Nucleocytoplasmic transport of active HER2 causes fractional escape from the DCIS-like state

To go beyond B2B1 cells and ErbB activation with synthetic ligands, we sought a very different cancer-relevant setting to test the role of nucleocytoplasmic transport. We conjectured that AP heterodimerization of ErbBs resembles the early response of HER2+ cancers that have become resistant to ErbB-targeted therapeutics. This context was mimicked by examining release from cytostatic doses of the dual ErbB inhibitor, lapatinib, in TM15 clone 6 (TM15c6) cells derived from an Erbb2-amplified tumor of a PMMTV-Cre;Erbb2+/V660E knockin mouse29,30,31. 3D cultures of TM15c6 cells were seeded in the presence of 2 µM lapatinib and grown for 6 days before lapatinib was washed out and outgrowth trajectories assessed. TM15c6 cultures did not exhibit DEs with all-or-none penetrance like those of nontransformed B2B1 cells, but lapatinib release caused significant decreases in circularity among the upper 35th percentile of outgrowths by size (Fig. 3a). Circularity is a scaled area-to-perimeter ratio that does not depend on geometric size, and we saw little evidence of any confounding area-circularity relationship in TM15c6 cultures for up to 5 days after lapatinib release (Supplementary Fig. 3a). These data established a foundation for genetic perturbations of Cse1l in TM15c6 cells.

We began with inducible ectopic expression of human CSE1L in TM15c6 cells (Supplementary Fig. 3b), using the same experimental paradigm as in B2B1 cells but substituting lapatinib release for AP addition. Relative to the pre-release circularity of each genotype, there was a reduction in circularity when CSE1L was co-induced with the washout of lapatinib (Fig. 3b). The CSE1L result was not heavily dependent on the upper percentile of outgrowths considered (Supplementary Fig. 3c). Reciprocally, inducible knockdown of Cse1l severely attenuated the loss of circularity from lapatinib release, and this phenotype was reverted upon co-induction of human CSE1L as an RNAi-resistant addback (Fig. 3c, d and Supplementary Fig. 3d). Therefore, acute reactivation of ErbB signaling in TM15c6 mammary carcinoma cells elicits a Cse1l-dependent multi-cell phenotype that is reminiscent of DEs.

DCIS escape in vivo entails many additional barriers that are not reflected in 3D culture32. As a surrogate, we intraductally administered TM15c6 cells in a paired design comparing the macroscopic tumor characteristics of inducible shCse1l to contralateral control33 after release from cytostatic doses of lapatinib (see Methods). Cse1l knockdown did not discernibly influence intraductal colonization or lesion stasis caused by 6 days of lapatinib treatment (Supplementary Fig. 3e). At 2 weeks after lapatinib release, tumor bioluminescence was comparable between groups and both genotypes had invaded the stroma by this time. However, we posited that the tortuosity of the final tumors would reflect DE characteristics of TM15c6 cells at earlier stages. When glands were excised and tumors imaged ex vivo for higher resolution assessment of macroscale organization, we found that shCse1l tumors were consistently more circular than paired controls (Fig. 3e). Across the cohort, tumor circularity did not relate significantly to tumor mass (Supplementary Fig. 3f), excluding differences in net growth as a confounder. The TM15c6 study—with its change in species (human to mouse), ErbB activator (AP to lapatinib release), and cellular context (in vitro to in vivo)—strengthens the general conclusion that CSE1L/Cse1l drives DE-like phenotypes at the onset of persistent ErbB signaling.

