Condensate Thresholds Hidden in Bulk Transcriptomes

Many proteins that organize signaling and RNA processing contain intrinsically disordered regions (IDRs), stretches of protein with no fixed 3D shape. At high enough concentration these floppy segments can make the protein demix into liquid droplets inside the cell, called condensates. Droplets form only above a critical concentration, so condensate-driven activity behaves like a switch in each cell. Bulk RNA sequencing, the standard clinical measurement, averages expression over thousands of cells. For a switch-like process the average is the wrong summary. A tissue where 10% of cells sit far above the threshold and a tissue where every cell sits just below it can have the same mean expression and opposite condensate activity. Cellular heterogeneity therefore hides biology in bulk data in a predictable way: it erases the part of the expression distribution that lies above the threshold.

We propose to test this idea and use it. First, in large single-cell atlases, we ask whether the expression of condensate-forming IDR proteins relates to their downstream programs in a threshold-like way, and whether this happens more often than for matched folded proteins. Second, we ask whether the hidden quantity, the fraction of cells above threshold, can be recovered from bulk data by methods that reconstruct whole single-cell count distributions rather than only cell-type proportions. Third, we ask whether that recovered fraction adds predictive signal in clinical bulk cohorts. In this proposal the IDR biology supplies a specific nonlinearity, and the heterogeneity-masking direction supplies the measurement problem and the tools for addressing it.

Core Claim

Let $x_i$ be the per-cell expression of an IDR scaffold gene in cell $i$, and let $c$ be an effective threshold. If downstream activity per cell is approximately $g(x_i) = \beta\,\max(0, x_i - c)$, or a sigmoid in $x_i$, then the tissue-level activity is

$$ \bar{A} = \mathbb{E}_i[g(x_i)] \neq g(\mathbb{E}_i[x_i]), $$

and the gap between the two sides grows with the cell-to-cell variance of $x$ near $c$. This is Jensen's inequality applied to a convex or switch-like function. The statistic that matters is the above-threshold fraction $\phi = P(x_i > c)$, together with the expected excess above $c$. Neither can be computed from the bulk mean alone.

We make two claims. The first is biological: this threshold structure appears in single-cell data for IDR proteins with predicted condensate localization at a higher rate than for expression-matched folded proteins. The second is methodological: distribution-level deconvolution recovers $\phi$ from bulk data, but proportion-level deconvolution does not, whenever a meaningful share of the variance in $x$ lies within cell types rather than between them.

Existing work does not contain this claim. A Functional Map of the Human Intrinsically Disordered Proteome classifies about 19,000 human IDRs by 147 sequence-level physical features and predicts condensate localization (AUC 0.66–0.73), but it works at the level of genes, not cells. Phase separation of insulin receptor substrate 1 drives the formation of insulin/IGF-1 signalosomes shows that the disordered tail of IRS-1 forms condensates at roughly 0.2 µM and drives AKT/ERK signaling, but only in uniform cell lines. Bulk deconvolution methods estimate cell-type composition or reconstruct counts, but none connects its output to a biophysical threshold.

Research Program

Stage one: single-cell discovery. Using the 3CA atlas (about 2.7 million cells from 1,631 tumor samples) and the pancreatic islet datasets of Xin et al. and Segerstolpe et al., we fit count-aware models within each cell type. Each model relates a scaffold gene's expression to a downstream program score. The comparison is between a linear model and a hinge or sigmoid model, selected on held-out donors.

Downstream programs come from curated pathway target sets. Three examples:

  • IRS-1 and insulin/AKT targets in islets.
  • NPM1 and ribosome biogenesis.
  • Stress-granule scaffolds and stress-response genes.

To avoid circularity, any gene that directly co-varies with the scaffold, such as a known co-regulated paralog, is removed from its program. A program score is also computed on a random split of genes not used during model selection.

Stage two: enrichment. Scaffolds are ranked by FAIDR-predicted condensate localization from the IDR-ome paper. We then test whether the strength of threshold structure is higher for these IDR proteins than for folded proteins matched on mean expression, dropout rate and cell-type specificity. This is the step where the IDR hypothesis can fail.

