Keratin filaments and soluble vimentin interact with isoform specificity to organize cell signaling

Abstract

The keratin intermediate filament (IF) cytoskeleton provides mechanical support to load-bearing epithelial cells, but its composition varies depending on cell type and tissue context. This variation, the large size of the keratin protein family, and emerging noncanonical functions of individual keratins suggest that the keratin cytoskeleton may have additional, more complicated roles. We report a surprising interaction between the keratin IF cytoskeleton in epithelial cells and soluble oligomers of vimentin, the principal IF component in mesenchymal cells. Rather than forming a cytoskeletal network, soluble vimentin functions as a recruitment factor, organizing regulatory kinases and non-muscle myosin on keratin filaments to promote signaling. Soluble vimentin and keratin filaments interact through the intrinsically disordered head domain of each protein. Because vimentin preferentially interacts with keratin 6A-containing filaments, increasing keratin 6A expression, as occurs during skin wound healing and cancer development, triggers this myosin-activating signaling cascade. These results demonstrate how varying composition, soluble pools, and client-scaffold interactions through intrinsically disordered regions allow IFs to organize cellular signals in parallel to their canonical mechanical role, supporting context-dependent cell and tissue regulation.

Introduction

The intermediate filament (IF) cytoskeleton functions as a mechanical scaffold allowing cells to resist internal and external forces (Block et al. 2015; Kechagia et al. 2024). This property is critical for load-bearing epithelial tissues, where the keratin IF cytoskeleton is linked between adjacent cells through desmosomes (Hatzfeld et al. 2017). Mutations in many keratin genes cause diseases characterized by tissue fragility, particularly affecting the skin, which is regularly exposed to mechanical stress (Omary 2009; Toivola et al. 2015). Furthermore, in both epithelial and non-epithelial cells, the IF cytoskeleton provides mechanical protection to the nucleus and other cellular structures (Guo et al. 2025).

Belying this seemingly simple mechanical role, the IF cytoskeleton is constructed from an exceptionally large protein family. In total, the human genome codes for more than seventy distinct cytoplasmic IF proteins, fifty-four of which are keratins (Oshima 2007; Moll et al. 2008; Coelho-Rato et al. 2024). Keratins assemble as obligate heteropolymers containing at least one type-I and one type-II keratin isoform, with the specific mixture of keratins expressed varying depending on tissue, anatomic location, and cell state (Moll et al. 2008). Non-epithelial cells form their IF cytoskeletons from other IF proteins, including vimentin in many mesenchymal cells, which may form homopolymers or heteropolymers depending on the cell type (Oshima 2007; Coelho-Rato et al. 2024). Why such a large protein family with context-specific combinatorial expression is required to construct the IF cytoskeleton remains incompletely understood, and there has been considerable interest in additional non-mechanical or context-specific roles for IF proteins (Redmond and Coulombe 2020; Sjoqvist et al. 2021; Coelho-Rato et al. 2023).

IFs have now been implicated in a wide range of functions, from lineage specification and differentiation to growth control (Redmond and Coulombe 2020; Sjoqvist et al. 2021; Coelho-Rato et al. 2023). However, determining whether purported functions of individual IF proteins are separable from the more general role of the IF cytoskeleton remains experimentally and conceptually challenging. Particularly in epithelial cells, weakening the IF cytoskeleton produces adhesion and barrier defects with significant downstream consequences (Vassar et al. 1991). Mechanical weakness likely complicates interpretation of most, if not all, IF loss-of-function models, whether through gene knockouts, which deplete IF bulk, or mutations which impair IF polymerization in a dominant-negative fashion. More generally, because IF expression is cell-type dependent, perturbing a single IF gene may simply highlight the particular importance of the IF cytoskeleton to the affected cell type rather than explain why those cells construct their IF cytoskeleton from one specific IF protein instead of another.

Adding further complexity, some IF proteins are thought to function both inside and outside the IF cytoskeleton network. Predominantly cytoplasmic IF proteins have been detected in the nucleus (Coulombe et al. 2024), at the cell surface (Patteson et al. 2020), and even in cell secretions (Thalla and Lautenschlager 2023). The precise macromolecular state of these presumably non-filamentous and non-cytoskeletal IF proteins is not fully defined, nor is it known whether they exchange or interact with IF polymers in the cytoskeleton. The degree to which extra-cytoskeletal IF functions may be shared among different IF proteins also remains unclear.

The challenge of disentangling potential non-canonical IF functions is readily apparent in the context of skin wound healing, where keratinocytes increase expression of “wound-associated” keratin isoforms, keratin 6A (K6A), K6B, K6C, K16, and K17 (Weiss et al. 1984; Mansbridge and Knapp 1987; Patel et al. 2006). Similar keratin expression changes occur in other contexts with increased epidermal remodeling, ranging from psoriasis to keratinocyte carcinomas. Largely based on knockout studies, multiple functions of wound-associated keratins have been identified (Kim et al. 2006; Depianto et al. 2010; Tam et al. 2012; Lessard et al. 2013; Chung et al. 2015; Hobbs et al. 2015; Maruthappu et al. 2017; Yan et al. 2020; Nair et al. 2021; Cohen et al. 2026), but there is conflicting evidence regarding their direct contribution to re-epithelialization and keratinocyte migration (Wojcik et al. 2000; Wong et al. 2000; Wojcik et al. 2001; Wong and Coulombe 2003; Rotty and Coulombe 2012; Kroger et al. 2013; Loschke et al. 2016; Wang et al. 2018). K6 loss potentiated keratinocyte migration in some (Wong and Coulombe 2003; Rotty and Coulombe 2012; Wang et al. 2018), but not all (Wojcik et al. 2000; Wojcik et al. 2001), knockout studies. Conversely, in a broad keratin knockout-reconstitution system, re-expression of K6 and other wound-associated keratins failed to slow migration compared to re-expression of steady-state keratins (Kroger et al. 2013; Loschke et al. 2016). Notably, increased migration was linked to loss of mechanical stability and disrupted cell adhesion in both systems (Kroger et al. 2013; Loschke et al. 2016; Wang et al. 2018).

Because genetic models involve large changes in keratin expression, we reasoned that a more minimally perturbing approach could address these discrepancies. We engineered keratinocytes with tuned expression ratios of wound-associated K6A and its closest steady-state paralog, K5, while maintaining total keratin abundance (Nanes et al. 2024). In a skin organoid model, modestly increased K6A promoted keratinocyte migration and wound closure without disrupting mechanical stability and cell-cell adhesion. Rather, K6A activated a phosphorylation signaling cascade leading to increased actomyosin contractility, an effect mediated not through the structural keratin rod domain, but through the intrinsically disordered head domain (Nanes et al. 2024). These results showed that the IF cytoskeleton can influence cell behavior independently of its canonical role supporting cell and tissue stability. How changing keratin isoform composition uncoupled a signaling cascade from keratin mechanical scaffold function, a question with relevance beyond the wound healing context, remained open.

Here, we now report the mechanism by which K6A and K5 differentially organize cellular signals to activate myosin. We found that keratin filaments organized a proximity-dependent signaling module promoting myosin regulatory light chain (RLC) phosphorylation. Surprisingly, this pathway relied on a soluble pool of another IF protein, vimentin, to recruit regulatory kinases into the module. Soluble vimentin and polymerized keratin filaments interacted through the intrinsically disordered head domains of each protein. Tuning the keratin isoform mixture modulated the strength of the vimentin–keratin interaction and, therefore, activity of the pathway. These results reveal significant complexity in the ability of IF proteins to organize cellular signals in parallel to their mechanical scaffold function, supporting context-dependent regulation through variation in IF cytoskeleton composition, soluble IF proteins, and intrinsically disordered region (IDR)-mediated interactions.

Results

K6A recruits vimentin to keratin filaments

Various signaling roles for individual keratin isoforms have been postulated (Redmond and Coulombe 2020; Di Russo et al. 2023), but a true isoform-tuned signaling function implies keratin isoform-preferential interactions, not only cell state changes associated with broad IF cytoskeleton disruption. Reliable identification of keratin filament binding partners has been challenging for several reasons. First, keratins are highly expressed in epithelial cells (Sun and Green 1978), which increases the risk of identifying non-specific associations. Keratins have large intrinsically disordered regions (IDRs) (Zhou et al. 2023), further compounding the risk of non-specific binding. Conversely, epidermal keratin filaments are highly insoluble; the harsh, highly chaotropic buffers capable of completely solubilizing keratin filaments likely disrupt interactions between keratins and other proteins and denature antibodies used for immunoprecipitation (Wang et al. 2016; MilliporeSigma 2021). While methods have been proposed to immunoprecipitate keratins using milder detergents (Lowthert et al. 1995; Wang et al. 2016), it remains unclear if they reliably capture keratin filaments versus a minor population of soluble keratin monomers or smaller aggregates.

To overcome these challenges, we built upon our minimal perturbation model, with keratinocyte cell lines engineered to express modestly increased levels of different keratin isoforms, including K6A and K5 (Nanes et al. 2024). We previously confirmed that these shifts in keratin isoform mixture do not differentially affect the expression ratio of other keratin isoforms or disrupt the overall mechanical function or filament network of the IF cytoskeleton (Nanes et al. 2024). We created K5^high and K6A^high cell lines with the increased keratin isoform fused to a promiscuous biotin ligase (Fig. 1A) (Branon et al. 2018). This allowed us to perform a proteomics screen comparing biotinylated proteins from each cell line, identifying in a differential manner proteins preferentially associated with keratin filaments in K5^high cells versus K6A^high cells (Fig. 1B–D). These proteins were implicated in a variety of biologic processes, including cytoskeletal organization (Figures 1B and S1).

Figure 1. K6A recruits vimentin to keratin filaments.

A. Illustration of the proximity biotin labeling approach to identify proteins differentially associated with intermediate filaments depending on their keratin isoform composition.

B. Most enriched Gene Ontology Biologic Process labels among proteins differentially associated with keratin filaments between K5^high and K6A^high cells, defined as greater than 3-fold differential association and P_adj < .05. See also Supplemental Fig. S1.

C. Volcano plot highlighting proteins differentially associated with intermediate filaments in K5^high versus K6A^high cells.

D. Differential association compared to maximum protein abundance across all samples. Proteins at the top-center of the plot are highly enriched near intermediate filaments regardless of keratin expression. Proteins at the bottom corners of the plot are preferentially associated with intermediate filaments depending on keratin isoform expression, but only at low levels. B-D, n = 2 independent replicates per group.

We previously found that K6A activated myosin by increasing regulatory light chain (RLC) phosphorylation, while the total amount of RLC near keratin filaments remained unchanged (Nanes et al. 2024). Consistent with this, proximity proteomics detected non-muscle myosin heavy chain (MYH9), essential light chain (MYL6), and RLC (MYL12B) associated with keratin filaments at similar levels in K6A^high and K5^high cells (Fig. 1D). This result was also consistent with the previously reported finding that MYH9 interacts with both K5 and K6A in vitro, although in that study MYH9 was destabilized in keratinocytes isolated from a K6-null mouse, a difference which could be due to decreased total keratin bulk in the knockout model rather than K6-specific stabilization per se (Wang et al. 2018). Further supporting a non-isoform-preferential interaction, non-muscle myosin was found to interact with keratins in cells expressing simple epithelial keratins K8 and K18 rather than epidermal keratins (Kwan et al. 2015). On its own, myosin interaction with IFs regardless of composition is not sufficient to explain increased RLC phosphorylation and enhanced contractility in K6A^high cells relative to K5^high cells.

To our great surprise, among the proteins with overall high keratin filament association, vimentin showed the highest differential association between K6A^high and K5^high cells (Fig. 1C–D). Vimentin is the principal IF protein expressed in most mesenchymal cells, but it is not highly expressed and does not form a filament network in keratinocytes, although vimentin expression has previously been reported in cultured keratinocytes (Richard et al. 1990) and skin explants (Biddle and Spandau 1996; Velez-delValle et al. 2016). Thus, we performed a series of experiments to validate the plausibility of keratin–vimentin interactions.

Vimentin is expressed in cultured keratinocytes, but does not assemble a cytoskeletal network

We first confirmed that our keratinocyte cell lines did express low levels of vimentin, with modestly increased vimentin in K6A^high cells compared to K5^high cells (Supplemental Fig. S2A). However, organized vimentin structures could only be detected by immunofluorescence in a very small fraction of cells (~0.1%; Supplemental Fig. S2B–D). The large majority of keratinocytes displayed no distinct vimentin structures, only a diffuse or speckled pattern, with no obvious difference between K5^high and K6A^high cells (Fig. 2, left). Even in the few cells with an organized vimentin immunofluorescence pattern, vimentin did not assemble into well-formed bundled filament networks, but rather formed smaller aggregates that did not align with keratin filaments (Fig. 2, right; Supplemental Fig. S2E). Similar structures have previously been termed squiggles (Prahlad et al. 1998; Yoon et al. 1998) and may represent condensates (Martinez-Cenalmor et al. 2024). Thus, we concluded that vimentin was present predominantly outside of a stable cytoskeletal network and distributed in the cytoplasm. We refer to this state as the soluble vimentin pool.

Figure 2. Vimentin does not form a filament network in keratinocytes.

Immunofluorescence imaging of vimentin (Vim) in monolayers of K5^high and K6A^high cells. Keratins are detected by their associated fluorescent tags (K5-G, K5-mNeonGreen; K6A-R, K6A-mRuby2). Cell monolayers were scratched to create a wound, as flatter cells at the wound edge facilitated clear imaging of the keratin filament network. Diffuse or speckled vimentin staining patterns (left columns) were typical. Cells displaying vimentin squiggles or filaments were extremely rare (right columns). See also Supplemental Fig. S2.

