ELMO2 is a therapeutic vulnerability in mesenchymal-like and drug-resistant non-small cell lung cancer

Abstract

Epithelial-mesenchymal transition drives tumor metastasis and therapeutic resistance, yet few treatments have been developed that target this process. Here, we show that ELMO2 represents a specific vulnerability in mesenchymal-like cells. ELMO2 suppression induces excessive autophagy and cell death via FAK activity inhibition. We identify ELMO3 as a functional paralog that compensates for ELMO2 loss, establishing a synthetic lethal interaction. The epithelial-mesenchymal transition core regulator ZEB1 represses ELMO3 transcription in mesenchymal-like cells, rendering them sensitive to ELMO2 blockade. ELMO3 is significantly downregulated in epithelial-mesenchymal transition-associated EGFR inhibitor-resistant cells. Furthermore, the survival of these resistant, mesenchymal-like cells depends on ELMO2/FAK signaling. Through structure-based screening, we identify C52, a small-molecule ELMO2 inhibitor that effectively kills ELMO3-low lung cancer cells and EGFR inhibitor-resistant cells. Our study uncovers an ELMO2-ELMO3 synthetic lethal interaction and establishes ELMO2 as a potential therapeutic target for mesenchymal-like cancer and drug-resistant non-small cell lung cancer.

Epithelial-mesenchymal transition (EMT) is a common mechanism for acquired resistance to EGFR tyrosine kinase inhibitors (TKIs). Here, the authors discover that ZEB1-mediated repression of ELMO3 creates a synthetic-lethal dependence on ELMO2, and that inhibiting ELMO2 selectively kills mesenchymal-like cells and EGFR TKIs–resistant cancer cells.

Introduction

Cancer is a leading cause of death globally, and resistance to treatment is a key factor in the failure of therapy for most patients. Epithelial-Mesenchymal Transition (EMT) is a biological process in which epithelial cells lose their polarity and tight intercellular connections, transforming into mesenchymal cells with enhanced migratory ability, invasiveness, and resistance to apoptosis^1–5. Cancer cells acquire invasiveness and metastatic potential through EMT, enabling them to break through the basement membrane and spread to other tissues and organs. EMT is crucial in understanding the mechanisms of cancer metastasis^6–9. In addition, EMT plays a pivotal role in mediating resistance to chemotherapy and targeted therapies^10–14. Given the important role that EMT plays in tumors, developing therapeutic strategies targeting resistant cells with an EMT phenotype is clinically significant¹⁵. The main approaches to target the EMT program for therapy include preventing EMT induction, reversing the process, and selectively targeting mesenchymal-like cells^16,17. Further understanding of the mechanisms linking EMT to resistance may provide insights and strategies to overcome treatment resistance in cancer.

Epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors (TKIs) are effective therapy against advanced non-small cell lung cancer (NSCLC) with activating EGFR mutations. Compared to those receiving standard chemotherapy, patients treated with EGFR TKIs exhibit prolonged progression-free survival and improved quality of life. The development of targeted therapies for EGFR has contributed to the recent decreases in lung cancer mortality^18–20. However, resistance to EGFR TKIs inevitably arises²¹. Key mechanisms of resistance can generally be categorized into EGFR-dependent mechanisms, such as the acquisition of secondary EGFR mutations like T790M, and EGFR-independent mechanisms including the activation of alternative receptor tyrosine kinases like MET, epithelial-mesenchymal transition (EMT), and small cell transformation²². Among patients treated with first- and second- generation EGFR inhibitors such as erlotinib or gefitinib, approximately 50 % develop therapeutic resistance due to the acquisition of secondary EGFR T790M mutations. These patients can benefit from the treatment with the third-generation EGFR inhibitor osimertinib^23–25. However, in the absence of secondary EGFR resistance mutations, such as T790M, or after use of drugs effective against EGFR T790M mutation, EGFR-TKI resistance occurs primarily through EGFR-independent mechanisms^21,26–29. Therefore, there is an urgent unmet clinical need for effective treatments to overcome acquired resistance to EGFR-TKIs. Restoration of an epithelial phenotype in NSCLC cell lines restored sensitivity to the EGFR inhibitor gefitinib³⁰. Developing effective treatments to overcome EMT-related resistance to EGFR-TKIs is critical for lung cancer patient survival.

Here, we investigated the differential vulnerability between mesenchymal-like cells and epithelial-like cells and identified ELMO2 as a specific target in ELMO3 low mesenchymal-like cells. ZEB1, a core EMT transcription factor, confers the dependency on ELMO2 in mesenchymal-like cancer cells by repressing the transcription of ELMO3 without affecting ELMO2 expression. Suppression of ELMO2 induced autophagy-dependent cell death through the regulation of FAK phosphorylation in ELMO3 low mesenchymal-like cells. ELMO2 disruption is a therapeutic vulnerability in EGFR-TKIs resistant cells as these cells underwent EMT and significantly down-regulated ELMO3 expression. We identified an ELMO2 small molecule inhibitor C52, from a structure-based screen. C52 significantly decreased the viability of EGFR-TKIs resistant cells, similar to ELMO2 knockdown and served as a potential therapeutic agent in NSCLC.

Results

ELMO2 is essential for cell survival in mesenchymal-like lung cancer cells

To identify genes that are selectively essential in mesenchymal-like cells, we used the DepMap database³¹ to select candidates that are required for the survival of mesenchymal-like cells, but not epithelial-like cells. We compared the gene effects on cell survival between mesenchymal-like cells and epithelial-like cells and found that ELMO2 was one of the top differential dependency gene in mesenchymal-like cells (Fig. 1A–C). Top 5 differential candidates also include GPX4 and FGFR1, which have been reported to be important for the survival of mesenchymal-like cells, indicating the reliability of the analysis^32–34. To determine whether ELMO2 is required for cell survival in mesenchymal-like cells, we knocked down ELMO2 expression in both mesenchymal-like cells and epithelial-like cells. The expression of ELMO2 following ELMO2 knockdown was determined by immunoblotting (Supplementary Fig. S1A). Mesenchymal-like cell lines NCI-H1299, A549, NCI-H1703, 95D and NCI-H1792 expressed mesenchymal marker vimentin, whereas epithelial-like cell lines NCI-H1975, PC-9, Calu-3, HCC827 and NCI-H358 cell lines expressed epithelial marker E-cadherin with little or no vimentin expression (Fig. 1D). ELMO2 expression showed no significant difference between mesenchymal-like cells and epithelial-like cells (Fig. 1D). Knockdown of ELMO2 in mesenchymal-like cells induced cell death (Fig. 1E, G, Supplementary Fig. S1B), but had no effect on epithelial-like cells (Fig. 1F, G, Supplementary Fig. S1C). Thus, ELMO2 knockdown selectively kills lung cancer cells with a mesenchymal phenotype.

Fig. 1. ELMO2 is essential for mesenchymal-like cell survival.

A Scatter plot showing the mean gene effects in epithelial-like cancer cells (n = 189) and mesenchymal-like cancer cells (n = 142). Each point represents mean gene effect for a gene in epithelial-like cells (y-axis) and mesenchymal-like cells (x-axis). B Plot showing the difference in mean gene effects in mesenchymal-like cancer cells (n = 142) and epithelial-like cancer cells (n = 189) (value in mesenchymal-like cells minus value in epithelial-like cells). The top five genes with the largest differences are indicated. C The mean gene effect score of ELMO2 is shown for mesenchymal-like cancer cells (n = 140) and epithelial-like cancer cells (n = 187). The P value is included in Source Data file. D Immunoblotting of ELMO2, epithelial marker E-cadherin, and mesenchymal marker vimentin in human NSCLC cell lines. The samples derive from the same experiment but different gels for E-cadherin, vimentin and another for ELMO2 and β-actin were processed in parallel. Data are representative of two independent experiments. Knockdown of ELMO2 causes cell death in mesenchymal-like cells (E), but not in epithelial-like cells (F) using clonogenic assay. Data in (E, F) are representative of three independent experiments. G Cell growth detected by CCK-8 was suppressed by ELMO2 knockdown in mesenchymal-like cells but not in epithelial-like cells. Data are mean ± SEM, two-way ANOVA with Šidák’s multiple comparisons test, n  =  3 independent experiments. Source data are provided as a Source Data file.

ELMO2 knockdown causes autophagy-dependent cell death in mesenchymal-like cells

ELMO2 knockdown caused dramatic morphological changes in mesenchymal-like cells (Fig. 2A, B). Cells were highly vacuolated following ELMO2 knockdown, a hallmark of autophagy (Fig. 2A, B). We treated ELMO2 knockdown cells with autophagy inhibitor Bafilomycin A1 or chloroquine, apoptosis inhibitor Z-VAD, necrosis inhibitor Necrostatin-1. The autophagy inhibitors rescued the cell death phenotype induced by ELMO2 knockdown, whereas the apoptosis and necrosis inhibitors failed to rescue (Fig. 2C, D, Supplementary Fig. S2A, B). Because of the potential off-target effects of pharmacological inhibitors, we employed genetic perturbation by knocking down key autophagy genes ULK1 and ATG5 to validate the autophagy phenotype (Supplementary Fig. S2C). Knockdown of ULK1 or ATG5 significantly rescued cell death, confirming that ELMO2 knockdown in mesenchymal-like lung cancer cells induced autophagy-dependent cell death (ADCD) (Fig. 2E, Supplementary Fig. S2D). MDC is a fluorescent dye that specifically labels autophagic vacuoles by ion trapping and binding to membrane lipids, thus is commonly used for the detection of autophagy. The MDC assay showed that ELMO2 knockdown caused autophagy in human lung cancer cells NCI-H1299 (Supplementary Fig. S2E). The autophagy phenotype was further confirmed using GFP-LC3 fluorescence and immunoblotting analysis of autophagy-related proteins (Fig. 2F–H). ELMO2 knockdown in NCI-H1299 and A549 cells increased the number of LC3 puncta (Fig. 2F, G), decreased p62 protein expression, and increased LC3-II/LC3-I ratio (Fig. 2H). Autophagic flux was measured using the tandem fluorescent mRFP-GFP-LC3 system. Due to the low pH inside lysosomes, GFP fluorescence is quenched while mRFP fluorescence remains stable, resulting in autophagosomes being labeled in yellow and autolysosomes in red. In ELMO2 knockdown cells, more red-only puncta were observed, indicating ELMO2 knockdown significantly increased the autophagy flux (Fig. 2I–K, Supplementary Fig. S2F–H). These results demonstrate that ELMO2 knockdown in mesenchymal-like lung cancer cells induces ADCD.