CSE1L-associated network states inhibit import of classical NLSs

The results with CSE1L, NUP37, NPIPB11, and KPNB1 in B2B1 cells (Fig. 2) justified examining other known regulators of nucleocytoplasmic transport. The pleiotropic exportin XPO1 was detected as a biologically variable transcript by stochastic profiling in AP-treated cells but at a reduced expression frequency and with bulk changes that were muted compared to CSE1L (Supplementary Fig. 4a and Supplementary Data 3). Although not measurably heterogeneous, the GTPase RAN and its binding partners (RANBP1 and RANBP3) were retained or elevated upon prolonged ErbB heterodimerization (Supplementary Fig. 4b–d). Conversely, factors regulating RanGTP hydrolysis (RANGAP1) and RanGDP exchange (RCC1 and NUTF2) were not significantly altered, but independent bulk replicates of B2B1 cultures exhibited surprising dispersion (Supplementary Fig. 4e–g). We clustered the bulk transcript profiles by replicate and noted covariations among KPNA2–KPNB1, CSE1L–XPO1, and RANBP1 that were nominally offsetting and thus difficult to interpret (Supplementary Fig. 4h). These observations suggested a need to examine nucleocytoplasmic transport at the systems level to determine the aggregate effect of the changes observed.

Drawing from earlier efforts34,35, we built a multicompartment computational model encoding nuclear import and export modules with the Ran transport cycle (Fig. 4a). A complete description of the model construction, scope, and parameter estimates is available in Supplementary Note 1 and Supplementary Data 4. For consistency, the model retains common-name designations from before34,35: CAS for CSE1L, Imp-α as a lumped parameter for KPNA1 through KPNA7, Imp-β for KPNB1, NTF2 for NUTF2, and CRM1 for XPO1. Changes in model parameters were not permitted unless a typographical error was discovered or the original parameter could not be definitively identified (Supplementary Note 1). To consider rate-limiting roles for NUP37 or NPIPB11, the model was elaborated with two classes of nuclear pores, one of which was selective to proteins harboring specific classes of nuclear localization sequences. However, this change proved inconsequential given the estimated number of nuclear pores per cell (Supplementary Note 1). The abundance of all protein species in the model was calibrated to B2B1 cells by quantitative immunoblotting against HeLa cells used in the prior work34,35 to arrive at a draft model for validation (Supplementary Note 1 and Supplementary Data 4).

We tested the fidelity of the calibrated model by designing a series of inducible tandem cargo reporters harboring fluorescence proteins, epitope tags, and recognized sequences for import and export (Supplementary Fig. 4i). Reporters were individually induced at a few million copies per B2B1 cell and triply localized by immunostaining to mitigate the effect of cleavage products; then, images were aggregated and segmented for estimating nuclear-to-cytoplasmic (N/C) ratios at steady state in single cells (Supplementary Fig. 4i–k). All reporters contained an offsetting nuclear export sequence (NES) to distinguish the potency of each class of nuclear localization sequence (NLS)—a monopartite NLS from simian virus 40 T antigen (NLSSV40Tag), a bipartite NLS from Cbp80 (2xNLSCbp80), or the Imp-β-binding domain of KPNA2 (IBBKPNA2). NLSSV40Tag and 2xNLSCbp80 are recognized by Imp-α:Imp-β dimers with different affinity (KD,SV40Tag ~ 32 nM36, KD,Cbp80 ~ 2.5 nM37), whereas IBBKPNA2 is imported after binding Imp-β alone35 (KD,IBB ~ 6.5 nM36). We found that the tandem cargo reporters yielded different predictions of N/C ratio in the model, and these predictions compared favorably to measured values averaged across the population (Fig. 4b–d and Supplementary Fig. 4l). Because no recalibration was needed for any tandem cargo reporter, the experimental tests validated the model for provisional use.