Vimentin forms small oligomers in cultured keratinocytes

This raised the question of what macromolecular state vimentin might adopt if not fully polymerized and incorporated into the IF cytoskeleton. Both vimentin and keratin filaments assemble in a hierarchical manner (Fig. 3A) (Herrmann and Aebi 2016). Individual IF proteins dimerize in parallel (vimentin homodimerizes; type I and type II keratins heterodimerize), then two dimers combine into antiparallel tetramers, which are the building blocks of filaments (Eibauer et al. 2024). Soluble IF proteins are difficult to observe in cells, but vimentin tetramers have previously been isolated from cultured cell lysates (Soellner et al. 1985; Robert et al. 2015).

Figure 3. Vimentin forms small oligomers in keratinocytes.

A. Illustration of intermediate filament assembly units. Vimentin forms homodimers, while type-I and type-II keratins form heterodimers.

B. Western blot analysis of keratinocyte lysates prepared with modified buffers to preserve vimentin oligomers. Left panel: Samples were denatured under standard conditions in the presence of a reducing agent. Center and right panels: Samples prepared with decreased heating in the absence of reducing agents. Carrots indicate predicted sizes of vimentin oligomers: Open black, monomer; Open grey, dimer; Filled grey, trimer; Filled black, tetramer. Buffers (see Methods): RIPA, modified radioimmunoprecipitation buffer; RLB, modified reporter lysis buffer; Urea, 6.5M urea buffer without DTT. β-ME, 2-mercaptoethanol.

C. Western blot analysis of keratinocyte lysates prepared following covalent crosslinking (see Methods). Left panel: Sample prepared without reducing agents to preserve DSP crosslinks. Center and right panels: Samples heated at the temperatures and durations indicated in the presence of a reducing agent. Right panel: Prior to heating, 6.5M urea was added to the sample. Carrots indicate predicted sizes of vimentin oligomers as in (B).

D-E. Vimentin immunoprecipitation (IP) from keratinocytes, followed by Western blot analysis to detect post-translational modification by phosphorylation (D) or citrullination (E). To enhance potential phosphorylation detection (D), cells were treated with orthovanadate (Na₃VO₄) phosphatase inhibitor or vehicle control prior to lysate preparation. Inputs were processed with reduced heating and without reducing agents, preserving some vimentin tetramers (filled carrots), while only monomers (open carrots) were detected in IP samples, which were eluted with reducing agents. IgG, isotype control; Phos.-Y, pan-phosphotyrosine; Phos.-S/T, pan-phosphoserine/threonine; Citrul, anti-modified citrulline (see Methods).

In keratinocyte lysates prepared under standard conditions, including heating to 95°C with a reducing agent to fully denature peptides for Western blotting, nearly all vimentin was detected in a single band near 57 kD, the expected monomer molecular weight (Fig. 3B, left). However, omitting reducing agents revealed an additional 230 kD band consistent with tetrameric vimentin (Fig. 3B, center). This complex was not detected in lysates prepared in 6.5 M urea (Fig. 3B, center). Decreased heating, possibly resulting in incomplete denaturation, increased the proportion of vimentin detected in the putative tetramer band and also revealed weaker bands near 115 kD, consistent with vimentin dimers, and 170 kD, which could represent vimentin trimers, a degradation product, or some other vimentin-containing complex (Fig. 3B, right).

To further characterize these complexes, we performed covalent crosslinking with dithiobis(succinimidyl propionate) (DSP) just prior to preparing keratinocyte lysates. Without reducing agents, which cleave DSP crosslinks, nearly all vimentin was detected at 230 kD and higher molecular weight bands (Fig. 3C, left). While a weak 170 kD complex was also detected, the 57 kD monomer band was completely depleted (Fig. 3C, left). Adding back a reducing agent depleted the higher molecular weight bands while strengthening the 230 kD, 170 kD, and 115 kD bands (Fig. 3C, center), but only adding both a reducing agent and 6.5 M urea recovered the 57 kD band (Fig. 3C, right).

While there are limitations to these results, including large insoluble fractions left over by gentle lysis conditions, potential for vimentin complex formation after cell lysis, and inability to definitively distinguish vimentin complexes with other proteins from vimentin multimers, the results are most consistent with a vimentin population maintained as small oligomers, including tetramers (Soellner et al. 1985).

In vitro, both vimentin and keratin filament assembly is driven solely by ionic composition of the buffer (Herrmann and Aebi 2016). Therefore, we next asked whether there was some factor, other than low concentration, that might explain the maintenance of vimentin as small oligomers in cells. Posttranslational modifications, including phosphorylation and citrullination, are linked to active regulation of vimentin filament disassembly (Chou et al. 1990; Barberis et al. 2009; Helfand et al. 2011) as well as extracellular secretion of vimentin that may serve as an inflammatory trigger in various pathological conditions (Mor-Vaknin et al. 2003; Parvanian et al. 2023a; Parvanian et al. 2023b). We found that vimentin in keratinocytes was both phosphorylated and citrullinated (Fig. 3D–E), likely contributing to its soluble state.

Co-expression of keratins and vimentin in vivo

Although vimentin has typically been regarded as unimportant or absent in epithelial tissues (Oshima 2007), including the epidermis, the possibility that soluble vimentin might engage in isoform preferential interactions with keratin filaments prompted us to more carefully consider the tissue contexts in which such interactions might occur. We focused on skin wound healing and keratinocyte carcinomas, two contexts with increased epidermal remodeling where expression of wound-associated keratin isoforms, including K6A, is significantly increased (Weiss et al. 1984; Mansbridge and Knapp 1987; Patel et al. 2006; Hameetman et al. 2013). Furthermore, both wound healing and skin cancer may demonstrate features of epithelial to mesenchymal transition, which is often marked by increased vimentin expression (Mendez et al. 2010).

Vimentin expression is frequently reported in poorly differentiated or invasive variants of cutaneous squamous cell carcinoma (cSCC), some of which display other mesenchymal characteristics (Iyer and Leong 1992; Smith et al. 1992; Patterson 2025). Interestingly, vimentin expression has also been reported in cSCCs arising from chronic wounds and burn scars (Ikegawa et al. 1989; Kim et al. 1992). By immunofluorescence, we identified scattered vimentin-positive cells in the interior of two well-differentiated cSCCs (Fig. 4A; Supplemental Fig. S3). Notably, while vimentin-positive cSCC keratinocytes were near the deep margins, they retained an overall epithelioid morphology, including contact with surrounding keratinocytes, and did not invade singly into the surrounding stroma (Fig. 4A; Supplemental Fig. S3).

Figure 4. Co-expression of keratins and vimentin in vivo.

A. Hematoxylin and eosin (H&E) staining and immunofluorescence images of sections from a cutaneous squamous cell carcinoma (cSCC). The approximate location of the immunofluorescence images near the deep margin of the tumor is indicated. See Supplemental Fig. S3 for an additional example.

B-C. Tissue expansion imaging of a cSCC (see Methods). B, Cluster of tumor keratinocytes with filamentous K6 immunofluorescence signal and weaker, predominantly non-filamentous vimentin signal. Arrowheads in small insets, immature vimentin filaments, indicated in both vimentin and K6 panels for positional correlation. Filled arrowheads indicate vimentin structures following the keratin filament pattern; the open arrowhead in inset (3) indicates a vimentin structure located in a gap between keratin filaments. C, Tumor-infiltrating mesenchymal cells with robust filamentous vimentin signal, but no K6 signal. Scale bars indicate pre-expansion distances.

D. Proportion of keratinocytes with detectable vimentin transcript in a published single-cell transcriptomics dataset of acute skin wounds (GSE241132). Each line represents serial samples following wounding. n = 3 patients. See also Supplemental Fig. S5A.

To overcome the technical limitations of standard immunofluorescence on paraffin-embedded tissue and better characterize the subcellular organization of keratins and vimentin in cSCC keratinocytes, we performed expansion imaging (Chen et al. 2015; Gambarotto et al. 2021) on an additional section from one of the tumors. The enhanced spatial resolution afforded by tissue expansion revealed strikingly different distributions of K6 and vimentin in double-positive cells (Fig. 4B). While K6 formed the expected filament network throughout the cytoplasm, vimentin was much less well organized. On close inspection, we did identify few immature filament-like vimentin structures (Fig. 4B, arrowheads in small insets). Interestingly, some, but not all, of these more organized vimentin structures appeared to colocalize with keratin filaments. As expected, vimentin in keratin-negative mesenchymal cells was both more abundant and more clearly organized into a filamentous pattern (Fig. 4C). We therefore concluded that both co-expression and interaction of keratins and vimentin are plausible, at least in the context of keratinocyte carcinomas.

In contrast, we did not identify a clear keratinocyte vimentin immunofluorescence signal in epidermis adjacent to the ulcerated surface of one of the SCCs (Supplemental Fig. S4A), the edge of a healing skin biopsy wound (Supplemental Fig. S4B), or at the migrating edge of an epidermal organoid wound healing model (Nanes et al. 2024) (Supplemental Fig. S4C). If anything, the vimentin signal was slightly stronger moving away from the ulcerated SCC, opposite the K6A signal gradient, although we could not definitively distinguish this diffuse vimentin signal from background (Supplemental Fig. S4A). Analysis of a published single-cell RNA sequencing (scRNAseq) dataset from acute skin wounds (Liu et al. 2025) did show an increased proportion of keratinocytes with detectable vimentin transcript seven days after injury (Fig. 4D), although vimentin transcript levels were low and did not differ systematically between migratory and non-migratory keratinocyte clusters (Supplemental Fig. S5A). Our previous reanalysis of bulk RNAseq data from acute and chronic wounds found detectable but unchanged vimentin transcript levels in wound and control samples, but this analysis could not distinguish keratinocytes from other cell types (Iglesias-Bartolome et al. 2018; Ramirez et al. 2018; Nanes et al. 2024).

The discrepancy between immunofluorescence and transcriptomics data may simply reflect the relative insensitivity of immunofluorescence for low-abundance diffusely cytoplasmic proteins. Indeed, the small pool of soluble vimentin in our keratinocyte cell lines was readily detectable biochemically (Fig. 3; Supplemental Fig. S2A), but more difficult to distinguish by immunofluorescence (Fig. 2; Supplemental Fig. S2B–E). We found that vimentin transcript was detected in keratinocytes in five other published scRNAseq datasets, albeit with wide variability between studies (He et al. 2020; Sole-Boldo et al. 2020; Deng et al. 2021; Shiu et al. 2022; Wiedemann et al. 2023) (Supplemental Fig. S5B), indicating that keratinocytes may express low levels of vimentin even in the absence of wounding. Consistent with the transcriptomics data, we found that multiple published proteomics datasets detected vimentin protein in the epidermis (Dyring-Andersen et al. 2020; Li et al. 2022; Nordmann et al. 2023; Solberg et al. 2023), as well as in isolated primary keratinocytes (Dyring-Andersen et al. 2020) and an unrelated keratinocyte cell line (Zaver et al. 2023) (Supplemental Fig. S6). While some level of contamination from other cell types cannot be completely ruled out, these studies used a variety of techniques to separate epidermal keratinocytes from other cells, including laser capture microdissection specifically selecting individual cells positive for keratin expression by immunofluorescence (Nordmann et al. 2023). Thus, while the extent of vimentin–keratin interaction in tissue remains uncertain, it may be broader than generally appreciated.

Soluble vimentin interacts with keratins in cells

We next sought biochemical confirmation of the keratin–vimentin interaction identified by proximity proteomics (Fig. 1C–D). As an initial test, we performed pulldowns from keratinocyte cell lines expressing epitope-tagged vimentin. To prevent vimentin from establishing a filament network at higher concentrations, we expressed a vimentin mutant (vimentin^Y117L) that inhibits polymerization beyond small aggregates (Meier et al. 2009). We confirmed that wild-type vimentin organized into a filament network when exogenously expressed, while vimentin^Y117L did not (Fig. 5A–B). Both endogenous K5 and K6A co-immunoprecipitated with vimentin^Y117L, supporting the plausibility of keratin–vimentin interactions (Fig. 5C).

Figure 5. Soluble vimentin interacts with keratins in cells.

A. Western blot analyses of keratinocyte cell lines expressing only endogenous vimentin (WT), exogenous wild-type vimentin (Vim^WT-FLAG) or exogenous non-filamentous vimentin (Vim^Y117L-FLAG). Open carrot, tagged vimentin monomers.

B. Immunofluorescence of exogenously expressed wild-type vimentin (Vim^WT-FLAG) or non-filamentous vimentin (Vim^Y117L-FLAG) in keratinocytes. Exogenous vimentin is detected using the FLAG epitope.

C. Immunoprecipitation (IP) of non-filamentous vimentin (Vim^Y117L-FLAG) from keratinocytes using anti-FLAG antibodies. Wild-type (WT) keratinocytes, which do not express tagged vimentin, were used as a control.

D. Fluorescence image of a keratinocyte expressing the K5 head domain tagged with mRuby3 and a mitochondrial surface anchor (K5^h-mito) and full-length K5 tagged with mNeonGreen. Note that K5^h-mito is excluded from the keratin filament network.

E. Immunoprecipitation (IP) of non-filamentous vimentin (Vim^Y117L-FLAG) from keratinocytes also expressing mitochondria-anchored K5 or K6A head domains (K5^h-mRb-mito, K6A^h-mRb-mito) using anti-FLAG antibodies. An antibody against mRuby (mRb) is used to detect the keratin head domains by Western Blot.

F. Live fluorescence imaging of keratinocytes expressing K6A tagged with mRuby3 (K6A-mRb) and non-filamentous vimentin tagged with mNeonGreen (Vim^Y117L-mNG). The insets track two vimentin droplets that have become entangled with keratin filament bundles. Arrowheads, keratin–vimentin interaction sites. See also Movies 1–3.

G-H. Crosslinked keratin immunoprecipitation (IP) from cells expressing FLAG-tagged keratins (K5-FLAG and K6A-FLAG). G, Diagram of experiment with images of the large insoluble fraction remaining following crosslinking and lysis with an IP-compatible buffer (see Methods). H, Western blot analysis of the soluble fraction (Input) and IP with anti-FLAG beads. Ig HC, immunoglobulin heavy chain.

