Transcriptome-wide mRNA condensation precedes stress granule formation and excludes stress-induced transcripts

Stress-induced condensation of mRNA and proteins into stress granules is conserved across eukaryotes, yet the function, formation mechanisms, and relation to well-studied conserved transcriptional responses remain largely unresolved. Stress-induced exposure of ribosome-free mRNA following translational shutoff is thought to cause condensation by allowing new multivalent RNA-dependent interactions, with RNA length and associated interaction capacity driving increased condensation. Here we show that, in striking contrast, virtually all mRNA species condense in response to multiple unrelated stresses in budding yeast, length plays a minor role, and instead, stress-induced transcripts are preferentially excluded from condensates, enabling their selective translation. Using both endogenous genes and reporter constructs, we show that translation initiation blockade, rather than resulting ribosome-free RNA, causes condensation. These translation initiation-inhibited condensates (TIICs) are biochemically detectable even when stress granules, defined as microscopically visible foci, are absent or blocked. TIICs occur in unstressed yeast cells, and, during stress, grow before the appearance of visible stress granules. Stress-induced transcripts are excluded from TIICs primarily due to the timing of their expression, rather than their sequence features. Together, our results reveal a simple system by which cells redirect translational activity to newly synthesized transcripts during stress, with broad implications for cellular regulation in changing conditions.


Let
be the length of an mRNA under consideration.We seek the functional form of , the proportion in the supernatant after centrifugation, which will depend on many factors.Several of these are experimental and we assume they do not change across samples or mRNAs, such as the spin speed , the sample height , the spin time , and the viscosity .Terminal velocity for a particle with mass and hydration radius is: We assume that both mass and hydration radius scale consistently with length on average.In the case of mass, we assume proportional scaling: a constant (average) molecular weight per unit length deriving from nucleotides and bound proteins.In the case of hydration radius, our scaling assumption is consistent with standard approaches (Yoffe et al. 2008) , for the radius of gyration scales as with for naked ssRNA due to secondary structure.The hydration radius will be proportional to assuming a constant density.The scaling relationship will also depend on the geometry of the vessel.
Assuming a constant mass per nucleotide and protein binding per unit length, the mass of an mRNP scales linearly with .Therefore, the leading term above will tend to scale with : with proportionality constant taking care of the specific conversion factors (e.g. the molecular weight per nucleotide) and with summarizing the scaling.
Then the proportion of the particle species remaining in the supernatant is the proportion of the tube which is more than from the bottom.That is: from the bottom.That is: We can then summarize the dependencies with two constants, as follows.
That is, two parameters should be sufficient to describe the average behavior of free (unclustered) mRNAs of length .A simple test of this model arises from a particular prediction: by rearranging, we see that which predicts a linear relationship between and .Indeed, our data show this relationship (Figure S1B).

A model of mRNP clustering
To model between-mRNP interactions and their effect on , we consider two kinds of interactions, both assumed to be Poisson-distributed with rates which vary across conditions and potentially across transcripts.
The first interaction type is between nucleotides, including direct RNA-RNA interactions and interactions mediated by proteins but dependent on nucleotide content; these are expected to be length-dependent.Consequently, we model them by a per-nucleotide interaction rate , such that the probability that an mRNP has nucleotide-dependent interactions is .
The second interaction type is between molecules and is independent of sequence length, including interactions mediated by the 5' cap or the 3' end.We model these as a per-molecule interaction rate such that the probability that an mRNP has such interactions is .
These rates permit us to model the probability that an mRNP of a given length engages in interactions with other mRNPs, which may have varying lengths/sizes and may themselves also engage in additional interactions, making prediction of sedimentation highly complex.To proceed analytically, we make a simplifying assumption: that under the conditions we study, most mRNP clusters are sufficiently large to sediment completely ( ).Under this assumption, the observed proportion in the supernatant for an mRNA of length is given by the free-mRNP pSup multiplied by the probability that the mRNP is free.The latter probability is equal to the probability of per-nucleotide interactions and per-molecule interactions.Together, we have .
The mean behavior of transcripts in any particular experiment, in this model, can be derived from only four fitted parameters and the lengths of all transcripts.Fits of this model are shown in Fig. 1G, where the no-stress sample is modeled with no clustering ( , ) and stress samples are modeled with clustering (solid lines) and also in comparison to curves where the per-molecule interaction rate is zero ( ; dashed lines).

Bounds on condensate sizes
How can we put bounds on the average size of condensates, given an observed value of pSup?Any pSup value can be realized in a range of ways, bounded by two extremes.At one extreme, the molecular population is homogenous (all molecules of a particular type are in condensates of the same size), and all clusters have the same probability of remaining in the supernatant, which is pSup.At the other extreme, the molecular population is heterogeneous, with each species of condensate having some proportion in the supernatant pSup i and proportion of the molecular population q i .Order the indices i such that the lightest species (monomers) has i=1 and all condensates are ordered by their size (and hence in the descending order of their pSup) up to n .First, if n=2 , it is clear that the observed , simply because the average of  ≥  2 two numbers is great than or equal to the smaller of the two numbers.Second, if we define the average of the condensates' pSups as , we have such that in general, pSup is an upper bound on the  ≥   average condensate size.

Alternatives to condensation
We considered the possibility changes in mRNA length during stress, either due to splicing or changes in polyadenylation, explain these solubility changes.Fewer than 5% of yeast genes contain introns, so splicing cannot account for the transcriptome-scale effects we observe.Polyadenylation yields poly(A) tails of only ~50 nucleotides on average (3% of the median transcript length) which, in high-expression genes, grow slightly shorter upon heat stress (Tudek et al. 2021) , not longer as would be required to decrease pSup during stress.
What about an increase in mRNP mass due to changes in protein binding?While formally possible, this would not plausibly increase free mRNP sizes by more than tenfold as we observe during severe stress.We provisionally conclude that changes in the physical size of otherwise free mRNPs cannot account for apparent sedimentation changes.
What about stress-induced binding to a large sedimentable structure, such as the ER, which has been argued to be the site of substantial protein synthesis of cytosolic proteins (Reid and Nicchitta 2015) ?While we cannot rule all possibilities out, our data suggest this is not the case.Most importantly, substantial transcriptome-wide mRNA clustering occurs while poly(A)+ mRNA appears diffuse and primarily cytosolic (42°C, Fig. 1B/C), inconsistent with localization to the ER or another structure.Once clusters become visually resolvable-i.e., stress granules form-these are well-known to be distinct structures.

Simulation of complex condensation
To assess whether assumed effects produce observed phenomena, we turned to a detailed simulation of condensation and sedimentation (available along with other analytical code; see "Data and code availability" section in the main manuscript).Direct simulation of interactions, consequent formation of condensates in a fully heterogeneous population, and sedimentation of these heterogeneous condensates according to sedimentation theory by the model above, yields the same trends as in Fig. 1G (Fig. S1D,E).Fig. S1E shows a qualitatively excellent reproduction of the major trends in the biological heat-shock data (Fig. 1D), using only one parameter which varies during stress: the rate of interaction per molecule, which can arise e.g. through interactions mediated by the 5' or 3' ends of transcripts (see main text).Importantly, the simulation does not invoke a key assumption used in the analytical model above, which is that all condensates sediment completely.

Table S2 :
Plasmids used in this study