Recordanelles for transcriptome-wide single-cell RNA recording in human cells

Cells orchestrate development and adapt to intracellular and extracellular cues by dynamically regulating gene transcription. However, most current approaches to measure transcription are disruptive and capture only static snapshots¹. Metabolic or enzymatic labeling of nascent or tagged RNAs can add temporal resolution to transcriptomics but these approaches are constrained by a lack of continuous transcriptome sampling, indirect inference of transcriptional activity, substantial input requirements or limited scalability beyond a small number of RNA species^2,3,4,5,6,7. DNA-based molecular recording methods that encode transient information into DNA enable retrospective reconstruction of gene expression events, yet have been largely confined to prokaryotes and/or a small set of target genes^8,9. Overcoming these limitations requires approaches for nondisruptive, unbiased, transcriptome-wide recording in eukaryotic cells.

The ideal RNA recording system would recruit, protect and accommodate a representative sampling of the transcriptome within a storage archive throughout time and cell division without bias or altering cellular physiology. We conceived of a solution meeting these demands, which we term ‘recordanelle’ (RNA recording organelle)—a genetically encoded system that is inspired by the principles of biomolecular condensation^10,11,12 to drive the unbiased and scalable storage of endogenous cytoplasmic RNAs within an orthogonal and heritable organelle (Fig. 1). The recordanelle is created with a single protein harboring three components—a generic RNA-binding domain (RBD), a biomolecular condensate assembler domain and a mitochondrial membrane anchor. Together, the RBD and the assembler domain collect endogenous RNAs in a protective environment, which is stabilized by the mitochondrial membrane anchor at an easily isolatable structure through time and cell division (Supplementary Notes). While transiently expressed RNAs are lost in normal cells, they should be preserved within the recordanelles of engineered cells. RNA sequencing (RNA-seq) would then allow reconstruction of the gene expression history of recordanelle-containing cells.

Here, we confirm that mitochondrial-anchored recordanelles constructed with the phage-derived RBDs ms2 bacteriophage major capsid protein (MCP) or pp7 coat protein (PCP) efficiently recruit engineered RNAs harboring their targeted stem loops ms2 or pp7, respectively. Swapping out the sequence-specific phage RBDs for mRNA-binding domains demonstrates transcriptome-wide recruitment, faithful purification by fractionation and deep-sequencing-based transcriptome reconstruction. Lastly, we show that the best-performing recordanelle—recordanelle^V2 containing TOM20^1–70 as the mitochondrial membrane anchor, TAF15N as the assembler peptide and PABPC4∆C^161A as the mRNA-binding domain—records transient and dynamic transcriptional responses resulting from synthetic inducers or interferon (IFN) signaling in bulk samples and in single cells. Critically, these transient transcriptional responses cannot be observed in nonengineered cells and recordanelle expression and transcriptome recording over several days does not alter mitochondrial function or cellular physiology.

RBDs efficiently capture selected RNAs within recordanelles

Toward creating the recordanelle, we determined whether we could recruit specific RNAs within mitochondrial membrane-anchored biomolecular condensates without disrupting the transcriptome (Fig. 2a). For the RBD part of the recordanelle, we used MCP and PCP, which specifically bind ms2 and pp7 stem loops, respectively^13,14. For mitochondrial anchoring, we used the first 70 aa of the human mitochondrial protein TOM20, which has been described previously^15,16. For the assembler protein, we decided to test several intrinsically disordered regions (IDRs) of proteins that have been previously used to engineer synthetic condensates (DDX4C, EWSR1N, FUSN, hnRNPA1C, TAF15N and TDP43C)^17,18,19, as well as two full-length members of the FET family (FUS and EWSR1)^20,21,22, several IDR variants of FUS with distinct condensation behaviors^21,23, several IDR1 and NTF2 variants of the core stress granule protein G3BP1 (refs. ^24,25,26) and the protein UBQL2, which has been shown to form condensates immiscible with other cellular condensates²⁷ (Supplementary Notes). As controls, we created recordanelle constructs without an assembler (∆AP) or RBD.