Fig. 2. ELMO2 knockdown causes autophagy-dependent cell death in mesenchymal-like cells.

A Cells with ELMO2 knockdown displayed extensive cytoplasmic vacuolation. Scale bar, 100 μm. B Visualization of cytoplasmic vacuoles in NCI-H1299 cells stably expressing GFP and an ELMO2 shRNA or a control shRNA. Scale bar, 50 μm. C, D Relative cell viability of ELMO2 knockdown cells treated with autophagy inhibitor Bafilomycin A1 or chloroquine, or pan-caspase inhibitor Z-VAD, or necrosis inhibitor Necrostatin-1. Autophagy inhibitor Bafilomycin A1 or chloroquine rescued cell death in cells expressing an ELMO2 shRNA. Data are mean ± SEM, one-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. E Knockdown of ATG5 or ULK1 in NCI-H1299 and A549 recused the cell death induced by ELMO2 knockdown in a clonogenic assay. F ELMO2 knockdown caused the increase of punctate GFP-LC3 distribution in human lung cancer cells NCI-H1299 and A549 stably expressing GFP-tagged LC3 and an ELMO2 shRNA or a control shRNA. Scale bar, 50 μm. G The number of punctate GFP-LC3 is quantified. Data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. H Immunoblotting of p62 and LC3-II/LC3-I ratio in NCI-H1299 and A549 cells expressing an ELMO2 shRNA or a control shRNA. The samples derive from the same experiment, but different gels for ELMO2, p62, β-actin and another for LC3 were processed in parallel. Representative images of autophagy flux (I) and quantification of the number of LC3 puncta/cell (J) and the ratio of mRFP+GFP-/mRFP+GFP+ (K) in NCI-H1299 cells stably expressing mRFP-GFP-LC3 and an ELMO2 shRNA or a control shRNA. Scale bar, 20 μm. Data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. Data in (A, B, H) are representative of two independent experiments, and data in (E, F, I) are representative of three independent experiments. Source data are provided as a Source Data file.

FAK mediates the function of ELMO2 in autophagy-dependent cell death in mesenchymal-like cells

In order to identify molecules that mediate the function of ELMO2 in mesenchymal-like cells, we applied co-dependency analysis using DepMap database to identify the effectors correlated with ELMO2 dependency. Dependency of PTK2 (FAK-coding protein) and ELMO2 is highly correlated (Fig. 3A–C). ELMO2 knockdown decreased FAK phosphorylation at Y397 site and reduced FAK localization in cell focal adhesion complex in NCI-H1299 and A549 cells, but had no effect on cell focal adhesions (Fig. 3D, E, Supplementary Fig. S3A, B). The expression of FAK and FAK phosphorylation at site Y397 showed no significant differences between mesenchymal-like cells and epithelial-like cells (Supplementary Fig. S3C). FAK wildtype (FAK-WT), FAK-Y397F, and FAK-K454R (kinase-dead mutation) overexpression in NCI-H1299 and A549 cells did not rescue the cell death phenotype induced by ELMO2 knockdown (Fig. 3F, Supplementary Fig. S3D, E). However, the expression of the charge-inducing FAK-Y397E mutant, which mimics the electrostatic effect of phosphorylation, rescued the cell death phenotype (Fig. 3F, Supplementary Fig. S3D, E). Overexpression of FAK-WT, FAK-Y397E, FAK-Y397F or FAK-K454R had no effect on cell focal adhesions (Supplementary Fig. S3F). mRFP-GFP-LC3 assay also showed that FAK-Y397E expression rescued the autophagy flux (Fig. 3G). Expression of FAK-Y397E reversed the decrease of p62 and increase of LC3-II/LC3-I induced by ELMO2 knockdown, demonstrating the charged residue at FAK-397 mediated the function of ELMO2 in autophagy (Fig. 3H). Furthermore, endogenous reciprocal co-immunoprecipitation assays confirmed that ELMO2 and FAK physically interacted (Fig. 3I). ELMO2 overexpression increased FAK phosphorylation at the Y397 site (Fig. 3J). These results suggest that ELMO2 is sufficient and necessary for FAK phosphorylation and FAK-Y397 mediates the autophagy phenotype induced by ELMO2 knockdown in mesenchymal-like cells.

Fig. 3. FAK mediated the function of ELMO2 in autophagy-dependent cell death in mesenchymal-like cells.

A CRISPR co-dependency analysis of ELMO2 in DepMap. The volcano plot highlights the top ten genes with the strongest absolute correlations to ELMO2 CRISPR gene effect values. Two-tailed Pearson correlation test and corrected for multiple comparisons using the Benjamini-Hochberg method. B The scatter plot of the correlation coefficient between the gene effect values of ELMO2 and PTK2 in cancer cells. Each data point represents an individual cell line (n = 1150). The Spearman correlation coefficient (R) and the two-tailed P value are displayed on the graph. The solid line represents the linear regression fit, and the shaded area indicates the 95 % confidence interval. C The scatter plot of the correlation coefficient between the gene effect values of ELMO2 and PTK2 in non-small cell lung cancer cells. Each data point represents an individual cell line (n = 125). The Spearman correlation coefficient (R) and the two-tailed P value are displayed on the graph. The solid line represents the linear regression fit, and the shaded area indicates the 95 % confidence interval. D Immunoblotting of FAK phosphorylation in mesenchymal-like cancer cells following ELMO2 knockdown. ELMO2 knockdown decreased the phosphorylation of FAK-Y397. The samples derive from the same experiment but different gels for ELMO2, FAK, another for FAK-pY397 and another for β-actin were processed in parallel. E Immunofluorescence staining showed the subcellular localization of FAK. F Clonogenic assay of NCI-H1299 and A549 cells expressing an ELMO2 shRNA and a FAK-WT or FAK-Y397E, FAK-Y397F, FAK-K454R. FAK-Y397E but not FAK-WT, FAK-Y397F, FAK-K454R rescued the cell death phenotype caused by ELMO2 knockdown. G Puncta were measured by fluorescence in human lung cancer cell NCI-H1299 stably expressing mRFP-GFP-LC3. FAK-Y397E but not FAK-WT expression in NCI-H1299 cells rescued autophagy flux caused by ELMO2 knockdown. H Immunoblotting of ELMO2, FAK, phosphorylation of FAK Y397, p62 and LC3 in NCI-H1299 and A549 expressing an ELMO2 shRNA with or without FAK-Y397E expression. The samples derive from the same experiment, but different gels for ELMO2, p62, another for FAK, β-actin, another for FAK-pY397 and another for LC3 were processed in parallel. I Co-immunoprecipitation of ELMO2 and FAK. ELMO2 interacted with FAK. The samples derive from the same experiment but different gels for ELMO2 and another for FAK were processed in parallel. J Immunoblotting of FAK and phosphorylation of FAK Y397 in NCI-H1299 and A549 cells overexpressing ELMO2. ELMO2 expression increased the phosphorylation of FAK Y397. The samples derive from the same experiment but different gels for ELMO2, FAK-pY397, another for FAK and β-actin were processed in parallel. Data in (D, E, G–J) are representative of two independent experiments, and data in (F) are representative of three independent experiments. IP immunoprecipitation, IB Immunoblotting, WT wild-type, OE overexpression. Source data are provided as a Source Data file.