Next, we predicted the steady-state N/C ratio as a readout of transport capacity for the different cargo classes when species abundances (initial conditions) of the model were perturbed combinatorially. Biologically realistic combinations were gleaned from the transcriptomic data collected among single B2B1 outgrowths during stochastic profiling (Fig. 4e). We found that presence of an NES had very little impact on the change of any reporter, consistent with its somewhat weaker potency38 and the limiting abundance of the CRM1 cofactor RanBP339 in B2B1 cells (Supplementary Note 1). The predicted outgrowth-to-outgrowth patterns of NLSSV40Tag and 2xNLSCbp80 were also quite similar, owing to their same connectivity in the model (Fig. 4a, e). A cluster of perturbations with high predicted N/C ratio for these reporters was notable (Fig. 4e, dashed), and thus we stratified at the upper 35th percentile to ask whether any initial conditions were significantly associated. The analysis identified a significant positive association with Imp-α, but not Imp-β, and a negative association with CAS/CSE1L (Fig. 4f–h). By contrast, IBBKPNA2 engaged the network differently35 and showed its own sensitivity to combinatorial perturbations, including a positive association for Imp-β and a negative association with CRM1/XPO1 (Supplementary Fig. 4m, n). The modeling results suggested a natural antagonism between karyopherins that depended on the type of cargo transported.

CSE1L–NUP37 proximity interactomes contain ErbBs

It was centrally important to identify cargo that could be directly relevant to the DE phenotype. Because transport interactions are very transient, we developed a proximity-labeling approach focused on CSE1L and NUP37—two confirmed effectors that strongly perturb DE penetrance (Fig. 2a, c). Although CSE1L and NUP37 transcripts are coexpressed broadly, there are no reports of binary or tertiary interactions involving these proteins. Thus, novel proximity interactions shared by CSE1L and NUP37 should strongly suggest conduits for the DE phenotype.

We scaled up from 3D culture to 15-cm plates by growing B2B1 cells as domes overlaid with dilute matrigel, a format that mimics 3D transcriptional changes for up to 10 days40. Lines expressing CSE1L or NUP37 fused to a promiscuous biotin ligase (BirA*) were moderately induced for 24 h (Supplementary Fig. 5a, b), followed by heterodimerization for 24 h and then biotin labeling for 24 h (Fig. 5a). Due to the overlay format, solubilized extracts contained many adsorptive extracellular matrix components that required a more-stringent two-step purification of biotinylated target proteins (Fig. 5a). Sequential selection with anti-biotin and streptavidin yielded many proximity interactions detectable by mass spectrometry (MS) after SDS-PAGE purification of eluted products (Fig. 5b, c and Supplementary Data 5). The MS data contained multiple known interactors for CSE1L and NUP37 previously identified by co-immunoprecipitation, crosslinking, or yeast two-hybrid (Fig. 5c). We specifically confirmed literature-derived interactions by immunoblotting anti-biotin immunoprecipitates from independent samples: KPNA2, RAN, and p53 for CSE1L (Fig. 5d, e) and Sec13, NUP160, and KPNA2 for NUP37 (Fig. 5f, g). Proteins sequestered from CSE1L or NUP37 were not detectably biotinylated, verifying the approach.

For uncovering novel proximity interactions in the MS data, we used a contaminant repository41 to identify proteins that rarely appear in BirA* datasets (Fig. 5c and Supplementary Data 5). Among the dozens of proteins that passed the criterion, roughly half overlapped between the CSE1L and NUP37 datasets (P < 10−13 by hypergeometric test). Strikingly, both studies identified EGFR, ERBB2/Erbb2, and the EGFR endocytic-trafficking adaptor SH3KBP142 as stringent proximity interactions (Fig. 5c and Supplementary Data 5). We independently verified the MS result by immunoblots for the EGFR and Erbb2 chimeras, quantifying extents of proximity labeling in the range of known interactions (Fig. 5d–g). CSE1L was previously reported in MS data from EGFR immunoprecipitates43; we confirmed this interaction but found that Erbb2 was labeled much more efficiently than EGFR (Fig. 5d, e). The same was not true in B2B1 cells expressing BirA*-NUP37, where EGFR and Erbb2 were labeled equally well (Fig. 5f, g). This ruled out differences in chimera abundance as an explanation for the Erbb2-biased CSE1L result. CSE1L strongly increases DE penetrance (Fig. 2a), and EGFR homodimers lacking Erbb2 do not elicit a DE phenotype9. The data suggested that nucleocytoplasmic transport of Erbb2-containing dimers themselves might dictate penetrance of the phenotype they evoke.

