Aberrant cytoplasmic localization of MLH1 characterizes a cell population that seeds breast cancer recurrence Authors: Aloran Mazumder, Jerry Dewitt, Elena Oropeza, Nindo Punturi, Daniel Lozano, Megha Raghunathan, Jonathan Piscitelli, Elham Sajjadi, Elena GueriniRocco, Konstantinos Venetis, Mariia Ivanova, Eltjona Mane, Marianna Dercole, Alberto Concardi, Nicola Fusco, Carol Manhart, Matthew Bainbridge, Svasti Haricharan Published: 2025 ABSTRACT AbstractEstrogen receptor-positive breast cancer remains a leading cause of cancer-related death in women, with mortality largely driven by late recurrence of treatment-resistant disease. Loss of MLH1 promotes resistance to estrogen-targeting therapies by uncoupling cell cycle progression from estrogen regulation. Here, we show that even when MLH1 is abundantly expressed, aberrant cytoplasmic localization in a subset of tumor cells drives endocrine therapy resistance by enabling estrogen-independent growth. This resistance arises from failure to undergo robust cell cycle arrest in response to endocrine therapy, creating acute dependency on CDK4/6 activity. Consequently, CDK4/6 inhibitors induce strong regression in cells with cytoplasmic MLH1 compared to cells with nuclear MLH1. As cytoplasmic localization occurs in ~11% of ER+ patients, it represents a contributor to MLH1 dysregulation. Incorporating cytoplasmic MLH1 localization into diagnostics could guide the use of CDK4/6 inhibitors in this hard-to-treat subset.Subject terms: Breast cancer, Diagnostic markers Introduction Hormone receptor-positive breast cancer patients, commonly termed estrogen receptor positive (ER+), comprise one of the largest cohorts of cancer patients globally. Almost 25% of ER+ breast cancer patients who are initially responsive to standard endocrine-targeting therapies are mismatch repair (MMR) deficient at the time of disease relapse1,2. MMR is a tumor suppressor pathway commonly dysregulated across cancer types, principally through the loss of function of one of four principal genes: MLH1, MSH2, MSH6, or PMS23. Tumors completely deficient for MMR, generally through germline mutations in one of the four genes detailed above, display a molecular phenotype characterized by high numbers of somatic point mutations and genetic instability of numerous microsatellite repeat sequences throughout the genome, referred to as microsatellite instability4,5. MMR loss has independent prognostic value and predicts response to standard therapies across cancer types6. Although best studied in the context of germline mutation of MMR genes, i.e., Lynch syndrome, MMR loss also occurs somatically in tumors through promoter hypermethylation, mutation and/or loss of expression of MMR genes1,7,8. Given its prognostic and predictive value, MMR loss is routinely diagnosed in the clinic either by measuring microsatellite instability and tumor mutation burden, or by immunohistochemistry to assess the presence or absence of MMR proteins3,9. However, its utility as a diagnostic differs based on cancer type10. ER+ breast cancer does not fall under the Lynch syndrome spectrum of cancers as it rarely associates with germline mutations in MMR genes11 although somatic loss of gene expression of MLH1 in ER+ breast cancers induces resistance to endocrine therapy7,12,13. Since somatic loss of MLH1 is detectable in ~25% of endocrine treatment-resistant ER+ breast cancer patients, it is an important driver of poor patient outcome2,7. Studies that investigate the functional impact of MLH1 loss in cancer thus far have largely focused on effects on maintaining MMR14,15. However, we and others have shown that a key consequence of partial, somatic MLH1 loss in ER+ breast cancer is ATM/Chk2-mediated cell cycle dysregulation rather than decreased mismatch repair2,7,16,17. The functional consequence of somatic MLH1 dysregulation in cancer cells on cell cycle regulation and endocrine treatment-resistant growth remains understudied18,19. Endocrine treatment resistance is a major clinical challenge in treating ER+ breast cancer patients and a primary cause of breast cancer-related death12. Sub-clonal diversity is a key contributor to endocrine treatment resistance20,21. Indeed, there is strong evidence in the literature demonstrating that acquired drug resistance in ER+ breast cancer is largely driven by drug-refractory sub-populations (subclones) within these heterogeneous tumors22–25. Therefore, understanding which sub-clonal populations can drive endocrine treatment resistance is a clinically relevant question that could potentially prevent deaths of thousands of ER+ breast cancer patients every year around the world. About 20% of ER + /HER2− breast cancer patients are intrinsically resistant to endocrine therapy while ~40% of initially responsive patients acquire resistance over time26. For patients who are endocrine treatment resistant, CDK4/6 inhibition through administration of palbociclib, ribociclib and abemaciclib has emerged as one of few targeted therapeutic strategies with quantifiable benefit27. However, although the combination of CDK4/6 inhibitors and endocrine therapy improves median progression-free survival in metastatic endocrine treatment resistant patients, not all patients respond27. This is likely because many patients who initially benefit from CDK4/6 inhibitor treatment eventually develop resistance, while several studies suggest that a proportion of patients never benefit from this combination at all28,29. Most successful targeted therapies have complementary diagnostic biomarkers to identify responsive patient populations30. There are no such clinically available biomarkers for predicting response to CDK4/6 inhibitors. In this sense, CDK4/6 inhibitors are a targeted yet “biomarker-untargeted” therapy12,31. Because CDK4/6 inhibitors are expensive, require chronic administration, and have many associated side effects, identifying predictive biomarkers for their use without undue physical and financial burden to patients is an important area of research32–34. Ataxia-telangiectasia mutated (ATM)/Checkpoint kinase (CHK)2 is a cell cycle checkpoint complex that controls CDK4/6 activity at the G1/S checkpoint35. Failure of ER+ breast cancer cells to activate ATM/CHK2 in response to endocrine therapy, through somatic loss of DNA repair proteins like MLH1, preserves CDK4/6 activity inducing resistance to standard endocrine therapies7,31. The ATM-dependent G1 checkpoint is the only cell cycle phase that is sensitive to extrinsic mitogenic cues such as hormones like estrogen36. It is not surprising, therefore, that germline mutation of ATM and/or CHEK2 predisposes incidence of ER+ breast cancer37,38 and mutations in CHEK2 specifically associate with poor prognostic ER+ breast cancer diagnoses37,39. As germline mutations in ATM and CHEK2 are rare37,40,41, somatic loss of proteins like MLH1 that mediate activation of this checkpoint, which are far more common events in patient tumors10,42–44, become critical drivers of endocrine treatment response, as well as CDK4/6 inhibitor sensitivity. Here we show, using multiple model systems and patient data that cellular niches with aberrant cytoplasmic localization of MLH1, detectable in ~11% of ER+ breast cancer patients, seed resistance to endocrine therapy and responsiveness to CDK4/6 inhibitors. These findings suggest a potential predictive diagnostic for first-line combinatorial administration of endocrine therapy and CDK4/6 inhibitors in a substantial, currently at-risk subset of breast cancer patients. Results Subcellular localization of MLH1 predicts endocrine treatment resistance in breast cancer patients To understand the association of MLH1 loss of function with breast cancer outcomes, we assessed MLH1 using immunohistochemistry in a cohort of 1022 breast cancer tumor biopsies from 828 (ER + /HER2–), 83 (HER2+) and 111 (triple negative) patients from the European Institute of Oncology (Fig. 1A: IEO dataset)1. We confirmed complete loss of MLH1 in ~5% of ER + /HER2– breast cancer samples, but also, of note, found that ~11% of tumors have abundant protein levels but remain cytoplasmic with little or no nuclear localization of MLH1 (Figs. 1B and S1A). Similar frequencies of complete loss vs cytoplasmic localization of MLH1 are observable in the other breast cancer subtypes (~11% with cytoplasmic MLH1 and 5% deficient for MLH1; Fig. S1B). However, ER + /HER2– patient tumors are four times as likely to have high levels of cytoplasmic MLH1 (score>2) as HER2+ and triple negative tumors combined (Fig. S1C–E), although this difference must be cautiously interpreted given the smaller sample size for HER2+ and triple negative tumors in the dataset. Similar to complete loss of MLH1, cytoplasmic MLH1 presence in ER + /HER2– breast cancer patient tumors correlates with decreased time to progression (p = 0.03): 11% of patients with cytoplasmic MLH1 progress on endocrine treatment within 5 years of diagnosis compared to only 2% of patients with nuclear localization of MLH1 (Fig. 1C). We also assessed the efficiency of mismatch repair in these tumors by quantifying microsatellite instability in five tumors categorized as cytoplasmic MLH1, five tumors categorized as MLH1 low, and three tumors categorized as nuclear MLH1. None of the tumors in any group tested positive for microsatellite instability, suggesting that residual MLH1 in these tumors is sufficient to protect MMR efficacy, in accordance with the literature1,10,45. Cytoplasmic localization of MLH1 in ER + /HER2- patient tumors associates with poor outcome. A Schematic describing the IEO patient tumor dataset. B Pie chart indicating proportion of patient tumors with MLH1 staining characterized as cytoplasmic (<5% nuclear positive cells and >10% cells with cytoplasmic staining), nuclear (>5% nuclear positive cells), or deficient (<10% of cells with either cytoplasmic or nuclear MLH1 positivity) along with representative images from three patient tumors in each group for cytoplasmic and nuclear MLH1 staining. C REMARK diagram and stacked column graphs comparing patient outcome based on MLH1 subcellular localization. Fisher’s Exact test determined p-value. For cytoplasmic vs nuclear MLH1 p = 0.03. Associated information on patient tumor dataset. D Proportion of PDXs with cytoplasmic and nuclear MLH1 in three independent PDX cohorts with representative images (MLH1, red and DAPI, blue as nuclear stain). More representative images in Figs. S2–3. Bar graphs quantifying percent viability from three biological replicates of three PDX derived organoids models after fulvestrant (E) or tamoxifen (F) treatment relative to vehicle-treated controls. Two-sided Student’s t test determined all p-values except for (C). For E HCI003 vs HCI017 p = 0.0010 and HCI007 vs HCI017 p = 0.017. For F HCI003 vs HCI017 p = 0.0056 and HCI007 vs HCI017 p = 0.00012.