Stage three: recovery from bulk. We build pseudo-bulk mixtures from held-out 3CA and islet samples, so the true $\phi$ is known for each mixture. We then estimate $\phi$ for each candidate scaffold three ways:

  • From the bulk mean alone.
  • From proportion deconvolution multiplied by reference cell-type means and variances.
  • From distribution-level deconvolution with Count Bridges, which models integer counts with Poisson birth-death bridges and reconstructs single-cell count profiles from aggregates.

We vary the share of $x$'s variance that lies within cell types. The key comparison is when the distribution-level method beats proportion-level estimates.

Stage four: utility. We add the recovered $\phi$ values to bulk models for immunotherapy response in the 330-sample melanoma cohorts and test whether they improve held-out prediction. The comparison is against the same model using mean expression and deconvolved composition.

Baselines

Naive reference: bulk mean. The naive reference is the bulk mean expression of each scaffold gene, used directly as the estimate of downstream activity and as the predictive feature. It is the "do nothing" option. Any value of our approach must show up as an improvement over it.

Proportion-level deconvolution. This is the strongest competitor for recovering $\phi$. We use MEAD, which corrects for measurement error across platforms and gives valid confidence intervals, and which reports competitive-to-superior accuracy against CIBERSORT, MuSiC and least squares on the same islet datasets we use. No numerical value is reported on Lacuna for it. We pair MEAD's proportions with per-type expression distributions from the reference to get an implied $\phi$.

If above-threshold cells are simply one cell type, this baseline should match ours. We will run it with the same references, gene panels and mixtures. Beating it means lower error on $\phi$ in mixtures with high within-type variance, with confidence intervals that do not overlap across mixture replicates.

Clinical baseline: POPPY. For the clinical task the strongest existing system is POPPY. It aligns bulk RNA-seq to a single-cell embedding space with ontology-guided contrastive learning, and it reports beating baselines on melanoma immunotherapy response. No reported AUROC is found on Lacuna for POPPY on these cohorts.

For a sense of scale from a different input type, SubNetDL (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer) reports a median AUROC of 0.77 on the MSK anti-PD-(L)1 lung cancer cohorts using mutation data. That number is not directly comparable to transcriptomic melanoma prediction.

We will retrain POPPY's predictor under our splits, with and without the $\phi$ features, using the same tuning budget. Beating it means a consistent AUROC gain across leave-one-trial-out folds.

Ablation: folded-protein control. The ablation that isolates the IDR ingredient replaces condensate-predicted IDR scaffolds with the matched folded-protein controls and computes $\phi$ the same way. If the gain survives this swap, our result is about generic nonlinearity in gene regulation, not condensates.

Why not just…?

Why not just deconvolve cell-type proportions and use cell-type means? This works if the threshold separates cell types. In that case bulk masking is just composition, and MEAD or POPPY already handle it. It is a real threat. Stage three settles it directly by measuring how much of each scaffold's variance lies within types and showing where the two approaches diverge. If within-type variance turns out to be small for the scaffolds that show thresholds, we will report that proportion methods are sufficient. That would be a useful negative result.

Why not just feed bulk expression into a flexible nonlinear predictor? A nonlinear model of the bulk mean still sees only the mean. Two tissues with the same mean and different variance are identical to it unless other genes happen to encode that variance. We include a gradient-boosted model on full bulk profiles as an additional baseline in stages three and four. If it recovers $\phi$ as well as Count Bridges, the explicit distributional step is unnecessary, and we will say so.

Why not ignore disorder and look for threshold genes generically? Threshold responses can arise from cooperative transcription-factor binding with no condensate involved. The folded-protein control in stage two exists for this reason. Disorder matters to this proposal only if condensate-predicted IDR proteins show threshold structure more often. Structure-based tools would not help define that set: IDiom notes that AlphaFold-style models are biased toward folded domains. This is why we use sequence-feature predictions from FAIDR instead.

Decisive Checks and Risks

The program is decisive at three points:

  • Stage two: threshold models must win on held-out donors, and they must win more often for IDR scaffolds than for matched controls.
  • Stage three: Count Bridges must beat MEAD-based estimates of $\phi$, and the gap must widen as within-type variance grows.
  • Stage four: $\phi$ must add held-out AUROC beyond POPPY.

The main risk is that mRNA is a poor stand-in for protein concentration, which blurs any true threshold. Dropout in single-cell data blurs it further. Both effects bias the study toward finding no threshold, so a positive result would be conservative. Where islet datasets allow, we will check the IRS-1 case against known insulin-pathway activity as a positive control.