To characterize the localization of recordanelles and target RNAs, we performed transient transfection experiments in HEK293T cells containing a reporter construct expressing mCherry fused to either ms2 (mCherry::ms2) or pp7 (mCherry::pp7) stem loops (Fig. 2a). Using immunofluorescence (IF) and RNA fluorescence in situ hybridization (FISH), we found that the recordanelles faithfully localize to the mitochondrial surface and specifically recruit the target RNAs (Fig. 2b, Extended Data Fig. 1 and Supplementary Video 1). For instance, using FISH probes targeting the mCherry sequence, we observed a strong enrichment of mCherry::ms2 RNA in MCP-containing but not PCP-containing recordanelles. Swapping mCherry::ms2 for mCherry::pp7 led to mCherry RNA localization exclusively within PCP-containing recordanelles. This demonstrates efficient and specific recruitment of engineered RNAs within recordanelles inside living cells.

To quantify RNA enrichment efficiency across recordanelles, we transiently transfected HEK293T cells with a dual reporter, as well as constructs containing a fixed MCP as the RBD and TOM20^1–70 as the anchor but with varying assemblers. The dual-reporter construct expressed mCherry and EGFP fused to ms2 (mCherry::ms2) and pp7 (EGFP::pp7) stem loops, respectively, enabling the precise quantification of targeted RNAs within the recordanelle compared to a nonspecific RNA expressed from the same construct. Using cellular fractionation and qPCR, we observed an enrichment of mCherry::ms2 RNAs within the recordanelles (that is, mitochondrial cell fraction) with an efficiency ranging from 3-fold to 11-fold compared to a mock control (Fig. 2c and Extended Data Fig. 2a). We also observed enrichment of target RNAs when using a ∆AP control construct. Some recordanelle architectures (for example, FUS^QG and FUSN), however, seem to exhibit a higher mean mCherry::ms2 enrichment (not statistically significant). On the basis of this small-scale screen, we selected several lead candidates for subsequent analysis (FUS^QG, FUSN, EWSR1 and TAF15N).

We also observed nonspecific enrichment of EGFP::pp7 or the endogenous housekeeping mRNA YWHAZ when using a subset of constructs (for example, FUS^6D and G3BP1), which could arise because of nonspecific RNA binding through the intrinsic RBDs of certain assemblers. Consequently, we additionally assessed the selectivity of the MCP-based recordanelles by calculating the ratio of mCherry::ms2 to EGFP::pp7 mRNAs and found values ranging from 1.2-fold to 6-fold. We found several promising candidate assemblers and, for four of these (EWSR1, FUS^QG, FUSN and TAF15N), we tested how their specificity changed when MCP was replaced with PCP and whether they exhibit nonspecific binding in absence of an RBD. Changing the RBD to PCP switched the RNA enrichment from mCherry:ms2 to EGFP::pp7, indicating that the recordanelle can be tuned by altering the RBD (Fig. 2c). Furthermore, in the absence of an RBD, we observed nonspecific RNA binding in the case of EWSR1 and FUS^QG, which we suspect is because of these full-length proteins containing their own RBDs. However, recordanelles with the assemblers TAF15N or FUSN—comprising only IDRs without the C-terminal RBDs—appeared to lack this nonspecific binding activity. Together, this indicates that the recordanelle architecture components must be carefully selected and combined to balance enrichment efficiency and selectivity.

To further characterize RNA recruitment efficiency and specificity, as well as assess the impact of recordanelle expression on the cellular transcriptome, we transiently transfected HEK293T cells with FUS^QG-based recordanelle constructs that exhibited the highest selectivity and performed RNA-seq on full cells and the mitochondrial fraction (Fig. 2d). With MCP-based or PCP-based recordanelles, we did not observe any differentially expressed genes (DEGs) at the full cell level compared to ∆RBD; however, within the recordanelle fraction, the targeted RNA (that is, mCherry::ms2 or EGFP::pp7) was among the top (MCP) or the only (PCP) significantly enriched gene(s). For the MCP-containing recordanelle, we additionally observed one enriched predicted lncRNA and two depleted genes. While we attribute these subtle changes to technical noise, we cannot rule out that the enrichment of the lncRNA is caused by a so far undescribed binding to MCP.