DE penetrance is inhibited by ErbB trafficking from the plasma membrane

Proximity labeling of EGFR–Erbb2 by BirA*-CSE1L or BirA*-NUP37 was not detectably altered by AP (Fig. 5d–g), excluding transport-protein binding as the regulatory step for ErbB-induced DEs. After ligand activation at the plasma membrane, ErbBs are internalized within a few minutes and can traffic to the nucleus in less than a half an hour44. We cultured B2B1 cells on coverslips and observed transient nuclear accumulation of colocalized EGFR–Erbb2 after 15 and 30 min of heterodimerization, which subsided at 1 h and disappeared at 4 h as the receptor was downregulated (Fig. 5h). Interestingly, after recovery in the presence of AP for 24 h, we quantified sustained EGFR–Erbb2 immunoreactivity in the nucleus that was nearly as intense as at early times after stimulation. This long-term nuclear localization coincided with the DE decision point (Fig. 1d, e), raising the possibility it could be relevant to phenotype penetrance. We tested the impact of ErbB internalization by mildly reducing clathrin heavy chain (CLTC) in B2B1 cells to impair clathrin-mediated endocytosis (Supplementary Fig. 5c). CLTC knockdown did not alter abundance of the ErbB chimeras, but CLTC-deficient cells exhibited significant increases in the frequency of DEs (Fig. 5i and Supplementary Fig. 5d), indicating that internalized EGFR–Erbb2 complexes are repressive.

EGFR and Erbb2 chimeras are recognized as classical Imp-α:Imp-β cargo

Recruitment of ErbB family members to the nucleus occurs when a polybasic juxtamembrane segment within the ErbB sequence acts as a tripartite NLS for transport44,45. We returned to the systems model of nucleocytoplasmic transport (Fig. 4a) and asked how steady-state transport of different cargoes depended on the abundance of proteins in the network (Fig. 6a–e and Supplementary Fig. 6a–f). For the B2B1 model, NES-harboring cargoes were distinguished only when RanBP3 abundances were very high (Fig. 6a), allowing us to ignore consideration of ErbB–CRM1 interactions46 in this setting. Further simplifying the analysis, we found that the sensitivities of monopartite (NLSSV40Tag) and bipartite (2xNLSCbp80) NLS cargoes were identical except for an offset in N/C ratios from differences in affinity for Imp-α:Imp-β (Supplementary Data 4). By contrast, IBBKPNA2 cargo was more sensitive to CAS/CSE1L and Ran abundance, and transport qualitatively differed in response to the abundance of Imp-α and Imp-β (Fig. 6b–e). ErbB-family receptors bind Imp-β46 but evidence for or against interactions with Imp-α was lacking. Systems-level analysis of ErbB nucleocytoplasmic transport thus required clarification of whether EGFR–Erbb2 heterodimers acted as classical NLS-like cargo or instead as non-classical IBB-like cargo19.

To verify EGFR–Erbb2 interactions with Imp-β and test for interactions with Imp-α, we probed for epitope-tagged karyopherins in EGFR and EGFR–Erbb2 immunoprecipitates. B2B1 cells modestly expressing inducible KPNA2 (115 ± 10% of endogenous) or KPNB1 (80 ± 9% of endogenous) were probed for epitope tag after immunoprecipitation of the EGFR chimera with or without prior AP treatment. We observed robust increases in co-immunoprecipitating Erbb2 chimera upon heterodimerization, as expected, and both KPNA2 and KPNB1 were consistently detectable in all immunoprecipitates (Fig. 6f, g). After scaling for target enrichment, there was no discernible difference in the level of KPNA2 vs. KPNB1 co-immunoprecipitation (Fig. 6g), suggesting the two karyopherins bind with roughly equal stoichiometry.