*** p ≤ 0.0001, ** p ≤ 0.01, * p ≤ 0.05. For E, F data are presented as mean values ± SD. All schematics were made using BioRender. https://BioRender.com/t5duv9a (Fig. 1A), https://BioRender.com/s998ll3 (Fig. 1B), https://BioRender.com/tlvm9x7 (Fig. 1C). Scale bars = 50 µm (B) and 20 µm (D). Abbreviations: IEO Institute of European Oncology, cMLH1 cytoplasmic MLH1, dMLH1 MLH1 deficient, and nMLH1 nuclear MLH1, DAPI 4’,6-diamidino-2-phenylindole, PDXs patient-derived xenografts, from HCI Huntsman Cancer Institute, BCM Baylor College of Medicine, and ST Xenostart, Treatments, Veh vehicle, Fulvs fulvestrant, and 4-OHT 4-hydroxy tamoxifen, n.s. not significant. Source data for all figures available with the paper. To test whether this substantial frequency of cytoplasmic localization of MLH1 is preserved in experimental model systems, we screened panels of ER + /HER2– breast cancer patient-derived xenografts (PDXs) from three independent collections (Huntsman Cancer Institute (HCI), Baylor College of Medicine (BCM) and XenoStart (ST)) using immunofluorescence for MLH1. In each of these panels, we found even higher frequency of cytoplasmic MLH1 localization, with 40% of HCI, 50% of BCM and 43% of ST PDXs, respectively showing significant cytoplasmic MLH1 localization with minimal nuclear localization of MLH1 (Fig. 1D, additional representative images from all PDXs in Figs. S2–3). To test the association between cytoplasmic MLH1 and response to endocrine therapy in PDX cohorts, we developed PDX-derived organoids (PDxO) and conducted ex vivo growth assays. We found significant endocrine treatment resistance in both HCI lines with cytoplasmic MLH1 (HCI003 and HCI007) demonstrating <25% growth inhibition after administration of either fulvestrant (Fig. 1E) or tamoxifen (Fig. 1F). In contrast, PDxOs with substantive nuclear localization of MLH1 (HCI017) are strongly responsive to both endocrine therapies demonstrating 60% growth inhibition (Fig. 1E, F; Fulv: p = 0.0010, tamoxifen: p = 0.0056). Sub-cellular localization of MLH1 drives endocrine treatment resistance in breast cancer cells To develop an easily manipulable experimental model system with consistent cytoplasmic localization of MLH1, we tested the effect of somatic MLH1 mutations from ER + /HER2− breast cancer patient samples on MLH1 subcellular localization (Fig. 2A). From all detectable mutations in MLH1 in ER + /HER2− breast cancer patient samples in public datasets, predominantly variants of uncertain significance, we selected mutations from each protein-coding domain of the gene (Fig. 2A). Three of the selected mutations are missense (G55V, E199K and D584H) and three are nonsense (E414*, E439* and E717*) mutations (Fig. 2A). A patient-derived MLH1 missense mutation induces cytoplasmic protein localization. A Schematic showing selected mutations in MLH1 detected in ER + /HER2− patient tumors. B Representative western blotting from three biological replicates assessing levels of MLH1 protein in MCF7 cells. C Representative images of MLH1 staining (red) with DAPI (blue) as nuclear counterstain with and without fulvestrant treatment in MCF7 cells with associated quantification from three biological replicates. For WT (Veh vs Fulvs p = 0.0041), For G55V (Veh vs Fulvs p = 0.026), For Fulvs (WT vs G55V p = 0.0014). D Representative western blotting from three biological replicates for MLH1 protein in MCF7 cells expressing shRNA against MLH1, sh control (shLuc) and cells expressing sh-resistant mutant or wild type constructs following cell fractionation into nuclear and cytoplasmic compartments with accompanying quantification as a dot plot from three biological replicates (shLuc vs G55V p = 0.0086). Alphafold structures of the MutL (MLH1-PMS2) complex with either wildtype (E) or G55V mutant (F) MLH1. G Representative western blotting from three biological replicates in MCF7 cells with indicated constructs assessing levels of PMS2 bound to MLH1 following immunoprecipitation pull down of MLH1 with accompanying quantification as a dot plot from three biological replicates. For (shLuc vs G55V p = 0.036) H Representative images of MLH1 staining in MCF7 cells after PMS2 knockdown (knockdown validated in Fig. S5D) and with and without fulvestrant treatment with accompanying quantification presented as bar graph from three biological replicates. For (Veh vs Fulvs P = 0.00002), for (Fulvs vs siPMS2+Fulvs p = 0.00014). Related cell growth data in Fig. S5E. I Stacked graph showing proportion of patients with indicated strength of nuclear or cytoplasmic MLH1 staining using immunohistochemistry. J–L Representative images of MLH1 and PMS2 staining in serial sections of PMS2 low patient tumors in IEO dataset. Inset with dotted lines show PMS2 or MLH1 staining in PMS2 low (K) vs PMS2 retained (L) areas. Representative images from three patient tumors. Two-sided Student’s t test determined all p-values except for (I), where Pearson’s Chi-squared test was used. *** p ≤ 0.0001, ** p ≤ 0.01, * p ≤ 0.05. For all graphs, data are presented as mean values ± SD. Schematic (A) was made using BioRender https://BioRender.com/se4m144. Scale bar = 20 µm (C, H) 100 µm (J–L). Abbreviations: wildtype WT wildtype, NT non-transfected, DAPI (4’,6-diamidino-2-phenylindole), WCL whole cell lysate, Cyto cytoplasmic lysate, Nuc nuclear lysate, sh short hairpin, si short interfering, lu shLuc control, 55 G55V, H3 histone 3, in input lysate, Ig anti-immunoglobulin G, M1 anti-MLH1, Treatments, Veh vehicle, Fulvs fulvestrant. Source data for all figures available with the paper. We then transfected two ER + /HER2− breast cancer cell lines (MCF7 and T47D) stably expressing shRNA against MLH1 (shMLH1) with sh-resistant constructs of each mutant MLH1 cDNA (MLH1mut) or the wildtype (MLH1WT) as control. Protein levels of missense MLH1mut constructs are comparable to that of MLH1WT constructs in both transiently (Fig. 2B) and stably transfected (Fig. S4A) cells. On the other hand, as expected, cell lines with nonsense MLH1mut constructs express lower levels of MLH1 analogous to shMLH1 cells (Figs. 2B and S4A) likely because they undergo nonsense-mediated decay. Next, we tested sub-cellular localization of each of these mutant proteins pre and post-endocrine treatment. As expected, MLH1WT cells have nuclear MLH1 in both MCF7 (Fig. 2C and individual channels in Fig. S4B, C) and T47D (Fig. S5A, B) cells, especially after endocrine treatment (Figs. 2C and S4C, S5B) as previously demonstrated7. Two out of three nonsense MLH1mut (E414* and E439*) have levels of MLH1 that are undetectable by immunofluorescence. The remaining nonsense MLH1mut (E717*) and two out of three missense MLH1mut, E199K and D584H also show significant nuclear MLH1 presence, especially after administration of endocrine treatment. In accordance with these immunofluorescence data, western blotting using an antibody that recognizes the N-terminal domain of the MLH1 protein detected a truncated protein band for E717* but not for the other two truncating mutations (Fig. S4D). However, a missense mutation in the ATPase domain of MLH1, MLH1G55V, appears unable to localize to the nucleus either pre or post-endocrine treatment, remaining abundantly cytoplasmic (MCF7: Figs. 2C and S4B, C: T47D; Fig. S5A, B). In fact, only 12% of MCF7G55V and 14% of T47DG55V cells have detectable nuclear MLH1 after administration of endocrine treatment compared to ~60% of MCF7WT (p = 0.0014, Fig. 2C) and ~50% of T47DWT (p = 0.004, Fig. S5C) cells. To orthogonally validate this finding, we conducted a biochemical cell fractionation assay in MCF7 cells to measure subcellular expression of MLH1 in both cytoplasmic and nuclear compartments. As anticipated, MLH1WT exhibits robust nuclear expression, which is comparable to the shLuc control (Fig. 2D). In contrast, MLH1G55V is predominantly expressed in the cytoplasm, demonstrating ~1.2-fold increase in its nuclear: cytoplasmic ratio relative to both shLuc and MLH1WT. As expected, no detectable protein expression is observed in the shMLH1 control. Collectively, our immunofluorescence and cell fractionation assays indicate that MLH1G55V aberrantly localizes to the cytoplasm. The MLH1G55V mutation is located within a well-ordered region of the protein possessing ATPase activity (Fig. 2A). Both MLH1 and PMS2 are homologs consisting of ordered amino-terminal ATPase domains, long intrinsically disordered regions which contain nuclear localization signals and globular carboxy-terminal domains46–50. MLH1 and PMS2 constitutively dimerize at the carboxy-terminal domains but also through their ATPase domains