Importantly, we did not observe any alteration of the full cell transcriptome when comparing recordanelle-containing cells to a mock transfection control (Extended Data Fig. 2c), indicating that the expression of the recordanelle does not perturb the cell under these conditions. In sum, we show that recordanelles can be engineered to efficiently, specifically and tunably recruit specific target RNAs without disrupting the transcriptome.

Capturing endogenous transcripts within recordanelles with poly(A) binders

We next set out to extend the use of recordanelles from specific to generalized RNA recruitment by swapping the stem-loop-binding MCP/PCP RBDs for mRNA binders (Fig. 3a). Mammalian mRNAs are hallmarked by the 5′ cap and the poly(A) tail; therefore, we designed several different recordanelles with RBDs targeting these structures, including (1) human EIF4E to bind the 5′ cap²⁸; (2) human LARP1, LARP4A, PABPN1, PABPC1 and PABPC4, yeast Pab1 and Nab2 and Escherichia coli Hfq and CspE to bind the poly(A) tail^{29,30,31,32,33,34}; (3) catalytically deficient PARN^28A,30A to bind the 5′ cap and the poly(A) tail³⁵; (4) two versions of the mRNA nuclear export receptor Nfx1, containing only its RBDs or also the RRM, which are known to nonspecifically bind RNAs³⁶; and (5) several versions of the poly(A)-binding protein PABPC4 (Supplementary Notes).

Just as we fixed the MCP/PCP RBDs when optimizing the assembler, we took the same approach here and fixed the assembler while varying the mRNA binders. To avoid any potential recordanelle saturation or nonspecific RNA binding from the FUS^QG-based assembler, we initially used TAF15N as it exhibited moderate specific RNA enrichment and showed no nonspecific binding in previous assays without additional RBDs (Fig. 2c), providing a clean baseline for assessing the impact of the mRNA RBD on recordanelle performance.

We first used IF to assess whether recordanelles faithfully localized to the mitochondrial membrane. While one recordanelle did not localize as anticipated (PABPN1), all others colocalized with a mitochondrial marker (Fig. 3b and Extended Data Fig. 3a,b). We next tested for general RNA binding using mCherry, EGFP and YWHAZ recruitment within the mitochondrial fraction of dual-reporter-expressing HEK293T cells as a proxy (Fig. 3c and Extended Data Fig. 3c,d). When comparing the recordanelles to control constructs lacking RBDs, we observed a range of recruitment behaviors. In several cases, the recordanelle constructs either weakly or did not enrich mRNA within the mitochondrial fraction and we did not characterize these further. However, the recordanelles based on LARP1/4A and variants of PABPC1/4 showed consistent and efficient mRNA recruitment within the mitochondrial fraction up to 15-fold.

Encouraged by the generalized mRNA recruitment capacity of the recordanelles, we next tested combinations of top-performing assemblers (FUS^QG, FUSN, EWSR1 and TAF15N) and RBDs (PABPC4∆C and PABPC4). Upon quantification of reporter and YWHAZ mRNAs within the mitochondrial and full cell fractions of transiently transfected HEK293T cells, we found that all recordanelle constructs could recruit mRNA within the mitochondrial fraction (Extended Data Fig. 3e,f). Moreover, this recruitment behavior was dependent on the RBD, indicating that recruitment was acting through generalized mRNA binding of the RBD rather than nonspecific binding because of the assembler. Whereas FUS^QG was the most efficient at specific RNA recruitment when coupled with the stem-loop binders MCP and PCP, in these generalized mRNA recruitment experiments, TAF15N exhibited the highest binding of the reporter RNAs and the housekeeping RNA (YWHAZ). We speculate that this could be because of the RBD influencing the biomolecular condensation behavior of the assembler, which has been shown previously in the context of PABPC1 (ref. ³⁷). This suggests that recordanelle performance depends on interactions between their component parts and, thus, combinatorial testing of recordanelle architectures could yield emergent properties.

