Co-Translocation Networks: Weighted Networks of Nuclear Localization in Arrested Cells

Gene co-expression networks group genes whose messenger RNA levels rise and fall together, then treat each group as a "program" the cell is running. This works poorly for the cell cycle, and it probably works worse when a drug halts that cycle. The largest single-cell map of cell-cycle proteins found that 85% of proteins that fluctuate with the cycle have transcripts that do not fluctuate (Spatiotemporal dissection of the cell cycle with single-cell proteogenomics). Much of the regulation happens after the RNA is made: proteins are modified, degraded, or moved into and out of the nucleus. This proposal keeps the network method and changes what it measures. The variable becomes each protein's nuclear fraction, meaning the share of its signal inside the nucleus, measured by imaging. We then ask which proteins enter and leave the nucleus together, and how a drug-induced cell cycle arrest rewires those groups.

The core claim is that arrest programs are written more clearly in coordinated nuclear translocation than in transcript co-expression. Three things follow if it holds. Modules built from nuclear fractions should not reappear in transcript networks from the same wells. The nuclear transport machinery should sit near the centre of arrest-specific localization modules even though it rarely ranks as a hub in transcript analyses. And localization-module summaries taken at the end of drug treatment should predict whether a population resumes dividing after the drug is removed better than transcript modules do. All three are testable with fixed-cell imaging plus single-cell RNA sequencing from the same wells.

Background and Positioning

Weighted gene co-expression network analysis (WGCNA) turns pairwise correlations $s_{ij}$ into soft edge weights $A_{ij}=s_{ij}^{\beta}$. The power $\beta$ shrinks weak correlations without cutting them off at an arbitrary threshold. It then groups genes by how many neighbours they share (topological overlap) and summarizes each group by its first principal component, the "module eigengene." Session Introduction motivates this style of analysis as a way to compress high-dimensional omics into functional units. CENPL, ISG20L2, LSM4, MRPL3 are four novel hub genes is the closest cell-cycle application: it combines WGCNA with rank aggregation over eight datasets and recovers mitosis-enriched hub genes with diagnostic AUC up to 0.948. That paper measures only mRNA and never asks where the encoded proteins sit or what happens under arrest.

Two studies link translocation to the cell cycle mechanistically. Both are single pathways without a network view, and they motivate this proposal. Coupling of Nuclear Translocation to Cell Size shows three things across more than 400,000 cells:

  • Whole-cell YAP/TAZ becomes diluted as cells grow.
  • Nuclear YAP/TAZ nonetheless stays constant because import speeds up.
  • Blocking CDK4/6, the kinases that push cells from G1 into DNA synthesis, breaks this coupling.

That study models one protein with supervised partial least squares regression. PAK6 promotes homologous-recombination shows the kinase PAK6 entering the nucleus after DNA damage to activate the ATR/CHK1 checkpoint. It traces one linear cascade. ZBTB34 illustrates how such translocations are usually established: one protein at a time, by mutating its nuclear localization signal and co-immunoprecipitating its importins.

Two further studies constrain the design. The proteogenomic map above warns that chemical synchronization distorts checkpoint behaviour, which pushes us to use single-cell pseudotime rather than synchronized populations. Pseudotime here means ordering unsynchronized cells along the cycle using fluorescent phase reporters. Library size-stabilized metacells shows that unequal library sizes create false co-expression edges. Arrested cells differ systematically in size and RNA content from cycling cells, so every network here is built on metacells that are size-stabilized. Metacells are small pools of similar cells.

Imaging-to-omics models are neighbours rather than competitors:

Neither measures where specific proteins are located, and neither builds modules. Static predictors of where a protein lives, such as models trained on CAPSUL, output one compartment per protein. They cannot represent a protein whose location changes with cycle phase or drug.

Research Program

Stage one: reanalysis of existing data, no new wet-lab work. We reanalyse the FUCCI U-2 OS resource from the proteogenomic map. FUCCI is a pair of fluorescent reporters that mark G1 versus S/G2 in live cells. The resource covers 1,219 proteins imaged by immunofluorescence across 273,709 cells, plus 1,152 single-cell transcriptomes.

Each image stains one protein, so correlations cannot be computed within a single cell. Instead, correlations are computed across pseudotime metacells. We bin interphase cells along FUCCI pseudotime into bins with equal cell counts, computed separately per protein. For protein $p$ in metacell $m$, the nuclear fraction is

$$f_{pm}=\frac{N_{pm}}{N_{pm}+C_{pm}},$$

where $N$ and $C$ are total nuclear and cytoplasmic intensities. The nucleus is segmented from the DNA stain and the cytoplasm from the cell boundary. Using a ratio cancels most of the effect of cell size and total expression, which is the imaging analogue of library-size stabilization. Mitotic cells are excluded because the nuclear envelope breaks down during mitosis and the fraction is undefined. We then run WGCNA on the matrix of $f_{pm}$ values and a matched transcript WGCNA on stabilized metacells of the RNA data. Finally, we test whether modules are preserved in each direction using standard preservation statistics.

The number of pseudotime bins limits statistical power. The soft-threshold power and the minimum module size will therefore be chosen by stability under bootstrap resampling of bins, not by the scale-free fit alone.