upon ATP binding. ATP binding to MLH1 facilitates large conformational rearrangements in the heterodimer through intrinsically disordered regions51. Gly55 is a critical residue in the MLH1 ATPase domain, so it is possible that mutation of this residue disrupts ATP binding and consequent conformational changes relayed through the intrinsically disordered regions. To evaluate whether the mutation potentially perturbs the structure of the globular amino-terminal domain, we first used AlphaFold3-Multimer to predict the structure of the MLH1-PMS2 complex with and without this mutation, but did not detect alterations in the predicted interface (Fig. 2E, F). We next tested whether the G55V mutation prevents binding of MLH1 to PMS2 using an immunoprecipitation pull-down assay. Our observations reveal strong PMS2 enrichment in complex with MLH1 in both shLuc control and MLH1WT cells (Fig. 2G). However, MLH1G55V exhibits significantly lower PMS2 enrichment compared to either shLuc or MLH1WT cells (Fig. 2G, p = 0.036). As orthogonal validation, we then examined the impact of lowering PMS2 levels on sub-cellular localization of MLH1. We lowered PMS2 levels in MCF7 cells utilizing siRNA and assessed nuclear localization of MLH1 using immunofluorescence both before and after endocrine treatment. We found that lowering PMS2 levels in these cells (knockdown validated in Fig. S5D) induces aberrant cytoplasmic localization of MLH1 (Fig. 2H) and causes endocrine therapy resistance (Fig. S5E) similarly to MLH1G55V, indicating that the sub-cellular localization of MLH1 is contingent upon availability of PMS2. This observation also validates in the IEO dataset, where patient tumors categorized as PMS2 low demonstrate pronounced cytoplasmic MLH1 expression (Fig. 2I–K), whereas those that are categorized as PMS2 proficient have strong nuclear MLH1 staining (Figs. 2L and S5F). Collectively, these data suggest that perturbation of MLH1-PMS2 binding, either through mutation or through low availability of PMS2 causes aberrant cytoplasmic localization of MLH1. We next tested whether any of these selected MLH1mut alter response to endocrine therapy using MCF7 and T47D shMLH1 cells either transiently (Fig. 3A) or stably (Fig. 3B) expressing the six mutant constructs, or the wild type as control and treated with two common classes of endocrine therapy: fulvestrant or tamoxifen. We observed no significant difference in baseline growth for any of these mutations whether expressed in MCF7 (Fig. S6A) or T47D (Fig. S6B) cells. However, with each of the 6 mutant MLH1 constructs, we observed varying levels of endocrine treatment resistance relative to isogenic MLH1WT cells. As expected, control isogenic shMLH1 cells demonstrate robust endocrine treatment-resistant growth relative to MLH1WT cells (Fig. 3A, B). Of all the mutations, MLH1G55V, most significantly and consistently induces resistance to endocrine treatment, even surpassing shMLH1 cells (Fig. 3A, B). Both MCF7G55V and T47DG55V cells continue to grow at 85-90% efficiency of vehicle-treated controls in presence of either type of endocrine treatment unlike MCF7WT and T47DWT counterparts which fall to ~50% growth efficiency after endocrine therapy (Fig. 3A, B). Lastly, we tested whether MLH1G55V-induced cytoplasmic localization of MLH1 also causes loss of dependence on estrogen signaling for growth using estrogen deprivation assays in serum-starved conditions. In MCF7WT cells, estrogen deprivation decreases growth to 22% of that observed in estrogen-supplemented cells, while MCF7G55V cells retain 61% growth efficiency (Fig. 3C; p = 0.003). In T47DWT cells, similarly, estrogen deprivation decreases growth to 4% of that seen in estrogen-supplemented counterparts, while T47DG55V cells remain at 19% growth efficiency (Fig. S6C; p = 0.003). As expected, both MCF7G55V and T47DG55V cells remain resistant to fulvestrant and tamoxifen even in serum-starved conditions where estrogen is the primary growth stimulus, unlike MCF7WT (Fig. 3C) and T47DWT counterparts (Fig. S6C). A patient-derived MLH1 missense mutation induces endocrine therapy resistance in 2D and 3D models. Heatmaps from three biological replicates representing growth of indicated cell lines stably expressing shRNA against MLH1 treated with indicated treatments (100 nM) after transient (A) or stable (B) expression of sh-resistant mutant and wild type constructs. C Bar graph from three biological replicates showing growth relative to day 0 of MCF7 (WT) and (G55V) cells in charcoal stripped serum with indicated treatments. For -E2 (WT vs G55V p = 0.003), for +E2 (WT vs G55V p = 0.019), for +E2+Fulvs (WT vs G55V p = 0.0006), For +E2 + 4-OHT (WT vs G55V p = 0.000002). Representative images of 3D spheroid growth of T47D cells with and without fulvestrant treatment (D), tamoxifen treatment (E), or in charcoal stripped serum with (+E2) or without (−E2) 10 nM beta-estradiol treatment (E2) (F) with accompanying box plots showing quantification of spheroid size. Each dot represents spheroid from an independent experiment (n = 4). For D WT (Veh vs Fulvs p = 0.0000010), G55V (Veh vs Fulvs p = 0.0017), Fulvs (WT vs G55V p = 0.00018). For E WT (Veh vs 4-OHT p = 0.0000028), G55V (Veh vs 4-OHT p = 0.0010), 4-OHT (WT vs G55V p = 0.000057). For F WT (−E2 vs +E2 p = 0.000029), G55V (−E2 vs +E2 p = 0.38), −E2 (WT vs G55V p = 0.00013). All experiments (A–F) were performed >2 times. Scale bar = 50 µm. Two-sided Student’s t test determined p-values. *** p ≤ 0.0001, ** p ≤ 0.01, * p ≤ 0.05. For bar graphs, data are presented as mean values ± SD. Box plots show median, quartiles, minima, maxima and outliers at 1.5x IQR. Abbreviations: WT wild type, rel. relative, sh short hairpin, A.U. arbitrary unit, Treatments, Veh vehicle, Fulvs fulvestrant, 4-OHT 4-hydroxy tamoxifen, E2 beta-estradiol. Source data for all figures available with the paper. Since MLH1G55V drives endocrine treatment resistance in two cell lines, we also studied its effect on 3D growth using T47DG55V cells (stable expression). T47DG55V spheroids retain ~60% 3D growth efficiency after fulvestrant treatment, similarly to T47D cells with shMLH1 (Fig. S6D), and distinct from T47DWT spheroids, which fall to ~35% (Fig. 3D; p = 0.00018). These results are reproducible with tamoxifen treatment (Fig. 3E; p = 0.00005). The estrogen independent growth phenotype is also more pronounced in 3D growth assays where estrogen deprivation has little to no detectable impact on the growth of T47DG55V (p = 0.3) spheroids while T47DWT spheroids decrease significantly (p = 0.00003) to ~22% of the size of estrogen supplemented controls (Fig. 3F). These observations validate in vivo (Fig. 4A), where mice with mammary fat pad xenograft tumors of MCF7WT cells treated with fulvestrant show tumor regression to ~3% of the size of tumors in vehicle treated mice (p = 0.003; Fig. 4B), while MCF7G55V tumors remain at 35% the size of vehicle-treated counterparts (Fig. 4C). In keeping with fulvestrant resistant tumor growth, MCF7G55V tumors from fulvestrant-treated mice continue proliferating at comparable rates to those from vehicle-treated mice, unlike MCF7WT tumors, which retain only 27% of their proliferative capacity (p = 0.001), as assessed by Ki67 positivity (Fig. S6E). As with the 3D growth experiments, this endocrine treatment resistance was also observable using estrogen deprivation in vivo: estrogen deprivation of mice xenografted with MCF7 cells with endogenous wildtype MLH1 causes tumor shrinkage to 90% of estrogen supplemented controls (Fig. 4D; p = 0.0002), while having no effect on tumors from MCF7 shMLH1 cells with either empty vector (Fig. 4E; p = 0.18) or MLH1G55V (Fig. 4F; p = 0.92). A patient-derived MLH1 missense mutation induces endocrine therapy resistance in vivo. A Schematic of the experimental design for (B–F). Line graphs of tumor volume over time for indicated groups with and without fulvestrant treatment (B–C) and with and without estrogen supplementation in drinking water (D–F). Each line represents an individual tumor. p = 0.0030 (B), 0.04 (C) and 0.00017 (D). Slope of the tumor for each group was calculated and Two-sided Student’s t test was performed on the slope to determine p-values. *** p ≤ 0.0001, ** p ≤ 0.01, * p ≤ 0.05. Schematic was made using BioRender https://BioRender.com/yn6q66g. Abbreviations: WT wild type, vol volume, mm3 millimeter cubed, n.s. not significant, Treatments, Veh vehicle, Fulvs fulvestrant, E2 beta-estradiol. Source data for all figures available with the paper. Together, these findings support a causal role for aberrant cytoplasmic localization of MLH1 caused by MLH1G55V in endocrine treatment resistance and estrogen-independent growth despite robust and stable expression of the mutant protein. Sub-clonal cell populations with cytoplasmic MLH1 seed endocrine treatment resistance We identified MLH1G55V in a treatment-naïve, primary ER + /HER2− breast cancer patient sample NeoPalAna at a low variant allele fraction (VAF) of 3.38% (Fig. 5A, inset)52. After treatment with anastrozole (an aromatase inhibitor, with a similar mechanism of action to estrogen deprivation) for 4 weeks this patient responded to treatment with a 62% decrease in proliferating cells in her post-treatment compared to pre-treatment biopsies, assessed using immunohistochemistry for proliferation marker Ki67 (Fig. 5A). Of note, the tumor did not reach the 2.75% complete cell cycle arrest (CCA) Ki67 positivity that associates with sensitivity to endocrine treatment and long-term decreased risk of recurrence53 although the proliferative inhibition in response to endocrine therapy in this patient is comparable to that of other endocrine therapy responsive patients in the trial whose tumors