progression.

The four genes whose k/d inhibited LRRC15 expression became the basis for further exploration into clinical tumor databases. We first correlated expression of this five-genes with overall survival in four mesenchymal-derived cancers from the Cancer Genome Atlas (TCGA): mesothelioma (n = 82), gastrointestinal stromal tumor (n = 384), sarcoma (n = 262), and uveal melanoma (n = 78). A significantly lower overall survival in patients with high expression of this gene signature corresponded with all four tumor types (SI Appendix, Fig. S5). In cohorts of patients from different tumor entities including breast cancer (29), glioblastoma (30), and skin cutaneous melanoma (31), our signature acted as a prognostic marker only in breast cancer patients (HR: 2.50, 95% CI 1.66 to 3.77, P = <0.001) (SI Appendix, Fig. S6). We then further explored the TCGA dataset to assess each gene’s correlation with LRRC15 across TCGA tumor types to better understand whether LRRC15 and the potential activators coexpressed in patient tumors (32).

In the TCGA datasets, LRRC15 was highly correlated with MMP2 (r = 0.77, P < 0.0005) and SPARC (r = 0.79, P < 0.0005), and modestly correlated with TGFβR2 (r = 0.35, P < 0.0005) in lung squamous cell carcinoma (LUSC) (Fig. 5D). Given the strong correlation of 3 out of 4 genes within LUSC tumors, along with LRRC15’s known role in anti-PD1 immunotherapy resistance, we applied our signature to a small anti-PD1 immunotherapy cohort involving LUSC patients to evaluate whether it could predict immunotherapy resistance (34). Stratifying patients by median expression of the five-gene set showed that LRRC15 together with its four regulators predicted worse survival in LUSC (P = 0.044, n = 13; Fig. 5E). The signature validated in an anti-PDI-treated ccRCC cohort (NCT01358721, NCT01354431, NCT01668784) (35), where high expression of MMP2, SPARC, TGFβR2, WNT5B, and LRRC15 predicted worse 2-y survival (P = 0.035, n = 24; Fig. 5F). Removing LRRC15 abolished the prediction, and LRRC15 alone was likewise nonpredictive (SI Appendix, Fig. S8A). Prognostic accuracy therefore depends on coexpression of all five genes rather than on any single component.

We evaluated a third cohort from metastatic urothelial cancer (mUC) patients treated with atezolizumab profiled for CD8+ T cell distribution (NCT02951767, NCT02108652) (36, 37). Pretreatment tumors were classified as “infiltrated” (CD8+ T cells within the tumor), “excluded” (CD8+ T cells confined to stroma), or “desert” (CD8+ T cells absent). Because LRRC15+ CAFs have been implicated in T cell exclusion and exhaustion, we tested the five gene sets in immune-excluded tumors. High expression predicted unfavorable prognosis in this subgroup (P = 0.012, 113 patients; Fig. 5G) but not in infiltrated or desert tumors, and LRRC15 alone was again nonpredictive (SI Appendix, Figs. S7B and S8C). The signature therefore tracks response to immunotherapy specifically in the immune-excluded phenotype, consistent with Dominguez et al. in the IMVIGOR210 trial (14), and reflects the pathobiological output of TGFβ–LRRC15 axis signaling. Immune-excluded tumors plausibly harbor expanded LRRC15+ CAFs and tumor cell populations that both form a physical barrier and impose immunomodulatory effects, positioning this gene set as a candidate resistance biomarker therapeutic entry point.

Discussion

Deciphering the intricate molecular pathways, genomic determinants, and TME factors that regulate biomarker dynamics is pivotal for refining targeted diagnostic and therapeutic approaches in oncology. Detailed knowledge of these factors enables the stratification of patient populations most likely to respond to specific treatments, thereby enhancing clinical efficacy (38, 39). The absence of pathobiological understanding is a major contributor to the substantial attrition rates in clinical oncology trials, where nearly half of failures are attributed to lack of established relevance of the target to the disease phenotype (38). With the surge in oncology target and biomarker discovery, the translation of these findings into clinically beneficial entities remains elusive, highlighting the critical need for comprehensive molecular profiling to fully realize the promise of personalized medicine in cancer care.

