Mkrn2^(−/−) (Mkrn2 KO) mice (aged 2–3 months, FVB background) were bought from cyagen biosciences company (Suzhou, China). These mice were then backcrossed for 10 generations to our C57BL/6 J strain. This Mkrn2 strain was maintained in C57 background. Mkrn2 knockout mice were generated by sibmating Mkrn2 heterozygotes. Genomic DNAs isolated from the tails were genotyped with indicated primers (Table 1) by PCR as described (Xu et al., 2018; Liu et al., 2024). Mice were housed in a specific-pathogen-free (SPF) facility with 12-h light/dark cycle and ad libitum access to food and water. Per cage was housed three to five mice by genotype. All behavioral experiments were performed during the light cycle. All animal studies were performed strictly in accordance with the instructions of the Institutional Animal Care and Use Committee (IACUC) at Shanghai Institute of Biochemistry and Cell Biology, CAS. Female and male mice were both used for experiments in this study, with C57BL/6 genetic background (SLAC, China).

The human embryonic kidney cell line HEK293T were purchased from the cell bank of Chinese Academy of Sciences (CAS). The human neuroblastoma cell line SHSY5Y was purchased from the American Type Culture Collection (ATCC, Manassas, United States). All cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM, high glucose), supplemented with 10% fetal bovine serum (FBS) and penicillin/streptomycin (all from Gibco, United States) in a 37 °C humidified atmosphere of 5% CO2. Plasmids used in this study were transfected into cells using polyethylenimine (Sigma) according to the manufacturer’s instructions. Briefly, cells were seeded into 6-well-plates, transfections were done when cells to be 90% confluent. 3 μg plasmids DNA and 9 μL polyethylenimine were diluted in 100 μL Opti-MEM Medium (Gibco, United States) respectively, incubated at room temperature for 5 min, and then mixed together for 20 min before added to cells. 48 h after trasfection, cells were harvested and prepared for Immunoprecipitation.

HEK293 or SHSY5Y-based mono or double-allele knockout cell lines were generated by genome editing using the CRISPR/Cas9 system (Li et al., 2020; Komor et al., 2017; Hsu et al., 2013). Five sgRNAs targeting MKRN2 was designed as previously reported (Li et al., 2020; Komor et al., 2017; Hsu et al., 2013) and three sgRNAs targeting CSDE1, as listed below (target sequences are lowercase):

Twenty-four hours after transfection, cells were screened by 2 μg/mL puromycin (Beyotime, China) for 7 days. Single-cell colonies were picked, amplified, and confirmed by immunoblotting analysis. We then extracted genomic DNAs, amplified the specific target sequences, and performed Sanger sequencing to verify the aimed edits in the genomes to established the stably transfected cell lines.

HEK293 or SHSY5Y control and MKRN2 KO cells were transfected with EGFP-MKRN2 or mCherry-CSDE1 or GFP-G3BP1 plasmids. After 24 h of expression, cells were digested and then spread on a 29 mm dish with 20 mm glass-bottom well. 48 h after transfection, cells were stressed with 100 mM sodium arsenite for 20–30 min at 37 °C before observation for Phase separation. Images were captured using Olympus SpinSR and Zeiss LSM900 confocal laser-scanning microscope with a 60x objective. Phase diagrams were constructed by measuring fluorescence intensity in each cell and assessing the presence of SGs using G3BP1 as a marker (Yang et al., 2020), using Fiji software.

In vitro LLPS experiments were performed at room temperature unless otherwise indicated. Purified EGFP-MKRN2 and mcherry-CSDE1 proteins were diluted with a phase separation buffer containing 25 mM Tris–HCl (pH = 7.4), 150 mM KCl, 2.5% Glycerol, 2% PEG20000, and 0.5 mM DTT, and the mixtures were incubated for 5 min to induce phase separation (Alberti et al., 2018; Liu et al., 2024). LLPS of CSDE1 and MKRN2 were induced by addition of indicated concentrations of PEG20000 or mRNAs. The samples were mixed in a 29 mm dish with 10 mm glass bottom well right before observation. Proteins were diluted to various salt (50–600 mM KCl) and protein (300–1600 nΜ) concentrations and observed using a Zeiss LSM900 confocal laser-scanning microscope equipped with a x2.5 or x20 objective. All imaged were captured within 5 min after LLPS induction. For LLPS experiments involving mRNA addition, total RNA was first isolated from HEK293 cells using TRNzol Universal RNA Reagent (Yeasen, China). The isolated RNA was then reverse-transcribed into cDNA. Selected target mRNAs, identified through prior RIP-seq analysis, were synthesized via in vitro transcription from the resulting cDNA.

Protein and mRNA concentrations were determined by measuring absorbance at 280 nm using a Nano-300 spectrophotometer (Allsheng, China) and Nanodrop (Thermo Fisher Scientific).

FRAP analysis was performed on the Olympus SpinSR confocal laser-scanning microscope. GFP or mcherry signals in regions of interest (ROI) were fully photobleached by using a 405-nm laser. Fluorescence intensity of ROI between pre-bleached and at the start of recovery after bleaching was recorded by microscope. FRAP data were analyzed using GraphPad Prism 9.

The plasmids containing CSDE1 and MKRN2 were amplified from the cDNA of human HEK293T cells using specific primers in Table 1, and inserted into pCDNA3.0, pET28a or pGEX4T-1 vectors that digested with BamH I or Xho I, and recombination with amplified fragments using the ClonExpress II One Step Cloning Kit (Vazyme, China) according to the manufacturer’s instructions. Five sgRNAs for MKRN2 were synthesized as oligos (Tingke, Shanghai, China), annealed and inserted into the lentiCRISPR v2 vector, transformation into Stbl3 bacteria, and all constructs were confirmed by sequencing.