had no detectable mutation in/dysregulation of MLH1 (Fig. 5A). These findings appear to contradict our experimental data showing a causal role for MLH1G55V in inducing estrogen-independent growth and endocrine treatment resistance (Fig. 4). However, initial reduction in bulk tumor size may not be predictive of ultimate outcome, since it cannot detect treatment response of sub-clonal MLH1G55V cells within the tumor that potentially seed recurrence. Cytoplasmic MLH1 cells persist after administration of endocrine therapy. A Schematic showing variant allele frequency of MLH1G55V (inset) and pre- and post-endocrine treatment Ki67 for all patients in NeoPalAna dataset. B Schematic model of sub-clonal drivers of persistent tumor growth. C Representative flow cytometry images of MCF7G55V (GFP+) and MCF7WT (untagged) cells treated as indicated with accompanying quantification from three biological replicates. Gating strategy for flow is provided in the Supplementary Figs. D Dot plots quantifying proportions of GFP + MCF7G55V and T47DG55V lines grown with indicated treatments from three biological replicates. For MCF7 day 5 (−E2 vs +E2 p = 0.034), day 7 (−E2 vs +E2 p = 0.004). For T47D day 3 (−E2 vs +E2 p = 0.044), day 5 (−E2 vs +E2 p = 0.014), day 7 (−E2 vs +E2 p = 0.0002). Associated data in Fig. S7A–D. E Representative images of 3D spheroids with indicated treatments with accompanying quantification presented as box plot, each dot represents an independent spheroid from a biological replicate (n = 4). For −E2 (Week 2 vs Week 8 p = 0.012), Week 8 (−E2 vs +E2 p = 0.010). Associated data in Fig. S7E–G. F Schematic of in vivo experimental design. G Line graph of tumor volumes of MCF7G55V (GFP+) and MCF7WT cells mixed in 1:9 ratio. Each line represents an individual tumor. H Proportion of MCF7G55V (GFP+) cells from three tumors from each group at harvest quantified using immunofluorescence. For day 48 (−E2 vs +E2 p = 0.000012). I Quantification of Ki67 in GFP+ (MLH1G55V) and GFP- (MLH1WT) cells in indicated conditions from three tumors in each group. For (−E2 p = 0.00017). Representative images in Fig. S7H. Two-sided Students t test was performed to determine p-values. *** p ≤ 0.0001, ** p ≤ 0.01 and * p ≤ 0.05. For graphs, data are presented as mean values ± SD. Box plots show median, quartiles, minima, maxima and outliers at 1.5x IQR. Schematics were made using BioRender https://BioRender.com/x6mukle (Fig. 5A), https://BioRender.com/e5fs30c (Fig. 5B), https://BioRender.com/qqigtzy (Fig. 5F). Scale bar = 100 µm. Abbreviations: WT wild type, CCA complete cell cycle arrest, Dx at diagnosis, Tx treatments, FSC-A forward scatter, GFP green fluorescent protein, Prop proportion, vol volume, mm3 millimeter cubed, Treatments, E2 beta-estradiol. Source data for all figures available with the paper. To test this hypothesis, we experimentally replicated the sub-clonality observed in patient tumors by admixing untagged MCF7WT/T47DWT and GFP-tagged MCF7G55V/T47DG55V isogenic cells at a 9:1 ratio, respectively (Fig. 5B). In flow cytometry assays, MCF7G55V cells are ~4-fold enriched for viability over MCF7WT cells after estrogen deprivation but show no enrichment in estrogen-supplemented control conditions (Fig. 5C; p = 0.02). We confirmed these findings with orthogonal 2D growth assays where we immunohistochemically quantified GFP+ cell populations over 7 days in response to estrogen deprivation relative to estrogen-supplemented controls. Similarly to flow cytometry results, MCF7G55V (p = 0.004) and T47DG55V (p = 0.0002) cells grow from ~10% to constitute ~50% of the total cell population in estrogen deprived conditions but remain at ~10% sub-clonality over time in estrogen-supplemented control groups (Fig. 5D). These data suggests that G55V-bearing cells have little or no growth advantage intrinsically but assume adaptive growth advantage when exposed to endocrine interventions. We confirmed these observations in response to fulvestrant (MCF7: Fig. S7A, T47D: Fig. S7C) and tamoxifen (MCF7: Fig. S7B, T47D: Fig. S7D) treatment. These findings also validate in 3D assays. After 6 weeks of estrogen deprivation, overall spheroid size significantly decreases relative to estrogen-supplemented controls (Fig. S7E), although the effect size (2-fold) is less than that seen for homogeneously T47DWT spheroids (Fig. 3F: 6-fold). However, the proportion of T47DG55V cells more than doubles (p=0.01) in estrogen-deprived conditions while remaining constant in estrogen-supplemented controls (Fig.5E). Similarly, the T47DG55V proportion increases 2-fold with fulvestrant (p=0.00008) and 3-fold with tamoxifen (p=0.000003) treatment, relative to vehicle controls (Fig.S7F, G). We then replicated sub-clonality in vivo by xenografting mixed populations of MCF7WT (untagged): MCF7G55V (GFP-tagged) cells at a 9:1 ratio to form tumors in the mammary fat pads of nude mice (Fig. 5F). Mice were initially supplemented with physiological levels of beta-estradiol until tumors were large enough to randomize into treatment groups: estrogen-deprived or continued estrogen supplementation. As expected, estrogen deprivation significantly decreases bulk tumor volume (p = 0.01) relative to estrogen-supplemented controls, although the response is relatively muted (Fig. 5G). However, proportion of MCF7G55V cells (assessed by quantification of the GFP tag) increases to ~100% of the tumor composition in estrogen-deprived mice, while remaining at ~20% tumor composition in the estrogen-supplemented control arm (Fig. 5H: p = 0.00001). These observations are supported by assessment of tumor proliferation, where MCF7G55V tumor cells in estrogen-deprived mice maintain proliferation at rates comparable to MCF7G55V tumor cells in estrogen-supplemented control mice (p = 0.52), while MCF7WT counterparts within the same tumor significantly inhibit proliferation (p = 0.0002) in estrogen-deprived vs -supplemented mice (Fig. 5I, representative images: Fig. S7H). These findings together suggest that sub-clonal cell populations with cytoplasmic localization of MLH1, induced by missense mutations like MLH1G55V, in patient tumors may escape standard endocrine monotherapy. Because they are sub-clonal, however, they likely remain undetected for extended periods of time while seeding recurrence. Cytoplasmic MLH1 reduces Chk2 activation, leading to endocrine therapy resistance We previously reported that MLH1 proficiency is a prerequisite for Chk2 activation in ER + /HER2– breast cancer cells in response to endocrine therapy, which is essential for responsiveness to treatment7. Therefore, we next tested whether cytoplasmic localization of MLH1 is analogous to MLH1 deficiency in preventing robust Chk2 activation in response to endocrine therapy using the MLH1G55V cell line experimental model systems and in our panel of PDXs with cytoplasmic localization of MLH1. While MCF7WT cells register a 16-fold increase in cells with detectable nuclear pChk2 positivity upon fulvestrant treatment (p = 0.004), MCF7G55V cells have no detectable increase (Fig. 6A). Endocrine therapy administration in ER + /HER2– breast cancer cells results in a proliferative G1 block regulated by ATM/Chk2 checkpoint activation. These results validate in T47D cells as well (Fig. S8A, B). To test whether this block is detectable in MLH1G55V cells, we assessed expression of S-phase marker PCNA in MCF7G55V and MCF7WT xenograft tumors grown in mice treated with either vehicle or fulvestrant. MCF7WT tumors have ~70% decrease in S-phase (PCNA+) cells in response to fulvestrant treatment, suggesting robust G1 arrest, while MCF7G55V tumors have only a~20% decrease (Fig. 6B; p = 0.006). Cytoplasmic MLH1 prevents Chk2 activation despite abundant and stable expression. Representative immunofluorescence photomicrographs showing pChk2 (A) in cells grown in 2D or PCNA (B) in xenografted tumors (red) with a DAPI (blue) nuclear counterstain in indicated cells with indicated treatments. Accompanying quantification is presented as bar graphs from three biological replicates (A) and three tumors in each group (B). For A Fulvs (WT vs G55V p = 0.0042). For B WT (Veh vs Fulvs p = 0.00012), G55V (Veh vs Fulvs p = 0.038) and Fulvs (WT vs G55V p = 0.0058). Associated data in T47D cells in Fig. S8A-B. C Bar graphs from three biological replicates showing growth of indicated cells after indicated treatments for 36 h relative to vehicle-treated controls. For Fulvs (WT vs G55V p = 0.0020). Associated data in T47D cells is presented in Fig. S8C. D Representative images of T47DWT and T47DG55V spheroids after indicated treatments with accompanying quantification. Each dot represents an individual spheroid from a biological replicate (n = 5). For Fulvs/Veh (WT vs G55V p = 0.00030) and Fulvs+DIM/Fulvs (WT vs G55V p = 0.0027). E Regression analysis of cytoplasmic MLH1 positivity plotted against nuclear pChk2 positivity in HCI and BCM PDX lines p = 0.0003. F Dot plot representing quantification of cells positive for cytoplasmic localization of MLH1 and nuclear localization of pChk2 in each PDX tumor of the ST panel p = 0.0007. Each symbol represents an individual PDX. Experiments were performed 3 times. Associated representative images in Fig. S8. Two-sided Student’s t test determined all p-values. *** p ≤ 0.0001, ** p ≤ 0.01 and * p ≤ 0.05. For all graphs, data are presented as mean values ± SD. Vertical line depicts median value in (F). Scale bars = 20 µm (A, B) and 50 µM (D). Abbreviations: WT wild type, DAPI 4’,6-diamidino-2-phenylindole, pChk2 phosphorylated Chk2, PCNA proliferating cell nuclear antigen, cMLH1 cytoplasmic MLH1, nMLH1 nuclear Mlh1, PDXs patient derived xenografts, HCI PDXs from Huntsman Cancer Institute, BCM PDXs from Baylor College of