I-K. Crosslinked cell fractionation from K5^high and K6A^high cells expressing FLAG-tagged keratins. I, Schematic diagram of experiment. Following covalent crosslinking and lysis with a 6.5M urea buffer, lysates were separated into soluble and insoluble fractions (see Methods). J, Western blot analysis of the insoluble pellets and soluble supernatants (Sup.). K, Quantification of the fraction of vimentin found in the pellet for each sample. n = 4 samples per group.

Using chimeric keratins with domains swapped between different isoforms, we previously found that the myosin activating function of K6A was mediated specifically through its intrinsically disordered head domain (Supplemental Fig. S7) (Nanes et al. 2024). We therefore asked whether keratin head domains could interact directly with vimentin. We created expression constructs consisting of K5 or K6A head domains fused to a fluorescent protein and a mitochondrial surface anchor (Pistor et al. 1994; Benjamin et al. 2010). When expressed in cells, these keratin head domain constructs are excluded from the keratin filament cytoskeleton and concentrated on the cytoplasmic face of the mitochondrial outer membrane (Fig. 5D), allowing us to interrogate keratin head domain interactions entirely outside the context of keratin filament polymerization. By co-immunoprecipitation from cells expressing both vimentin^Y117L and mitochondria-anchored keratin head domain peptides, we confirmed that both K6A and K5 head domains interact with non-filamentous vimentin (Fig. 5E).

While co-immunoprecipitation of full-length keratins and keratin head domain peptides with non-filamentous vimentin supports the plausibility of keratin–vimentin interactions in cells, these experiments cannot determine whether vimentin interacts with intact polymerized keratin filaments, the predominant macromolecular state of keratins in keratinocytes. Furthermore, standard co-immunoprecipitation experiments cannot exclude the possibility of keratin–vimentin complexes forming in solution following cell lysis. To overcome these concerns, we performed live imaging of cells expressing keratins and vimentin^Y117L with fluorescent protein tags. As has been previously reported (Basu et al. 2025), tagged vimentin^Y117L formed droplets with dynamic behavior consistent with condensates (Fig. 5F; Movies 1–3). We observed vimentin^Y117L droplets become entangled with keratin filament bundles (Fig. 5F, arrowheads; Movies 1–3), supporting an interaction between non-filamentous vimentin and keratins assembled into filaments. Due to the heterogeneous morphology and behavior of vimentin^Y117L droplets, we were not able to determine if their interactions with keratin filament bundles differed between K6A^high (Movies 1 and 2) and K5^high cells (Movie 3).

Download video file ^{(1.6MB, mp4)}

Download video file ^{(1.8MB, mp4)}

Download video file ^{(2.5MB, mp4)}

To further confirm endogenous soluble vimentin interaction with keratin filaments biochemically, we first attempted to immunoprecipitate keratin from cells expressing K5 or K6A with high affinity epitope tags, using covalent crosslinking to stabilize both keratin filaments and potentially transient keratin–vimentin complexes (Fig. 5G). This approach was inconclusive, with very limited keratin solubilized in immunoprecipitation-compatible lysis buffer or recovered in the pulldown (Fig. 5G–H). However, we took advantage of the seemingly inconvenient property of limited keratin filament solubility to perform a cell fractionation experiment. Because covalent crosslinking prevented full keratin solubilization even in 6.5 M urea, we reasoned that otherwise soluble vimentin would be pulled into the insoluble fraction if it interacted with, and was therefore crosslinked to, insoluble keratin filaments (Fig. 5I). We found that endogenous vimentin was indeed present in the keratin filament pellets, and although this fractionation is only partially specific, there was an increased proportion of vimentin in the insoluble fraction from K6A^high cells compared to K5^high cells (Fig. 5I–K).

Non-filamentous vimentin interacts directly with keratin filaments in vitro

To further validate our biochemical and microscopy analyses of keratin–vimentin interactions, we next sought to visualize direct vimentin association with keratin filaments in vitro using negative stain electron microscopy (EM). We recombinantly synthesized and purified type-I and type-II keratin isoforms. K14 was mixed with either K5, K6A, or a chimeric keratin consisting of the head domain of K6A fused to the rod and tail domains of K5 (K6A^h5^rt), which we previously found promoted myosin activation similarly to full-length K6A (Nanes et al. 2024). All three type-I–type-II keratin pairs polymerized into filament networks visible by EM (Fig. 6A, left column). To test the ability of non-filamentous vimentin to interact directly with keratin filaments, we added vimentin head domain containing peptides to each keratin mixture prior to filament assembly (see Methods). We focused on the vimentin head domain for two reasons. First, the vimentin head domain alone cannot form coiled-coil dimers with other IF proteins, excluding canonical subunit-like incorporation into keratin filaments as a mode of interaction (Herrmann and Aebi 2016). Second, given recent interest in the potential for IDR-mediated interactions to nucleate molecular assemblies (Banani et al. 2017), as well as the ability of keratin head domains to interact with vimentin (Fig. 5E), we reasoned that the intrinsically disordered vimentin head domain might be similarly relevant.

Figure 6. Non-filamentous vimentin interacts with keratin filaments in vitro.

A. Negative stain electron microscopy of recombinantly synthesized and in vitro polymerized keratin filaments in the presence or absence of vimentin fragments (Vim^h-eGFP, vimentin head domain joined to eGFP; Vim^h+c1, vimentin head domain with coil 1 of the rod domain). K14, a type-I keratin, was polymerized with type-II keratins K5, K6A, or a chimera with the head domain of K6A fused to the rod and tail domains of K5 (K6A^h5^rt; see Methods). Arrowheads, vimentin fragments decorating keratin filaments. See also Supplemental Fig. S8A.

B-D. Co-sedimentation analysis of in vitro polymerized keratin filaments and unpolymerized vimentin. B, Schematic diagram of experiment. Fluorescently labelled K14 (K14–488) and either K5, K6A, or K6A^h5^rt chimera were first polymerized separately in vitro, then mixed with unpolymerized fluorescently labeled vimentin (Vim-Cy3) and centrifuged at low speed to pellet intermediate filament polymers (see Methods). C, Fluorescence and Coomassie-stain images of gel electrophoresis of the pellets. Upper and lower gels represent independent experimental replicates. Note that each experimental replicate contains two K5/K14 and vimentin control samples, which were averaged for quantification. D, Vimentin to keratin ratio in pellets, after subtracting the average amount of pelleted vimentin in the absence of keratin. See also Supplemental Fig. S8B.

When combined with keratin pairs and allowed to assemble, vimentin head domains (Vim^h-eGFP) decorated keratin filaments at irregular intervals (Fig. 6A, center column), confirming a direct interaction. Adding Vim^h-eGFP somewhat decreased the density of keratin filaments, although this may reflect the large excess of vimentin fragments (1.5 μM) compared to keratin pairs (0.6 μM). In contrast, vimentin fragments containing the head domain and coil 1 of the rod domain (Vim^h+c1) significantly disrupted keratin filament formation (Fig. 6A, right column), consistent with the known effect of coil 1 truncations.(Chernyatina et al. 2012) The remaining short keratin filaments were heavily decorated with Vim^h+c1 (Fig. 6A). Interestingly, keratin filament disruption by Vim^h+c1 was noticeably more severe for K6A and K6A^h5^rt containing filaments compared to K5 containing filaments (Fig. 6A), consistent with an isoform-preferential interaction. We confirmed that neither Vim^h-eGFP or Vim^h+c1 formed filaments or large aggregates in the absence of keratins (Supplemental Fig. S8A).

We next combined pre-assembled recombinantly polymerized keratin filaments with recombinant full-length vimentin. For this experiment, vimentin was maintained in an unpolymerized state through titration of salt concentration in the buffer (see Methods). In a low-speed co-sedimentation assay, unpolymerized vimentin co-sedimented with keratin filaments, with the amount of co-sedimentation increased with keratin filaments containing either full-length K6A or the K6A^h5^rt chimera compared to K5 (Fig. 6B–D; Supplemental Fig. S8B). Thus, the isoform-preferential aspect of keratin–vimentin interaction depends on the keratin head domain and likely requires polymerized keratin filaments.

Validation of isoform-preferential keratin–vimentin interactions in situ

Turning back to the cellular context, we independently confirmed the isoform-preferential keratin–vimentin interaction using a proximity ligation assay (PLA), an imaging-based approach employing two oligonucleotide-conjugated antibodies, rolling-circle amplification, and in-situ hybridization to sensitively detect co-localization between two proteins in situ (Fredriksson et al. 2002). To increase robustness of the PLA analysis, we performed experiments on mosaic monolayers of cells with different keratin expression mixtures, allowing side-by-side comparison within each sample (Nanes et al. 2024) (Fig. 7A). Consistent with the proteomic, biochemistry, and in vitro assays, we identified increased association between vimentin and keratins in K6A^high versus K5^high cells, regardless of whether K14, K5, or K17 was used as the bait (Fig. 7B–D). Preferential interaction depending on overall keratin isoform balance of the cell, regardless the specific keratin isoform bait, indicates that vimentin interacts with keratins in the context of polymerized filaments, where all expressed keratin isoforms are closely mixed, not with keratin monomers. Furthermore, the increased association between vimentin and keratin filaments in K6A^high cells was a population effect, not driven by a small number of outlier cells which might represent those with organized vimentin structures distinguishable by immunofluorescence (Fig. 7B–D; Supplemental Fig. S2B–D). We also verified that keratin–vimentin interactions were increased in cells expressing the K6A^h5^rt chimera, confirming the importance of the K6A head domain to the isoform-preferential vimentin interaction (Fig. 7E). Finally, we found that keratin–vimentin interactions were increased in cells expressing increased K6B, a closely related paralog of K6A whose head domain differs by only two amino acids (Supplemental Fig. S9).

Figure 7. Validation of isoform-preferential keratin–vimentin interactions in situ.

A. Schematic diagram of a proximity ligation assay (PLA) experiment for detecting keratin–vimentin interaction in intact cells. A mosaic of K5^high and K6A^high cells allows side-by-side comparison of PLA signal in cells with different keratin isoform mixtures, enhancing robustness of the approach (see Methods).

B-D. PLA detecting colocalization between vimentin and K14 (B), K5 (C), or K17 (D), comparing K5^high and K6A^high cells. Left: Annotated images indicating cell type (green outline, K5^high; red outline, K6A^high), PLA detections (white diamonds), and PLA detection density (heatmap). Right: Plot of PLA detection density per cell. Outlined dots, connected between groups, mean values of experimental replicates. Data were analyzed using mixed effects models (see Methods). B, n = 806–975 cells per group in 48 images. C, n = 103–113 cells per group in 17 images. D, n = 84–105 cells per group in 18 images.

E. PLA detecting colocalization between vimentin and K14, comparing cells expressing K5 or the K6A^h5^rt chimera. Annotations as in (B), except red outline indicates K6A^h5^rt cells. n = 108–127 cells per group in 48 images.

F. Fluorescence imaging of keratinocytes expressing mRuby2-tagged K6A incubated in vehicle or 1,6-hexanediol at the indicated concentration.

G. PLA detecting colocalization between vimentin and K14, comparing K5^high and K6A^high cells treated with 1% 1,6-hexanediol. Annotations as in (B). n = 546–602 cells per group in 46 images.

Given involvement of both keratin (Figures 5E, 6A–D, and 7E) and vimentin (Fig. 6A) intrinsically disordered head domains, we asked how disrupting IDR-mediated interactions would affect keratin–vimentin association. 1,6-hexanediol (1,6-HD) is an aliphatic alcohol which dissolves IDR-mediated biomolecular condensates while preserving other protein-protein interactions (Basu et al. 2025). At high concentrations (8%), 1,6-HD fully disrupts IF assembly, but even lower concentrations (≤1%) are reported to disrupt IDR condensates, including those formed by vimentin^Y117L (Basu et al. 2025). We confirmed that 1% 1,6-HD did not disrupt the organization of keratin filaments in keratinocytes (Fig. 7F). Nonetheless, 1% 1,6-HD significantly decreased interaction between vimentin and keratins as measured by PLA (Fig. 7G). Together, these data led us to the conclusion that a soluble pool of vimentin interacts with keratin filaments, and preferentially with keratin filaments containing a greater proportion of K6A.

Soluble vimentin recruits myosin-activating kinases to K6A-containing filaments

How could isoform-preferential association between soluble vimentin and K6A-containing keratin filaments promote actomyosin contractility? We considered the possibility that soluble vimentin acts as a recruitment factor for kinases, bringing them into the proximity of keratin filaments where they are more likely to interact with RLC, with a bias toward filaments with high K6A content (Fig. 8A). There is precedence for such a role, as soluble vimentin has been implicated as a transport adaptor for MAP kinase in neurons (Perlson et al. 2005). Furthermore, vimentin was previously identified as a phosphorylation target for ROCK, triggering vimentin filament disassembly, and implying at least transient ROCK–vimentin binding (Goto et al. 1998). ROCK also targets RLC (Sellers et al. 1981). We detected modest colocalization between vimentin and ROCK by PLA (Supplemental Fig. S10A), as well as frequent colocalization between vimentin and MLCK (Supplemental Fig. S10B), another RLC activator (Scholey et al. 1980). Neither ROCK nor MLCK were differentially associated with vimentin in K6A^high cells compared to K5^high cells (Supplemental Fig. S10A–B). However, once differentially recruited to K6A-containing keratin filaments by vimentin, these kinases would be organized near other keratin-interacting proteins, including RLC (Figures 1D and 8A).

Figure 8. Soluble vimentin recruits myosin-activating kinases to K6A-containing filaments.

A. Diagram of the proposed model. Soluble vimentin (Vim) shuttles regulatory kinases preferentially to K6A-containing filaments, where they phosphorylate myosin regulatory light chain (RLC).