GST tagged CSDE1 or His6- tagged MKRN2 were expressed in BL21 E. coli cells. After isopropyl-β-d-thiogalactopyranoside (IPTG, Sangon, China) induction, E. coli cells were pelleted by centrifuge, lysed in PBS buffer and incubated with glutathione or Ni²⁺ NTA agarose beads (Smart-Lifesciences and NanoMicro, China) to enrich the GST − or His6- tagged proteins, followed by elution with 20 mM reduced L-glutathione or 400 mM imidazole (Sangon) that dissolved in PBS buffer, and then dialysis in PBS buffer supplemented with final concentration of 25% glycerol before being aliquoted and preserved at −80 °C. The fluorescent proteins for LLPS followed the same procedures.

E1(UBA1), E2 (UBCH5A) and HA-UB were added to the first multiple cloning site (MCS) of the PACYC vector and MKRN2 was added to the second multiple cloning site to construct the plasmid pACYC-HA-UB- UBA1—UBCH5A -MKRN2, which was co-transformed with pET28a-CSDE1-His6 in BL21 E. coli competent cells and screened with ampicillin plus chloramphenicol antibiotics. After induction with IPTG, the cells were lysed with 8 M urea lysis buffer (50 mM Tris-Cl, 50 mM Na₂HPO₄, 300 mM NaCl, 8 M urea, 0.5% NP-40 and 20 mM imidazole, pH 8.0) and then incubated with Ni-NTA agarose beads for 4 h at room temperature. The immunoprecipitates were washed three times with 8 M urea lysis buffer and denatured in 2x SDS protein loading buffer at 100 °C for 10 min before immunoblotting analysis. The CSDE1 proteins recovered from the E. coli ubiquitination system were subjected to mass spectrometry analysis for ubiquitination site mapping as previously described (Xu et al., 2018).

Procedures for MS analysis were as previously described (Xu et al., 2018). In brief, the protein pellet was concentrated in 8 M urea, 100 mM Tris-Cl (pH 8.5), followed by TCEP reduction, NEM alkylation, and trypsin digestion. Peptides were separated and analyzed by the EASY-nLC system (Thermo Fisher) and the Q Exactive mass spectrometer (Thermo Fisher), respectively. Ubiquitylation modification sites were determined with Thermo Proteome Discoverer 2.1 (Thermo Fisher) and searched in Uniprot Human database.¹

Purified GST or GST-MKRN2 (30 μg), CSDE1- His6 proteins (30 μg) and Glutathione Sepharose4B (Sangon) were incubated at 4 °C overnight in 800 μL GST pull-down buffer (20 mM Tris–HCl, 100 mM NaCl, 1 mM EDTA, 5 mM MgCl2, 1 mM DTT, 0.5%NP- 40, 10 μg/mL of BSA, pH 7.4). These beads are then pelletized and washed three times with a pull-down buffer. The recovered immunoprecipitates were boiled in 2xSDS-PAGE protein loading buffer for 10 min, and then subjected to immunoblotting analysis with the indicated antibodies (Jia et al., 2020).

Exogenous or endogenous protein-expressing cells were lysed in IP buffer (50 mM Tris-Cl, pH 7.5, 150 mM NaCl, 5 mM EDTA, 1% NP-40, 0.1% SDS, 0.5% Sodium deoxychlate) and CO-IP buffer (50 mM Tris–HCl, pH7.5, 150 mM NaCl, 5 mM EDTA and 1% NP-40) supplemented with protease inhibitor cocktail (LabEAD) (Li et al., 2020). Resuspended cells were lysed using a probe sonicator with 20% amplitude, 1 s pulse-on/2 s pulse-off cycles with total duration ≥1 min on ice and centrifuged at 15,000 rpm for 10 min at 4 °C; the supernatants were collected, and then incubated with specific antibodies or anti-Flag M2 affinity gels (Sigma, United States) at 4 °C overnight. The beads are pelletized and washed 3 times with IP buffer. The antibodies used were as follows: normal rabbit IgG (1:5,000 dilution, 2729S, Cell signaling), anti-MKRN2 (1:2,000, A12219, ABclonal), and anti-CSDE1 (1:2,000 dilution, 13319-1-AP, Proteintech). The immunoprecipitates were enriched and denatured at 99 °C for 10 min in 2xSDS-PAGE loading buffer. The immunoprecipitates and inputs were subjected to SDS-PAGE and then were transferred to PVDF membranes (Bio-Rad, United States). The membranes were blocked with 10% skimmed milk at RT for 1 h, and were immunoblotted with the specified antibodies as follows: anti-MKRN2, anti-CSDE1 and anti-Flag (1:1,000 dilution, 20543-1-AP, Proteintech), anti-Myc (1:1,000 dilution, 16286-1-AP, Proteintech), anti-HA (1:1,000 dilution, HRP-81290, Proteintech), anti-His (1:2000 dilution, 10001-0-AP, Proteintech), anti-GST (1:10,000 dilution, HRP-66001, Proteintech), and anti-GAPDH (1:5,000 dilution, 60004-1-Ig, Proteintech). Then the membranes were incubated with HRP-conjugated goat anti-mouse IgG (1:5,000 dilution; SA00001-1, Proteintech) or goat anti-rabbit IgG (1:5,000 dilution; SA00001-2, Proteintech) secondary antibodies for 1 h at room temperature. The protein–antibody complex was detected by a chemiluminescence detection system via an ECL chemiluminescence detection kit, and the signals were visualized using a Tanon 5,200 Imaging System (Tanon, Shanghai, China).