ELMO2 and ELMO3 have synthetic lethal interaction

In order to identify candidates involved in differential ELMO2 dependency in mesenchymal-like cells, we performed a pan-cancer analysis of ELMO2 dependency and gene expression (Fig. 4A). We found that ELMO3 expression is inversely correlated with ELMO2 dependency (Fig. 4A). ELMO3 expression was highly correlated with epithelial markers ESRP1 and CDH1 (E-cadherin) (Fig. 4B, C), which raises the possibility that ELMO3 expression is responsible for the differential dependency of ELMO2 in mesenchymal-like cells. ELMO3, but not ELMO2, was inversely correlated with the expression of mesenchymal biomarker vimentin in clinical lung cancer tissues (Fig. 4D, E). ELMO3 expression was determined in mesenchymal-like and epithelial-like lung cancer cells (Fig. 4F, G). ELMO3 expression is significantly lower in mesenchymal-like cells than in epithelial-like cells (Fig. 4F, G). Knockdown of ELMO2 significantly reduced cell viability in ELMO3-low mesenchymal-like cells compared to ELMO3-high epithelial-like cells (Fig. 4H). Analysis of DepMap data suggested that the ELMO2 dependency was inversely correlated with ELMO3 expression in NSCLC (Fig. 4I). To determine whether ELMO3 compensated the function of ELMO2 on cell viability, we overexpressed ELMO3 in ELMO3-low mesenchymal-like cell lines NCI-H1299 and A549 (Supplementary Fig. S4A). ELMO3 expression rescued the cell death phenotype induced by ELMO2 knockdown (Fig. 4J, Supplementary Fig. S4B, C). Knocking down both ELMO3 and ELMO2 in epithelial-like cell line PC-9 and NCI-H1975 significantly reduced cell viability, whereas knocking down ELMO2 or ELMO3 alone did not affect cell viability (Fig. 4K, Supplementary Fig. S4D). Furthermore, ELMO2 and ELMO3 double knockdown in PC-9 decreased the phosphorylation of FAK-Y397 and increased autophagy (Fig. 4L). In contrast, knockdown ELMO2 or ELMO3 alone had no effect on FAK phosphorylation and autophagy (Fig. 4L). To further investigate the molecular mechanism underlying this functional compensation, we determined the interaction of ELMO2/ELMO3 with FAK (Fig. 4M). Immunoprecipitation of endogenous FAK resulted in the co-precipitation of ELMO2 or ELMO3, suggesting that ELMO3 can functionally compensate for ELMO2 by engaging similar downstream signaling pathways through direct interaction with FAK in epithelial-like cellular contexts (Fig. 4M). These results indicate a synthetic lethal interaction between ELMO2 and ELMO3, identifying ELMO2 as a genetic vulnerability in ELMO3-low cancer cells.

Fig. 4. ELMO2 and ELMO3 have synthetic lethal interaction.

(A) Correlation between mRNA expression and ELMO2 gene effect score in mesenchymal-like cells and epithelial-like cells. Two-tailed Pearson correlation test and corrected for multiple comparisons using the Benjamini-Hochberg method. ELMO3 expression is highly correlated with epithelial markers ESRP1 (B) and CDH1 (E-cadherin) (C). Each data point represents an individual cell line (n = 1517). The Spearman correlation coefficient (R) and exact two-tailed P value are displayed on the graph. D Immunoblotting of ELMO3, ELMO2, vimentin, and E-cadherin in human lung cancer tissues. The samples derive from the same experiment but different gels for ELMO2, another for vimentin, E-cadherin, β-actin and another for ELMO3 were processed in parallel. E Representative immunohistochemistry staining of E-cadherin, vimentin, ELMO3, and ELMO2 in human lung cancer tissues. Scale bar, 50 μm. Consistent results were observed across three biological replicates. F ELMO3 expression was determined by qRT-PCR. Data are presented as mean ± SEM from three independent experiments for each individual cell line. G ELMO3 expression was determined by immunoblotting in mesenchymal-like cells and epithelial-like cells of NSCLC. The samples derive from the same experiment but different gels for ELMO3 and another for β-actin were processed in parallel. H Relative viability of ELMO3-high and ELMO3-low cell lines expressing an ELMO2 shRNA. Data are mean ± SEM, comparison between ELMO3-low and ELMO3-high groups was performed using a two-tailed Mann-Whitney U test, n  =  3 independent experiments. The P value is included in Source Data file. I Correlation between ELMO2 gene effect and ELMO3 expression in NSCLC cell lines. Each data point represents an individual cell line (n = 95). The Spearman correlation coefficient (R) and the two-tailed P value are displayed on the graph. The solid line represents the linear regression fit, and the shaded area indicates the 95 % confidence interval. J Clonogenic assay of NCI-H1299 and A549 cells expressing an ELMO2 shRNA and an ELMO3 cDNA or a control vector. ELMO3 expression rescued cell death phenotype caused by ELMO2 knockdown. K Clonogenic assay of epithelial-like cell lines PC-9 and NCI-H1975 expressing an ELMO2 shRNA, or an ELMO3 shRNA or both ELMO2 and ELMO3 shRNAs. L Immunoblotting of p62, LC3, FAK, and phosphorylation of FAK Y397 in PC-9 cells expressing an ELMO2 shRNA, or an ELMO3 shRNA or both ELMO2 and ELMO3 shRNAs. Double knockdown of ELMO2 and ELMO3 increased autophagy and decreased the phosphorylation of FAK Y397. The samples derive from the same experiment but different gels for ELMO2, another for ELMO3, another for FAK, another for p62, FAK-pY397, β-actin and another for LC3 were processed in parallel. M Co-immunoprecipitation of FAK and ELMO2 or ELMO3. FAK interacted with ELMO2 or ELMO3. The samples derive from the same experiment but different gels for ELMO2, another for ELMO3 and another for FAK were processed in parallel. N Tumor growth of A549 cells transplanted in mice. Each data point represents mean tumor volume ± SD, n = 5 biological replicates. Two-way ANOVA with Šidák’s multiple comparisons test. Data in (D, G, L, M) are representative of two independent experiments, and data in (J, K) are representative of three independent experiments. OE overexpression, IP immunoprecipitation, IB Immunoblotting, iKD inducible knockdown. Source data are provided as a Source Data file.

To determine the effects of ELMO2 knockdown on tumor growth in vivo, we constructed doxycycline-inducible ELMO2 knockdown cancer cells, and transplanted them into mice (Supplementary Fig. S4E). ELMO2 knockdown induced by doxycycline suppressed tumor growth in vivo (Fig. 4N). ULK1 knockdown or ELMO3 overexpression, or FAK-Y397E overexpression rescued the suppressive effect of ELMO2 knockdown in mice (Fig. 4N). These results suggest that ELMO3 compensates the function of ELMO2, FAK phosphorylation mediates the function of ELMO2 in autophagy, and autophagy is responsible for cell death phenotype induced by ELMO2 knockdown.

ZEB1 suppresses ELMO3 expression to render mesenchymal-like cancer cell dependency on ELMO2

To investigate the mechanism of differential ELMO3 expression in mesenchymal-like cells and epithelial-like cells, we performed co-expression analysis utilizing the DepMap database to identify potential effectors associated with ELMO3 expression. The expression of ZEB1, a key transcription factor in the EMT process, is significantly reversely correlated with ELMO3 expression (Fig. 5A), consistent with previous reports that ELMO3 is a ZEB1-responsive gene in NSCLC cell lines³⁵. Knockdown of ZEB1 significantly increased ELMO3 but not ELMO2 expression in mesenchymal-like NCI-H1299 and A549 cells (Fig. 5B, C). It has been reported that ZEB1 overexpression induced a mesenchymal phenotype and EGFR inhibitor resistance³⁶. We established an epithelial-like HCC827 cell line expressing ZEB1 cDNA and analyzed the expression of ELMO2 and ELMO3 (Fig. 5D, E). ZEB1 overexpression significantly down-regulated ELMO3 expression in HCC827 cells (Fig. 5D, E). In contrast, ELMO2 expression remained unchanged upon ZEB1 overexpression in HCC827 cells (Fig. 5D, E). To determine whether ZEB1 regulates ELMO3 transcription, we cloned the promoter region of ELMO3 upstream of a luciferase gene in a reporter plasmid and co-transfected the ZEB1 cDNA construct and the ELMO3 reporter construct into HCC827 cells. ZEB1 expression suppressed the luciferase signal of the reporter plasmid (Fig. 5F). To further investigate whether ZEB1 directly binds to ELMO3 promoter, we performed chromatin immunoprecipitation (ChIP) analysis using a ZEB1 antibody. The ELMO3 promoter region was significantly enriched compared to a control IgG ChIP as assessed by quantitative PCR, further demonstrating that ELMO3 is a direct target of ZEB1 (Fig. 5G). To determine whether the induction of EMT is sufficient to alter ELMO3 expression, we induced EMT using the transforming growth factor Beta (TGF-β), which has been shown as an EMT inducer³⁷. TGF-β treatment increased ZEB1 expression, and decreased ELMO3 but not ELMO2 expression in epithelial-like cell line NCI-H1975 and HCC827 (Fig. 5H, I), suggesting that EMT induction is sufficient to alter ZEB1 and ELMO3 expression. To investigate whether ZEB1 is required in cell death induced by ELMO2 knockdown, we performed rescue experiments using ZEB1 knockdown in mesenchymal-like cells. ZEB1 knockdown rescued the cell death phenotype induced by ELMO2 knockdown in NCI-H1299 and A549 (Fig. 5J, Supplementary Fig. S5A). Furthermore, ZEB1 knockdown in NCI-H1299 and A549 reversed the decrease of FAK phosphorylation at Y397 site and p62, as well as reversed the increase of LC3-II/LC3-I induced by ELMO2 knockdown (Fig. 5K). Conversely, ELMO2 knockdown in epithelial-like HCC827 cells expressing ZEB1 cDNA induced a cell death phenotype (Fig. 5L , Supplementary Fig. S5B), a decrease of FAK phosphorylation at Y397 site and p62, as well as an increase of LC3-II/LC3-I ratio (Fig. 5M). These results suggest that the suppression of ELMO3 expression in mesenchymal-like cancer by ZEB1 renders cells sensitive to ELMO2 blockade. To further investigate the role of ELMO2 and ELMO3 in EMT, we examined the expression of epithelial marker E-cadherin and mesenchymal marker vimentin following ELMO2 knockdown in mesenchymal-like NCI-H1299 cells and ELMO3 knockdown in epithelial-like NCI-H1975 cells. Neither knockdown of ELMO2 in NCI-H1299 cells nor knockdown of ELMO3 in NCI-H1975 cells led to significant changes in E-cadherin or vimentin expression (Supplementary Fig. S5C, D), indicating that ELMO2 and ELMO3 do not directly regulate EMT. Overexpression of ELMO2 showed no significant promotion of cell migration or invasion in transwell assays (Supplementary Fig. S5E, F). Injection of tumor cells overexpressing ELMO2 or a control vector in the mouse model revealed no significant difference of metastatic nodules in the lungs (Supplementary Fig. S5G). Together, these data support the conclusion that EMT status regulates ELMO3 expression.