As a complementary approach, we evaluated potential interactions by proximity ligation assay (PLA). We confirmed assay specificity with positive and negative controls, as well as heterodimerizer-regulated proximity of EGFR and Erbb2 chimeras (Fig. 6h and Supplementary Fig. 6g, h). Above-background proximity ligation was observed for all ErbB-karyopherin pairs, with some signals approaching the saturated spot detection of the KPNA2:KPNB1 positive control (Fig. 6i, j). PLA spots often increased after extended heterodimerization from the increased stability of ErbB chimeras with AP (e.g., Fig. 5d). Together, we conclude that ErbBs behave like classical NLS cargo by binding an Imp-α:Imp-β heterodimer.

Nuclear ErbBs engage a miR-205–KPNA1 negative-feedback circuit

Retrograde ErbB transport to the nucleus has been delineated mechanistically, but its consequences are multifaceted and contextual7. Nuclear ErbBs require their tyrosine kinase activity47 and are reported to act as transcriptional coactivators48 and chromatin-binding factors44,47,48. Therefore, we asked whether specific genomic loci were bound constitutively or inducibly to EGFR–Erbb2. By confocal imaging of 3D cryosections, we found EGFR–Erbb2 colocalized in the nucleus, and nuclear immunoreactivity increased significantly with the addition of AP for 24 h (Fig. 7a and Supplementary Fig. 7a). To preserve this 3D context as best as possible, input chromatin was prepared from B2B1 cells cultured in the overlay format for 1 week ± AP added on the sixth day. ErbB chimeras were immunoprecipitated by their respective C-terminal epitope tags and processed for chromatin immunoprecipitation sequencing (ChIP-seq) to identify enriched loci above naive IgG controls. For two different anti-epitope antibodies against the Erbb2 chimera, we immunoprecipitated comparatively little chromatin and were unable to detect any reliable enrichment peaks. This result is consistent with previous work involving ERBB2-amplified cells48 and suggests that the C terminus is mostly obscured in the nucleus after fixation. ChIP-seq of B2B1 samples was successful when immunoprecipitating EGFR chimeras, and we recurrently identified enriched binding peaks for 26 loci without AP and 51 loci with AP, 17 of which were common to both conditions (Fig. 7b and Supplementary Data 6).

As reported49,50, EGFR occupancy often occurred along the gene body of highly expressed genes (TXNIP, NEAT1) and transcription factors of the AP1 (JUNB) and KLF (KLF5) families. The AP-independent loci contained multiple genes previously linked to 3D cell-to-cell heterogeneity in MCF10A-5E cells: KRT5, NFKBIA, and SOD220,40,51. These transcripts showed various time- and AP-dependent patterns of change, but none had been nominated as DE candidates by the earlier frequency-matching approach (Supplementary Fig. 7b–d and Supplementary Data 3). Instead, our attention turned to the MIR205HG locus encoding miR-205 (Fig. 7b), an epithelial microRNA frequently misregulated in cancer52. Although the earlier transcriptomic data could not be used to analyze miRNA abundance, miR-205 is known to be reduced by chronic ERBB2 signaling through changes in DNA methylation53. We confirmed miR-205 downregulation in B2B1 cells after 3 days of AP treatment but noted that the change was time dependent, reversing and becoming more variable when measured 2 days earlier (Supplementary Fig. 7e). Further, miR-205 abundance significantly increased when B2B1 cells were acutely heterodimerized in overlay cultures that improve uniformity of the response (Fig. 7c). As with other loci, EGFR occupancy covered the MIR205HG promoter and length of the gene body (Fig. 7d). We used ChIP-qPCR to confirm specificity and AP-independence of binding (Fig. 7e); presumably, effects on the MIR205HG locus change when bound EGFR is co-associated with active Erbb2 (Fig. 7c). Analogously, EGFR–Erbb2 heterodimers strongly elicit a DE phenotype, whereas EGFR homodimers do not9,10.