The initial identification of LRRC15’s regulation by TGFβ in MSCs under prolonged supraphysiological cytokine levels (19) prompted a reevaluation of its translational relevance. In addition, LRRC15 has been associated with increased chemotherapy resistance, metastasis, and lower overall survival in immunohistochemical analysis of LRRC15 in OS patient samples (10). To address this, our study aimed to investigate LRRC15 regulation under shorter durations and physiologically relevant concentrations of TGFβ. As tumors progress, CAFs and cancer cells increasingly exhibit behaviors and interactions that enhance tumor complexity and resilience. Both cell types are responsive to extracellular molecules such as growth factors and cytokines, with cancer cells commonly inducing expression changes in CAFs via paracrine signaling, which can lead to selective CAF expansion (15, 40). Furthermore, CAFs maintain tumor-promoting properties independent of direct interactions with cancer cells (16). This convergence of characteristics guided our focus to LRRC15+ MSC-derived cancer cells for deeper insights.

Our investigation revealed two distinct cellular responses to TGFβ following cytokine withdrawal: One subset of cell lines demonstrated LRRC15 upregulation, suggesting an adaptive mechanism, while another remained unaffected. Integrative analysis using scRNAseq, chemical compound screening, and siRNA k/d pinpointed TGFβR2, WNT5B, SPARC, and MMP2 as key mediators of TGFβ–LRRC15 activity. Although TGFBR2’s regulation of LRRC15 is expected because of its involvement in the canonical TGFβ signaling pathway, its consistent emergence as a top hit across our scRNA-seq, LOPAC screening, and siRNA validation serves as a robust internal positive control that bolsters confidence in our findings with MMP2, SPARC, and WNT5B, noncanonical regulators absent from prior LRRC15-associated signatures (14) which collectively predict immunotherapy resistance and disease progression in diverse patient cohorts investigated in the present study.

In patient tissue databases, the impact of these genes on TGFβ–LRRC15 activation was observed across a wide range of solid tumors (SI Appendix, Fig. S5), including those demonstrating expression exclusively in CAFs but not in cancer cells. Moreover, a significant correlation emerged with a previously identified LRRC15-containing 11-gene panel (SI Appendix, Fig. S9), which was obtained by screening of CAFs in TGFβ-driven TME in PDAC patients (14). Our signature correlated with CAF transcriptomic profiles and their role in driving CAF phenotypes (TGFBR2) (13) and TGFβ signaling in fibroblasts (SPARC, MMP2) (41, 42). This suggests parallel functional roles for LRRC15 in cancer cells and CAFs, warranting future studies to extend our findings to LRRC15+ CAF. LRRC15, along with the four genes regulating its expression, differentiated survival rates and disease progression were characterized by the exclusion of CD8+ T cells and resistance to ICTs. However, we did not find significant prognostic value in tumors categorized as immune-infiltrated or immune-desert. These variations may be attributed to immunological mechanisms driven by non-LRRC15 related mechanisms. Among the four genes within the TGFβ–LRRC15 axis, TGFβR2 has previously been established as a regulator of LRRC15+ CAF formation (13). Tumor fibrosis is a common consequence of aberrant TGFβ signaling, contributing to immune cell exclusion in the stroma-rich ECM and poor drug delivery (17). Several of these ECM components were significantly coexpressed with LRRC15, such as matrix metalloproteinases (MMP2, MMP14) and collagen genes (COL1A1, COL1A2). Further, MMP2 plays a critical role in cancer cells by promoting angiogenesis and cell growth, and by facilitating collagen degradation in the TME, thereby aiding tumor invasion. This highlights its dual function in promoting cancer progression and supporting CAF populations. In parallel, SPARC, a secreted extracellular matrix protein, has been shown to boost collagen production and induce canonical TGFβ signaling pathways (42). While further research is necessary to delineate its function in tumor tissues, studies indicate that k/d in pterygium fibroblasts diminishes TGFβ signaling and MMP2 expression (41), potentially leading to a similar effect on LRRC15+ CAF and cancer cell interactions. Future studies extending these findings to noncancer stromal cells could illuminate how these genes modulate LRRC15 in the broader tumor ecosystem, potentially revealing therapeutic applications.