Transcriptome-wide recording with recordanelles

To deeply characterize recordanelle component combinations, we decided to go beyond using the reporter system and examine generalized mRNA recruitment on a larger unbiased scale with bulk RNA-seq. Specifically, we transiently transfected HEK293T cells with recordanelles based on combinations of top-performing assemblers (FUS^QG and TAF15N) and RBDs (PABPC4∆C, PABPC4∆C^161A, PABPC4 and LARP1) followed by RNA-seq on full cells and the recordanelle. As here we would expect to bind a substantial fraction of the transcriptome, we used synthetic RNA spike-ins to enable accurate global quantification of RNA enrichment³⁸. When comparing each recordanelle to a mock transfection, we observed enrichment in a large portion of the transcriptome within the recordanelle, which was indicated by a positive shift in the log₂ fold change (log₂FC) density plots (Fig. 3d and Extended Data Fig. 4). In the case of FUS^QG-based recordanelles, some transcripts are enriched in the mitochondrial fraction in control constructs without an RBD, confirming the nonspecific RNA binding behavior of FUS^QG observed previously (Fig. 2c), whereas RNA recruitment within TAF15N-based recordanelles was dependent on the presence of the mRNA binder.

To quantify the global enrichment in RNAs within the recordanelle, we set a stringent threshold corresponding to an FC greater than 1.5 and a false discovery rate (FDR) less than 0.01, which revealed that 64% (FUS^QG) and 90% (TAF15N) of genes expressed within HEK293T cells were enriched within the recordanelles compared to a nonbinding control. Importantly, this does not indicate that the recordanelles are bound to 64–90% of all RNA in the cell but, rather, that 64–90% of transcripts are represented in the recordanelles. In contrast, control constructs without an RBD showed an enrichment less than 0.1%. An analysis of DEGs further confirms transcriptome-scale enrichment of mRNAs within the recordanelle (Fig. 3e and Extended Data Fig. 4). Given that TAF15N-based recordanelles showed superior transcriptome-scale binding, we selected them as our lead constructs in subsequent experiments.

This global enrichment occurs without altering the full cell transcriptome, likely because of only a small sampling of the cellular cytoplasmic pool being recruited to the recordanelle. In the presence of a PABPC4-containing or PABPC4∆C-containing recordanelle, we observe a high concordance between the whole cell and mitochondrial fraction (Extended Data Fig. 5), indicating that the enriched RNAs reflect the cytoplasmic RNA content of a cell. The few depleted or enriched transcripts within the mitochondrial fraction compared to the whole cell correspond to nuclear (for example, the lncRNA Xist) or mitochondrial RNAs, respectively, which can be expected from a recorder enriching cytoplasmic RNAs.

Taken together, we have shown that several recordanelles architectures can capture a large fraction of the cytoplasmic transcriptome without altering global RNA levels in the cell.

Recordanelles record transient gene expression events

We have thus far shown that recordanelles can recruit and accommodate a representative sampling of the transcriptome. We next aimed to assess whether the recordanelles could protect (that is, ‘record’) stored information throughout time. To test this, we developed a reporter cell line expressing mCherry::ms2 mRNA under a doxycycline (Dox) inducible promoter. In this cell line, when exposed to a transient Dox stimulus, mCherry::ms2 is robustly expressed but then returns to baseline levels after 1 day (Fig. 4a and Extended Data Fig. 6). By quantifying the recovery of mCherry RNA from one day to the next, we can assess the capacity of recordanelles to protect RNAs throughout time.