Fig. 5. ZEB1 suppresses ELMO3 transcription to confer mesenchymal-like cancer cell dependency on ELMO2.

A Correlation of ZEB1 expression and ELMO3 expression. Each data point represents an individual cell line (n = 1684). The Spearman correlation coefficient (R) and the two-tailed P value are displayed on the graph. B The expression of ELMO2, ELMO3, and ZEB1 was determined by qRT-PCR in NCI-H1299 and A549 expressing a ZEB1 shRNA or a control vector. Data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. C ZEB1, ELMO3, ELMO2, vimentin and E-cadherin expression were determined by immunoblotting in NCI-H1299 and A549 expressing a ZEB1 shRNA or a control vector. The samples derive from the same experiment but different gels for ELMO2, β-actin, another for E-cadherin, vimentin, another for ELMO3 and another for ZEB1 were processed in parallel. The expression of ELMO2, ELMO3 and ZEB1 was determined by qRT-PCR (D) and immunoblotting (E) in HCC827 cells expressing a ZEB1 cDNA or a control vector. The samples derive from the same experiment, but different gels for ELMO3, β-actin, another for ELMO2 and another for ZEB1 were processed in parallel. In (D), data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. F ZEB1 overexpression repressed luciferase activity in ELMO3 promoter reporter assays. Data are mean ± SEM, one-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. G Enrichment of ELMO3 promoter in a chromatin immunoprecipitation assay using ZEB1 antibody or a control IgG. Data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. The expression of ELMO2, ELMO3 and ZEB1 was determined by qRT-PCR (H) and immunoblotting (I) in NCI-H1975 and HCC827 cell lines with or without TGF-β (10 ng/ml) treatment. The samples derive from the same experiment, but different gels for ELMO3, β-actin, another for ELMO2 and another for ZEB1 were processed in parallel. In panel H, data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. J Cell viability of NCI-H1299 and A549 cells expressing an ELMO2 shRNA, or both ELMO2 and ZEB1 shRNAs. Data are mean ± SEM, two-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. K Immunoblotting of p62, LC3, FAK, and phosphorylation of FAK Y397 in NCI-H1299 and A549 cells expressing an ELMO2 shRNA, a ZEB1 shRNA, or both ELMO2 and ZEB1 shRNAs. The samples derive from the same experiment but different gels for ELMO2, another for FAK-pY397, β-actin, another for FAK, another for ZEB1, and another for p62 and LC3 were processed in parallel. L Cell viability of HCC827 expressing an ELMO2 shRNA and a ZEB1 cDNA or a control vector. Data are mean ± SEM, two-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. M Immunoblotting of p62, LC3, FAK, and phosphorylation of FAK Y397 in HCC827 expressing an ELMO2 shRNA and a ZEB1 cDNA or a control vector. The samples derive from the same experiment but different gels for ELMO2, β-actin, another for FAK-pY397, another for FAK, another for ZEB1, and another for p62 and LC3 were processed in parallel. Data in (C, E, I, K, M) are representative of two independent experiments. OE overexpression, KD knockdown. Source data are provided as a Source Data file.

ELMO2 inhibitor C52 induces mesenchymal-like cell death

We applied a structure-based screen to identify small molecules that target ELMO2 and identified a series of small molecules that are likely interaction with ELMO2. We evaluated the effects of top 63 commercially available candidate small molecules on the viability of NCI-H1299 cells and identified six compounds (C6, C43, C3, C40, C52 and C56) that exhibited over 50 % inhibition at 10 µM (Fig. 6A). We next assessed their selectivity in mesenchymal-like and epithelial-like cells, and identified C52 specifically suppressed the viability of mesenchymal-like cells compared with epithelial-like cells (Fig. 6B, Supplementary Fig. S6A–E). The structure of C52 is shown in Fig. 6C. C6, C43, C3, C40 and C56 showed no differential activity in mesenchymal-like and epithelial-like cells, as well as in ELMO3 over-expression cell lines (Supplementary Fig. S6A–E). To explore the molecular basis of C52–ELMO2 interaction, we analyzed the docking-predicted binding conformation of C52 within a predicted binding pocket of ELMO2. As shown in Fig. 6D, C52 was positioned within a hydrophobic cleft surrounded by residues TYR-287, GLN-290, VAL-291, PHE-294, ILE-420, LEU-421, GLN-422, ARG-431, HIS-435, LEU-485, PRO-486, and LYS-488. Multiple hydrophobic contacts were observed between C52 and surrounding nonpolar residues, including HIS-435, which may contribute to binding stabilization. In addition, predicted hydrogen bond interactions were detected between C52 and ARG-431, together with a predicted π–π interaction between the aromatic ring of C52 and the imidazole ring of HIS-435. These interactions were identified using PLIP (v2.3.0) and visualized in PyMOL, supporting the structural plausibility of the proposed binding mode. The structure shown represents a computational docking prediction rather than an experimentally determined X-ray or cryo-EM structure. Isothermal titration calorimetry (ITC) studies demonstrated that C52 binds to ELMO2 with a dissociation constant (KD) of 1.0 µM ± 0.17 µM, whereas no binding was observed with the ELMO2 His435Ala (H435A) mutant, suggesting the critical role of HIS-435 in the interaction (Fig. 6E, F). We also functionally validated the on-target activity of C52 using a genetic rescue assay. Depletion of endogenous ELMO2 via 3’ UTR-targeting shRNA was followed by reconstitution with either wild-type ELMO2 cDNA lacking 3’UTR or a C52 binding-deficient ELMO2-H435A mutant (Supplementary Fig. S7A). Cells expressing a vector control underwent cell death upon endogenous ELMO2 knockdown, confirming the essential survival function of ELMO2 (Supplementary Fig. S7B). In contrast, cells expressing either wild-type or H435A mutant ELMO2 were fully viable following knockdown, demonstrating that the H435A mutation does not impair the core survival function of ELMO2 (Supplementary Fig. S7B). Expression of wild-type ELMO2 restored the sensitivity of cells to C52-induced cell death (Fig. 6G). In contrast, cells expressing the ELMO2-H435A mutant were resistant to C52 treatment (Fig. 6G). C52 showed no toxicity in normal lung cell lines HBE and MRC-5 (Supplementary Fig. S7C). ELMO3 overexpression rescued the cell death phenotype induced by C52 treatment (Fig. 6H), suggesting that the activity of C52 is dependent on ELMO3 expression. Similar to ELMO2 knockdown, C52 suppressed FAK phosphorylation in a concentration-dependent (Fig. 6I) and time-dependent manner (Fig. 6J). As expected, C52 induced autophagy-dependent cell death in NCI-H1299 and A549, similar to the phenotype observed upon ELMO2 knockdown (Fig. 6K). These results suggest that the mechanisms of C52 affecting cell viability is similar to those of ELMO2-knockdown. To determine the effects of C52 in clinical samples, we tested its effects in patient-derived lung cancer organoids. The expression of ELMO3 was lower in mesenchymal-like organoids 1 and 2 than in epithelial-like 3 and 4 (Fig. 6L), consistent with the previous observations in lung cancer tissue samples and cell lines (Fig. 4D, G). C52 treatment induced cell death in organoids 1 and 2, but not in 3 and 4 (Fig. 6M, Supplementary Fig. S7D). Defactinib, a potent FAK inhibitor (FAKi), induced cell death in organoid 1 and 2 (Supplementary Fig. S7E). We further observed the potential synergistic effects between C52 and FAK inhibitors. These results demonstrate that Defactinib suppresses lung cancer cell proliferation (Supplementary Fig. S8A), and exhibits a significant synergistic effect with C52 (Supplementary Fig. S8B–D).

Fig. 6. ELMO2 inhibitor C52 induced cell death of ELMO3-low cells and patient-derived lung cancer organoids.