Medicine, Rel. relative, A.U. arbitrary unit, Treatments, Veh vehicle, Fulvs fulvestrant, DIM di-indolyl methane. Source data for all figures available with the paper. This striking discrepancy in Chk2 activation and associated G1-S transition suggests that cytoplasmic MLH1 localization induced by MLH1G55V impairs Chk2 activation leading to a failed G1-S checkpoint block in response to endocrine treatment. To test this hypothesis functionally, we used small molecule Chk2 activator, di-indolyl methane (DIM) to exogenously activate Chk2 in MLH1G55V cells54. DIM, unlike fulvestrant treatment induces robust Chk2 activation comparable to that induced by fulvestrant in MLH1WT cells (MCF7: Fig. 6A, T47D: Fig. S8A, B). In validation of our hypothesis that blunted Chk2 activation mediates endocrine treatment resistance in MLH1G55V cells, administration of DIM rescues sensitivity to fulvestrant in 2D growth assays (MCF7: Fig. 6C, T47D: Fig. S8C). This observation also validates in 3D assays where addition of DIM to fulvestrant decreases spheroid size of T47DG55V cells by 71% when compared to fulvestrant alone, phenocopying the effect of fulvestrant monotherapy on T47DWT cells (Fig. 6D; p = 0.0003). We then investigated whether cytoplasmic localization of MLH1, independent of G55V mutation, prevents Chk2 activation to drive endocrine treatment resistance. We screened our panels of MLH1WT ER + /HER2- breast cancer PDXs for pChk2 using immunofluorescence. First, combining HCI and BCM panels (n = 4 HCI; n = 6 BCM), we conducted an agnostic regression analysis between percent cells with cytoplasmic (but not nuclear) MLH1 and percent cells with nuclear pChk2 in each PDX tumor. In this analysis, cytoplasmic MLH1 presence inversely correlates with pChk2 positivity across PDX lines (Fig. 6E with representative images in Fig. S8D–F). We independently validated this correlation in a third ER + /HER2- PDX panel (ST, n = 16). We categorized PDXs from this panel as nuclear or cytoplasmic MLH1 as described in Materials and Methods (representative images in Figs. 1D and S2–3). We found that nuclear MLH1 PDXs have ~60% nuclear pChk2 positivity, while cytoplasmic MLH1 PDXs show <2% positivity for nuclear pChk2 (Fig. 6F; p = 0.0007). Together these data strongly support a role for cytoplasmic MLH1 in mediating cell cycle dysregulation that precludes response to standard endocrine therapies in ER + /HER2− breast cancer patients. Cytoplasmic MLH1 induces CDK4/6 dependence in ER + /HER2– breast cancer cells and patient tumors Using MCF7G55V cells to represent cells with aberrant cytoplasmic MLH1 localization, we next tested whether aberrant cytoplasmic localization of MLH1 alters other key G1/S-phase cell cycle regulators consequent to Chk2 dysregulation (Fig. 7A). Gene expression of G1 cyclin-dependent kinase CDK6 increases 3-fold (p = 0.01), and early G1 cyclin CCND1 1.6-fold in MCF7G55V cells upon endocrine treatment (p = 0.01) compared to MCF7WT cells (Fig. 7B). Additionally, MCF7G55V cells downregulate CDKN1B the G1 inhibitor directly downstream of Chk2, relative to MCF7WT cells (Fig. 7B; p = 0.003). To test whether upregulation of G1 CDKs by cytoplasmic MLH1 induces cellular dependence on these CDKs for endocrine therapy-resistant growth, we knocked down CDK6 in MCF7G55V (knockdown validated in Fig. 7C) and T47DG55V (knockdown validated in Fig. S9A) cells using siRNA. Knocking down CDK6 in these cells restores sensitivity to endocrine treatment (MCF7: Fig. 7D, T47D: Fig. S9B), suggesting that aberrant cytoplasmic localization of MLH1 induces dependence on G1 CDKs for endocrine therapy resistance in ER + /HER2– breast cancer cells. Cytoplasmic MLH1 predisposes sensitivity to CDK4/6 inhibition in ER + /HER2- breast cancer. A Schematic showing how MLH1 activates Chk2 and inhibits CDK4/6 activity. B qRT-PCR analysis of MCF7 cells of indicated genotypes for relative expression of specified genes following fulvestrant treatment from three biological replicates. For CDK6 (WT vs G55V p = 0.012), CDKN1B (WT vs G55V p = 0.0028), CCND1 (WT vs G55V p = 0.010). C Western blotting confirming CDK6 knockdown in MCF7 cells. Representative blot from three biological replicate. D, E Bar graph showing growth of MCF7WT, MCF7G55V and MCF7 shMLH1cells with respective treatments from three independent biological replicate. Bar graphs describing relative spheroid size grown from T47DWT, T47DG55V and T47D shMLH1 cells (F) and specified PDxO (G) in response to palbociclib treatment. For F each dot represents spheroid from an independent biological replicate (n = 5). For G each dot represents a biological replicate (n = 4). All experiments were performed >2 times. For D Fulvs (WT vs G55V p = 0.00002) and (WT vs shMLH1 p = 0.0019). For E Fulvs (WT vs G55V p = 0.00084, WT vs shMLH1 p = 0.0014), Palbo (WT vs G55V p = 0.0022, WT vs shMLH1 p = 0.010). For F WT (Veh vs Palbo p = 0.00013), G55V (Veh vs Palbo p = 0.000019), Palbo (WT vs G55V p = 0.000004). For G (HCI003 vs HCI017 p = 0.000012, HCI007 vs HCI017 p = 0.000012). H, I Line graphs of tumor volume over time for indicated groups with and without palbociclib treatment each line represents an individual tumor. Waterfall plots depicting CDK4/6 inhibitor response expressed as percent growth of treated tumors relative to vehicle-treated controls in PDX lines in response to (J) palbociclib (p = 0.011) (K) abemaciclib (p = 0.06). For graphs, data are presented as mean values ± SD. Schematic was made using BioRender https://BioRender.com/aok2eve. Two-sided Student’s t test determined p values for (B–I) and Fisher’s exact test determined p values for (J, K). Abbreviations: WT wild type, nMLH1 nuclear MLH1, cMLH1 cytoplasmic MLH1, pChk2 phosphorylated Chk2, Tx treatments, si small interfering, sh short hairpin, n.s. not significant, rel relative, vol volume, mm3 millimeter cubed, PDXs patient-derived xenografts, HCI PDXs from Huntsman Cancer Institute, ST PDXs from Xenostart, T/C treatment/control, Treatments, Fulvs fulvestrant, Palbo palbociclib. Source data for all figures available with the paper. We next tested whether CDK6 dependence in MLH1G55V cells results in sensitivity to CDK4/6 inhibitors. Both MCF7G55V and T47DG55V cells show 40-60% growth inhibition in contrast to the 5-10% growth inhibition of MCF7WT and T47DWT counterparts in response to CDK4/6 inhibitors (palbociclib— MCF7: Fig. 7E, T47D: Fig. S9C and abemaciclib—MCF7: Fig. S9D, T47D: Fig. S9E). As expected shMLH1 cells respond similarly to G55V cells (MCF7: Fig. 7D, E, T47D: Fig. S9B). These results validate in 3D spheroid assays where T47DG55V cells are significantly (p = 0.000004) more susceptible to palbociclib treatment than T47DWT cells (Fig. 7F). We also tested whether this susceptibility to CDK4/6 inhibitors extends to non-mutational cytoplasmic MLH1 localization by assessing organoid growth in two ER + /HER2- PDXs with cytoplasmic MLH1 localization relative to one with nuclear localization (Fig. 7G). Both HCI003 and HCI007 (with cytoplasmic MLH1) have >50% growth inhibition after monotherapy with palbociclib (p = 0.00001) while HCI017 (with nuclear MLH1) shows <10% growth inhibition (p = 0.26). Similar to the organoid data, MCF7G55V xenograft tumors regress in response to monotherapy with palbociclib in vivo, showing >95% inhibition in growth compared to MCF7WT tumors which show a 45% inhibition in tumor volume (Fig. 7H, I; p = 0.002). These data also validate in vivo using PDXs with cytoplasmic MLH1, which are enriched for responders to palbociclib (Fig. 7J) and abemaciclib (Fig. 7K) relative to those with nuclear MLH1. Taken together, our data suggest that cytoplasmic localization of MLH1 can be induced by some point mutations in the ATPase domain of MLH1 that disrupt heterodimerization of MLH1 and PMS2, loss of PMS2 expression, and potentially through point mutations in PMS2 that also disrupt heterodimerization with MLH1 (Fig. 8A). These multiple mechanisms that contribute to cytoplasmic localization of MLH1 account for ~11% of all ER+ breast cancer patients, making its predictive properties of clinical significance. Therefore, we next questioned whether our findings translate to patient data as well. In NeoPalAna (Fig. 8B), we identified two patient tumors of interest: one with MLH1G55V at 3% VAF and another with MLH1Q48C, a mutation in the same ATPase domain as G55V that is detected only after administration of endocrine treatment to the patient with 4% VAF (Fig. 8B, VAFs in inset). While these tumors appear responsive to initial endocrine treatment in terms of proliferative inhibition as measured by Ki67 positivity, both tumors show significant further inhibition, 80% and 84%, respectively, in proliferation after palbociclib treatment (Fig. 8B). In fact, only in response to CDK4/6 inhibitors do these tumors achieve the 2.7% CCA proliferation index that associates with long-term response to endocrine treatment (Fig. 8B). This is in contrast to patient tumors with MLH1WT, which demonstrate 70% growth inhibition with endocrine treatment alone and little further proliferative inhibition (17%) with addition of palbociclib, many of them failing to achieve CCA (Fig. 8B). To test whether the causal role of MLH1G55V in inducing aberrant cytoplasmic localization of MLH1 and thereby endocrine treatment resistance extends to other mutations in the ATPase domain of MLH1, we introduced the Q48C MLH1 mutant into MCF7 shMLH1 cells. We found that, similarly to MLH1G55V, MLH1Q48C causes aberrant cytoplasmic localization of MLH1 both before and after endocrine treatment, in contrast to MLH1WT cells (Fig. 8C). As observed in MLH1G55V cells, MLH1Q48C cells demonstrate considerable resistance to endocrine therapy (Fig. 8D p = 0.002) but show