miR-205 posttranscriptionally regulates many genes52, but one compelling target is the Imp-α family member KPNA154. The 3' UTR of human KPNA1 contains three predicted55 binding sites for miR-205, the most proximal of which is broadly conserved, and we confirmed that miR-205 suppresses endogenous KPNA1 (Fig. 7f and Supplementary Fig. 7f)54. By immunofluorescence, KPNA1 protein expression was highly mosaic among single cells in B2B1 outgrowths (Fig. 7g). Importantly, KPNA1 protein abundance declined significantly upon addition of AP heterodimerizer without detectable changes in KPNA1 mRNA, implicating a posttranscriptional mechanism (Fig. 7h and Supplementary Fig. 7g). When B2B1 cells were engineered with inducible ectopic KPNA1 lacking the native 3' UTR targeted by miR-205, we found that heterodimerizer-induced DEs were almost completely abolished (Fig. 7i). This result starkly contrasted earlier results with KPNA2 that showed no effect upon ectopic expression (Supplementary Fig. 2i). Imp-α isoform switching has been linked to cell fate determination in other settings56, and the miR-205–KPNA1 circuit induced by nuclear ErbBs added an isoform-specific negative feedback to the regulation of DEs.

Nucleocytoplasmic ErbBs give rise to species-specific ultrasensitivity

The earlier systems model of nucleocytoplasmic transport predicted opposite effects of Imp-α and CAS/CSE1L abundance on the steady-state accumulation of classical cargo. Elevating Imp-α (e.g., KPNA1) increased N/C ratio of cargo and decreased DE penetrance (Figs. 4f and 7i), whereas high CAS/CSE1L decreased N/C ratio and increased DE penetrance (Figs. 2a and 4h). We thus modeled internalized ErbB heterodimers as classical cargo with an unknown affinity for Imp-α:Imp-β, equating high C/N [ = (N/C)−1] ratio at steady state with the DE phenotype (Fig. 7j), presumably due to canonical signaling from cytoplasmic complexes. Imp-α and CAS/CSE1L abundances reciprocally affected C/N ratio across a wide range of binding strengths (Fig. 7j). However, model predictions were more sensitive to Imp-α than CSE1L except for very high abundances, and the changes in C/N ratio were too gradual to explain how two DE states could coexist in the same population of cells.

We next considered the two antagonistic feedbacks elaborated before: (1) Active ErbBs internalize and traffic toward the nucleus, but they also induce CSE1L, which is predicted to inhibit ErbB nuclear transport by competing for available Imp-α (Figs. 2g, 4a, and 5h). (2) In parallel, nuclear ErbB heterodimers repress KPNA1 through miR-205, thereby reducing nuclear import, exacerbating the CSE1L competition with the Imp-α that remains, and increasing canonical cytoplasmic ErbB signaling (Fig. 7f–i and Supplementary Fig. 1b and 7f). After combining transcriptional, posttranscriptional, and cell-biological circuits, the overall ErbB feedback architecture (Fig. 7k, upper) is reminiscent of a synthetic-biology toggle switch capable of mixed single-cell behavior.

To determine whether cells were operating close to a mixed regime, we added proportional feedbacks from cytoplasmic ErbBs to CAS/CSE1L (positive) and from nuclear ErbBs to Imp-α (negative; Fig. 7k, upper). Repeating the simulations with the revised feedback architecture completely changed the dependence of C/N ratio on Imp-α and CAS/CSE1L. Sensitivity to overall Imp-α levels was severely blunted, consistent with results for KPNA2 (Supplementary Fig. 2i), and the response to CSE1L became switchlike over a narrow range of expression (Fig. 7k, lower). We further noted that CSE1L abundance heterogenizes upon ErbB activation, with a range of expression that more than doubled after 24 h of AP addition (Supplementary Fig. 7h). The revised model indicates that very small initial differences in CSE1L are sufficient to yield divergent outcomes due to the nucleocytoplasmic transport properties of ErbB (Supplementary Fig. 7h, inset).