A The cell viability of NCI-H1299 cells after treatment with virtual screening candidates. B Cell viability of mesenchymal-like cells NCI-H1299, A549, NCI-H1703, 95D, NCI-H1792 and epithelial-like cells PC-9, NCI-H1975, HCC827, NCI-H358, Calu-3 treated with C52. Each data point represents mean ± SEM, n = 3 independent replications. C Chemical structure of ELMO2 inhibitor C52. D Docking-predicted binding mode of C52 within the ELMO2 binding pocket. The overall structure of ELMO2 is shown in cartoon representation (cyan), with C52 depicted in stick representation (orange). The inset provides a close-up view of the binding pocket, highlighting interacting residues including TYR-287, GLN-290, VAL-291, PHE-294, ILE-420, LEU-421, GLN-422, ARG-431, HIS-435, LEU-485, PRO-486, and LYS-488. A blue solid line indicates a predicted hydrogen bond, and gray dashed lines denote hydrophobic contacts between C52 and surrounding residues. A green dashed line represents a π–π interaction between C52 and HIS-435. E ITC titrations of C52 to ELMO2. (F) ITC titrations of C52 to ELMO2 H435A mutant. G Cell viability of NCI-H1299 cells expressing a wild-type ELMO2 cDNA or ELMO2 H435A cDNA or a control vector treated with C52. Data are mean ± SEM, two-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. H Cell viability of NCI-H1299 cells expressing an ELMO3 cDNA or a control vector treated with C52. ELMO3 overexpression rescued the cell death caused by C52 treatment. Each data point represents mean ± SEM, n = 3 independent replications. I Immunoblotting of FAK and FAK Y397 phosphorylation in human lung cancer NCI-H1299 and A549 cells treated with C52 with 5 or 10uM. The samples derive from the same experiment but different gels for FAK, β-actin and another for FAK-pY397 were processed in parallel. J Immunoblotting of FAK and FAK Y397 phosphorylation at various time points in NCI-H1299 cells treated with C52. The samples derive from the same experiment but different gels for FAK, β-actin and another for FAK-pY397 were processed in parallel. K Immunoblotting of p62 and LC3 expression in NCI-H1299 and A549 cells treated with C52. The samples derive from the same experiment but different gels for LC3, β-actin and another for p62 were processed in parallel. L Immunoblotting of ELMO3, ELMO2, vimentin and E-cadherin expression in patient-derived lung cancer organoids. The samples derive from the same experiment but different gels for ELMO3, β-actin, another for ELMO2, vimentin and another for E-cadherin were processed in parallel. M Viability of patient-derived lung cancer organoids following C52 treatment. Data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. N Tumor growth of mesenchymal-like A549 transplantation in mice treated with C52 or control. Each data point represents mean tumor volume ± SD, n = 6 biological replicates. Two-way ANOVA with Šidák’s multiple comparisons test. O Tumor growth of epithelial-like PC-9 transplantation in mice treated with C52 or control. Each data point represents mean tumor volume ± SD, n = 6 biological replicates. Two-way ANOVA with Šidák’s multiple comparisons test. Data in (E, F, I–L) are representative of two independent experiments. Source data are provided as a Source Data file.

We determined whether C52 can served as a potential safe and effective small molecular therapeutic inhibitor for tumors in vivo. Dose-escalation study showed C52 exhibited a dose-dependent antitumor response (Supplementary Fig. S9A). C52 treatment produced no observable toxicity in heart, liver, spleen, lung and kidney tissues upon histopathological assessment (Supplementary Fig. S9B). C52 significantly suppressed tumor growth in mesenchymal-like A549 xenograft mice (Fig. 6N), and showed no significant growth inhibition in epithelial-like PC-9 xenograft mice (Fig. 6O). Furthermore, C52 failed to suppress the growth of tumors expressing the ELMO2-H435A mutant in vivo (Supplementary Fig. S9C).

ELMO2 is a therapeutic vulnerability in EMT-associated EGFR-TKI resistant cells

EMT was found to be responsible for drug-resistance in lung cancer¹¹. To determine whether ELMO2 is a therapeutic vulnerability in EMT-associated drug resistance cells, we generated EGFR inhibitor Osimertinib-resistant cell lines NCI-H1975-OR and HCC827-OR (Fig. 7A). Osimertinib-resistant cell lines demonstrated EMT phenotypes, with a decrease in epithelial marker E-cadherin expression and an increase in mesenchymal marker vimentin expression, indicating that these cells had undergone EMT (Fig. 7B). ZEB1 expression was significantly elevated in resistant cell lines compared to parental cell lines (Supplementary Fig. S10A, B). We then determined the expression of ELMO3 and ELMO2 in NCI-H1975-OR, HCC827-OR, and their parental cell lines. ELMO3 expression significantly decreased in resistant cell lines compared with their respective parental cell lines, whereas ELMO2 expression showed no difference (Fig. 7C, D). In addition, analysis of publicly available transcriptomic data from EGFR TKI-resistant patient-derived cell lines that had acquired EMT-associated resistance showed a significant decrease of ELMO3 expression compared to EGFR TKI-sensitive or non-EMT-associated resistant patient-derived cell lines (Fig. 7E)³⁸. Consistent with the result of analysis of patient-derived cell lines, ELMO3 expression was significantly downregulated in EGFR-TKI-resistant clinical samples compared with the pre-treatment samples (Fig. 7F). To further validate these findings at the protein level, we performed immunohistochemical analysis on EMT-associated Osimertinib-resistant tumor tissues (Fig. 7G). The expression of epithelial marker E-cadherin and mesenchymal marker vimentin was determined by immunoblotting in these pre-treatment and post-resistant clinical samples (Supplementary Fig. S10C). The immunohistochemical results revealed that ELMO3 expression was significantly downregulated in EMT associated Osimertinib-resistant tumor tissues, whereas ELMO2 expression showed no significant change (Fig. 7G, H). ELMO2 knockdown resulted in cell death in resistant cell lines, but not in parental cell lines (Fig. 7I, Supplementary Fig. S10D). Consistent with the findings from the genetic suppression of ELMO2, pharmacological targeting of ELMO2 with C52 exhibited a significant increase of sensitivity in NCI-H1975-OR and HCC827-OR than their parental cell lines (Fig. 7J). In NCI-H1975-OR and HCC827-OR cell lines, genetic suppression of ELMO2 or pharmacological targeting with C52, and treatment with FAKi induced autophagy and reduced FAK phosphorylation at Y397 site (Fig. 7K, L, Supplementary Fig. S10E). We investigated the effects of C52 on tumor growth in NCI-H1975 and NCI-H1975-OR xenograft mice and found that C52 treatment significantly inhibited tumor growth in NCI-H1975-OR xenograft mice, but had no effect in NCI-H1975 xenograft mice (Fig. 7M). These results suggest that ELMO2 is a potential therapeutic target in EGFR-TKI-resistant lung cancer, and C52 exhibits activities in resistant cells.

Fig. 7. ELMO2 is a therapeutic vulnerability in EMT-associated EGFR-TKI resistant cells.

A Cell viability of NCI-H1975 and NCI-H1975-OR, HCC827 and HCC827-OR treated with EGFR inhibitor Osimertinib. Each data point represents mean ± SEM, n = 3 independent replications. B Immunoblotting of epithelial marker E-cadherin and mesenchymal marker vimentin in EGFR inhibitor Osimertinib-resistant cell lines NCI-H1975-OR, HCC827-OR and their parental cell lines. ELMO2 and ELMO3 expression were determined by qRT-PCR (C) or immunoblotting (D) in EGFR inhibitor Osimertinib-resistant cell lines NCI-H1975-OR, HCC827-OR and their parental cell lines. ELMO3 expression was significantly decreased in NCI-H1975-OR and HCC827-OR compared to their parental cell lines. The samples derive from the same experiment but different gels for ELMO3, β-actin and another for ELMO2 were processed in parallel. In (C), data are mean ± SEM, unpaired two-tailed Student’s t test, n  =  3 independent experiments. E The expression of ELMO2, ELMO3 and EMT associated genes in patient-derived cell lines (n = 8 biologically independent cell lines) of EGFR-TKI resistance. F ELMO3 expression in paired pre-treatment and Osimertinib-refractory clinical samples (GEO databases). Paired two-tailed Student’s t test, n = 9 biological replicates. G Representative immunohistochemistry staining of ELMO3 and ELMO2 in paired pre-treatment and Osimertinib-resistant clinical samples. Scale bar, 100 μm. Consistent results were observed across three biological replicates. H ELMO2 and ELMO3 IHC quantification in pre-treatment and Osimertinib-resistant clinical samples. Average optical density (AOD) value was calculated as [AOD = Integrated Optical Density (IOD)/Area]. Paired two-tailed Student’s t test, n = 3 biological replicates. I Relative cell viability of NCI-H1975 and NCI-H1975-OR, HCC827 and HCC827-OR expressing an ELMO2 shRNA or a control vector. Data are mean ± SEM, two-way ANOVA with Tukey’s multiple comparisons test, n  =  3 independent experiments. J Cell viability of NCI-H1975-OR, HCC827-OR, and parental cell lines treated with ELMO2 inhibitor C52. Each data point represents mean ± SEM, n = 3 independent replications. K Immunoblotting of p62, LC3, FAK, and phosphorylation of FAK Y397 in EGFR inhibitor Osimertinib-resistant cell lines NCI-H1975-OR, HCC827-OR, and their parental cell lines expressing an ELMO2 shRNA, or a vector control. The samples derive from the same experiment but different gels for ELMO2, β-actin, another for p62, another for FAK-pY397, LC3 and another for FAK were processed in parallel. L Immunoblotting of p62, LC3, FAK, and phosphorylation of FAK Y397 in EGFR inhibitor Osimertinib-resistant cell lines NCI-H1975-OR, HCC827-OR, and their parental cell lines with or without C52 treatment. The samples derive from the same experiment but different gels for p62, LC3, another for FAK-pY397, β-actin and another for FAK were processed in parallel. M Tumor growth curves of NCI-H1975 and NCI-H1975-OR transplantation in mice treated with C52 or control. Each data point represents mean tumor volume ± SD, n = 6 biological replicates. Two-way ANOVA with Šidák’s multiple comparisons test. Data in (B, D, K, L) are representative of two independent experiments. Source data are provided as a Source Data file.

Discussion

Resistance to EGFR inhibitors is common in lung cancer patients. Few options are available for the treatment of resistance. In this study, we identified ELMO2 as a potential therapeutic target in resistant cells. We found that ELMO3-low EGFR-TKI-resistant cells exhibit an EMT phenotype and are dependent on ELMO2. ELMO2 inhibition in these cells induced cell death. We identified an ELMO2 inhibitor, C52, from a structure-based in silico screen that demonstrated efficacy in the induction of cell death in resistant cells.