RNA immunoprecipitation was performed as previously described (Keene et al., 2006) using MKRN2 knockout SH-SY5Y cells. Briefly, cells transfected with Flag, CSDE1-Flag, and MKRN2-Myc plasmids were collected after 48 h, and the cells pellet were resuspended with an approximately equal volume of polysome lysis buffer[100 mM KCl, 5 mM MgCl₂, 10 mM HEPES (pH 7.0), 0.5% NP40, 1 mM DTT, 100 units/mL RNase inhibitor (Vazyme), 400 μM vanadyl ribonucleoside complexes (VRC, NEB, United States) and Protease inhibitor cocktail (1:100, Roche)]. Incubated on ice for 10 min, centrifuged at 15,000 g for 15 min to clear lysate of large particles. Anti-FLAG M2 Affinity gels (Sigma, United States) were washed with NT2 buffer (50 mM Tris–HCl, 150 mM NaCl, 1 mM MgCl2, 0.05% NP-40 pH 7.4) before using for immunoprecipitation. 50 μL of pre-cleared supernatant lysate were saved as the total cellular mRNA (input), and the rest of RNA-protein complexes were immunoprecipitated with anti-FLAG M2 Affinity gels (Sigma, United States) in 850 μL of NT2 buffer supplemented with 100 units/mL RNase inhibitor, 400 μM VCR, 10 μL of 100 mM DTT and 20 mM EDTA, and incubated for 4 h at 4 °C with tumbling end over end. Beads were washed with 1.0 mL of ice-cold NT2 buffer for five times supplemented with 0.5 M urea and 0.1% SDS, and then resuspended the beads in 100 μL of NT2 buffer supplemented with 30 μg of proteinase K (Roche) to release the RNP components. RNAs were extracted using Trizol (Invitrogen), and subjected to RIP-seq analysis.

Total RNAs were extracted from cells using a Trizol reagent (Tiangen, China). Complementary DNA (cDNA) was synthesized using HiScript III RT SuperMix for qPCR (+gDNA wiper) (Vazyme #R323). Quantitative PCR (qPCR) assay was performed to determine the relative abundances of selected mRNAs using specific primers in Table 2, stained by SYBR Green (Tiangen, China) on the Roche LightCycler 480 II (384 well) real-time PCR system. The relative abundances of mRNAs were normalized to that of ATP2A1 mRNAs, using the 2^−∆∆Ct method as previously described (Xu et al., 2018). All experiments were repeated three times.

All results have been obtained from at least three independent experiments, data of experiments are expressed as mean ± standard error of the mean (SEM) and analyzed by two-tailed unpaired t-test and one-way ANOVA with Tukey’s post-hoc test using GraphPad Prism 9. (GraphPad Software Inc., United States). ^∗p < 0.05 was considered to be significantly different, ^∗∗p < 0.01 was considered to be very significantly different, ∗∗∗p < 0.001 was considered to be highly significantly different, and ∗∗∗∗p < 0.0001 was considered to be extremely significantly different.

To identify potential substrates of MKRN2, we performed coimmunoprecipitation (Co-IP) followed by mass spectrometry and identified CSDE1 as a candidate substrate. Co-IP assays confirmed that both endogenous and exogenously expressed MKRN2 interact with CSDE1 in SH-SY5Y cells (Figures 1A,B). Fluorescence colocalization analysis further demonstrated strong cytoplasmic co-localization of MKRN2 and CSDE1 (Supplementary Figure 1A).

To determine whether MKRN2 mediates the ubiquitination of CSDE1, we co-expressed Myc-MKRN2 and Flag-CSDE1 in HEK293T cells. Co-IP analysis confirmed their interaction and revealed that MKRN2 promotes CSDE1 ubiquitination (Figure 1C). Deletion of the RING finger domain in MKRN2 significantly impaired its ability to ubiquitinate CSDE1, both in HEK293T cells and in a reconstituted E. coli ubiquitination system (Figure 1D; Supplementary Figure 1B), underscoring the essential role of this domain. Notably, CSDE1 ubiquitination was largely dependent on MKRN2, with minimal contribution from E1 or E2 enzymes alone (Supplementary Figure 1B).

To further investigate the functional role of MKRN2 in CSDE1 ubiquitination, we generated MKRN2 knockout (KO) cell lines in both SH-SY5Y and HEK293 cells (Supplementary Figures 1C,D). Immunoprecipitation analyses revealed a marked reduction in CSDE1 ubiquitination in MKRN2-KO cells, accompanied by a modest increase in CSDE1 protein levels (Figure 1E; Supplementary Figure 1E). Re-expression of MKRN2 in KO cells partially restored CSDE1 ubiquitination, confirming that MKRN2 positively regulates this process. The residual ubiquitination observed in MKRN2-deficient cells suggests the involvement of additional E3 ligases or regulatory mechanisms.

To determine whether the MKRN2–CSDE1 interaction is direct, we performed in vitro GST pull-down assays using recombinant GST-MKRN2 and CSDE1-His₆ proteins. These assays confirmed direct binding between MKRN2 and CSDE1, independent of other cellular factors (Supplementary Figure 1F).