B. Proximity ligation assay (PLA) detecting colocalization between ROCK and K14, comparing K5^high and K6A^high cells. Left: Annotated images indicating cell type (green outline, K5^high; red outline, K6A^high), PLA detections (white diamonds), and PLA detection density (heatmap). Right: Plot of PLA detection density per cell. Outlined dots, connected between groups, mean values of experimental replicates. Data were analyzed using mixed effects models (see Methods). n = 321–353 cells per group in 66 images.

C. Traction force microscopy (TFM) of K5^high and K6A^high cells treated with vehicle, a ROCK inhibitor (GSK269962A), or 1% 1,6-hexanediol (1,6-HD). Annotated images: Reconstructed traction force vectors (arrows) on a heatmap of force magnitude and a watermark of keratin filaments, which was used to segment the cell border (rings). Plot: Comparison of strain energy densities between cell lines and treatment groups. Outlined dots, connected between groups, mean values of experimental replicates in separate batches of TFM substrates. Vehicle, n = 61–68 cells per group; GSK269962A, n = 30–34 cells per group; 1,6-HD, n = 24–29 cells per group.

D. TFM of keratinocytes expressing only endogenous vimentin (WT), exogenous wild-type vimentin (Vim^WT-FLAG) or exogenous non-filamentous vimentin (Vim^Y117L-FLAG). Annotations and plot as in (C), except watermark is a cytoplasmic marker. ***, P < 10⁻³. n = 40 cells per group.

E. Western blot analysis of vimentin knockdown by shRNA (shVim#1; shVim#2; non-targeting control, shScrm) in keratinocytes expressing a chimeric keratin containing the head domain of K6A joined to the rod and tail of K5 (K6A^h5^rt) or full-length K5 (K5^high). Dashed line separates different brightness adjustments for different baseline vimentin levels. See also Supplemental Fig. S11 for an alternate display.

F. TFM of K6A^h5^rt cells expressing shVim#1, shVim#2, or shScrm. Annotations and plot as in (C). ***, P < 10⁻³; **, P < .01. n = 29–43 cells per group.

G. TFM of K6A^h5^rt or K5^high cells expressing shVim#2 or shScrm. Annotations and plot as in (C). ***, P < 10⁻³. n = 18–21 cells per group.

H. PLA detecting myosin activation by colocalization of phosphorylated myosin regulatory light chain (pRLC) and non-muscle myosin heavy chain (MYH9), comparing K5^high and K6A^high cells expressing shVim#2 or shScrm. Left: Annotated images as in (B). Right: Plot of PLA detection density per cell. Data were analyzed using a mixed effects model (see Methods). ***, P < 10⁻³. n = 67–113 cells per group in 16–24 images per shRNA condition.

I. Keratin filament network dynamics (see Methods) in K6A^h5^rt cells and K5^high cells upon expression of shVim#2 or shScrm. See also Supplemental Fig. S12 for filament images, segmentations, and dynamics score maps. ***, P < 10⁻³; **, P < .01; *, P < .05. n = 39–49 cells per group.

Interestingly, ROCK association with keratins was modestly but reproducibly increased in K6A^high cells compared to K5^high cells as measured by PLA (Fig. 8B), although a differential interaction was not captured by proximity proteomics. We therefore tested the effect of ROCK inhibition on K6A-promoted RLC phosphorylation and its direct consequence, cell contractility. Inhibiting ROCK substantially blocked both RLC phosphorylation (Supplemental Fig. S10C) and cell contractility (Fig. 8C) in K5^high and K6A^high keratinocytes, results consistent with our model. Our model also predicts that disruption of the IDR-mediated interaction between soluble vimentin and keratin filaments should prevent K6A from triggering myosin activation. In fact, we found that treatment with 1% 1,6-HD to disrupt keratin–vimentin interactions (Fig. 7F–G) blocked cellular force generation almost to the degree of ROCK inhibition (Fig. 8C). However, since even low concentrations of 1,6-HD may interfere with kinase activity (Duster et al. 2021), this result may be nonspecific.

Because our model implies a central role for soluble vimentin mediating the activation of myosin and increased force generation with increased expression of K6A (Fig. 8A), we directly tested this prediction using both gain-of-function and loss-of-function approaches. First, we measured cellular force generation in keratinocytes with increased expression of either wild-type vimentin or non-filament-forming vimentin^Y117L. Both wild-type and mutant vimentin increased contractile force generation compared to controls (Fig. 8D). Furthermore, there was a trend toward increased contractility in cells expressing non-filamentous vimentin^Y117L compared to wild-type vimentin, although this difference did not reach statistical significance (Fig. 8D).

As a complimentary approach, we knocked down vimentin in keratinocyte cell lines with shRNA (Fig. 8E; Supplemental Fig. S11). Using traction force microscopy, we found that vimentin depletion significantly decreased force generation in K6A^h5^rt cells, as predicted by our model (Fig. 8F). Compared to the effect in K6A^h5^rt cells, vimentin depletion only slightly decreased traction force generation in K5^high cells, reflecting the decreased dependence of cellular force generation on vimentin recruitment of myosin-activating kinases in cells with lower levels of K6A (Fig. 8G). We also directly confirmed that vimentin depletion abrogated the difference in myosin activation between K6A^high and K5^high cells by measuring colocalization of phosphorylated RLC with non-muscle myosin heavy chain by PLA (Fig. 8H).

Finally, we tested the effect of vimentin knockdown on dynamic intracellular movement of the keratin IF network. We previously found that intracellular displacement of keratin filament bundles on a time scale of several minutes was decreased in K6A^high and K6A^h5^rt cells compared to K5^high cells (Nanes et al. 2024). While the mechanism of this effect is not fully clear, it was reversed by non-muscle myosin or ROCK inhibition, indicating that it is downstream of myosin activation increased by K6A expression (Nanes et al. 2024). Whether decreased keratin filament dynamics in K6A^high cells results from a direct change in motor protein action against keratin filaments, an indirect change in the cellular environment such as obstruction by actin stress fibers (Pegoraro et al. 2017; Smoler et al. 2020), or another cause, keratin filament dynamics represent an orthogonal measurement downstream of myosin activation to confirm the importance of vimentin as a myosin-activating signaling recruitment factor. Consistent with our model, we found that, compared to a non-targeting shRNA control, vimentin knockdown indeed returned filament dynamics in K6A^h5^rt cells back to the level seen in K5^high cells (Fig. 8I; Supplemental Fig. S12). These results confirm that increased myosin activation, increased traction force generation, and decreased filament dynamics caused by K6A all depend on soluble vimentin acting as a recruitment factor for molecular signals.

Discussion

Our data reveal two mechanisms by which keratins and vimentin co-modulate a cellular signaling pathway. First, isoform-preferential interactions allow changes in keratin cytoskeleton composition to influence the set of proteins recruited to the IF scaffold. In other contexts, certain keratins and vimentin are thought to influence cell signaling by alternatively stabilizing or serving as a localization sink for different effector proteins (Paramio et al. 2001; Kim et al. 2006; Wang et al. 2012; Kroger et al. 2013; Chung et al. 2015; Loschke et al. 2016; Redmond and Coulombe 2020; Cohen et al. 2026; Redmond et al. 2026). The mechanism here is more subtle. Because proteins associated with a 1-dimensional filament are more likely to interact than those freely diffusing through the 3-dimensional cytoplasm, isoform-preferential interactions allow the IF cytoskeleton to function as a tunable signaling organizer as different proteins are recruited to filaments. In this case, the outcome of changing IF cytoskeleton composition is increased myosin activation, not abundance, driving cellular force generation. Crucially for epithelial cells with a keratin IF cytoskeleton, switching the keratin isoform expression mixture can modulate IF signaling activity without disrupting the overall IF network, thus preserving its critical mechanical scaffold role.

Importantly, we used a minimally perturbing gain-of-function approach to identify keratin isoform-preferential interactions modulating cell signaling without broadly disrupting or depleting the keratin cytoskeleton. Because we directly compared cells with different ratios of K5 and K6A, but equal total keratin, we conclude that the observed effects were truly driven by isoform composition rather than general depletion of IFs. Broad loss-of-function perturbations, including keratin gene knockouts and mutations with dominant-negative effects on keratin filament polymerization, have identified many important IF functions, but cannot necessarily distinguish functions of a specific keratin isoform from functions that depend on the keratin cytoskeleton in general.

Second, and perhaps most surprisingly, soluble pools of even lowly expressed IF proteins may function as recruitment factors for signaling molecules. To date, IF proteins have been studied predominantly within the context of their prominent filament networks, while their more difficult to observe soluble states are less well understood. Nonetheless, IF protein functions apart from the IF cytoskeleton are increasingly emerging (Parvanian et al. 2023b). In this study, we found not only that a soluble pool of vimentin modulated cell signaling, but that it did so in cooperation with polymerized keratin filaments. Keratin and vimentin filament networks are thought to be antagonistic, with cells generally forming one IF cytoskeleton or the other (Kuburich et al. 2022), although there are notable exceptions (Osborn et al. 1980; Yoon et al. 2001). We speculate that in epithelial cells, a combination of abundant keratin filaments and post-translational modifications stabilize any vimentin that is expressed in a soluble state in order to support its role as a recruitment factor for signaling molecules. Whether soluble vimentin serves a signaling role throughout epithelial tissues or only in limited contexts such as cancer, where vimentin expression is increased, remains an open question. Intriguingly, one of the most prominent phenotypes of the vimentin-null mouse is a skin wound healing defect (Eckes et al. 2000). While this phenotype has been attributed to defects in dermal fibroblasts (Cheng et al. 2016), our results raise the possibility that keratinocytes themselves may contribute to the phenotype, with vimentin unavailable to recruit regulatory kinases to keratin IFs enriched in wound-associated K6A.

Although the precise structural nature of the interaction between soluble vimentin and keratin filaments remains to be determined, our data indicate that it is likely mediated through the intrinsically disordered head domains of each protein. Recent structural studies indicate that in mature IF polymers, head domains are concentrated within the filament lumen (Weber et al. 2021; Eibauer et al. 2024; Beales et al. 2026). Despite their interior location, IF head domains may still interact with other filament-associated proteins through gaps along the filament lattice, particularly in the case of keratin filaments, which display profound structural heterogeneity (Weber et al. 2021). Furthermore, a static model of luminal head domain accessibility may be misleading. Dynamic tetramer exchange occurs along the length of polymerized IFs (Colakoglu and Brown 2009; Noding et al. 2014; Tran et al. 2025), producing at least transient luminal exposure. In addition to organizing signaling assemblies, IDR-mediated interaction between tetramer head domains and the filament lumen could provide a mechanism to explain how new tetramers are incorporated mid-filament (Zhou et al. 2021).

Luminal organization of keratin head domains also offers an intriguing potential mechanism for preferential interaction of soluble vimentin with K6A-enriched keratin filaments. We detected K6A-preferential vimentin interactions with polymerized keratin filaments using multiple approaches both in cells (Figures 1, 5I–K, 7A–E) and in vitro (Fig. 6). However, other approaches either biochemically enriching for soluble keratins (Fig. 5C) or explicitly targeting interactions outside of the filament context (Fig. 5E) did not detect a vimentin interaction preference for K6A versus K5. Although there are other possible explanations, including technical differences in assay sensitivity, this pattern suggests that keratin filament polymerization is required for a K6A-preferential vimentin interaction. In the polymerized filament context, keratin isoform-driven changes in luminal accessibility could produce isoform-preferential interactions even in the absence of direct binding affinity differences.

The implications of a cellular signaling circuit organized by IDR-mediated interactions with keratin head domains are potentially quite broad. In some respects, the keratin–vimentin–myosin signaling module operates like a membraneless organelle (MLO), where IDR-mediated or other multivalent low-affinity interactions recruit client molecules to a scaffold and spatially organize multiple cellular functions (Banani et al. 2017; Lyon et al. 2021). In this framework, keratin filaments form the scaffold, while soluble vimentin and its associated regulatory kinases represent the clients. As with MLOs, a subtle change in scaffold composition—K6A to K5 ratio—modulates client recruitment and signaling function. However, MLOs are typically defined as phase-separated biomolecular condensates, while polymerized IFs rely on specific high-affinity protein interactions (Eibauer et al. 2024). Nonetheless, concentrating disordered IF head domains in the filament lumen establishes the same potential for IDR-mediated condensate-like client recruitment, a mechanism very likely underlying the interaction between soluble vimentin and keratin filaments through their respective head domains (Lin et al. 2016; Zhou et al. 2021). This raises the intriguing possibility that the IFs might recruit other IDR-containing proteins, and, by virtue of their linear structure, promote biochemical interactions through dimensionality reduction. Notably, IDR-mediated recruitment and dimensionality reduction have been leveraged to engineer a synthetic microtubule-based MLO (Reinkemeier and Lemke 2022), supporting the plausibility of such a model. In epithelial cells, keratin isoform diversity establishes a thus far undiscussed potential for cell regulation through complex interaction networks organizing a variety of biochemical processes on the IF cytoskeleton, akin to the myosin-activating module organized by K6A-containing filaments identified here. Even subtle isoform-specific differences in keratin head domain interaction affinities or composition-driven changes in IF lumen accessibility could serve to fine-tune multiple signaling circuits, contributing to the establishment of distinct epithelial cell types and states. If this is the case, the characteristic scaffold structure of the IF cytoskeleton could be as important for organizing cellular signals as it is for mechanical support.