EMT has been shown to play a critical role in the resistance to multiple therapies in cancer, including EGFR-TKI³⁹. Recently, targeting the EMT process with agents such as NP137, a netrin-1 inhibitor, and TGF-β inhibitors, in combination with traditional chemotherapy or targeted therapies, has gained attention¹⁷. Our study identified a vulnerability in EGFR inhibitor-resistant lung cancer cells. ELMO2/ELMO3 is synthetic lethal in these resistant cells where the differential expression of ELMO3 in mesenchymal-like cells versus epithelial-like cells rendered ELMO2 selectively essential in mesenchymal-like resistant cells. ELMO3 expression was significantly decreased in EGFR inhibitor Osimertinib-resistant cell lines compared with their parental cell lines, which made it possible to induce cell death in EGFR-TKI-resistant cell lines by disrupting ELMO2 functions. ELMO2 regulated autophagy through the function of FAK, which is essential in cancer cells. Although FAK inhibitors have been developed, their therapeutic potential is limited because of toxicity in normal tissues. Our results identified a pathway to disrupt FAK functions and potentially reduce or eliminate the toxicity of FAK disruption in normal cells because of ELMO3 expression in these cells.

The results of our study further demonstrated that the relationship between EMT and drug resistance is complex and significant. ELMO3 expression was highly correlated with EMT-associated biomarkers such as ESRP1 and CDH1. ELMO3 was transcriptionally regulated by EMT core transcription factor ZEB1, and these alterations that specifically appear in mesenchymal-like drug-resistant cells provide potential therapeutic targets that can be selectively disrupted with limited toxicity.

We selected a region of ELMO2 predicted to interact with FAK based on its structure for the small molecule screen. The compound C52 discovered from the screen demonstrated its effects and molecular mechanisms similar to ELMO2 knockdown in human lung cancer cells and organoids. These results suggest that ELMO2 is a druggable therapeutic target and the disruption of its interaction with FAK is potentially achievable. Our study identified a therapeutic strategy to overcome EGFR-TKI resistance and a biomarker for patient stratification in future clinical trials. Targeting EMT-related pathways holds promise for enhancing treatment responses and improving survival outcomes in lung cancer patients.

Methods

Ethical statement

The use of patient samples in this study was approved by the Institutional Review Board of Shanghai East Hospital, Tongji University (K22-041Y). All animal experiments were performed in compliance with the ethical regulations and were approved by the Institutional Animal Care and Use Committee of Zhongshan Hospital, Fudan University (B2025-046R).

Cell lines and cell culture

The NCI-H1299, A549, NCI-H1703, 95D, NCI-H1792, PC-9, NCI-H1975, Calu-3, HCC827, NCI-H358, HBE, and MRC-5 cell lines were purchased from the Cell Bank of the China Academy of Sciences (Shanghai, China). All cell lines were grown at 37 °C and 5 % CO₂, cultured in medium supplemented with 10 % fetal bovine serum (FBS), 100 IU/ml penicillin, and 100 μg/ml streptomycin. NCI-H1299, A549, NCI-H1703, NCI-H1792, NCI-H1975, PC-9, HCC827 and NCI-H358 were cultured in RPMI-1640 (Gibco); 95D, Calu-3, HBE and MRC-5 cell lines were cultured in DMEM (Gibco).

Derivation of Osimertinib-resistant cell lines

To create resistance lines, two EGFR-mutant, Osimertinib-sensitive human lung cancer cell lines, NCI-H1975 (EGFR L858R/T790M) and HCC827 (EGFR del19), were subjected to chronic, repeated exposure to stepwise-increased concentration of Osimertinib, starting at the IC₃₀. Fresh drug was added every 48–72 h. The established Osimertinib-resistant (OR) cell lines NCI-H1975-OR and HCC827-OR were maintained in Osimertinib at a concentration of 1 μM.

Human LUAD samples

Tumor specimens were collected from patients diagnosed with LUAD who underwent surgical resection at Shanghai East Hospital, Tongji University. The expression levels of E-cadherin, vimentin, ELMO2, and ELMO3 were assessed using western blot analysis or immunohistochemical staining. Fresh tumor tissues were utilized for the establishment of patient-derived organoid cultures. Written informed consent for the use of tissue samples was obtained from all participating patients.

DepMap data mining

CRISPR gene effect data (Chronos) and gene expression data were downloaded from the 2024Q4 DepMap release using the Broad Institute’s DepMap portal. In the DepMap CRISPR gene effect data, the median value for non-essential genes is 0, while for essential genes, it is −1. Genes with values closer to −1 are considered more essential. A lower Chronos score indicates a higher likelihood that the gene is essential for the cell line. A strongly-selective gene is identified as a gene whose dependency has the product of skewness and kurtosis in the top quartile and has at least one dependent cell line.

DepMap mesenchymal-like and epithelial-like analysis

Cell lines were characterized as being mesenchymal-like cells (ssGSEA Hallmark Epithelial-Mesenchymal Transition score > 0.25 and Growth Pattern is Adherent) and epithelial-like cells (ssGSEA Hallmark Epithelial-Mesenchymal Transition score < −0.33 and Growth Pattern is Adherent) based on DepMap annotations. 189 cells were classified as epithelial-like cells, and 142 cells were classified as mesenchymal-like cells. The mean gene effect for 3540 strongly selective genes was assessed in both groups, yielding two values per gene: one for epithelial-like and another for mesenchymal-like cells. Additionally, the difference between these mean gene effects was then calculated, resulting in a gene list with greater dependency in mesenchymal-like cells.

DepMap genetic co-dependency and expression analysis

Using the DepMap CRISPR gene effect data (Chronos) and gene expression data, we identified genes whose expression correlates with ELMO2’s gene effect, and those co-dependent on ELMO2 by calculating correlation coefficients.

Reagents

Small molecular inhibitor C52 and an additional 62 candidate inhibitors were purchased from Topscience. Osimertinib (S7297), Necrostatin-1 (S8037), and Defactinib (S7654) were purchased from Selleck Chemicals. Bafilomycin A1(HY-100558), Chloroquine (HY-17589A), Z-VAD (HY-164388), TGF-β (HY-P7118), and Doxycycline (HY-N0565) were purchased from MedChemExpress.

shRNAs and plasmid constructs

shRNAs used in this study include shELMO2-1(5’-GAGATGGCCCATCAGCTATAT-3’), shELMO2-2 (5’-GAACTGAGGAGGATTGCATTT-3’) and shELMO2 (3’UTR) (5’-CTTGAAGTGACTACCTGTATC-3’); shELMO3 (5’-TGGCCTCGGACATGATCTTTG-3’); shATG5 (5’-AGATTGAAGGATCAACTATTT-3’); shULK1 (5’-GCCCTTTGCGTTATATTGTAT-3’); shZEB1 (5’-CCTCTCTGAAAGAACACATTA-3’). These shRNAs were inserted into pLKO.1-puro. Additionally, shELMO2-1 was also cloned into the doxycycline-inducible vector Tet-pLKO-puro.

Full-length FAK, ELMO2, ELMO3, and ZEB1 cDNA (lacking each 3′ UTR) were amplified by PCR and cloned into the pCDH-CMV-MCS-EF1-Puro vector. ZEB1 cDNA was also cloned into the pCDNA3.1 vector. The plasmid of FAK site mutant Y397E, Y397F, K454R, and ELMO2 site mutant H435A was constructed using Mut Expression II Fast Mutagenesis Kit V2 (C214-01, Vazyme). The ELMO3 promoter was amplified by PCR and cloned into the pGL3-basic vector.

Lentiviral production

Lentiviral production was performed using HEK293T cells. Scrambled shRNA, shELMO2-1, shELMO2-2, shELMO2 (3’UTR), shELMO3, shULK1, shATG5, shZEB1, ELMO2-pCDH, ELMO2 H435A-pCDH, ELMO3-pCDH, FAK WT-pCDH, FAK Y397E-pCDH, FAK Y397F-pCDH, FAK K454R-pCDH, ZEB1-pCDH, and vector controls with the packing plasmids PMD2.G and PSPAX were transfected using Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instructions. Infected lentiviruses were collected 24 h and 48 h after transfection, then centrifuged to remove cell debris and filtered through 0.45 μm filters (Millipore).

Antibodies

All primary antibodies for the Western blot were diluted in blocking buffer and incubated overnight at 4 °C. ELMO2 (Abcam, 1:1000, ab181234), ELMO3 (Abcam 1:1000, ab219791), FAK (CST, 1:1000, #71433), phosphor-FAK(Tyr397) (CST, 1:1000, #8556), p62 (Proteintech, 1:1000, 18420-1-AP), LC3 (CST, 1:1000, #12741), Vimentin (CST, 1:1000, #5741), E-cadherin (CST, 1:1000, #3195), Vinculin (Proteintech, 1:1000, 26520-1-AP), ZEB1 (Proteintech, 1:1000, 21544-1-AP), beta-actin (Proteintech, 1:1000, 66009-1-Ig). Primary antibodies were detected with anti-mouse IgG (H + L) (Proteintech,1:10000, SA00001-1), anti-rabbit IgG (H + L) (CST,1:10000, #7074) and anti-Goat IgG (H + L) (Abcam,1:20000, ab6885).