In the current study, we identified several candidate regulatory genes, siRNA k/d of some, such as H2AFV, did not significantly affect TGFβ-induced LRRC15 expression in our cancer cell models. This may reflect epigenetic mechanisms and DNA methylation reprogramming (43) that complicate gene regulation (44), or residual protein that sustain H2AFV’s role within the TGFβ–LRRC15 axis despite siRNA transfection (45). More complete CRISPR/Cas9 gene editing and multi-omic assessment in additional cell lines are likely necessary to determine their involvement in TGFβ–LRRC15 activity (46). Although cancer cell models exhibited inducible LRRC15 expression, each originated from a distinct tissue type, with inherent differences in basal signaling and gene expression, which influence baseline omics profiles and the magnitude of TGFβ responsiveness. Despite the underlying differences, k/d of the four candidate genes attenuated TGFβ-induced LRRC15 across multiple models, supporting their role as pathway regulators. The persistence of pathway disruption across heterogeneous tissue backgrounds underscores the robustness of these regulatory nodes, while their relative contribution varies as a function of cellular context.

Our pioneering small molecule screens identified numerous compounds capable of disrupting the TGFβ–LRRC15 pathway, potentially providing alternative strategies to counter protumorigenic effects of TGFβ without directly inhibiting the pathway itself, a common dose-limiting factor that induce adverse off-target effects. For example, bosutinib, a Bcr-Abl tyrosine-kinase inhibitor under study in chronic myeloid leukemia as a combination therapy with anti-PD1 immunotherapy (33), Given the off-target effects associated with TGFβ inhibitors, complementary inhibition of immunosuppressive pathways and DNA damage response processes related to TGFβ are an unmet clinical need. In addition, the significant downregulation of DNA damage response processes in TGFβ-inducible cell lines may expose a therapeutic vulnerability. The positive hits identified from our drug screening, particularly those supported by both in vitro and in silico screens, could be further explored for such applications in the future.

Two well-characterized ALK5/TGFβR1 inhibitors (SB-525334 and RepSox) emerged independently from the unbiased LOPAC1280 screen as TGFβ–LRRC15 axis antagonists, functioning as internal positive controls that confirm on-target screen performance and pathway-specific signal detection (Fig. 1E and SI Appendix, Table S1). Orthogonal pharmacological validation by galunisertib (a TGFβR1 inhibitor) in three responsive cell lines (Fig. 1C) further confirms pathway specificity. The compound screen functioned not as a standalone biological claim but as one of three orthogonal hit-discovery approaches; only candidate genes that converged across the compound screen, the scRNA-seq differential expression analysis, and the SCING/hdWGCNA network inference were prioritized for siRNA functional validation. Downstream pharmacological development of individual screen hits, including dose–response characterization and target-engagement profiling, is beyond the scope of the present manuscript and is the subject of separate ongoing investigations.

Upon exploring LRRC15 as a biomarker for therapeutic resistance, our analysis of TCGA sarcoma datasets suggested that LRRC15 alone is insufficient as a prognostic biomarker for aggressive disease and overall survival and is not predictive of therapeutic outcomes in immunotherapy trials (SI Appendix, Fig. S7C). These observations contrast with prior clinical findings in an osteosarcoma cohort where LRRC15 was identified as a prognostic marker for therapy resistance and reduced survival (10). This discrepancy highlights that LRRC15 expression alone may not solely drive tumor aggressiveness. Integrating the results of this study, we hypothesize that LRRC15 functions primarily as a bystander protein, indicative of an extensive network of TGFβ-driven pathobiological processes prevalent in the microenvironment of aggressive, immunoresistant tumors.

We have previously demonstrated that LRRC15 can effectively be applied for radiotheranostics (18). This strategy facilitates the selective delivery of diagnostic and cytotoxic radionuclides to specifically image and ablate both cancer cells and CAFs. In addition to the ablation of LRRC15+ cell populations, tumors treated with 177Lu-LRRC15-RIT exhibited downregulation of SPARC, MMP2, TGFβR2, and WNT5B (SI Appendix, Fig. S4). These findings partially confirm that LRRC15 is coexpressed with other crucial factors driving tumor progression and resistance, potentially providing comprehensive genomic profiling to guide clinical decision-making and patient stratification, similar to the precision medicine advancements of the NCI-MATCH and ComboMATCH trials (47, 48).