MKRN2 contains four conserved C3H zinc fingers, a Cys-His motif, and a C3HC4 RING domain, and has been implicated as a putative ribonucleoprotein (William et al., 2010). CSDE1 comprises five cold shock domains (CSD1–5) interspersed with unstructured linker regions and is known to function as an RNA-binding protein (Guo et al., 2019; Gangfuss et al., 2021). To map the domains responsible for MKRN2-mediated ubiquitination of CSDE1, we generated a series of truncated mutants for both proteins (Supplementary Figure 1G): four MKRN2 truncations (GST-MKRN2-T1 to T4) and three CSDE1 truncations (CSDE1-His₆-T1 to T3). GST pull-down assays showed that GST-MKRN2-T1, which contains the first three C3H zinc fingers, exhibited strong binding to CSDE1 (Figure 1F), indicating that the N-terminal C3H region of MKRN2 is primarily responsible for the interaction. Notably, in some experimental replicates including the representative data shown, the T4 fragment of MKRN2 also demonstrated a detectable interaction with full-length CSDE1. However, this interaction proved less consistent and robust compared to that mediated by the T1 fragment, and it was not reproducibly observed in subsequent domain-mapping assays using isolated truncations.

Reciprocal binding assays using truncated CSDE1 constructs revealed that the N-terminal cold shock domains (CSD1–5) of CSDE1 are critical for binding to the C3H zinc finger region of MKRN2 (Figures 1G,H). Collectively, these data demonstrate that the interaction between MKRN2 and CSDE1 is mediated by the N-terminal C3H zinc fingers of MKRN2 and the N-terminal cold shock domains of CSDE1, and that this interaction is essential for MKRN2-dependent ubiquitination of CSDE1.

To define how MKRN2 governs CSDE1 through ubiquitination, we conducted high-resolution mass-spectrometry mapping of CSDE1 ubiquitin acceptor sites. Seven ubiquitinated lysine (K) were confidently identified (Table 3; Figure 1I). To pinpoint the residues preferentially targeted by MKRN2, we systematically replaced each lysine with arginine (K → R) and monitored ubiquitination by co-IP. Mutation of four residues-K81, K91, K208, and K727-produced the strongest reduction in CSDE1 ubiquitination (Figure 1I). Stepwise combination of these mutations (double K81/91R, triple K81/91/208R and quadruple K81/91/208/727R) led to a progressive loss of ubiquitination that plateaued with the quadruple mutant (Figure 1J), establishing K81, K91, K208, and K727 as the principal MKRN2-directed ubiquitination sites on CSDE1.

CSDE1 is an RNA-binding protein that couples translation to mRNA turnover (Chang et al., 2004) and serves as one of seven core scaffolds of the stress-granule (SG) network, indispensable for efficient SG assembly (Youn et al., 2018; Guo et al., 2019). Emerging evidence shows that liquid–liquid phase separation (LLPS) drives the biogenesis of membrane-less organelles and is particularly important for synaptic organization (Chen et al., 2020; Gomez et al., 2021). Conversely, pathological LLPS is increasingly implicated in neurodevelopmental and neurodegenerative diseases (Milovanovic et al., 2018; Zeng et al., 2016; Zeng et al., 2018; Shih et al., 2022; Mu et al., 2024; Zbinden et al., 2020; Ambadipudi et al., 2017; Ray et al., 2020; Ambadipudi et al., 2017). These considerations prompted us to interrogate the phase-separation behavior of CSDE1 and its E3 ligase MKRN2.

We first generated EGFP–MKRN2 and mCherry-CSDE1 reporters (Supplementary Figure 2A). Under basal culture conditions only rare cells displayed micron-scale condensates (data not shown). We therefore challenged HEK293 or SH-SY5Y cells with 0.5 mM sodium arsenite to provoke oxidative stress. Within 30–60 min both proteins nucleated abundant cytoplasmic droplets that extensively overlapped (Figures 2A,E; Supplementary Figure 2D). Quantification revealed that co-expression of MKRN2 reduced the number of CSDE1-positive condensates per cell (Figure 2B), whereas CSDE1 co-expression increased MKRN2 droplet number (Figure 2F), implying that MKRN2 restricts CSDE1-driven SG formation while CSDE1 reciprocally stabilizes MKRN2 condensates.

To verify that these structures are bona fide LLPS assemblies we co-transfected GFP–G3BP1, the archetypal SG nucleator (Yang et al., 2020). G3BP1 colocalized almost completely with CSDE1 puncta (Supplementary Figure 2F). Time-lapse imaging showed fusion events and relaxation into rounded droplets, confirming liquid-like character. Droplet size and number varied markedly between cells, underscoring cellular heterogeneity in LLPS capacity.

We next exploited CRISPR KO lines to test genetic dependencies. Loss of MKRN2 elevated CSDE1 condensate density; re-expression of wild-type MKRN2, but not its RING-deletion mutant, restored control levels (Figures 2C,D; Supplementary Figure 2E). Conversely, CSDE1 deletion diminished MKRN2 droplet formation, and rescue with wild-type CSDE1, but not with the ubiquitination-defective 4KR mutant (K81/91/208/727R), reversed the defect (Figures 2G,H). Thus, the MKRN2–CSDE1 interaction and CSDE1 ubiquitination are required for their reciprocal regulation during LLPS.

Because CSDE1 recruits G3BP1 to nascent SGs (Guo et al., 2019), we enumerated G3BP1 droplets in KO cells. CSDE1 absence increased G3BP1 puncta number, whereas MKRN2 loss had no significant effect (Supplementary Figures 2G,H), indicating that CSDE1 directly modulates G3BP1 partitioning into condensates.