Clonogenic assay

Cells were infected with the indicated lentiviruses. The medium was replaced with medium containing 2 μg/mL puromycin after 48 h. Medium was replaced with puromycin-free medium following puromycin selection. For crystal violet staining, cells were fixed with 10 % formalin for 20 min at room temperature, subsequently stained with 0.1 % crystal violet in 70 % ethanal for 30 min at room temperature with constant shaking. To remove unbound crystal violet, cells were washed with deionized water three times for 5 min each.

Cell viability assay

Cell viability was measured using the CCK-8 assay. Cells were seeded in 96-well plates at a density of 1000 cells per well and drug-treated on the following day. CCK-8 regent was added 72 h after the drug was added, and fluorescence at 450 nm was measured on the PerkinElmer EnVision 2105 Multimode Plate Reader.

Synergy analysis was performed using the Zero Interaction Potency (ZIP) model⁴⁰. Varying concentrations of individual drugs and their combinations were administered. After 72 h of treatment, cell viability was measured using the CCK-8 assay, and the results were ultimately analyzed using the SynergyFinder platform⁴¹.

Transwell migration and invasion assay

In vitro cell migration and invasion assays were performed as described previously⁴². Cells were serum-starved for 24 h, then seeded 5 ×10⁴ into the upper chamber of uncoated or Matrigel-coated transwell inserts (200 μL serum-free medium). The lower chamber contained 600 μL of medium with 10 % FBS. After 48 h incubation, cells on the lower membrane surface were fixed with 4 % PFA for 15 min, then stained with 0.5 % crystal violet, and visualized with an Olympus BX43F microscope.

Immunofluorescence staining and image acquisition

Cells were plated onto the glass slides in 12-wells for immunofluorescence staining. Cells were then fixed with 4 % PFA for 30 min, permeabilized with 0.1 % Triton-X100 for 30 min, and blocked with 5 % BSA for 1 h at room temperature. Primary and secondary antibodies were diluted in blocking buffer and incubated at 4 °C overnight and 1 h at room temperature, respectively. Glass slides were mounted with Vectashield Mounting Media with DAPI (H −1200, Vector Laboratories). Images were acquired using an Olympus BX43F microscope. The following antibodies were used: rabbit anti-FAK (1:300 dilution, #71433, CST), Vinculin (Proteintech, 1:500, 26520-1-AP), Alexa Fluor 488 goat anti-rabbit IgG (H + G) (1:500 dilution, A11008, Life Technologies). The focal adhesion density per cell was calculated according to the equation: FA density = Number of FAs in the region (cell) / Area of the region (cell)^43,44.

Monodansyl cadaverine (MDC) assay

For MDC staining, cells were washed one time with PBS and incubated with MDC in PBS at 37 °C for 30 min. After incubation, cells were washed three times with PBS. Cells were immediately analyzed under an Olympus BX43F microscope.

qRT-PCR

Total RNA was isolated using RNAeasy kit (Tiangen), and reverse transcribed with cDNA Reverse Transcription kit (Takara) according to the manufacturer’s instructions. Quantitative real-time PCR was performed using the QuantiTect SYBR green PCR master mix (Takara). All analyses were performed using an Applied Biosystems Q5 PCR machine. mRNA expression was normalized by the expression of beta-actin. The primer sequences were as follows: human beta-actin forward primer: 5’- CACCATTGGCAATGAGCGGTTC-3’ and reverse primer: 5’-AGGTCTTTGCGGATGTCCACGT-3’; ATG5 forward primer: 5’-GCAGATGGACAGTTGCACACAC-3’ and reverse primer: 5’-GAGGTGTTTCCAACATTGGCTCA-3’; ULK1 forward primer: 5’-GCAAGGACTCTTCCTGTGACAC-3’ and reverse primer: 5’-CCACTGCACATCAGGCTGTCTG-3’; ELMO3 forward primer: 5’-TAAGACCTGGAAGGAGATGCGG-3’ and reverse primer: 5’-CGCATTCACCTTGGTTCGGAAG-3’; ELMO2 forward primer: 5’- ATAGCCGAGGAAATCACCGTGG-3’ and reverse primer: 5’- GTCGTTTGTCCTCAGGAGCCTT-3’; ZEB1 forward primer: 5’- AACTGCTGGGAGGATGACAC-3’ and reverse primer: 5’- TCCTGCTTCATCTGCCTGA-3’.

Formation and quantification of exogenous GFP-LC3 vacuoles and autophagy flux analysis

Cells were infected with lentivirus carrying the GFP-LC3 plasmid. These cells were then transduced with shRNA lentiviruses. Autophagy flux in cells was measured upon the transduction with lentivirus carrying the mRFP-GFP-LC3 tandem construct. Cells were plated on coverslips following the infection with shRNA lentiviruses. Fluorescence was recorded with an Olympus BX43F microscope. Puncta positive for green and red correspond to autophagosomes, and puncta only positive for red correspond to autolysosomes.

Co-immunoprecipitation (co-IP) assay

Cells were harvested and washed with PBS before being lysed with IP lysis buffer with 1 mM PMSF. Whole cell lysates obtained by centrifugation were incubated with antibodies at 4 °C. After washing with IP buffer three times, the immunocomplexes were separated by SDS-PAGE. Immunoblotting was performed following standard procedures.

Luciferase reporter assay

Cells were co-transfected with Renilla luciferase plasmid (pRL-TK) with pGL3 basic luciferase vector or pGL3-ELMO3 (contains ELMO3 promoter fragment), and pCDNA3.1 vector or pCDNA3.1-ZEB1 using Lipofectamine 3000 (Invitrogen). After 48 h, the activities of Renilla and firefly luciferase were measured by Dual-Luciferase Reporter Assay System according to the manufacturer’s instructions (Promega) on PerkinElmer EnVision 2105 Multimode Plate Reader. Firefly luciferase activity was normalized to Renilla luciferase signal to account for variations in transfection efficiency. The resulting values were then further normalized to the pGL3-Basic control.

Chromatin Immunoprecipitation (ChIP)

ChIP was conducted using a ChIP Assay Kit (Beyotime, P2078) according to the manufacturer’s protocol. Briefly, Cells were crosslinked with 1 % formaldehyde at 37 °C for 10 min, followed by quenching with 125 mM glycine for 5 min at room temperature. The cells were then lysed using SDS lysis buffer. DNA was sheared into a size of 200–1000 bp (Bioruptor Pico). A fraction of sheared chromatin was processed for crosslink reversal, protein K digestion, and phenol-chloroform purification. The purified DNA was then electrophoresed on a 1.2 % agarose gel for analysis. For chromatin immunoprecipitation, the protein-chromatin complex was incubated with protein A + G-coated agarose beads and anti-ZEB1 antibody overnight at 4 °C. Rabbit IgG was used as a negative control. The immunoprecipitated complexes were subjected to sequential washing steps, followed by DNA elution using elution buffer, and then incubated at 65 °C for 4 h to reverse formaldehyde cross-links. ChIP DNA was purified and amplified with qPCR. ChIPs were normalized to the input control (1 % of the initial chromatin).

Screening of small-molecule ELMO2 inhibitors

To identify potential inhibitors of ELMO2, we conducted a multi-stage virtual screening pipeline followed by molecular dynamics (MD) simulations. Full details of protein and ligand library preparation, docking workflow, and post-screening analyses are provided below.

Protein preparation

The crystal structure of human ELMO2 in complex with BAI1 (PDB ID: 6idx; residues 1–520; resolution 1.70 Å) was retrieved from the Protein Data Bank. This structure was selected due to its superior resolution and reliability among the available experimental structures. Chain A of 6idx, corresponding to ELMO2, was extracted. The Structure Preparation module in MOE (2022, Chemical Computing Group) was used for initial processing of the structure. The processing involved correction of structural issues, addition of missing hydrogen atoms, and reconstruction of incomplete side chains. The Protonate3D module of MOE was used to assign protonation states of ionizable residues at pH 7.4, considering local hydrogen-bonding networks and solvent exposure. The processed structure was subjected to restrained energy minimization to alleviate local steric clashes. The ligand-binding pocket was predicted using the ProteinsPlus server⁴⁵, and the docking grid was centered at (x = –59.085, y = 6.101, z = 4.997) with box dimensions of 24 × 19.5 × 20.25 Å. Final docking preparation was performed using AutoDock Tools (v1.5.6), including the addition of hydrogen atoms, assignment of Gasteiger charges, merging of non-polar hydrogens, and conversion to PDBQT format. All crystallographic water molecules were removed because no conserved structural waters were identified within the binding pocket.

Ligand library preparation

A chemically diverse compound library comprising 12.79 million small molecules was assembled from Topscience (https://www.tsbiochem.com/), which aggregates multiple major commercial vendors, including TargetMol, ChemDiv, ChemBridge, Life Chemicals, Specs, Vitas-M, Analyticon, Enamine, InterBioScreen, Maybridge, Bionet-Key Organics, and Alinda Chemicals. Ligands were preprocessed using Open Babel (v3.1.0) to enumerate possible stereoisomers and common tautomers, assign protonation states at physiological pH (7.4) based on pKa predictions, generate 3D conformations, and perform energy minimization using the MMFF94 force field to yield low-energy geometries suitable for docking⁴⁶. The resulting structures were converted to PDB format for downstream processing. Final docking-ready PDBQT files were generated using the Meeko toolkit (v0.5.0, default parameters), which integrates RDKit-based cheminformatics with AutoDock atom typing to ensure consistent torsional definitions, Gasteiger charge assignment, and rotatable bond handling. All processed ligands were organized into a VirtualFlow-compatible hierarchical directory structure to enable reproducible, large-scale, parallelized virtual screening⁴⁷. The ligand organization scripts were obtained from the official VirtualFlow repository⁴⁷.