significantly greater sensitivity to both palbociclib (Fig. 8E p = 0.0007) and abemaciclib (Fig. 8F p = 0.0006) treatments when compared to MLH1WT cells. We discovered another missense mutation in MLH1, specifically MLH1E34K, located within the same ATPase domain in a second clinical trial, NCT03007979 (Fig. 8G). In vitro, similarly to MLH1G55V and MLH1Q48C, MLH1E34K causes aberrant cytoplasmic localization of MLH1, both pre and post-endocrine treatment (Fig. 8H). Again, similarly to the other two ATPase domain missense mutations, MLH1E34K induces considerable resistance to fulvestrant treatment, demonstrating only ~25% growth inhibition in contrast to ~50% for MLH1WT (Fig. 8I p = 0.0008). Conversely, MLH1E34K induces significant responsiveness to both palbociclib (Fig. 8J: p = 0.00009) and abemaciclib (Fig. 8K: p = 0.0023) treatments when compared to MLH1WT. In the clinical trial in which the patient with MLH1E34K was identified, ER + /HER2− breast cancer patients were administered serial cycles of CDK4/6 inhibitors, and their circulating tumor DNA was sequenced55. In corroboration of our in vitro data, MLH1E34K decreases in VAF after 3 cycles of CDK4/6 inhibition from 2.5% to undetectable unlike the VAF of RB1 deletion, a known driver of resistance to palbociclib, which increases after exposure to treatment from 1% to 14% (Fig. 8L, M). In this clinical trial, patient tumors were also independently categorized as either proficient or deficient in MMR based on high blood tumor mutation burden and microsatellite instability. Using this categorization, we found that the patient with MLH1E34K, as well as other patients categorized as MMR deficient, show a significant decrease in total circulating tumor DNA load with repeated exposure to palbociclib in contrast to patients who are proficient for MMR or have RB1 variants (Fig. 8M). Alternate mechanisms that promote cytoplasmic MLH1 and predict response to CDK4/6 inhibitors in breast cancer. A Schematic showing alternate mechanisms that can lead to cytoplasmic MLH1 localization in breast cancer. Schematic describing additional MLH1 mutations identified in patient tumors from the NeoPalAna (B) and NCT3007979 (G) trials. Representative images of MLH1 staining (red) with DAPI as nuclear counterstain with and without fulvestrant treatment in MCF7Q48C (C) and MCF7E34K (H) relative to MCF7WT cells. Accompanying quantification is presented as bar graph from three biological replicates. Bar graphs from three biological replicates showing 2D growth of MCF7Q48C (D–F) or MCF7E34K (I–K) cells with indicated treatments. For C WT (Veh vs Fulvs p = 0.00084), Q48C (Veh vs Fulvs p = 0.026) and Fulvs (WT vs Q48C p = 0.0016). For D WT (Veh vs Fulvs p = 0.0017), Q48C (Veh vs Fulvs p = 0.0013) and Fulvs (WT vs Q48C p = 0.0027). For E WT (Veh vs Palbo p = 0.014), Q48C (Veh vs Palbo p = 0.0024) and Palbo (WT vs Q48C p = 0.00070). For F WT (Veh vs Abema p = 0.057), Q48C (Veh vs Abema p = 0.0032) and Abema (WT vs Q48C p = 0.00060). For H WT (Veh vs Fulvs p = 0.00011) and Fulvs (WT vs E34K p = 0.00033). For I WT (Veh vs Fulvs p = 0.00037), E34K (Veh vs Fulvs p = 0.0033) and Fulvs (WT vs E34K p = 0.00088). For J WT (Veh vs Palbo p = 0.0064), E34K (Veh vs Palbo p = 0.00025) and Palbo (WT vs E34K p = 0.000092). For K WT (Veh vs Abema p = 0.035), E34K (Veh vs Abema p = 0.00092) and Abema (WT vs E34K p = 0.0023). VAF of RB1 deletion and MLH1E34K at baseline and after 3 cycles of CDK4/6 inhibitor treatment (L) and circulating tumor fraction in MMR-deficient, MMR proficient and RB1-deleted patient blood samples (M) in NCT03007979. Two-sided Student’s t test determined all p-values. *** p ≤ 0.0001, ** p ≤ 0.01 and * p ≤ 0.05. For graphs, data are presented as mean values ± SD. Schematics were made using BioRender https://BioRender.com/0z0p9kx (Fig. 8A), https://BioRender.com/ic2sdri (Fig. 8B) https://BioRender.com/6z6yy8w (Fig. 8G). Scale bars = 20 µm. Abbreviations: cMLH1 cytoplasmic MLH1, nMLH1 nuclear MLH1, dMLH1 MLH1 deficient, MMR mismatch repair, dMMR deficient MMR, pMMR proficient MMR, Dx at diagnosis, Tx treatments, VAF variant allele frequency, NPA NeoPalAna, NCT Clinical trial NCT3007979, CCA complete cell cycle arrest, RB1 Retinoblastoma, mut mutation, del deletion, BL baseline, Prog progression, ctDNA circulating tumor DNA, 1-3x 1 to 3 cycles, CDK4/6i CDK4/6 inhibitors, DAPI 4’,6-diamidino-2-phenylindole, Treatments, Veh vehicle, Fulvs fulvestrant, Palbo palbociclib, Abema abemaciclib. Source data for all figures available with the paper. Discussion The impact of our findings could be substantial given that cytoplasmic MLH1, caused by MLH1 mutation as well as other non-mutational events, occurs in ~11% of ER + /HER2− breast cancer patients (~22,000 patients a year in the US alone) (Fig. 9). Currently, although MMR loss is a recognized clinical diagnostic for many cancer types56,57, existing diagnostic assays either quantify complete absence of MMR proteins using immunohistochemistry, or genomic instability by assessing tumor mutation burden or microsatellite instability56. However, these routine clinical diagnostics in their current iteration neither detect sub-clonal MLH1 mutations nor classify aberrant cytoplasmic localization of MLH1. Sub-clonal MLH1 mutations are unlikely to alter genomic instability of the tumor sufficiently to be detected in assays that sequence a bulk tumor sample, and sub-cellular localization of MLH1 is not currently quantified by immunohistochemical diagnostics. However, results presented above suggest that aberrant cytoplasmic localization of MLH1 may serve as a refined prognostic marker of resistance to first-line standard endocrine therapies, which along with complete absence of MLH1 could stratify 20–30% of ER + /HER2− breast cancer patients for personalized therapeutic combinations. Given the common use of MLH1 immunohistochemistry as a clinical diagnostic, it is feasible to propose the extended use of this assay to detect cytoplasmic localization of MLH1 in addition to total protein levels. In the absence of standardized guidelines for the evaluation of cytoplasmic MLH1 expression, a provisional scoring system may be proposed to categorize tumors dichotomously. Tumors could be classified as cMLH1-negative (score 0) when there is a complete absence of cytoplasmic staining, and as cMLH1-positive (scores 1–3) when cytoplasmic staining is detectable in ≥10% of tumor cells. Among positive cases, a score of 3 corresponds to strong and diffuse cytoplasmic expression in >90% of tumor cells, whereas scores 1 and 2 reflect lower or intermediate staining levels. This system provides a framework for incorporating cMLH1 assessment into diagnostic and predictive workflows. However, further refinement is needed to determine the biological and clinical relevance of concurrent nuclear and cytoplasmic MLH1 expression. Future studies should investigate whether these distinct subcellular localization patterns confer differential predictive value, particularly in the context of response to endocrine therapy combined with CDK4/6 inhibition. Given that endocrine treatment resistance is a critical clinical problem affecting almost a third of women who are diagnosed with ER + /HER2− breast cancer across the world, such efforts would appear warranted. Working model. 11% of ER + /HER2− breast cancer patients have cytoplasmic MLH1 at diagnosis. These patients are at increased risk of eventual resistance to endocrine therapy and, therefore, disease relapse. However, administering a combination of CDK4/6 inhibitors along with endocrine therapy in the primary setting may prevent the outgrowth of resistant subpopulations of cells and result in favorable outcomes. Schematic made using BioRender https://BioRender.com/qpzcjqj. We have previously reported that the somatic loss of mismatch repair proteins observed in breast cancer patients often does not correlate with the traditional indicators of reduced MMR efficacy, such as microsatellite instability, which is typically seen in classical Lynch syndrome cancers like colorectal cancer. This is in accordance with multiple studies in the literature suggesting that residual levels of MLH1 are sufficient to maintain MMR efficacy but disrupt cell cycle functions58–60. Accordingly, somatic MLH1 loss, which is commonly found in 12–15% of ER + /HER2− patients but does not associate with either high mutation burden or microsatellite instability, renders them resistant to standard endocrine therapy, while making them responsive to CDK4/6 inhibitors45. Our data clearly indicate that cytoplasmic MLH1, which is observed at a similar frequency to complete MLH1 loss, exhibits comparable behavior in determining treatment responsiveness. However, due to incomplete understanding of the biological implications of cytoplasmic MLH1, it has not emerged as a clinically actionable diagnostic marker. In the context of MLH1 ATPase domain mutations, as well, previous studies have focused on the impact of these mutations in binding to damaged DNA and other downstream repair proteins have failed to consider impacts on subcellular localization of MLH1 and cell cycle regulation, emphasizing the importance of modeling mutations identified in patient tumors in experimental model systems that reflect the complete cellular context without focusing exclusively on DNA repair outputs61,62. Considering that the increase in cytoplasmic MLH1 enhances sensitivity to alternative therapeutic strategies, such as CDK4/6 inhibition, there is a compelling rationale to further validate these findings in larger datasets, across various breast cancer subtypes and different cancer types. The question of sub-clonal persister-like populations as seeds of endocrine treatment