To probe droplet dynamics we performed FRAP on mCherry–CSDE1 puncta in HEK293 cells. Bleached regions recovered ~70% fluorescence within 90 s (Figures 2I–L), validating liquid-like properties. Recovery half-time was significantly shorter in MKRN2-KO cells (t½ ≈ 25 s) than in controls (t½ ≈ 45 s), and re-expression of MKRN2 lengthened t½ toward baseline (Figures 2K,L). Thus, MKRN2 attenuates CSDE1 molecular mobility within condensates, most likely by ubiquitin-dependent cross-linking or surface charge modulation.

Collectively, these data establish that MKRN2 and CSDE1 undergo oxidative-stress-induced LLPS, reciprocally regulate each other’s droplet formation, and that MKRN2-mediated ubiquitination fine-tunes condensate dynamics-potentially influencing the spatial translational control of CSDE1-bound mRNAs.

We next asked whether CSDE1 and MKRN2 can phase-separate in a minimal cell-free system. Prokaryotic expression vectors encoding mCherry–CSDE1 and EGFP–MKRN2 were constructed (Supplementary Figure 2I) and the fluorescent proteins were purified to homogeneity. Both proteins spontaneously formed spherical, liquid-like droplets under crowding conditions. After systematically varying buffer composition and PEG-20000 concentration, we selected 2% PEG as the standard crowding agent (Supplementary Figures 2J,K).

Titration of protein concentration revealed a steep, positive relationship between input concentration and droplet number for each protein alone (Figures 2M–O). Raising KCl from 50 to 600 mM produced a mild, non-significant increase in droplet count (Supplementary Figures 2L,M), indicating that electrostatic interactions contribute weakly to condensation under these conditions.

Strikingly, when both proteins were present simultaneously, the droplet formation of both CSDE1 and MKRN2 was strongly suppressed (≈ 60% reduction) (Figure 2P). Importantly, the two proteins never co-localized into the same droplets in vitro; instead, MKRN2 remained diffuse or formed separate, smaller foci that excluded CSDE1. Thus, in the minimal system, MKRN2 acts as a potent, dose-dependent antagonist of CSDE1 condensation rather than a co-partitioner.

The discrepancy with the co-condensation observed in cells implies that additional cellular factors—RNA, post-translational modifications, or auxiliary proteins (Ju Lee et al., 2017; Panas et al., 2016)—are required to license mixed CSDE1–MKRN2 droplets. Collectively, the cell-free data establish that MKRN2 can autonomously suppress CSDE1 phase separation through direct binding and/or steric interference, providing a biochemical basis for the ubiquitin-independent arm of the regulatory circuitry described in vivo.

We have confirmed that mutating the ubiquitination sites of CSDE1 reduces MKRN2 ubiquitination (Figures 1F,G). Truncated variants of the two proteins also display distinct interaction profiles (Figures 1H,I).

To dissect how these modifications affect phase separation, we generated fluorescently tagged plasmids bearing site-specific mutations or truncations (Supplementary Figure 3A,B). Single lysine-to-arginine substitutions at any of the seven ubiquitination sites did not block CSDE1 condensation (Figure 3A), yet mutations at four major sites (K81, K91, K208, K727) doubled condensate number in HEK293 cells (Figure 3B), identifying them as the principal MKRN2 targets. These single mutations impair MKRN2 ubiquitination activity without abolishing condensation, phenocopying MKRN2 knockout (Figure 2D).

In stark contrast, combined mutations (double K81R/K91R, triple K81R/K91R/K208R, or quadruple K81R/K91R/K208R/K727R) completely eliminated CSDE1 phase separation. The protein lost cytoplasmic confinement and became dispersed throughout the cytoplasm and nucleus (Figure 3C), a mislocalization not seen in CSDE1-null cells. We infer that compromised ubiquitination perturbs CSDE1 structure—possibly exposing cryptic NLS/NES motifs (Köhler and Hurt, 2007) or weakening RNA binding (Guo et al., 2020)—thereby abrogating both localization and condensation. Co-expression of these multi-site mutants also reduced MKRN2 condensates (Figure 3D), mirroring the consequences of CSDE1 depletion and confirming that proper ubiquitination is indispensable for CSDE1 structural integrity and phase-separation competence.

Truncations of either protein produced similar mislocalization: cytoplasmic fragments became diffuse and nuclear signal appeared (Figures 3E,G). These fragments formed ≤2 condensates per cell, some nuclear. Truncated MKRN2 elevated wild-type CSDE1 condensate number (Figure 3F), recapitulating MKRN2 knockout (Figure 2D). Even the isolated C3H zinc-finger domain, though sufficient for CSDE1 binding (Figure 1J), was mislocalized and failed to condense, indicating that any structural deletion disrupts MKRN2 function. Conversely, truncated CSDE1 suppressed wild-type MKRN2 condensation (Figure 3H), resembling CSDE1 knockout (Figure 2H). The CSDE1-T3 fragment, which lacks both the cold-shock domain and ubiquitination sites, neither altered condensate number nor interacted detectably with MKRN2 (Figure 1I), underscoring that the CSD and intact ubiquitination sites are required for productive phase separation.

Collectively, our data establish a reciprocal regulatory circuit: MKRN2 ubiquitinates and destabilizes CSDE1 to limit condensate formation (RING-domain dependent), whereas CSDE1 promotes MKRN2 condensation. This creates a negative-feedback loop in which MKRN2 restrains its own positive regulator (Figure 3I). The structural integrity and ubiquitination status of both proteins are thus interlocked determinants of their phase behavior and subcellular distribution.