Materials and Methods

Cell culture

Human keratinocyte cell line Ker-CT (ATCC CRL-4048), derived from human foreskin keratinocytes through expression of HTERT and CDK4 (Ramirez et al. 2003; Vaughan et al. 2009), was a gift from Dr. Jerry Shay (UT Southwestern Medical Center). Keratinocytes were grown on tissue-culture treated plastic coated with bovine type I collagen (PureCol, Advanced Biomatrix #5005) by pre-incubating dishes in a 100 μg/mL collagen solution for 30 minutes. Keratinocytes were maintained in Keratinocyte Serum-Free Medium (K-SFM; Gibco #17005042) with the included 5 ng/mL human recombinant epidermal growth factor and 50 mg/mL bovine pituitary extract supplements according to the manufacturer’s instructions in a humidified incubator at 37 °C and 5% CO₂. To induce desmosome formation and support differentiation, we developed a keratinocyte Differentiation Medium based on a simplification of the classic Rheinwald and Green formula (Rheinwald and Green 1975; Rasmussen et al. 2013). This medium was composed of fully-supplemented K-SFM mixed in equal parts with DMEM/F12 (Gibco #11320033) and additionally supplemented with 2% fetal bovine serum (FBS; Sigma F0926). Timing of medium change from K-SFM to Differentiation Medium for each applicable experiment is described below. Based on an estimated calcium concentration of 3.75mM in FBS (Blankenship and Heitman 2005), the calcium concentration of the Differentiation Medium was 0.65mM.

HEK293 cells (a gift from H. Choe) used for generation of lentiviral particles were maintained in DMEM (Gibco #11995–065) supplemented with 10% FBS and antibiotic–antimycotic reagent (Gibco #15240–062) in a humidified incubator at 37 °C and 5% CO₂. hTERT RPE-1 cells were obtained from ATCC (CRL-4000) and maintained in equal parts DMEM/F12 supplemented with 10% FBS in a humidified incubator at 37 °C and 5% CO₂. All cell lines were periodically tested for mycoplasma using a PCR-based Genlantis Mycoscope Detection Kit (MY01100).

Plasmids

Primer and synthetic DNA sequences (Integrated DNA Technologies) used to assemble the following constructs are provided in Supplemental Table S1. Construction of pLVX-KRT5-mNG-IRES-Puro, pLVX-KRT6A-mRb-IRES-Neo, pLVX-K6h5r5t-mNG, pLVX-KRT5-mRb-IRES-Puro, pLVX-K5-FLAG-IRES-mNeonGreen, and pLVX-K6A-FLAG-IRES-mNeonGreen was described previously (Nanes et al. 2024).

Proximity biotin labeling baits

The sequence of miniTurbo, a small promiscuous biotin ligase (Branon et al. 2018), was synthesized with a C-terminal FLAG tag and flanking XbaI and BamHI restriction sites, which were used to insert the fragment into the pLVX-IRES-Puro lentiviral expression vector (Clontech). The KRT5 and KRT6A sequences were amplified by PCR from pLVX-KRT5-mNG-IRES-Puro and pLVX-KRT6A-mRb-IRES-Neo and inserted in-frame and N-terminal to the biotin ligase with a short peptide linker using seamless cloning (HiFi DNA Assembly Master Mix, New England Biolabs), creating pLVX-KRT5-miniTurbo-FLAG-IRES-Puro and pLVX-KRT6A-miniTurbo-FLAG-IRES-Puro respectively. To create a marker construct for these experiments (pLVX-EGFP), the EGFP sequence from pEGFP-C1 (Clontech) was amplified by PCR and inserted into pLVX-IRES-Puro using the SnaBI and BamHI restriction sites.

In-vitro expression

For in-vitro expression and fluorescent labeling of vimentin, the sequence encoding human full-length vimentin (amino acids 1–466) was modified to remove its single internal cysteine residue (C328A) and add a GGC tripeptide to the C-terminal end. For in-vitro expression of vimentin head domain constructs, the head domain (amino acids 1–93; vimentin^head) (Lin et al. 2016) followed by eGFP and a 6-histidine tag, or the head and coil-1 domain (amino acids 1–253; vimentin^head-c1) were assembled or amplified by PCR. Then the sequences encoding vimentin^C328A-GGC, wild-type vimentin, vimentin^head -eGFP-6His, vimentin^head-c1, KRT5, KRT6A, and K6A^h5^rt were cloned into pEt24d(+) (EMD Biosciences) by PCR.

Vimentin expression

The sequence of vimentin was isolated and a C-terminal FLAG tag added by two-step PCR, then substituted for the keratin sequence in pLVX-KRT5-FLAG-IRES-mNeonGreen using seamless cloning, creating pLVX-Vimentin-FLAG-IRES-mNeonGreen. pLVX-Vimentin^Y117L-FLAG-IRES-mNeonGreen was then created by site-directed mutagenesis (Q5 Site-Directed Mutagenesis Kit, New England Biolabs). pLVX-Vimentin^Y117L-mNeonGreen was similarly created by deletion of the FLAG-IRES sequence.

Keratin head domain expression

Keratin domains were defined based on NCBI Reference Sequence annotations (NP_000415.2, NP_005545.1). The head domain sequences from KRT5 (amino acids 1–167) and KRT6A (amino acids 1–162), as well as the sequence for mRuby3, were amplified by PCR from the keratin expression plasmids described above. The bacterial membrane anchor from Listeria monocytogenes ActA, which targets proteins to the cytoplasmic face of the mitochondrial outer membrane when expressed in mammalian cells,(Pistor et al. 1994; Benjamin et al. 2010) was amplified by PCR from pCAG-mCherry2-mito (Addgene #164164)(Tanida et al. 2020) by PCR. These sequences were then inserted into pLVX-IRES-Puro between the EcoRI and BamHI restriction sites using seamless cloning (HiFi DNA Assembly Master Mix, New England Biolabs) to create pLVX-K5h-mRuby3-mito and pLVX-K6Ah-mRuby3-mito respectively.

K6B expression

The KRT6B coding sequence was amplified from DNASU plasmid HsCD00040197 by PCR, then substituted for the KRT5 sequence in pLVX-KRT5-mRb-IRES-Puro by seamless cloning, creating pLVX-KRT6B-mRb-IRES-Puro.

Vimentin knockdown

pLKO.1-Scrambled (Addgene #136035) (Fischer et al. 2020) was modified to express H2B-mRuby3 for visual identification of transduced cells based on a nuclear fluorescence signal (Bajar et al. 2016). The H2B-mRuby3 sequence was synthesized and inserted into pLKO.1 in place of the existing PuroR sequence using BamHI and KpnI restriction sites, creating pLKO.1-H2B-mRb. Vimentin targeting sequences for shRNA constructs were selected from the Genetic Perturbation Platform Web Portal (Broad Institute; https://portals.broadinstitute.org/gpp/). Complementary primers were synthesized and annealed, then cloned into pLKO1-H2B-mRuby3 using AgeI and EcoRI restriction sites. The clones selected for shVim#1 (TRCN0000029120) and shVim#2 (TRCN0000029119) have, respectively, moderate and high predicted knockdown performance.

Lentivirus transduction and cell line selection

Lentiviral particles were generated using the pLVX system (Clontech) with packaging vectors psPAX2 and pMD2.G (Addgene plasmids #12260 and #12259). HEK293 cells were transfected with expression and packing plasmids following standard calcium phosphate or polyethylenimine (Polysciences #23966) protocols. Supernatant was collected two days after transfection, filtered through 0.45-μm mixed cellulose esters membrane syringe filters (Fisher Scientific #09-720-005) and incubated on target cells overnight. After multiple days of culture and cell expansion, cells expressing the relevant fluorescent marker were collected on a FACS Aria II SORP flow cytometer equipped with 405 nm, 488 nm, 561 nm, and 633 nm lasers (Beckton Dickinson) and returned to culture.

Proximity biotin labeling and proteomics

Keratinocytes were transduced with lentivirus to express K5 or K6A linked to a small promiscuous biotin ligase (Branon et al. 2018). Because these constructs lack a selection marker, a mixed-population lentivirus was prepared by combining the tagged keratin constructs with an EGFP marker construct at a molar ratio of 100:1. Cells were selected by fluorescence activated cell sorting as described above. Each cell line was then grown to confluence in collagen-coated 10 cm tissue culture plates and switched from K-SFM to Differentiation Medium for 24 hours before beginning the experiment.

Two independent experimental replicates were performed. Cells were incubated with 100 μM biotin for 1 hour and lysed in JS buffer (50 mM Tris HCl pH7.5, 150 mM NaCl, 5 mM EGTA, 1.5 mM MgCl₂, 1% Glycerol and 1% Triton X-100) supplemented with protease and phosphatase inhibitors (Halt Protease and Phosphatase Inhibitor, Thermo). Cleared lysates were incubated with Protein A Sepharose beads to remove proteins with nonspecific binding to the beads. The resulting supernatant (1 mg) was incubated with 100 μl of streptavidin-coated magnetic beads (Dynabeads MyOne Streptavidin T1 beads, Invitrogen) overnight at 4 °C with rotation to enrich for biotinylated proteins. Beads were washed three times with ice-cold JS buffer and in-gel digested with trypsin for mass spectrometry analysis on an Orbitrap Fusion Lumos (Thermo Fisher). Peptide spectra were analyzed using Proteome Discover 2.4 (Thermo Fisher) and matched to the human protein database from Uniprot.

Proteomics data were analyzed using the DEP package for differential enrichment analysis of proteomics data in R (Zhang et al. 2018). Only proteins with no missing values in at least one of the two cell lines were included in the analysis. Abundance data was background-corrected and normalized by variance stabilizing transformation (Huber et al. 2002), and any remaining missing values were imputed as draws from a Gaussian distribution centered on quantile 0.005 of the observed values with standard deviation equal to the median standard deviation of the observed values (Gatto and Lilley 2012). Differential enrichment was estimated based on protein-wise linear models and empirical Bayes statistics (Ritchie et al. 2015), and local false discovery rates were estimated using the fdrtool package (Strimmer 2008b; Strimmer 2008a). Functional analysis was then performed using the clusterProfiler package (Yu et al. 2012). We searched for overrepresentation of proteins with at least 3-fold enrichment and adjusted P < .05 in the Gene Ontology Biological Process annotation set (Ashburner et al. 2000; Gene Ontology 2021) using Fisher’s exact test, with adjustment for multiple comparisons (Benjamini and Hochberg 1995).

Western blotting

Cells were grown to confluence and switched to Differentiation Medium 24 hours prior to sample collection. Because keratin filaments are highly insoluble, a high-ionic-strength urea buffer was used in place of a standard lysis buffer (Wang et al. 2016). The 6.5 M urea lysis buffer also contained 50 mM tris, 1 mM EGTA, 2 mM DTT, 50 mM sodium fluoride, 1 mM sodium vanadate, and Halt Protease Inhibitor Cocktail (ThermoFisher #78430), and was adjusted to pH 7.5. After washing twice with PBS, cells were collected into urea lysis buffer using a cell scraper, then incubated at 4 °C for 20 minutes on a shaker. The lysates were then sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) twice for 10 seconds at 20% power with a 10 second intervening rest. Any remaining insoluble material, which was typically negligible, was pelleted by centrifugation at 16.1×10³ g for 20 minutes at 4 °C, and the supernatant was transferred to a clean tube. Protein concentrations were measured using a BCA assay (ThermoFisher #23228). Laemmli sample buffer (BioRad #161–0747) with 2-mercaptoethanol (ThermoFisher #BP176) was added to the samples, which were then heated to 95 °C for 5 minutes.

Samples were run on 4–20% Mini-PROTEAN pre-cast gels (Biorad) and transferred to nitrocellulose membranes (ThermoFisher #88018). Once membranes were blocked with 5% bovine serum albumin (BSA; Equitech-Bio BAH65–0500) in Tris buffered saline (TBS) with 0.1% Tween 20 (TBST; ThermoFisher BP337) for 1 hour at room temperature, they were incubated in primary antibodies overnight at 4 °C (Supplemental Table S2). Membranes were washed with TBST and probed with secondary antibodies conjugated with horseradish peroxidase (Supplemental Table S2). Once they were washed again, bands were detected through enhanced chemiluminescence using freshly prepared substrate solution based on 100 mM Tris (ThermoFisher BP152) adjusted to pH 8.5 with hydrochloric acid (ThermoFisher A144S-212), 1.2 mM luminol (Sigma A8511), 2.5 mM p-coumaric acid (Sigma C9008) and 0.04% hydrogen peroxide (w/v, Alpha Aesar L14000) on a Syngene G:BOX imager equipped with a Synoptics 4.2-megapixel camera and controlled with Syngene Genesys software or on a BioRad ChemiDoc MP digital gel/blot imaging system.

Fluorescence imaging of cell monolayers

Cells were prepared and seeded onto collagen-coated 12 or 18 mm-diameter #1.5 coverslips in tissue culture dishes. For keratinocytes, after cells were grown to confluence, the culture medium was switched from K-SFM to Differentiation Medium. Other cells were maintained in growth medium. For experiments requiring creation of a wound edge, a line scratch was created in each coverslip using a 1000 μL pipette tip 4 hours after the medium change and the cultures were returned to the incubator to allow migration to occur. If indicated, cells were treated with different concentrations of 1,6-hexanediol (Sigma-Aldrich #240117) in Differentiation Medium for 5 minutes prior to fixation. 24 hours after the medium change, coverslips were removed from the incubator, washed 3 times with PBS, incubated in 4% paraformaldehyde (Electron Microscopy Sciences #15713) in PBS for 10 minutes, incubated in 0.1% Triton X-100 in PBS for 8 minutes, and washed 3 more times with PBS. For experiments using anti-FLAG antibody, samples were incubated in blocking solution (1% BSA with 22.5 μg/mL glycine in PBS) for 30 minutes at 37°C. Samples were incubated in primary antibody diluted in PBS for 30 minutes at 37 °C (Supplemental Table S2), washed 3 times with PBS, incubated in fluorophore-conjugated secondary antibody diluted in PBS for 30 minutes at 37 °C (Supplemental Table S2), and washed 3 more times with PBS. Finally, coverslips were mounted onto glass slides using Fluoromount-G (ThermoFisher 00-4959-52).

Slides were imaged using an inverted phase-contrast and epifluorescence Nikon ECLIPSE Ti microscope equipped with a motorized and programmable stage, a Nikon Perfect Focus System, a Hamamatsu Orca-Flash4.0 scientific CMOS camera, a SOLA solid state white-light excitation system, an OKO lab custom-built environmental chamber with temperature control and CO2 stage incubator, and Nikon Elements acquisition software. High-resolution images (Figures 2, S2E, 5B, 5D, and 7F) were acquired using a Plan Apo 60× 1.4NA objective, generating an effective pixel size of 108 nm. Low-resolution images (Fig. S2C) were acquired using a Plan Fluor 20× 0.45NA objective, generating an effective pixel size of 325 nm.