resistance and recurrence in ER + /HER2− breast cancer is a research area of intense focus. Previous, seminal studies have identified mutations in ESR1, the ER gene, and ERBB2, the HER2 gene as drivers of endocrine treatment resistance and disease recurrence63–66. In both cases, mutational fraction of these genes is low and often sub-clonal in primary ER + /HER2− breast cancer patients at diagnosis65,67 but these sub-clones can grow out in the presence of endocrine treatment to eventually cause recurrence, which is almost invariably fatal68. The importance of identifying these mutations early on has been emphasized by the presence of existing targeted therapies (CDK4/6 inhibitors in the case of ESR1 mutations67, and HER2 inhibitors in the case of ERBB2 mutations69) that could prevent recurrence and breast cancer-related death for these patients. Unfortunately, in both these cases the initial VAF is so low that it is undetected, or it is posited that the mutations occur de novo during endocrine therapy administration70. Mutations in MLH1 have a similar caveat in terms of clinical utility as they are also sub-clonal contributing to low VAF at diagnosis. In addition, mutations in MLH1 are often unknown and missense which makes determining their pathogenicity and clinical utility challenging. Assessing the functional consequences of individual MLH1 mutations is complex due to the multifaceted nature of the MLH1 protein. With its ability to bind multiple protein partners, including itself, form complex interactions with DNA, possessing both ATPase and endonuclease activities when dimerized with a nucleolytic MutL homolog, missense mutations in MLH1 can lead to various outcomes and impact its many functions differently. These range from protein instability to loss of specific functions associated with this multitude of activities15,71–77. We show here that missense mutations in the ATPase domain of MLH1, previously considered variants of unknown significance, may in fact impact the ability of MLH1 to regulate the cell cycle in breast cancer cells. We previously showed that complete loss of MLH1 delinks cell cycle regulation in ER+ breast cancer cells from their primary growth stimulus, serum estradiol, by preventing Chk2 activation7. This consequently results in resistance to endocrine treatments that work by targeting the estrogen-ER signaling axis7. We build on this previous finding here to show that even with abundant MLH1, aberrant cytoplasmic localization of the protein significantly decreases activation of Chk2 in response to endocrine therapy. Concordantly, adding a Chk2 activator (DIM) to cells with cytoplasmic MLH1, or inhibiting downstream effectors such as CDK4/6, rescues endocrine therapy responsiveness. Many studies in the literature have identified CHEK2 germline variants as specifically associating with increased incidence of ER + /HER2− tumors, but not other breast cancer subtypes37,78. MLH1 is necessary for efficient phosphorylation of Chk2 in response to DNA damage. In breast cancer cells, DNA damage activates the MMR pathway and promotes formation of protein scaffolds at damage sites, which help stabilize, activate, and amplify ATM/ATR-dependent signaling to Chk1 and Chk2. MLH1, in complex with PMS2, contributes to this signaling cascade—likely by facilitating recruitment or stabilization of ATM at damage sites79,80. Our data further support a unique role for Chk2 in mediating the cell cycle effects of estrogen stimulus for treatment responsiveness of luminal tumors. The downstream effect of loss of Chk2 activity is dependence on G1 cyclin-dependent kinases CDK4 and CDK6 for sustaining proliferation, and therefore a therapeutic opportunity for the use of CDK4/6 inhibitors. CDK4/6 inhibitors in the metastatic setting have demonstrated significant clinical benefit in ER + /HER2− patients32. However, the benefit of CDK4/6 inhibition as an adjuvant treatment in early-stage ER + /HER2− patients remains conflicting32,81–84. There is consensus within the cancer research community that adjuvant use of CDK4/6 inhibitors, which would greatly benefit a subset of patients, is only achievable with the identification of predictive biomarkers that can be assessed at diagnosis31,33,85,86. Currently, there are no specific predictive biomarkers for adjuvant use of CDK4/6 inhibitors, and clinicians rely on more generic indicators of proliferation, such as high Ki67, to guide decision-making on the administration of CDK4/6 inhibitors. However, the inaccuracy of Ki67 as a biomarker for administering CDK4/6 inhibitors to patients with ER + /HER2− breast cancer is widely acknowledged87,88. Given that immunohistochemistry-based staining for MLH1 is routinely used in the clinic for colorectal and endometrial cancer patients89,90, it may well prove feasible to administer this diagnostic for breast cancer patients along with standard immunohistochemical tests to characterize ER, PR and HER2 status. The data presented here provide rationale to initiate clinical trials using immunohistochemistry-based assessment of cytoplasmic MLH1 as a companion diagnostic to test the combinatorial efficacy of CDK4/6 inhibitors and endocrine therapies as first-line treatment in a substantive subset of ER + /HER2− patients. The use of aberrant cytoplasmic localization of MLH1 along with complete loss of MLH1 as a predictive marker for responsiveness to CDK4/6 inhibition has the potential to significantly improve breast cancer patient outcome for as many as 30% of patients likely to be resistant to current standard of care. Methods All research reported in this study was approved by the Institutional Review Board (IRB) and Institutional Animal Care and Use Committee (IACUC) of SBP and SDSU. Cell lines, mice, si/shRNA transfection and growth assays Cell lines were obtained from ATCC (2015) and maintained and validated as previously reported. Mycoplasma tests were performed on parent cell lines and stable cell lines in December 2023 with Lonza Mycoalert Plus Kit (Cat# LT07-710) as per the manufacturer’s instructions. Transient transfection with mutant MLH1 constructs (purchased from TWIST Biosciences) and siRNA against CDK4/6 (purchased from Sigma-Aldrich) was conducted using JetPrime PolyPlus transfection reagent. Stable cell lines were maintained in the presence of specified antibiotics at recommended concentrations. Knockdown and expression were validated using western blotting. Growth assays were conducted using CytoSmart cell counter. Growth assay results were plotted as fold change, normalized as specified. Three-dimensional growth assays were conducted over 6–8 weeks with twice-weekly drug treatments using standard protocols91. Images were captured when colonies were established and then treatment was administered, with images taken again every week. Fold change in area of colonies was calculated over time and represented as percent growth. For mouse experiments, all animals were housed in sterile cages and maintained under pathogen-free conditions. Mice were housed in four per cage in individually ventilated cage systems. Relative humidity in the rodent rooms is monitored and recorded daily by animal technicians. The animal facility was kept at a relative humidity of 30–70%, within the acceptable range for rodents. A standard 12-h light/12-h dark diurnal cycle was used, with the room temperature maintained at 72 °F. Tumor growth assays in vivo were carried out by injecting 5–7 × 106 MCF7 cells into the L4 mammary fat pad/mouse. Mice for MCF7 experiments were 4–6-week athymic nu/nu female mice (from Sanford Burnham Prebys animal facility) or (from Jackson laboratory). Tumors were randomized for treatment at ~100–150 mm3 volume for MCF7 xenografts. Mice were euthanized using CO₂ asphyxiation followed by cervical dislocation and tumors were harvested at <2 cm diameter and were embedded in paraffin blocks, OCT and snap frozen. All mouse experiments were performed in compliance with all relevant ethical regulations for animal testing and research, and all experiments conducted in the study received approval from Sanford Burnham Prebys (IACUC # 21-047) and San Diego State University (IACUC # 23-087). Inhibitors and agonists All drugs were maintained as stock solutions in DMSO, and stock solutions were stored at –80 and working stocks at –20° unless otherwise mentioned. 