Having established that MKRN2 and CSDE1 form an LLPS-dependent regulatory axis in both cellular and cell-free systems, and given that (i) CSDE1 is an established autism-spectrum-disorder (ASD) risk gene implicated in neuronal migration and differentiation (Guo et al., 2019; Kobayashi et al., 2013; Xia et al., 2014), (ii) germline Csde1 deletion causes cortical malformation and embryonic lethality in mice (Jia et al., 2025), (iii) MKRN2 inhibits neurogenesis in Xenopus (Yang et al., 2008), and (iv) dysregulated LLPS is increasingly linked to neurodevelopmental disorders (Shih et al., 2022; Mu et al., 2024; Zbinden et al., 2020; Ambadipudi et al., 2017; Ray et al., 2020; Ambadipudi et al., 2017), we asked whether Mkrn2 loss-of-function in mice would elicit behavioral phenotypes relevant to neuropsychiatry.

We acquired constitutive Mkrn2-knockout (KO) mice from Cyagen Biosciences and verified complete Mkrn2 deletion by PCR genotyping and by loss of MKRN2 protein in prefrontal-cortex lysates (Figures 4A,B). Body weight was recorded before every behavioral session to control for non-specific effects (Figure 4C). Animals were then subjected to four validated paradigms: (1) self-grooming for repetitive behavior (Supplementary Figure 4A), (2) open-field for locomotion and anxiety-like behavior (Supplementary Figures 4C–E), (3) three-chamber social interaction for sociability and social novelty preference (Figure 4D), and (4) rotarod for motor coordination (Supplementary Figure 4B). No genotype differences emerged in grooming duration, center time, total distance, or rotarod latency, indicating that MKRN2 deletion does not perturb baseline locomotion, anxiety, repetitive behavior, or motor learning.

Strikingly, however, Mkrn2-KO mice displayed markedly enhanced sociability. During phase I of the three-chamber assay both genotypes preferred the stranger mouse over the empty cup, but KO animals exhibited significantly greater interaction time. In phase II, KOs showed a robust preference for the novel stranger (stranger 2) versus the now-familiar stranger 1, whereas controls did not, revealing exaggerated social novelty seeking (Figure 4D). Qualitatively similar trends were present in females but did not reach statistical significance.

To dissect social behavior with higher granularity, we partnered with the Hu Ji Laboratory (ShanghaiTech University) and the Wei Pengfei Laboratory (Shenzhen Institute of Advanced Technology) to perform 3D video-tracking of unrestrained social encounters (Supplementary Figure 4F; Han et al., 2023; Huang et al., 2021). Sex-specific divergence was immediately apparent (Figures 4E,F; Supplementary Figure 4G). Female KOs displayed increased back-to-back rearing and enlarged inter-individual distance during rearing bouts-ethological signatures of social withdrawal that parallel the passive avoidance phenotype often observed in autistic girls (Dean et al., 2017; Mandy and Tchanturia, 2015). Conversely, male KOs exhibited excessive face-to-face hunching and elevated rear contact while in motion, reflecting abnormally intrusive approach behavior reminiscent of the “forced social proximity” and poor personal-space regulation seen in a subset of autistic boys (May et al., 2014).

Collectively, our data demonstrate that Mkrn2 deficiency in mice produces nuanced, sex-dimorphic alterations in social repertoire that mirror the qualitative heterogeneity of human ASD. The specificity of the phenotype-sparing motor, anxiety, and repetitive domains-positions Mkrn2 loss as a selective driver of social-circuit dysfunction rather than a global neurodevelopmental insult. The recapitulation of male-biased penetrance and female-specific expressivity further validates the model for mechanistic dissection of sex-specific pathways underlying ASD.

The observed social deficits in Mkrn2-KO mice prompted us to investigate the cellular mechanisms downstream of its substrate, CSDE1. Given CSDE1’s established role as an RNA-binding protein that drives stress-granule assembly (Youn et al., 2018) and selectively remodels the translatome (Chang et al., 2004; Guo et al., 2020; Kakumani et al., 2020), we asked whether its messenger-RNA partners also regulate, or are regulated by, CSDE1 phase separation. mRNA itself can act as a multivalent scaffold that accelerates RBP condensation, especially when translation stalls and unprocessed transcripts accumulate (Panas et al., 2016). We therefore set out to identify the CSDE1-bound transcriptome under conditions in which CSDE1 ubiquitination by MKRN2 is either intact or abolished.

We performed RIP-seq in MKRN2-knockout SH-SY5Y cells stably expressing Flag-tagged CSDE1. Three isogenic conditions were compared: (i) empty vector, (ii) CSDE1-Flag alone, and (iii) CSDE1-Flag plus MKRN2-Myc rescue. Anti-Flag immunoprecipitates were subjected to strand-specific library preparation and 2 × 150-bp paired-end sequencing (Genewiz). After stringent peak calling (CLIPper, FDR ≤ 0.01) and differential-binding analysis (DESeq2), ~ 20,000 transcripts showed CSDE1 occupancy that was reproducibly shifted ≥ 1.5-fold in the absence of MKRN2. Effect sizes were modest (|log₂FC| < 2.5 for 90% of targets), consistent with fine-tuning rather than on/off regulation.