Stage two: new data under arrest. We build a panel of about 150 proteins for cyclic multiplexed immunofluorescence in FUCCI U-2 OS cells. Because all panel proteins are measured in the same cell, correlations can also be computed within cells. The panel includes:

  • canonical translocators named in the file: YAP/TAZ, FoxO3a, NF-κB, β-catenin, Smad1, ATF6, PAK6, PKM2, hexokinase 2, calreticulin;
  • cycle and checkpoint regulators: p21, p27, CDC25A/B/C, cyclin D1, cyclin B1, CHK1;
  • transport machinery: importins α1/α3/α4/β1, XPO1, XPOT, Ran, RanGAP and representative nucleoporins;
  • about 40 cell-cycle-dependent proteins taken from the stage-one map, which serve as anchors.

The treatments are:

Sister wells go to single-cell RNA sequencing at the same timepoint. Replicate plates are washed out, and the fraction of cells that re-enter S phase is scored by EdU incorporation at 24 and 48 hours. EdU incorporation labels cells that are synthesizing DNA. The dose × duration grid gives roughly 60 conditions, each with a measured re-entry fraction.

Stage three: mechanism. For the top hubs of arrest-specific localization modules, we knock the hub down with siRNA during CDK4/6 arrest. We then ask whether the module stops acting as a coordinated unit, measured as a drop in its within-module mean topological overlap. A knockdown of a non-hub transport factor serves as the control.

Baselines

Phase composition only. The naive reference is the fraction of cells in G1, S and G2 read from FUCCI at the end of treatment. It is used to predict re-entry and to classify the arrest drug. Much of what any network "finds" could simply reflect arrest changing the mix of phases, so every model must beat this. There is no reported result on Lacuna for predicting re-entry from phase composition, so we establish the number ourselves using the same leave-one-condition-out splits as every other method.

Transcript WGCNA on stabilized metacells. This is the strongest existing competitor for "which co-regulated program marks this arrest." It follows the WGCNA procedure of the CENPL study with metacells built as in LSMetacell, computed on the sister-well RNA. Transcript eigengenes are passed to the same regularized regression as the localization eigengenes, with the same tuning budget and nested cross-validation. To check that our transcript pipeline is sound, we run it on scGeneScope (scGeneScope) for compound classification from single profiles. The relevant reported figures are PhenoSeq at 0.293 balanced accuracy for synthetic transcript embeddings versus 0.270 for raw imaging (PhenoSeq). A real-transcript pipeline should clear 0.293 on that split. If it does not, the pipeline is broken and the head-to-head comparison is void.

Per-protein differential localization. Each protein's nuclear fraction is tested between conditions, and the significant proteins feed a lasso with the same number of input features as there are localization modules. This isolates whether grouping proteins into modules adds anything over the individual localization measurements themselves.

Supervised partial least squares on nuclear fractions. This follows the YAP/TAZ study, which used partial least squares regression to predict nuclear YAP/TAZ from phosphorylation states. It is the strongest supervised use of the same data.

Whole-cell intensity network. This ablation uses the same images and the same WGCNA procedure, but on total per-cell intensity with no nuclear/cytoplasmic split. It isolates the compartment ingredient.

The proposal counts as beating the baselines if localization eigengenes do all of the following:

  • predict re-entry fraction better than phase composition and transcript eigengenes (higher held-out R² with non-overlapping bootstrap intervals);
  • match or exceed per-protein lasso and partial least squares while using fewer, interpretable features;
  • clearly beat the whole-cell ablation.

Why not just…?

Why not just run WGCNA on the RNA and correlate modules with arrest? The 85% transcript–protein decoupling and the median 7.7-hour lag between transcript and protein peaks, both reported in the proteogenomic map, argue that transcript modules will mostly track delayed or indirect effects. That remains an argument until tested. Stage one's two-way preservation test and stage two's head-to-head prediction of re-entry settle it directly. If transcript eigengenes predict re-entry as well as localization eigengenes, the central claim fails, and we will report that.

Why not just test each protein's nuclear fraction separately, as in differential expression? With about 150 proteins this is feasible and could be competitive. It is the most serious threat, so it is a named baseline with a matched feature budget. Modules earn their place only if they generalize better to held-out drugs. The concrete test is whether modules learned under CDK4/6 and temsirolimus arrest carry over to etoposide conditions, where individually selected proteins are more likely to be overfit to one drug.

Why not just synchronize populations and fractionate nuclei for Western blots? The proteogenomic map shows that synchronization itself stresses cells and distorts checkpoint behaviour. Fractionation also yields one number per protein per population, which leaves nothing to correlate across cells or across pseudotime. We keep one fractionation experiment, on the top three hubs, purely to validate the imaging segmentation.

Why not just predict localization from images with a general model like PhenoSeq or FORM? Those models use general stains or morphology. PhenoSeq's authors describe a ceiling on what morphology can reveal for drugs that barely change cell shape, and cell cycle arrest is often exactly that kind of change. Antibody-specific nuclear fractions measure the quantity of interest directly.

Proposed Experiments and Decisive Checks

The first decisive check comes from stage one: are localization modules from the FUCCI map preserved in the matched transcript network? A low preservation statistic would confirm that the two networks encode different programs. High preservation would mean the localization layer is largely redundant and the project should narrow to the translocators the transcript network misses.

The second check controls for phase composition. We compare arrested cells to DMSO cells at the same FUCCI pseudotime, for example CDK4/6-arrested G1 cells against early-G1 control cells. If the localization modules change under this matching, the change reflects a different program within the same phase rather than a shift in which phases are present.

The third check is the prediction of re-entry fraction across the roughly 60 dose × duration conditions, evaluated against every baseline above with conditions held out.

The fourth is the knockdown test of transport hubs from stage three. If knocking down a hub breaks the module's coordination while the matched control knockdown does not, co-translocation modules are shown to be run by the transport machinery rather than merely correlated with it.