4-hydroxy tamoxifen (cat# H7904) and fulvestrant (cat# I4409) were purchased from Sigma-Aldrich, and stocks were diluted to 10 mmol/L working stocks for all experiments other than dose curves, where specified concentrations were used. For all experiments, cells were treated 24 h after plating, and thereafter every 48 h until completion of the experiment. For mouse xenograft experiments, fulvestrant concentrations of 250 mg/kg body weight were used. Beta-estradiol was purchased from Sigma-Aldrich (cat# E8875), maintained in sterile, nuclease-free water, and diluted to obtain 10 mmol/L stocks for all experiments. For mouse xenograft experiments, 17 β-estradiol was added to the drinking water twice a week at a final concentration of 8 μg/mL (cat# E2758; Sigma). CHK2 activator (3, 3;-diindolyl methane, cat# sc-204624; Santa Cruz Biotechnology) was used at 100 μmol/L concentrations. Abemaciclib (LY2835219, cat# S7158; SelleckChem) and palbociclib (PD-0332991, cat# S1116; SelleckChem) were used at 2.5 μmol/L final concentrations for cell line assays and at 70 mg/kg/day in chow for tumor growth assays as described previously. Flow cytometry, immunostaining and microscopy Flow cytometry was conducted to determine differential growth effects caused by presence or absence of estrogen (E2) for cells harboring WT or mutant MLH1 when grown in co-culture. MCF7 cells with G55V mutation that were positive for GFP were plated at a 1:9 ratio with WT MLH1 MCF7 cells. Cells were analyzed at D0 and D7 post E2 or no E2 treatment. On D0 and D7 cells were trypsinized, washed and resuspended in 1x PBS + 2% FBS supplementation with DAPI as a viability marker and stored on ice until analyzed. GFP+ and GFP- populations of live cells were determined by flow cytometry using Aurora 5-laser Full Spectrum Analyzers (Aurora 5 L). Immunostaining was performed based on manufacturer’s instructions. Cells were washed in 1x PBS; fixed for 20 min at room temperature in 4% PFA; blocked for 1 h at room temperature in 5% goat serum and 1% Triton X-100 in 1x PBS; incubated with primary antibody overnight at 4° in 1% goat serum and 1% Triton X-100 in 1x PBS antibody diluent; incubated with secondary antibody in diluent for 1 h at RT; and then mounted with DAPI-containing mounting media (cat# P36935). Primary antibodies used for immunostaining of cells and PDX sections include MLH1 (Millipore sigma; cat# M8320; 1:200), pChk2 (cat# 2197), PCNA (cat# 2586) and Ki67 (cat# 9027). Fluorescent images were captured with an ECHO microscope and quantified with ImageJ. MLH1 and PMS2 immunohistochemistry were performed using the ES05 and EP51 clones, respectively, on an automated staining system (Dako Omnis, Agilent, Santa Clara, CA, USA). Microsatellite instability testing was conducted using the automated Idylla MSI Test (Biocartis NV, Mechelen, Belgium). qPCR and Western blotting 2 × 105 cells were seeded overnight per well in 6-well plates. Following treatment with fulvestrant 1 µM for 36 h, total RNA was extracted using RNeasy Mini Kit (Qiagen, Hilden, Germany). cDNA was synthesized using oligo(dT) and random primers (AB Bioscience, MA, USA), and qPCR analysis was performed with SYBR Green (Roche, NJ, USA). Primers for each were designed using NCBI primer blast and were ordered from Integrated DNA Technologies (IDT). 36B4 was used as internal control. Primer details are provided in supplementary table 1. Western blotting was conducted as previously described. All antibodies were purchased from Cell Signaling Technology and used at 1:1000 dilutions unless otherwise specified. Primary antibodies were incubated with the membrane overnight at 4 degrees and included MLH1 (cat# M8320), CDK6 (cat# D48S8) and GAPDH (cat# 2175, 1:5,000). Primary antibody used for PMS2 is from abcam (ab110638). Polyclonal primary MLH1 antibody used to detect truncated bands is from proteintech (11697-1-AP). 3D spheroids and sub-clonal assays For 3D spheroid assays, 4 × 103 T47D cells were plated in 96 well-plate pre-coated with 1.5% agar. For sub-clonal 3D assay, G55V (GFP) and WT T47D cells were mixed in 1:9 ratio. Spheroids were allowed to grow for 1 week followed by respective treatments for 4 to 6 weeks. For –E2/ + E2 experiments, 10 nM of estradiol was added in the +E2 condition biweekly. For all other drug assays, fresh drug or vehicle were added biweekly. Spheroid pictures were captured using ECHO microscope and quantification of size was done using image J software. For in vivo sub-clonal xenograft experiment, G55V (GFP) and WT MCF7 cells were mixed in 1:9 ratio. A total of 5 × 106 cells were injected into the left flank. Tumors were randomized for +E2 and –E2 group at ~100–150 mm3 volume. Tumor measurements were taken biweekly and bioluminescence pictures of mice were also captured using IVIS spectrum imaging system. Tumors were harvested at <2 cm diameter and were embedded in paraffin blocks, OCT and snap frozen. PDxO establishment and treatments For PDxO preparation, PDX tissue chunks were digested in a GentleMACS dissociator in warm advanced DMEM/F12 with GentleMACS human tumor dissociation enzymes and 10 μM Y-27632 added. After digestion, differential centrifugation was performed to enrich for organoids and deplete single cells. Organoids were embedded in 200-μl Matrigel domes, which were plated in six-well tissue culture plates onto a 50-μl Matrigel base layer. After a 5-min incubation period, plates were flipped, and Matrigel domes were solidified for 10 min before subtype-specific culture medium was added. For all breast cancer subtypes, 10 μM Y-27632 was added fresh to the PDxO base medium (Advanced DMEM/F12 with 5% FBS, 10 mM HEPES, 1× Glutamax, 1 μg ml–1 hydrocortisone, 50 μg ml–1 gentamicin and 10 ng ml–1 hEGF). Medium was exchanged every 3–4 days, and, once mature, cultures were passaged by incubating in dispase solution (20% FBS in dispase with Y-27632), followed by a wash step with base medium and a dissociation step in TrypLE Express. Single cells were seeded at 200,000–400,000 cells per dome. For drug treatments, Mature organoids were collected from culture using dispase treatment at 37 °C for 25 min. Five thousand to ten thousand cells (~50 organoids) were seeded per well in 96 tissue culture plates, each comprising a solidified 10 μl Matrigel base layer and 100 μl of PDxO medium supplemented with 5% (by volume) Matrigel. Media was replaced every 48 h with fresh drug. Cell titre Glo 3D was conducted to measure viability. Categorization of cytoplasmic vs nuclear MLH1 in PDXs and patient tumors PDXs were categorized as nuclear MLH1 if they had >5% cells with detectable nuclear localization of MLH1 and as cytoplasmic MLH1 if they had <5% cells with detectable nuclear localization of MLH1 but >10% of cells with detectable cytoplasmic localization of MLH1 (representative images in Fig. 1). Patient tumor biopsies were pathologically graded for cytoplasmic MLH1 on a scale of 0 to 3; Tumors are classified as cMLH1-negative (score 0) when there is a complete absence of cytoplasmic staining, and as cMLH1-positive (scores 1–3) when cytoplasmic staining is detectable in ≥10% of tumor cells. Among positive cases, a score of 3 corresponds to strong and diffuse cytoplasmic expression in >90% of tumor cells, whereas scores 1 and 2 reflect lower or intermediate staining levels. Tumor biopsies were categorized as pMLH1 if they had >5% cells with detectable MLH1 and dMLH1 if they had <5% cells with detectable MLH1. Statistical analysis Statistical significance between two groups was assessed by the unpaired two-sided Student’s t-test. All statistics were derived from at least three biological replicates as three independent experiments. For tumor growth rates, slopes were calculated and compared for statistical differences. For categorical data, a Fisher’s Exact or a Pearson’s Chi-squared test was used as indicated based on sample size. All experiments were conducted in triplicate, and each experiment was duplicated ≥2 times. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements Funding: This work was supported by the Komen CCR18548157, ACS RSG-22-094-01 and R37-CA270362-02 (to SH.); the California Institute for Regenerative Medicine, EDUC4-12813 (to AM.); T32 CA211036-05 (to JD.); NIGMS R35GM142651(to CM.); (KV) was supported by the Fondazione Umberto Veronesi; This work was partially supported by the Italian Ministry of Health through Ricerca Corrente 5 × 1000 funds; the Italian Ministry of Innovations via the Sustainable Growth Fund—Innovation Agreements under the Ministerial Decree of December 31, 2021, and the Director’s Decree of November 14, 2022 (2nd Call), Project No: F/350104/01-02/X60 and the Fondazione Umberto Veronesi (to NF.). Author contributions Data collection, curation, and analysis: A.M., J.D., E.O., N.P., D.L., M.R., J.P., E.S., E.G-R., K.V., M.I., E.M., M.D., A.C., N.F., C.M., M.B., and S.H. Conceptual oversight: A.M. and S.H. Preparing manuscript: A.M., J.D. and S.H. Writing and editing of manuscript: A.M., J.D., J.P., C.M., M.B. and S.H. Peer review Peer review information Nature Communications thanks Rachael Natrajan, Alessandro Sartori and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability Source data is provided with the manuscript. The IEO dataset analyzed in this study was previously published by (Fusco et al.)1. The NeoPalAna dataset was previously published by (Ma et al.)52. The NCT3007979 was previously published by (Krishnamurthy et al.)55. Original immunofluorescence and immunohistochemistry images associated with the manuscript are available upon request from the corresponding author. Source data are provided with this paper. Competing interests (E.G-R.) has received advisory fees, honoraria, travel accommodation and expenses, grants and/or non-financial support from AbbVie, AstraZeneca, Exact Sciences, GE HealthCare, GSK, Illumina, Lilly, Novartis, MSD, Pfizer, Roche, Sophia Genetics, Stemline (Menarini), Thermo Fisher Scientific. (KV) from MSD, Roche, AstraZeneca, and Johnson & Johnson. (MI) from Agilent Technologies and (NF) has received honoraria for consulting/advisory roles, participation in speaker bureaus, research support, and/or travel support from AbbVie, Alira Health, AstraZeneca, Daiichi Sankyo, Epredia, Exact Sciences, Gilead, GSK, Leica Biosystems, Lilly, Menarini Group, Merck, MSD, Novartis, Owkin, Pfizer, Pillar Biosciences, Roche, Sakura, Sysmex, ThermoFisher, and Veracyte. Rest of the authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-025-67257-8. References Associated Data Supplementary Materials Supplementary Information Reporting Summary Transparent Peer Review file Source Data Data Availability Statement Source data is provided with the manuscript. The IEO dataset analyzed in this study was previously published by (Fusco et al.)1. The NeoPalAna dataset was previously published by (Ma et al.)52. The NCT3007979 was previously published by (Krishnamurthy et al.)55. Original immunofluorescence and immunohistochemistry images associated with the manuscript are available upon request from the corresponding author. Source data are provided with this paper.