Cross-referencing these data with the SFARI autism gene database yielded 117 overlapped ASD genes (Supplementary Figure 5A). Recognizing that RIP-seq captures the CSDE1-containing complex (which may include both direct and MKRN2-mediated RNA associations), we integrated this list with neurobiological literature to prioritize candidates with high relevance. After applying filters (mean TPM ≥ 10, FDR ≤ 0.05, proteomic support), HNRNPUL2 and KDM4C were identified (Figure 5A). To broadly probe MKRN2–CSDE1 roles in ASD pathways, we expanded the list to 10 additional neurodevelopmentally relevant genes, including MARK1 (synaptogenesis) (Maussion et al., 2008; Kelly-Castro et al., 2024) and PAX6 (neurogenesis) (Kikkawa et al., 2019; Divya et al., 2016; Kawada et al., 2018), despite their lower statistical significance in the RIP-seq data (Figure 5A; Supplementary Figure 5A).

Targeted qPCR validation of these 12 candidates confirmed reproducible enrichment of HNRNPUL2, MARK1, and PAX6 in CSDE1 immunoprecipitates (Figure 5B; Supplementary Figure 5B). Notably, the direction of change in these focused assays did not always align with the initial genome-wide trends, highlighting that our RIP-seq data identify mRNAs whose association with the CSDE1 complex is subject to MKRN2-dependent regulation, rather than predicting the exact outcome of all downstream validations. Given their established neurobiological relevance—HNRNPUL2 as a high-confidence ASD gene and MARK1 as a key synaptogenesis regulator—and their reproducible association with CSDE1, we selected HNRNPUL2 and MARK1 for in-depth mechanistic investigation.

To visualize the transcripts in living cells, we inserted 24 × MS2 stem–loops into the 3′ UTR of each gene and co-expressed MCP-GFP together in control or MKRN2-knockout HEK293 cells. Live-cell imaging revealed that all four mRNAs formed discrete, spherical nuclear foci (Figure 5C; Supplementary Figure 5C). Quantitative single-molecule RNA-FISH showed a 1.7- and 2.1-fold increase in MARK1 and HNRNPUL2 focus number per nucleus in MKRN2-KO cells (p < 0.01) (Figures 5D,E), whereas KDM4C and PAX6 trends did not reach significance (Supplementary Figure 5D). Under these experimental conditions, CSDE1 was predominantly cytoplasmic (as shown in Figures 2A,E; Supplementary Figure 2D), indicating a spatial separation from the nuclear mRNA foci. These data demonstrate that MKRN2 deficiency specifically alters the subnuclear condensation state or abundance of HNRNPUL2 and MARK1 mRNAs.

We next asked whether these specific mRNAs could directly modulate the core biophysical property of the MKRN2-CSDE1 axis: biomolecular condensation. To address this, we reconstituted phase separation in vitro with purified components. Full-length MARK1 and HNRNPUL2 mRNAs were synthesized by T7 runoff transcription, capped, poly-adenylated, and gel-purified before use.

Titration of MARK1 mRNA (0–500 ng μL⁻¹) into 0.6 μM CSDE1 at 150 mM KCl revealed a biphasic response: condensate number rose steeply between 50 and 200 ng μL⁻¹, plateaued, and declined at ≥500 ng μL⁻¹, consistent with RNA-driven cross-linking followed by solubilization at excess RNA (Supplementary Figures 5E,F). MKRN2 droplet counts remained unchanged across the same interval (Supplementary Figure 5G), indicating that MARK1 mRNA does not act as a general phase-separation scaffold. All subsequent assays were performed at 200 ng μL⁻¹ RNA, the lowest concentration that saturated CSDE1 droplet yield.

Both MARK1 and HNRNPUL2 mRNAs specifically lowered the critical concentration of CSDE1 and increased droplet number at fixed protein concentration (Figures 5F,G). HNRNPUL2 mRNA was consistently more potent than MARK1 mRNA (1.6-fold vs. 1.3-fold increase; p = 0.045, one-way ANOVA). Neither transcript altered MKRN2 droplet formation (Figure 5H), confirming selectivity.

We then tested whether MKRN2 could antagonize this RNA-driven CSDE1 condensation (Figure 5I). In the absence of RNA, co-incubating CSDE1 and MKRN2 had no significant effect. Addition of MARK1 mRNA triggered robust CSDE1 droplets that were reduced by 42 ± 5% when MKRN2 was present (Figure 5J). MKRN2 droplet counts stayed constant (Figure 5K), suggesting ubiquitination of CSDE1—rather than direct RNA binding—dampens condensation. An identical experiment with HNRNPUL2 mRNA produced an even stronger suppression (58 ± 4% reduction) (Figure 5L), paralleling the higher basal potency of this transcript. Concomitantly, MKRN2 droplets were diminished only when CSDE1 was co-expressed (Figure 5M), indicating reciprocal exclusion from the RNA-rich phase.

Together, this in vitro reconstitution shows that (i) MARK1 and HNRNPUL2 mRNAs act as selective scaffolds for CSDE1 condensates, and (ii) MKRN2-mediated ubiquitination reverses RNA-driven phase separation, providing a direct mechanistic link between post-translational modification and condensate dynamics regulated by ASD-relevant mRNAs.

MKRN2 inhibits CSDE1 condensation and reduces its dynamics, whereas CSDE1 promotes MKRN2 droplet formation, suggesting co-regulation for maintaining stress granule homeostasis (Youn et al., 2018; Guo et al., 2019). The discrepancy between cellular co-condensation and in vitro suppression indicates that their interaction dynamically modulates the energy landscape of phase separation, potentially requiring cellular cofactors such as RNA (Ju Lee et al., 2017; Panas et al., 2016). This aligns with emerging roles of E3 ligases in regulating condensate dynamics via ubiquitination (Yasuda et al., 2020; Li et al., 2023). We propose that MKRN2 may sterically inhibit CSDE1 self-assembly through ubiquitination or form an alternative soluble complex. The RNA-rich cellular environment likely overcomes this suppression, enabling coordinated phase separation and highlighting the critical influence of biological context on phase separation regulation (Langdon et al., 2018).