To quantify vimentin filament detection rate, low-resolution images were captured from random locations in the interior of DAPI- and vimentin immunofluorescence-labeled monolayers. To enhance contrast, DAPI images were filtered using the histogram equalization function in ImageJ. Images were segmented using Cellpose version 2.0 with the pretrained “nuclei” model (Stringer et al. 2021) using the contrast-enhanced DAPI image as input, producing cell nucleus masks. Masks were expanded outward up to 20 pixels without overflowing into neighboring regions, and the median vimentin immunofluorescence signal within each mask was calculated. While this simple segmentation algorithm captured the central portion, but not the distal ends, of spindle-shaped RPE-1 cells, relatively uniform vimentin immunofluorescence signal throughout the cell allows these masks to provide reasonable estimates of the signal median (Fig. S2C). Cells were annotated as vimentin filament-positive or vimentin filament-negative based on comparison of the median vimentin signal to a threshold set by applying Rosin’s method across the distribution of all cells in the dataset (Rosin 2001). The rate of vimentin filament-positive cells then was calculated separately for each independent sample.

Characterization of vimentin oligomerization

Keratinocytes were grown to confluence and switched to Differentiation Medium 24 hours prior to lysate collection. We tested several lysis buffer recipes with the goal of at least partially preserving any vimentin oligomeric complexes: Modified 6.5M urea buffer as described above with DTT omitted; modified reporter lysis buffer (RLB) containing 20mM HEPES, 250mM NaCl, 10% Glycerol, 50mM NaF, 1mM Na₂VO₄, 1% Titron X100, and Halt Protease Inhibitor Cocktail (ThermoFisher, #78430), adjusted to pH 7.5; and modified RIPA containing 50mM HEPES, 5mM EDTA, 150mM NaCl, 0.1% SDS, 0.5% Deoxycholate, 1% Titron X100, and Halt Protease Inhibitor Cocktail, adjusted to pH 7.5. Cell lysates were collected in the applicable lysis buffer using a cell scraper, then sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) for 2 min at 20% power with an alternating 10 s on/off cycle at 4°C. Any remaining insoluble material was pelleted by centrifugation at 16.1×10³ g for 20 minutes at 4 °C, and the supernatant was transferred to a clean tube. Protein concentrations were measured by BCA assay (ThermoFisher #23228). Laemmli sample buffer (BioRad #161–0747) with or without 2-mercaptoethanol (ThermoFisher #BP176) was added to the samples, which were then heated to 90°C or 95°C for 3 to 7 minutes as indicated. Samples were then run on gels and analyzed as described in “Western blotting” above.

To more reliably preserve vimentin oligomers during lysate preparation, cells were washed with PBS, then incubated in 1 mM dithiobis(succinimidyl propionate) (DSP, ThermoFisher #22585) on ice for 1-hour to covalently crosslink interacting proteins. The reaction was quenched by incubating in 25 mM Tris, pH 7.5 on ice for 15-minutes. Cells were then washed with PBS and lysed in modified RIPA buffer, sonicated, and cleared by centrifugation, and heated in sample buffer with or without 2-mercaptoethanol as described above. Because the addition of 2-mercaptoethanol, which cleaves DSP crosslinks, failed to recover vimentin monomers on its own, a variation of this procedure also included addition of urea to reach a final concentration of 6.5M to the sample prior to heating. Samples were run on gels and analyzed as described in “Western blotting” above.

Detection of post-translational modifications

Keratinocytes were grown to confluence and switched to Differentiation Medium 24 hours prior to lysate collection. For experiments probing for phosphorylation, cells were treated with 10mM sodium orthovanadate (Sigma #S6508) for 2 hours immediately prior to lysis. Cells were collected in modified RIPA (described in “Characterization of vimentin oligomerization” above), sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) for 2 min at 20% power with an alternating 10 s on/off cycle at 4°C, and cleared by centrifugation at 16.1×10³ g for 20 minutes at 4 °C. Protein concentrations were measured by BCA assay. Vimentin was isolated from the cell lysates by immunoprecipitation. For each sample, lysate containing 500 μg protein was combined with 3 μL of anti-vimentin antibody (ThermoFisher #MA5–35320) or isotype control (ThermoFisher #02–6102) and 15 μL of Pierce Protein A/G Magnetic Beads (ThermoFisher #88802). The samples were then incubated with rotation for 2 hours to overnight at 4°C. Remaining lysate (input) was combined with Laemmli sample buffer (BioRad #161–0747) with 2-mercaptoethanol (ThermoFisher #BP176). Following incubation, the beads were washed three times with PBS, then bound protein was eluted directly in Laemmli sample buffer with 2-mercaptoethanol and 1mM DTT. All samples were denatured at 95 °C for 5 minutes and analyzed by Western blot. Phosphorylation was detected pan-phosphotyrosine or pan-phosphoserine/threonine antibodies (Supplemental Table S2). Citrullination was detected using an anti-citrulline (modified) detection kit (Sigma #17–347B) according to the manufacturer’s instructions.

Cutaneous squamous cell carcinoma and skin wound tissue

We obtained anonymized formalin-fixed paraffin-embedded (FFPE) tissue from two skin biopsies of well-differentiated squamous cell carcinomas and a skin excision which incidentally captured the healing wound edges from a prior diagnostic biopsy. Hematoxylin and eosin (H&E) stained sections on glass slides were imaged using a Hamamatsu NanoZoomer S60 whole-slide scanner with a 20× 0.75NA objective yielding an effective pixel size of 0.46 μm. We also obtained additional unstained sections on glass slides or as curls, which were processed for standard immunofluorescence or expansion as described below. This study was approved by the UT Southwestern Institutional Review Board.

Tissue immunofluorescence

In preparation for immunofluorescence, unstained sections were deparaffinized by 3 successive 5-minute incubations in xylene followed by rehydration in 3 successive 3-minute incubations in 100% ethanol and 2 successive 2-minute incubations in deionized water. For antigen retrieval, slides were placed in Tris-EDTA buffer (10 mM Tris base, 1 mM EDTA, 0.05% Tween 20, pH 9), heated to 95 °C in a microwave oven, and incubated until the buffer cooled to room temperature, approximately 45 minutes. Slides were rinsed three times in Hanks Balanced Salt Solution (HBSS) for 5 minutes each, and a PAP pen was used to circle the tissue sections. Slides were then incubated in blocking solution [5% goat serum (Abcam ab7481), 5% bovine serum albumin (Equitech-Bio BAH65–0500), and 0.5% Triton X-100 (Sigma X100) in PBS] for 1 hour. Primary antibodies (Supplemental Table S2) were diluted in PBS and incubated on the tissue sections overnight at 4 °C in a humidified chamber. Following primary antibody incubation, slides were washed 3 times with HBSS for 10 minutes each. Secondary antibodies (Supplemental Table S2) were diluted in PBS and incubated on the tissue sections for 60 minutes. Finally, the slides were washed 3 times with HBSS for 10 minutes each, and a small amount of Fluoromount-G with DAPI (ThermoFisher 00-4959-52) was used to affix coverslips to the slides.

Immunofluorescence slides were imaged using an inverted epifluorescence Nikon ECLIPSE Ti microscope equipped with a Zyla sCMOS camera (Andor), SOLA solid state white-light excitation system, and μManager acquisition software (Edelstein et al. 2010). Images were acquired using Plan Fluor 10× 0.3NA (650 nm/pixel), Plan Apo 20× 0.75NA (325 nm/pixel), and Plan Fluor 40× 1.3NA objectives (162 nm/pixel). To create large field of view images, partially overlapping images were stitched together pairwise using the Fiji Image Stitching Plugin (Preibisch et al. 2009).

Tissue expansion imaging

5-μm thick FFPE section curls were incubated at 60 °C in xylenes for 30 minutes twice, then carefully transferred to fresh, pre-warmed xylenes at 37 °C and incubated for another 30 minutes. Samples were then rehydrated through two 30-minute incubations in 100% ethanol, one 30-minute incubation in 70% ethanol, and one 30-minute incubation in 50% ethanol. Samples were washed three times briefly with 1×PBS, then transferred to pre-warmed antigen retrieval buffer (10 mM Tris base, 1 mM EDTA, 0.05% Tween 20, pH 9) at 95 °C and allowed to cool to room temperature (30 minutes). Samples were rinsed with 1×PBS three times and stored overnight at 4 °C. Samples were anchored in a solution of 2% acrylamide and 1.4% formaldehyde in 1×PBS for 5 hours at 37 °C without agitation. Each sample was then transferred to an 0.5 mL microcentrifuge tube containing 90 μl of monomer solution [MS; 23% w/v sodium acrylate (AmBeed A107105), 10% w/v acrylamide (Bio-Rad #1610140), 0.1% w/v N,N’-methylenebisacrylamide (Sigma #146072) in 1× PBS] with 5 μl of 10% TEMED (Bio-Rad #1610800; final concentration 0.5%) and left to infuse for 1 hour at 4 °C. After monomer infusion, skin sections were transferred to an ice-cooled (4°C) gelation chamber filled with fresh MS with 0.5% TEMED and 0.5% APS APS (Bio-Rad #1610700). Gels were allowed to polymerize for 1 hour at 37 °C in humidified conditions, then trimmed to remove excess gel. Trimmed hydrogels were incubated in denaturation buffer (200 mM SDS, 200 mM NaCl, 50 mM Tris-Base in deionized water) overnight at 95 °C until gels became transparent and the original tissue samples were no longer visible, indicating completion of the denaturation process. To enable full physical expansion and removal of residual denaturing agents, gels were transferred to deionized water for 90 minutes with the water exchanged every 30 minutes. Following enlargement, gels were measured to determine the expansion factor and processed for indirect immunofluorescence. Gels were shrunk with 1×PBS for 15 minutes, then blocked with 3% BSA in PBS for 1 hour at room temperature with agitation. Gels were incubated with primary antibodies (Supplemental Table S2) overnight at 4°C then washed three times with PBST (0.1% Tween-20 in PBS), and incubated with secondary antibodies (Supplemental Table S2) and DAPI overnight at 4 °C with agitation. After three washes with PBST, gels were incubated with Atto 565 NHS-ester (1:5,000; Sigma #72464) in 1×PBS for 1 hour with constant agitation. Finally, gels were washed briefly with 1×PBS and incubated in deionized water overnight to allow complete expansion and facilitate removal of unbound antibodies. Gels were placed on 35 mm glass-bottom dishes (Mattek P35G-1.5–20) and imaged with a Nikon CSU-W1 confocal microscope equipped with a Hamamatsu Orca Fusion Digital CMOS camera. Images were acquired with a Plan Apochromat Lambda D 20× 0.8NA objective using the 405 nm, 488 nm, 561 nm, and 640 nm laser lines, generating an effective pixel size in object space of 324 nm.

Live fluorescence imaging

Cells were seeded on collagen-coated #1.5 glass-bottom culture dishes (Cellvis D35-20-1.5-N) and grown to 25% confluence. Cells were imaged using an inverted phase-contrast and epifluorescence Nikon ECLIPSE Ti microscope equipped with a motorized and programmable stage, a Nikon Perfect Focus System, a Hamamatsu Orca-Flash4.0 scientific CMOS camera, a SOLA solid state white-light excitation system, an OKO lab custom-built environmental chamber with temperature control and CO2 stage incubator, and Nikon Elements acquisition software. The environment was maintained at 37 °C with 5% CO₂ during the imaging period. Images were acquired using a Nikon Plan Apo DIC H 100× 1.40NA objective generating an effective pixel size in object space of 65 nm. Annotated videos were prepared in ImageJ (Schneider et al. 2012) using the Scale Bar, Time Stamper (Schindelin et al. 2012), and draw_arrows_in_movies (Daetwyler et al. 2020) plugins and encoded using ffmpeg.

Single-cell RNA sequencing reanalysis

Processed single-cell RNA sequencing datasets were downloaded from the National Center for Biotechnology Information Gene Expression Omnibus and analyzed using Seurat version 4.4.0 in R (Hao et al. 2021).

For datasets from Sole-Boldo et al. (2020) (GSE130973) and Deng et al. (2021) (GSE163973), processed Seurat objects including normalization, clustering, dimensionality reduction, and cell type annotations were loaded directly from deposited R data files. These processed datasets were used as-is with only cosmetic changes to the annotated cell type names, and thus exactly match the original analyses.

Processed Seurat objects were not available for the other datasets, so read-count data were analyzed following the methods reported in their respective publications to the extent possible. Nevertheless, there are minor differences in the normalization and clustering results, possibly reflecting differences in software versions and CPU architecture which influence the pseudo-random number generator used for these algorithms.

For the data from Shiu et al. (2022) (GSE203262), cells with fewer than 200 features, more than 4,500 features, or more than 20% features mapped to mitochondrial genes were excluded. Samples were normalized and integrated using the sctransform routine with 2,000 variable features and the first 30 principal components. Clustering was done with a resolution parameter of 0.5.

For the data from Wiedemann et al. (2023) (GSE202352), one file could not be read (subject 3, hip dermis) and was excluded from the analysis. Cells with fewer than 300 features, more than 4,500 features, or more than 20% features mapped to mitochondrial genes were excluded. Samples were integrated using the sctransform routine with 2,000 variable features and the first 30 principal components. Clustering was done with a resolution parameter of 0.5.

For the data from He et al. (2020) (GSE147424), deposited data files were already filtered and log-normalized. Anchor-based integration was performed using canonical correlation analysis with 2,000 variable features and the first 20 principal components. Clustering was done with a resolution parameter of 0.5.