The transcriptomic downregulation of TGFβ–LRRC15 axis components (SPARC, MMP2, TGFβR2, WNT5B) observed following 177Lu-LRRC15-RIT (SI Appendix, Fig. S4; ref. 18) could in principle reflect either target-engagement-driven biological reprogramming or nonspecific cytotoxic effects of the β-emitter. Two independent lines of evidence support the target-mediated interpretation. First, external beam radiation is well established to upregulate, not downregulate, the TGFβ axis, with TGFβ mRNA and protein induction preceding radiation-induced fibrosis and reactive-oxygen-species-mediated cleavage of the latent TGFβ complex as the proposed mechanism (49–52). The directionality of the post-RIT signature reported here is therefore the opposite of what generic radiation cytotoxicity would predict, rendering the alternative hypothesis inconsistent with the established radiobiological literature. Second, recently published radioligand transcriptomic profiling has demonstrated that 177Lu-PSMA-I&T treatment produces target-pathway-dominated rather than emission-class-dominated transcriptomic responses in patient-derived prostate cancer xenografts, with response-correlated transcriptomic changes reflecting on-target pathway biology (53). This pattern is consistent with earlier 131I versus 211At thyroid radionuclide profiling that demonstrated absorbed-dose-dominated rather than emission-class-dominated clustering of differential gene expression (54). The convergence of i) directional opposition with the canonical EBRT TGFβ response and ii) the established principle of target-pathway dominance in receptor-targeted radionuclide therapy supports interpretation of the observed signature as a target-engagement-driven biological response, in which selective depletion of LRRC15-positive cells eliminates the cellular source of the LRRC15-axis transcriptional output.

In summary, by computationally integrating high-throughput small molecule screening data with scRNA sequencing, we identified TGFβR2, MMP2, SPARC, and WNT5B as key mediators driving LRRC15 upregulation in response to TGFβ signaling. These molecules, secreted by both cancer cells and CAFs, are critical in promoting EMT, during which LRRC15 expression is significantly elevated. Their coordinated activity shapes a fibrotic and immunosuppressive TME, predicting resistance to immune checkpoint inhibitors and correlating with aggressive cancer phenotypes. These findings elucidate the molecular architecture of the TGFβ–LRRC15 axis, revealing a cascade of effector molecules that present potential targets for the development of inhibitors and modulators to disrupt these pathways. This improved understanding may also provide insight into the mechanisms underlying immunoresistance, offering a basis for exploring diagnostic and therapeutic strategies to enhance the efficacy of immunotherapy.

Materials and Methods

A detailed description of materials, methods, and equipment from this study is provided in SI Appendix.

In Vitro Evaluation and Quantification of LRRC15.

Detailed descriptions of cell lines used, as well as LRRC15-based flow cytometry, confocal microscopy, and confocal microscopy-based screening methods are available in SI Appendix.

Silencer RNA.

Cells were transfected with gene-targeting silencer siRNAs (ThermoFisher, #4392420), treated 72 h later with 0 to 8 ng/mL TGFβ1 for 24 h, then fixed and stained for LRRC15 as described. Transfection efficiency was confirmed by loss of LRRC15 in all three siLRRC15 conditions. Plates were imaged on an ImageXpress Micro Confocal high-content microscope and analyzed in MetaXpress. LRRC15 expression was quantified as average integrated staining intensity per well, normalized to nontransfected TGFβ1-treated and untreated controls.

scRNAseq Analysis.

Single-cell quality control and preprocessing were performed in Seurat (parameters in SI Appendix). Cells with similar expression profiles were aggregated within each cell line using the hdWGCNA metacell approach (22) to reduce sparsity for differential expression and network modeling. Given the limited sample size (2 replicates per cell line), DEGs associated with TGFβ inducibility were identified by six one-vs.-opposite-group Wilcoxon rank sum tests, each comparing a single cell line to all lines in the opposing group; only DEGs consistent across all comparisons were retained, minimizing false positives driven by cell line-specific variation. Two DEG sets were carried forward: all genes passing FDR <5%, and large-effect genes with |log2FC| ≥ 0.5. Network and in silico drug repositioning analyses are detailed in SI Appendix.

Immunotherapy Trial Data Analysis.

Gene expression data for the atezolizumab mUC trials (NCT02951767, NCT02108652) (36, 37) were obtained from the IMvigor210CoreBiologies package (EGAD00001003977); LUSC (34) and RCC anti-PD1 (35) data were retrieved from GEO (GSE93157) and as a normalized expression matrix, respectively. Tumors were filtered to the top 75% by LRRC15 expression, and a five-gene score (LRRC15 plus four LRRC15-related genes) was computed with eigenWeightedMean (MultiGSEA). Kaplan–Meier curves were generated with survminer in R 4.4.1.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Raw and processed scRNAseq data have been deposited in the Gene Expression Omnibus (GSE290255) (55). Analysis code is available at https://github.com/XiaYangLabOrg/TGFB-LRRC15 (56).