The mislocalization of CSDE1 ubiquitination-site mutants underscores that ubiquitination is crucial for its proper cytoplasmic retention. Loss of ubiquitination may induce conformational changes, disrupt nuclear export, or interfere with cytoplasmic anchoring, leading to aberrant nuclear diffusion (Köhler and Hurt, 2007). This mislocalization directly compromises CSDE1’s phase separation capability, highlighting a primary mechanism by which MKRN2 regulates its function. While our work definitively establishes MKRN2 as the E3 ligase for CSDE1, the precise functional consequence—whether regulating protein stability via the proteasome or altering its interactome or activity—remains an exciting avenue for future study.

Our mapping reveals that the central region of MKRN2, encompassing multiple CCCH zinc finger motifs, is involved in CSDE1 binding. The T1 fragment (with three zinc fingers) is a major driver, while the T4 fragment (with one zinc finger) also showed binding capacity in initial assays (Laity et al., 2001). The inability to consistently validate binding with isolated T4 may reflect its loss of native conformation outside the full-protein context. This suggests that multiple motifs may cooperatively form a composite binding interface within the full-length protein, a hypothesis awaiting structural validation.

A key finding is the MKRN2-dependent regulation of specific ASD-relevant mRNAs, HNRNPUL2 and MARK1. Our RIP-seq data, which profile the CSDE1-containing ribonucleoprotein complex, showed that MKRN2 loss modestly reshapes its RNA content. However, targeted qPCR validation revealed a pronounced increase in CSDE1 association with these specific candidates, a trend corroborated by increased nuclear mRNA foci in MKRN2-KO cells. This apparent discrepancy likely arises from methodological differences: RIP-seq provides a population-averaged, static snapshot of complex composition, which can be influenced by indirect transcriptional changes and buffering within the RNP network. In contrast, hypothesis-driven validation may more sensitively capture the functional regulation of specific high-affinity interactions. Therefore, we prioritized candidates by integrating the RIP-seq dataset (as an indicator of complex membership) with independent criteria of established neurobiological relevance (e.g., SFARI database, synaptic function), rather than relying solely on statistical significance from the sequencing data alone. Notably, these mRNAs form nuclear foci, while CSDE1 is cytoplasmic under steady-state conditions. This spatial separation, coupled with their robust biochemical association, suggests a dynamic or mediated interaction model rather than stable co-condensation. It posits that the MKRN2-CSDE1 complex may function as a regulatory hub, where MKRN2 could influence CSDE1’s nucleocytoplasmic shuttling or its partnership with other nuclear RNA-binding proteins to engage target mRNAs.

The physiological relevance of this axis is underscored by sex-dimorphic social deficits in Mkrn2-KO mice, which resemble the heterogeneity of human ASD (Dean et al., 2017; Mandy and Tchanturia, 2015; May et al., 2014). Given CSDE1’s status as an ASD risk gene (Guo et al., 2019; Xia et al., 2014) and its regulation of ASD-associated mRNAs (e.g., MARK1, HNRNPUL2), dysregulation of the MKRN2-CSDE1 complex within neuronal condensates could disrupt synaptic plasticity and neural circuit function, core aspects of neurodevelopmental disorders (Guo et al., 2019; Zeng et al., 2016).

We acknowledge certain limitations. The RIP-seq data reflect the RNA content of the CSDE1 complex but cannot resolve direct versus indirect associations. Furthermore, while our quantitative analysis focused on condensate number as a key metric, a comprehensive biophysical characterization—including size distribution, density, and internal architecture—remains for future investigation. Our in vitro LLPS assays, while providing controlled mechanistic insights, use simplified conditions that may not fully recapitulate the cellular milieu. Future efforts should therefore focus on: (1) validating the spatial and functional dynamics of this axis in primary neurons in vivo; (2) determining the precise consequence of CSDE1 ubiquitination (e.g., on protein stability via degradation assays); and (3) elucidating the structural basis of the MKRN2-CSDE1 interaction. Such studies will be crucial for exploring this pathway as a potential therapeutic target.

Looking forward, the MKRN2-CSDE1 axis presents rich opportunities for both mechanistic dissection and therapeutic exploration. To move beyond correlation toward causation, future studies could employ advanced in vivo imaging (e.g., two-photon microscopy) to directly visualize the dynamics of this complex in living neurons during synaptic activity. Reconstitution experiments incorporating neuronal lysates or specific RNA cohorts could identify the bona fide cofactors enabling physiological co-condensation. Furthermore, the sex-dimorphic behavioral phenotypes in Mkrn2-KO mice provide a unique entry point to dissect sex-specific molecular underpinnings of ASD-related social circuits. Ultimately, this pathway’s dependence on LLPS makes it uniquely tractable for targeted modulation. High-throughput screening for small molecules that correct pathological phase separation driven by MKRN2 deficiency could yield novel therapeutic leads for neurodevelopmental disorders.

In summary, we identify the MKRN2-CSDE1 axis as a novel LLPS-coupled ubiquitination pathway that fine-tunes the condensation state of a key RBP and its association with specific neuronal mRNAs. Dysregulation of this axis disrupts granular dynamics and mRNA metabolism, contributing to the complex social behavioral phenotypes observed in ASD models, thereby highlighting a new target for mechanistic and therapeutic exploration in neurodevelopmental disorders.