For the human acute skin wound single-cell RNA sequencing data from Liu et al. (2025) (GSE241132), cells with fewer than 200 features, more than 4,500 features, or more than 20% features mapped to mitochondrial genes were excluded. Samples from each timepoint were normalized and integrated separately using the sctransform routine with 4,000 variable features and the first 40 principal components. Cell clusters and type annotations were assigned from the deposited metadata table, with cell type names cosmetically adjusted to match the other datasets.

Keratin immunoprecipitation

K5^high and K6A^high cell lines expressing FLAG-tagged keratin constructs were grown to confluence and switched to Differentiation Medium 24 hours prior to sample collection. These conditions were chosen to maximize the formation of stable keratin filament bundles. Cells were washed with PBS, then incubated in 1 mM dithiobis(succinimidyl propionate) (DSP, ThermoFisher #22585) on ice for 1-hour to covalently crosslink interacting proteins. The reaction was quenched by incubating in 25 mM Tris, pH 7.5 on ice for 15-minutes. Cells were then washed with PBS and lysed in an immunoprecipitation buffer reported to partially solubilize keratins (40 mM HEPES, 120 mM sodium chloride, 1 mM EDTA, 1% Triton X-100, 2% Empigen BB [Sigma-Aldrich #30326], and protease inhibitors [Halt Protease and Phosphatase Inhibitor, Thermo], pH 7.5) (Lowthert et al. 1995; Wang et al. 2016). After incubation on a shaker at 4 °C for 20-minutes, lysates were sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) twice for 10 seconds at 20% power with a 10 second intervening rest. The lysates were then centrifuged at 16.1×10³ g for 20 minutes at 4 °C, producing a substantial pellet (Fig. 5G). The supernatant was transferred to a clean tube and protein concentrations were measured using a BCA assay (ThermoFisher #23228).

For each sample, 150 μL of anti-FLAG M2 magnetic beads slurry (Millipore Sigma M8823) was prepared in a microcentrifuge tube by washing three times with tris-buffered saline (TBS; 50 mM Tris, 150 mM sodium chloride, pH 7.4) and resuspending in 50 μL of TBS. Lysate with 500 μg of protein was added to the beads and incubated with rotation at 4 °C for 2-hours. Remaining lysate (input) was combined with Laemmli sample buffer (BioRad #161–0747) with 2-mercaptoethanol (ThermoFisher #BP176). Following incubation, the beads were washed three times with TBS, then bound protein was eluted directly in Laemmli sample buffer with 2-mercaptoethanol. All samples were denatured at 95 °C for 5 minutes and analyzed by Western blot.

Vimentin immunoprecipitation

Wild-type keratinocytes, keratinocytes expressing FLAG-tagged vimentin^Y117L, or keratinocytes expressing mitochondrially-targeted keratin head domain fragments (K5^h-mRuby3-mito, K6A^h-mRuby3-mito) along with vimentin^Y117L were grown to confluence in K-SFM but were not switched to Differentiation Medium to maximize the amount of IF protein recoverable in immunoprecipitation-compatible lysis buffer. Cells were washed with PBS and lysed by scraping into RIPA buffer (50 mM Tris, 150 mM sodium chloride, 0.5 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate, 0.1% sodium dodecyl sulfate, 50 mM sodium floride, 1 mM sodium orthovanadate, and protease inhibitors [Halt Protease and Phosphatase Inhibitor, Thermo], pH 7.5). After incubation on a shaker at 4 °C for 20 minutes, lysates were sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) six times for 10 seconds at 20% power with a 10 second intervening rest. The lysates were then centrifuged at 16.1×10³ g for 20 minutes at 4 °C, and the supernatants transferred to clean tubes. Protein concentrations were measured using a BCA assay (ThermoFisher #23228). Additional RIPA buffer was added as needed to adjust the protein concentration of each sample to 833 μg/mL.

A portion of each sample was reserved, combined with Laemmli sample buffer (BioRad #161–0747) with 2-mercaptoethanol (ThermoFisher #BP176), denatured at 97 °C for 5 min, and used as input fractions for downstream analysis. For immunoprecipitation, 500 μL of lysate was transferred to a clean tube with mouse anti- FLAG antibody (Sigma #B3111; clone M2) and incubated overnight at 4 °C to allow antibody binding. If applicable, lysates without antibody were used as controls for non-specific binding to the beads. For each sample, 25 μL of Protein A/G magnetic beads slurry (ThermoFisher #88802) was prepared in a microcentrifuge tube by washing twice with Wash Buffer (25 mM Tris, 150 mM sodium chloride, 0.05% Tween-20, pH 7.5), then combined with pre-complexed lysate-antibody mixture or lysate-only control. Sample-beads mixtures were incubated with gentle mixing for one hour. Following incubation, the beads were washed three times with Wash Buffer and once with distilled water, then bound protein was eluted directly in Laemmli sample buffer with 2-mercaptoethanol. Samples were denatured at 97 °C for 5 minutes and analyzed by Western Blot.

Cell fractionation for crosslinked keratin filaments

K5^high and K6A^high cell lines expressing FLAG-tagged keratin constructs were grown to confluence and switched to Differentiation Medium 24 hours prior to sample collection. Cells were washed with PBS, then incubated in 1 mM dithiobis(succinimidyl propionate) (DSP, ThermoFisher #22585) on ice for 1-hour to covalently crosslink interacting proteins. The reaction was quenched by incubating in 25 mM Tris, pH 7.5 on ice for 15-minutes. Cells were then washed with PBS and lysed in 6.5 M urea buffer (50 mM tris, 1 mM EGTA, 50 mM sodium fluoride, 1 mM sodium vanadate, and Halt Protease Inhibitor Cocktail [ThermoFisher #78430], pH 7.5; the reducing agent DTT, which would cleave the DSP crosslinker, was excluded from the buffer).(Wang et al. 2016) After incubation on a shaker at 4 °C for 20-minutes, lysates were sonicated using a Fisherbrand Model 505 Sonic Dismembrator (ThermoFisher) twice for 10 seconds at 20% power with a 10 second intervening rest. The lysates were then centrifuged at 16.1×10³ g for 20 minutes at 4 °C. Unlike keratinocyte lysates prepared without crosslinking in 6.5 M urea buffer as described above in “Western blotting,” the crosslinking procedure resulted in a noticeable pellet. The supernatants, representing the soluble fraction, were transferred into clean tubes and combined with Laemmli sample buffer (BioRad #161–0747) with 2-mercaptoethanol (ThermoFisher #BP176). The pellets, representing the insoluble fraction, were washed in lysis buffer then dissolved directly in Laemmli sample buffer with 2-mercaptoethanol. All samples were denatured at 95 °C for 5 minutes and analyzed by Western blot.

In-vitro purification and reconstitution of intermediate filaments

Keratin or vimentin constructs in pEt24d(+) were expressed in Rosetta2 pLysS bacteria as inclusion bodies. The proteins were purified in multiple steps consisting of isolation of the inclusion bodies, solubilization in guanidinium chloride (GuHCl) for vimentin or in urea for keratins, and purification by size-exclusion chromatography. Briefly, the cell extract obtained after sonication of the frozen bacteria pellet in a buffer containing 50 mM Tris HCl pH 8, 200 mM NaCl, 25% Glycerol, 1 mM EDTA, 10 mg/ml lysozyme, 20 mM MgCl2, 8 ug/mL DNase 1, 40 ug/mL RNase A, 1% NP40, 1% Deoxycholic Acid sodium, and protease inhibitors (Roche cOmplete Protease Inhibitor Cocktail), was centrifuged for 30min at 12,000×g, at 4 °C. The inclusion bodies pellet was washed 3 times using cycles of resuspension/centrifugation (30 min at 12,000 g, at 4 °C) in a buffer containing 10 mM Tris HCl pH 8, 0.5% Triton X-100, 5 mM EDTA, 1.5 mM DTT, protease inhibitors (Roche cOmplete Protease Inhibitor Cocktail). As a last step the inclusion bodies pellet was washed in a buffer containing 10 mM Tris HCl pH 8, 1 mM EDTA, 1.5 mM DTT, and protease inhibitors (Roche cOmplete Protease Inhibitor Cocktail) (Herrmann et al. 2004).

For purification of vimentin^head-eGFP-6His, the supernatant of the cell extract produced as described above was supplemented with 2 M Guanidinium chloride and 20 mM imidazole, and adjusted to 500 mM NaCl and pH7.4. The peptide was purified by affinity chromatography using a His Trap FF (5 mL, Cytiva). The column was first washed with a buffer containing 20 mM Tris HCl pH 7.4, 20 mM imidazole, 500 mM NaCl, 5 mM 2-Mercaptoethanol, 0.1 mM PMSF, and 2 M Guanidinium chloride. This was followed by an additional wash using the same buffer composition, except that 2 M Urea was used instead of Guanidinium chloride. Bound proteins were eluted with an imidazole gradient from 20 to 500 mM, with protein elution occurring at approximately 155 mM imidazole.

For full-length intermediate filament constructs, the purified inclusion body pellets were further processed to produce renatured proteins and filaments. For vimentin, the inclusion body pellet was solubilized with 6M GuHCl in 10 mM TrisHCl pH 7.5, and clarified for 30 min at 10,000×g at 4 °C, and the vimentin-containing supernatant was collected. A size-exclusion chromatography using Superdex 200 Increase 10/300 GL (Cytiva) was performed and protein purity was checked using SDS-PAGE. For full-length vimentin, the protein concentration was adjusted to 0.2–0.4 mg/ml and renatured by serial dialysis of 30 min at 22 °C using buffers of decreasing concentration of GuHCl, followed by an overnight dialysis step at 4 °C in the GuHCl free buffer. The dialysis buffers were 5 mM Tris HCl pH 7.5, with 1 mM EDTA, 0.1 mM EGTA, 1 mM DTT, containing 4, 2, 0 M GuHCl. For filament reconstitution, the protein solution was dialyzed against a high-salt buffer containing 10 mM Tris HCl pH 7.5 with 100 mM KCl.

For keratins, the inclusion body pellets were solubilized with 6 M Urea, 10 mM 2-Mercaptoethanol in 50 mM Tris HCl pH 7.5, clarified for 30 min at 10,000×g, at 15 °C, and the keratin-containing supernatant was collected. A size-exclusion chromatography using Superdex 200 Increase 10/300 GL (Cytiva) was performed and protein purity was checked using SDS-PAGE. Equimolar mixes of type-I and type-II keratins were incubated for 30 min at 22 °C in the above buffer, and heterotypic complexes of type-I and type-II keratins were separated from the uncomplexed proteins using an anion-exchange chromatography (HiTrapQHP, 1 mL, Cytiva) with a 25-mL gradient of 0–200 mM GuHCl in above buffer. Keratin filaments were then reconstituted in vitro by successive dialysis in 6 M Urea, 10 mM 2-Mercaptoethanol in 25 mM Tris HCl pH 7.5 for 4 hours, 2 M Urea, 5 mM 2-Mercaptoethanol in 5 mM Tris HCl pH 7.5 for 2 hours, and 5 mM 2-Mercaptoethanol in 5 mM Tris HCl pH 7.5 overnight (Lee and Coulombe 2009).

For keratin filament assembly in the presence of vimentin head domain peptides, purified heterotypic complexes of type I and type II keratins (0.6 μM) were reconstituted as described above. Each vimentin head domain peptide (1.5 μM), equilibrated in 2M urea, was added at the onset of the final dialysis, after the keratins had been equilibrated in buffer containing 2 M urea. For the control samples without keratins, each vimentin head domain peptide, equilibrated in 2M urea, was subjected to renaturation by dialysis in 5 mM 2-Mercaptoethanol in 5 mM Tris HCl pH 7.5 for 2 hours, followed by overnight dialysis in the same buffer.

Fluorescent labeling of recombinant intermediate filament proteins

To produce fluorescently labelled keratin filaments, K14 was coupled to an Alexa fluor 488 C5-maleimide dye (Thermo Fisher scientific), complexed with a type-II keratin, and keratin filaments were reconstituted as described above. Fluorescently labelled vimentin was produced by coupling a Cy3b-maleimide dye (Cytiva) on the C-terminal cysteine of vimentin^C328A-GGC prior to renaturation as described above.

Briefly, an inclusion bodies pellet of K14 or vimentin was solubilized with 6 M GuHCl, 50 mM Phosphate buffer pH 7.0, and clarified for 30 min at 10,000×g, at 4 °C. The protein-containing supernatant was purified using a size-exclusion chromatography using Superdex 200 Increase 10/300 GL (Cytiva) in the same buffer. Purity was checked by SDS-PAGE. Pure protein fractions were labelled with either Alexa fluor 488 C5-maleimide dye or Cy3b-maleimide dye, and subsequently the free dye was removed using a size-exclusion chromatography using Superdex 200 Increase 10/300 GL (Cytiva) in 6 M Urea, 10 mM 2-Mercaptoethanol in 50 mM Tris HCl pH 7.5 buffer for K14-Alexa488, or in 6M GuHCl in 10 mM TrisHCl pH 7.5 for vimentin- Cy3b. For K14, Alexa488-labelled heterotypic complexes were produced by mixing the Alexa488-labelled K14 with either unlabeled K5, or K6A, or K6Ah5rt, as described above. To produce Alexa488-labelled keratin filaments, unlabeled heterotypic complexes were mixed with 18% of the Alexa488-labelled ones, and filaments were then reconstituted in vitro by successive dialysis, as described above. For vimentin, the unlabeled wild-type vimentin was mixed with 37% of vimentin^C328A-GGC-Cy3b. Note that the C328A substitution does not affect vimentin filament assembly (Winheim et al. 2011). The protein mix was then renatured as described above.

Negative stain electron microscopy

To assess the interaction of vimentin head domain peptides with keratin filaments and verify the ability of labelled keratin and vimentin to polymerize, samples were inspected by negative stain electron microscopy. 3 μl of the filament solution was applied onto glow-discharged carbon-coated EM grids (G2300, Plano GmbH), incubated for 1 minute, and fixed with 1% uranyl acetate. Excess stain was removed by blotting. Specimen were examined in an FEI Tecnai G2 Spirit transmission electron microscope with a voltage of 120 kV.