Docking workflow

Docking was conducted in four stages using the VirtualFlow pipeline⁴⁷.

Initial screening

An initial large-scale structure-based virtual screening was performed using the prepared ELMO2 structure and a chemically diverse compound library comprising 12.79 million small molecules. All ligands were docked into the defined binding pocket using qvina2 (exhaustiveness = 2, default parameters) to efficiently identify potential binders. The top 200,000 ligands with the most favorable predicted binding affinities were selected for subsequent focused docking (Supplementary Fig. S11A).

Focused docking

The top 200,000 ligands were re-docked into the same grid using AutoDock Vina (v1.2; exhaustiveness = 8, default parameters) and smina (v2016; scoring function = Vinardo, exhaustiveness = 8, default parameters) to enhance sampling thoroughness, thereby identifying reliable candidates. Ligands were ranked independently by each program, and the top 20,000 ligands were selected based on their average rank across both scoring functions, thereby providing internal cross-validation of the docking results (Supplementary Fig. S11B).

Molecular dynamics simulation

To evaluate receptor stability and flexibility, a 600-ns MD simulation of apo ELMO2 was performed using GROMACS (v2023.2) with the CHARMM36 force field. The initial system was constructed using CHARMM-GUI (https://www.charmm-gui.org), solvated in explicit TIP3P water, and neutralized with 0.15 M NaCl to reproduce physiological ionic conditions⁴⁸. The simulation was carried out under periodic boundary conditions in the NPT ensemble at 303 K. MD simulation data supporting this study are available at Zenodo (10.5281/zenodo.17384863), including the full 600-ns trajectory and corresponding topology files. The backbone RMSD stabilized after approximately 100 ns and fluctuated within 0.25 nm thereafter, indicating convergence and structural stability (Supplementary Fig. S12A). The residue-wise RMSF profile revealed that flexible regions were mainly located at surface loops, whereas the binding pocket remained stable throughout the simulation (Supplementary Fig. S12B). Clustering of the equilibrated trajectory yielded 11 representative receptor conformations that were subsequently used for ensemble docking (10.5281/zenodo.17520072).

Ensemble docking

To account for receptor flexibility, the top 20,000 ligands from the focused docking stage were docked against all 11 MD-derived receptor conformations using smina with the Vinardo scoring function. For each ligand, the docking affinity across conformations was calculated, and ligands exhibiting consistently favorable ranks were prioritized for further analysis.

The overall docking workflow, which combined the use of qvina2, AutoDock Vina, and smina—three widely used virtual screening engines employing distinct scoring functions and search algorithms—provided intrinsic cross-validation of binding predictions and enhanced the reliability of the virtual screening results.

Post-screening Analysis

The top 2,000 ligands from the ensemble docking stage were visually inspected to evaluate key binding interactions, including hydrogen bonds, hydrophobic contacts, and shape complementarity within the ELMO2 binding pocket. A subset of 63 compounds was selected and purchased from Topscience for in vitro biological evaluation. The structure-based docking dataset across all 11 MD-derived receptor conformations, supporting the validation of these ligands, is available at Zenodo (10.5281/zenodo.18771244).

Tumor model in mice

BALB/c nude mice were obtained from SLAC Laboratories (Shanghai, China). Mice were female, 6- to 8- week-old, and maintained under specific-pathogen-free conditions with a 12 h light/dark cycle at the Zhongshan Hospital animal facility. The ambient temperature was controlled at 22 ± 2 °C with 40–60 % relative humidity. All animal experiments were approved by the Institutional Animal Care and Use Committee of Zhongshan Hospital, Fudan University (B2025-046R) and were conducted in accordance with relevant ethical guidelines. The endpoints of the mouse experiments were determined prospectively according to the design of each individual study and the approved animal protocol. Mice were monitored regularly throughout the experiments for body weight, overall condition, posture, activity, and, where relevant, tumor progression. Animals were euthanized at the predefined endpoint of each experiment or earlier if humane endpoint criteria were reached. For tumor-bearing mice, humane endpoint criteria included institutionally approved limits on tumor burden, with a maximum longest tumor diameter of ≤1.5 cm (approximately ≤1,500–2,000 mm³), as well as tumor ulceration or necrosis, impaired mobility or feeding, marked body weight loss, and other persistent signs of distress.

To evaluate the impact of ELMO2 deficiency on tumor growth in vivo, A549 lung cancer cells expressing a Tet-pLKO-ELMO2 shRNA, or Tet-pLKO-ELMO2 shRNA and ULK1 shRNA, or ELMO3 cDNA, or FAK Y397E were trypsinized and washed three times in cold PBS, resuspended in PBS, aliquoted into 5 × 10⁶ cells per injection, and transplanted into mice subcutaneously. Doxycycline treatment was started once tumor reached approximately 200 mm³. Doxycycline was administered daily via oral gavage at a dose of 1 mg per mouse for the entire duration of the study.

To investigate the effect of C52 on tumor growth in vivo, we subcutaneously inoculated A549, PC-9, NCI-H1975, NCI-H1975-OR cells, and A549 cells with knockdown of endogenous ELMO2, which express either ELMO2-WT or ELMO2-H435A mutant, respectively. When the tumor volume reached approximately 200 mm³, the mouse models bearing subcutaneous tumors of each cell type were randomly divided into a control group and a C52-treated group. The C52-treated group received daily intraperitoneal injections of C52 at a dose of 25 mg/kg, while the control group was administered DMSO. For the dose-escalation study, C52 was administered at doses of 2.5 mg/kg, 10 mg/kg, or 25 mg/kg. Tumor size was measured, and tumor volume was calculated by the following formula: mm³ = (width x width x length) /2 [mm].

Organoids preparation and drug sensitivity tests

Fresh tumor tissues from NSCLC patients were washed twice using cold PBS, followed by enzymatic digestion using Trypsin-EDTA (Sigma-Aldrich). Cell pellets were washed twice using a cold washing medium and then resuspended in 75 % Matrigel (Corning). Matrigel domes were incubated for 10 min at 37 °C before the addition of the PDO culture media kit (MasterAim). Y-27632 (10 μmol/L) was added to the medium for the 3 days after seeding or passaging. For drug sensitivity tests, organoids were harvested, passaged, and seeded into 96-well cell culture plates. The organoid density was adjusted to 100 organoids/µL of Matrigel before seeding. 100 µL of the organoid culture medium was added to each well. After the diameter of organoids was up to 50–100 micrometers, the organoid culture medium was replaced with 100 µL of drug-containing culture medium and refreshed every three days. Organoids were photographed every day during drug treatment using the Zen software (German). The viability of organoids was tested using CellTiter-Lumi^TM Luminescent 3D cell viability assay kit (Beyotime Biotechnology). The fluorescence value was detected using a multifunctional enzyme-linked immunosorbent assay reader (Thermo Fisher).

Isothermal titration calorimetry (ITC)

Isothermal titration calorimetry (ITC) experiments were conducted using the MicroCal PEAQ-ITC system at 25 °C. The proteins employed in the ITC assay were full-length GST-ELMO2 and GST-ELMO2 H435A. Prior to the experiment, the proteins were centrifuged and degassed. In the ITC assay, a calorimetric cell containing protein (concentration: 20 μM) was titrated with the compound (concentration range: 200 μM) through 20 injections of 2 μL aliquots. The parameters, including the stoichiometry (N), enthalpy of the reaction (ΔH), and association constant (K), were determined using MicroCal PEAQ-ITC Analysis Software to fit the experimental data.

Statistical analyses

No statistical methods were used to predetermine sample size. The sample sizes were determined based on prior experience and established practices in cell biology and animal experimentation. No data were excluded from the analyses. Both cells and animals were randomly assigned to experimental groups. The investigators were kept blinded to group allocations during both the experimental procedures and the outcome assessments. Statistical analyses were performed using GraphPad Prism 9.1.1. Comparisons between two groups were performed using a two-tailed unpaired or paired Student’s t test as indicated in the figure legends. For multiple-group comparisons, statistical significance was assessed using either one-way ANOVA with Tukey’s post hoc test or two-way ANOVA with Šidák’s multiple comparisons test, as appropriate. A p-value < 0.05 was considered statistically significant. Correlation analyses were performed using Spearman’s rank correlation coefficient. Data were presented as mean ± standard error of the mean (SEM) or mean ± standard deviation (SD) as indicated in the figure legends. The number of biological replicates (n) is indicated in each figure legend, and exact p values are provided within the individual figures or Source Data file. The confidence intervals are provided in the Source Data file.

Reporting summary

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

Author contributions

Q.H., M.L., Y.X., F.W., and X.Y. designed the experiments. M.L., Y.X., Y.C., F.X., and X.T. performed the experiments. F.W. performed the virtual screening. M.L., J.H., Y.S., X.Y., F.Z., C.Z., X.C., M.W., and Q.H. helped with the discussion and interpretation of results. M.L. and Q.H. wrote the manuscript. All authors read and approved the paper. Q.H., Y.X. and F.W. provided funding.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.”

Data availability

The CRISPR gene effect data (Chronos) and gene expression data in this study were obtained at https://depmap.org/portal/. The molecular dynamics simulation data supporting this study are available at Zenodo (10.5281/zenodo.17384863).The representative protein conformations data are available at Zenodo (10.5281/zenodo.17520072). The structure-based docking and interaction analysis dataset for ELMO2 ligand validation is available at Zenodo (10.5281/zenodo.18771244). Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-72062-y.