In-vitro co-sedimentation of recombinant intermediate filament proteins

Keratin filaments were reconstituted in using purified heterotypic complexes of K5/K14, K6A/K14, or K6A^h5^rt/K14, 18% labelled with Alexa Fluor 488 dye as described above. Unpolymerized, renatured vimentin, 37% labelled with Cy3b dye, was prepared separately. The three types of keratin filaments were incubated alone or with the unpolymerized vimentin for 30 min at 22 °C, and subsequently centrifuged at 20000×g for 30 minutes. Pellets were analyzed on a 9% SDS-PAGE and quantified by densitometry. Alexa Fluor 488 and Cy3b fluorescence signals were imaged before Coomassie staining using FUSION FX7 Spectra (Vilber Smart Imaging) and quantified with Fiji software. Band fluorescence intensity measurements were made by integrating pixel densities in the band of interest and subtracting the contribution of the background. To quantify the ratio of vimentin co-sedimenting with keratin filaments, the band intensity of vimentin sedimented while incubated alone was subtracted from the total band intensity of vimentin sedimented in the presence of K5/K14, or K6A/K14, or K6A^h5^rt/K14 filaments, and subsequently divided by total band intensity of keratin in the corresponding pellet. Results were collected from two distinct experiments using the same batches of heterotypic complexes K5/K14, or K6A/K14, or K6A^h5^rt/K14, 18% labelled with Alexa Fluor 488 dye, and vimentin 37% labelled with Cy3b dye stored in urea or GuHCl, respectively.

Traction force microscopy

Silicone gel substrates for traction force microscopy (TFM) were prepared in #1.0 glass-bottom 35-mm tissue culture dishes (Cellvis D35-14-1.0-N). Prepolymers for high-refractive index (n = 1.49) 8-kPa gels were prepared by mixing components A and B of QGel 920 (Quantum Silicones) in a 1:1.15 ratio as previously described (Gutierrez et al. 2011; Han et al. 2015). The prepolymers were spread onto the coverglass bottoms of the dishes by spinning at 1,250 r.p.m for 30 seconds in a spin-coater (Laurell Technologies WS-650Mz-23NPPB), then cured at 100 °C for 2 hours, producing gels approximately 35-μm thick. The gels were then treated with 5% (3-aminopropyl)triethoxysilane (APTES; Sigma #A3648) in ethanol to functionalize their surfaces. To allow for visualization and tracking of substrate deformation, 40-nm carboxylate far-red fluorescent beads were covalently linked to the gel surfaces by incubating the gels under a suspension of the beads (1:5,000 dilution from the 5% stock suspension) in 40-mM HEPES, pH 8, with 0.01% 1-ethyl-3-(3-dimethylaminopropyl)carbodiimide (EDC; Thermo Scientific #22981) as a catalyst. To facilitate cell adhesion during cell seeding, the substrates were coated with fibronectin by incubation with 50-μg/ml of fibronectin and 0.1 mg/ml EDC in PBS for 15 minutes at room temperature. The coated dishes were washed three times with PBS, incubated with 0.1% bovine serum albumin in PBS for 30 minutes, and finally washed three more times with PBS.

To measure keratinocyte traction force generation, 4×10⁴ cells were seeded onto a substrate in Differentiation Medium, then returned to the incubator for 4 hours to allow attachment to occur. For knockdown experiments, cells were transduced with lentivirus to express targeting shRNA or a scrambled control 72 hours before seeding onto TFM substrates. For ROCK inhibition experiments, 1 μM GSK 269962A (Selleck Chemicals #S1474) was added to cells 1 hour prior to imaging. If indicated, cells were treated with 1% (weight/volume) 1,6-hexanediol (Sigma-Aldrich #240117) in Differentiation Medium for 5 minutes prior to imaging. Following attachment, cells were imaged using a GE DeltaVision OMX SR inverted microscope equipped with a motorized and programmable stage, three PCO sCMOS cameras, solid state lasers (405nm, 488nm, 568nm, and 640nm), and environment control. Images of the substrate beads and keratin tags were acquired in Ring-TIRF mode with a U Plan Apo 60× 1.49NA objective (generating an effective pixel size in object space of 80 nm), with imaging performed at 37 °C. Following initial imaging, during which cell positions were recorded, a prewarmed solution of dilute bleach was added to remove cells from the substrate, thereby allowing repeat imaging of the same locations with the substrate in its relaxed, strain-free state.

Image analysis was performed using the previously published u-inferforce package written in MATLAB (Han et al. 2015; Han et al. 2021). Briefly, the substrate deformation field was determined by tracking the displacement of each bead between the deformed and relaxed images using cross-correlation analysis. The displacement field, after outlier removal and stage drift correction, was then used to estimate the traction field over areas of interest surrounding each cell. Traction fields were estimated using the boundary element method (Han et al. 2015). While more computationally intensive than the more commonly used Fourier transform traction cytometry method, the boundary element method allows L1-norm rather than L2-norm regularization, which resolves force variation at shorter length scales (Han et al. 2015). A regularization parameter of 0.05 was selected based on L-curve analysis of a subset of images (Han et al. 2015). This parameter value was applied to all images analyzed to permit comparison of the resulting force fields. Cell area was segmented based on either the cytosolic fluorescent protein (cells transduced with pLVX-Vimentin-FLAG-IRES-mNeonGreen) or tagged keratin signals. Since the tagged keratin signal did not always clearly extend to the cell boundary, masks were dilated by 1.6 μm (20 pixels) when this signal was used, although we previously verified that the precise amount of dilation did not meaningfully affect the results (Nanes et al. 2024). Strain energy, which represents the mechanical work of the cell on the elastic substrate, was calculated as ½ × (displacement · traction), integrated over the segmented cell area (Butler et al. 2002).

Proximity ligation assay

Mixed populations of keratinocytes were prepared and seeded onto 8-chamber glass slides with removable wells (Lab-Tek, ThermoFisher #154534) or collagen-coated 12 mm-diameter #1.5 coverslips (Warner Instruments #64–0712) in tissue culture dishes. Cells were grown to approximately 50% confluence, then culture medium was switched from K-SFM to Differentiation Medium for 24 hours. For knockdown experiments, cells were transduced with lentivirus to express targeting shRNA (Supplemental Table S1) or a scrambled control 72 hours before processing. If indicated, cells were treated with 1% (weight/volume) 1,6-hexanediol (Sigma-Aldrich #240117) in Differentiation Medium for 5 minutes or 1 μM ROCK inhibitor GSK 269962A (Selleck Chemicals #S1474) in Differentiation Medium for 1 hour at 37 °C just prior to fixation. Slides were removed from the incubator and washed three times with PBS. For experiments using antibodies against keratins, the slides were incubated in methanol for 2 minutes. For other experiments, the slides were incubated in 4% paraformaldehyde in PBS for 10 minutes, then incubated in 0.1% Triton X-100 in PBS for 8 minutes. Slides were then washed twice with PBS and processed using Duolink Proximity Ligation Assay (PLA) reagents (In-Situ PLA Probe Anti-Mouse MINUS, Sigma DUO92004; In-Situ PLA Probe Anti-Rabbit PLUS, Sigma DUO92002; In-Situ Detection Reagents FarRed, Sigma DUO92013) according to the manufacturer’s in situ fluorescence protocol. Briefly, slides were incubated in the provided Blocking Solution for 60 minutes, incubated in primary antibodies diluted in the provided Antibody Dilutant solution for 30 minutes (Supplemental Table S2), incubated in PLA probe solution for 60 minutes, incubated in the ligation solution for 30 minutes, and incubated in the amplification solution for 100 minutes, with washes in the provided buffers between each step. Finally, a #1.5 cover glass was affixed to the slide using Duolink In-Situ Mounting Medium with DAPI (Sigma DUO82040).

Slides were imaged using an inverted phase-contrast and epifluorescence Nikon ECLIPSE Ti microscope equipped with a motorized and programmable stage, a Nikon Perfect Focus System, a Hamamatsu Orca-Flash4.0 scientific CMOS camera, a SOLA solid state white-light excitation system, an OKO lab custom-built environmental chamber with temperature control and CO2 stage incubator, and Nikon Elements acquisition software. Images were acquired on the tagged keratin, PLA reaction, and DAPI channels using a Plan Apo 60× 1.4NA objective (generating an effective pixel size in object space of 110 nm). For some experiments, z-stack images were captured with 5 planes at 300 nm intervals, and maximum intensity projections were used for downstream analysis. Image analysis was performed using custom Python and MATLAB scripts. Individual cells were segmented using Cellpose version 1.0 (Figures 7C–D, 8H, and S10A–B) or 2.0 (Figures 7B, E, G; 8B; S9; and S10C) with the pretrained “cyto” model (Stringer et al. 2021). A normalized sum of the tagged keratin fluorescence signals was used as input for the cytoplasm channel, and DAPI signal was used as input for the nuclear channel. The median tagged keratin fluorescence signal was calculated for each cell. Cells were annotated as expressing high or low levels of each keratin using a linear discriminator function developed based on thresholds set between bimodal peaks in the fluorescence signal distributions. Segmentations and classifications were manually reviewed, and any incorrectly segmented or classified cells were excluded from analysis. Cells that were poorly adhered to the slide or out of focus were also excluded. Two algorithms were used to identify PLA detection signals. For some experiments Figures 7C–D, 8H, and S10A–B), individual local intensity clusters in the scaled PLA image were detected using previously published MATLAB code combining wavelet denoising and multiscale products of wavelet coefficients (Olivo-Marin 2002; Aguet et al. 2013). For other experiments (Figures 7B, E, G; 8B; S9; and S10C), a Difference of Gaussians blob detector from the scikit-image Python library (van der Walt et al. 2014) was applied to the scaled PLA image. Finally, PLA detections were counted within each segmented cell mask, and the density of PLA detections was compared between cells with different keratin expression levels. Linear mixed effects models were fit using the lme4 package (Bates et al. 2015) with fixed effects at the cell level and random effects at the image level. P values were calculated using the Kenward-Roger method implemented by the emmeans and pbkrtest packages (Halekoh and Højsgaard 2014).

Keratin filament network dynamics analysis

Live imaging of cells expressing fluorescently tagged keratins, filament network segmentation, dynamics score calculations were performed as previously described (Nanes et al. 2024). In brief, cells were seeded on collagen-coated #1.5 glass-bottom culture dishes (Cellvis D35-20-1.5-N) and transduced with lentivirus to express targeting shRNA (Supplemental Table S1) or a scrambled control 72 hours before analysis. Culture medium was switched from K-SFM to Differentiation Medium 8 hours before analysis. cells were imaged using an inverted phase-contrast and epifluorescence Nikon ECLIPSE Ti microscope equipped with a motorized and programmable stage, a Nikon Perfect Focus System, a Hamamatsu Orca-Flash4.0 scientific CMOS camera, a SOLA solid state white-light excitation system, an OKO lab custom-built environmental chamber with temperature control and CO2 stage incubator, and Nikon Elements acquisition software. The environment was maintained at 37 °C with 5% CO₂ during the imaging period. Images were acquired every 1 minute for 4 minutes using a Plan Apo 60× 1.4NA objective, generating an effective pixel size in object space of 108 nm.

Images were analyzed using the previously published u-deliniate MATLAB package (Gan et al. 2016), which used a multi-scale steerable filter and iterative graph matching to connect fragments into complete filaments. To analyze filament network motion within cells, segmented filament networks at baseline were compared to filament networks segmented from images taken 1, 2, 3, or 4 minutes later. A local structural similarity score defined to incorporate both distance and orientation information was computed across each pair of filament networks. Each pixel in the first network was matched to the closest pixel in the second network within a defined radius, and vice versa. A radius of 2.2 μm was used for all studies here. For each matched pair, the difference in position and filament orientation was calculated and used to create distance and angle maps, which were spatially smoothed by a Gaussian filter, combined, and scaled so the final similarity score decreased monotonically from 1 to 0 as the distance increased to the search radius or the angle increased to π/3. The dynamics score was defined as 1 minus the combined similarity score, averaged over each cell mask area, excluding pixels without a filament match. The final dynamics score is 0 for identical networks and increases to 1 with increasing local differences in filament position or orientation. Because this analysis cannot reliably distinguish movement of filaments within cells from movement of the cells themselves, all image sequences were pre-screened to ensure that only stationary cells were included in the analysis. As a check, a quality score was computed for each cell, defined as the fraction of the cell mask area with a filament match, which was typically higher than 0.9. Image sequences of cells with low quality scores were reviewed by visual inspection. They tended to show failed filament segmentation due to low signal-to-noise ratio, debris, or cell movement, which was missed in the initial screen; these cells were excluded from subsequent analysis.

Protein alignment and disorder prediction

Peptide sequences for human keratin 5 (KRT5; P13647), keratin 6A (KRT6A; P02538), and vimentin (VIM; P08670) were downloaded from UniProt and aligned using MUSCLE version 5 (Edgar 2022). Domains were annotated based on the UniProt records for each protein, except that all head domains were adjusted to start at position 1, rather than position 2, for consistency. Local disorder prediction scores were calculated using AIUPred version 2 with default settings (Meszaros et al. 2018). Alignment display was prepared with Jalview (Waterhouse et al. 2009).

Statistical analysis

All statistical analysis was performed using R (R Foundation for Statistical Computing). Unless otherwise indicated, the Kruskal-Wallis rank sum test was used to evaluate non-parametric scaled data, with Dunn’s method for multiple comparisons used for experiments involving more than two groups (Dunn 1964). Two-tailed P values less than 0.05 are considered to be significant and are shown in figures or indicated in legends.

Data availability

Proximity proteomics data are available from the MassIVE repository (DOI: 10.25345/C52Z13139). Label-free quantitation results and analysis code related to the proximity proteomics experiment are available as Supplemental Data.