To assess mCherry::ms2 protection by TAF15N-based recordanelles, we transfected the reporter cell line with recordanelle constructs followed by transient Dox treatment and mCherry qPCR at different time points. We then quantified the mCherry::ms2 mRNA ‘transcript recovery’ by comparing the ratio between values on days 1 and 0. A high ratio indicates that reporter mRNA levels are similar on days 0 and 1, indicating that the transiently expressed transcript was stabilized. Inversely, a low ratio signifies that most reporter transcripts have disappeared. Constructs containing PCP, MCP, LARP1 or LARP4A led to a transcript recovery at the full cell level of around 0.07, which was similar to the mock control, indicating that these recordanelles do not protect mCherry RNA (Fig. 4b). However, the PABPC1/4-based recordanelles led to 2–4-fold higher transcript recovery after 1 day. Even after 2 days, it was still possible to detect the mCherry transcripts in cells expressing the PABPC1/4-based recordanelles (Extended Data Fig. 6). This indicates that certain recordanelle architectures can both recruit and protect RNA for at least 48 h.

One surprising result from this experiment was the fact that a subset of the recordanelles that we previously showed could recruit RNAs (LARP4A, LARP1 and MCP) did not protect mCherry::ms2 any better than a nonbinding PCP control. This was especially true for MCP-containing recordanelles that robustly bind mCherry::ms2. We speculate that this could be because of the ability of the RBD to protect the mRNA; for instance, whereas MCP binds before the poly(A) tail, PABPC4-based proteins directly bind the poly(A) tail and, thus, may shield the mRNAs against ribonucleases.

To further characterize RNA protection within the TAF15N-based recordanelles containing either PABPC4∆C (recordanelle^V1) or PABPC4∆C^161A (recordanelle^V2), we performed another mCherry::ms2 protection assay. This time, we harvested cells after 1 and 2 days and quantified mCherry::ms2 mRNA in both the whole cell and the recordanelle. In the control recordanelle^MCP condition, we observed no difference in mCherry::ms2 between cells transiently treated with Dox and the untreated negative control after 2 days, at both the full cell and the recordanelle level. In contrast, in the recordanelle^V1 and recordanelle^V2 conditions, we observed significant mCherry::ms2 protection at the full cell (P = 0.0009 and 0.0106, respectively) and mitochondrial (P < 0.0001) levels (Fig. 4c and Extended Data Fig. 6). These data confirm reporter RNA recording and protection, as well as indicate that, through recordanelle purification, we can substantially increase the sensitivity of detecting transient gene expression events. On day 0, when the reporter was maximally induced, mCherry::ms2 expression was higher in recordanelle^V1 but not recordanelle^V2 compared to the control recordanelle^MCP (Extended Data Fig. 6). Because recordanelle^V2 records this transiently expressed RNA without altering its initial expression level, it represents the superior architecture.

Recordanelles enable the reconstruction of transient immune signaling

Having established that recordanelles can record ectopically expressed RNAs under Dox induction, we next wanted to explore whether they could also record more complex and dynamic endogenous gene expression events. For this purpose, we chose IFNB signaling as it has a defined transcriptional response in cells, leading to the upregulation of well-known IFN-stimulated genes (ISGs). We started by confirming that HEK293T cells exposed to increasing concentrations of IFNB show a dose-dependent increase in ISG overexpression (Supplementary Fig. 1).

We then evaluated IFNB stimulus recording in cells expressing recordanelle^V1 or recordanelle^V2 (Fig. 5a,b and Extended Data Fig. 7). As a control, we expressed a recordanelle construct with the RBD MCP (recordanelle^MCP), which should not bind mRNAs (as no ms2-tagged transcripts are present in this experiment). After 20 h of IFNB treatment in cells expressing recordanelles, for all conditions, we observed robust expression of the two representative ISGs—IFIT1 and IFI44L. We then removed IFNB and analyzed the ISG signal at the full cell and mitochondrial level 2 days later. Under control conditions, the ISG signal decayed completely and was indistinguishable from a control that had never been exposed to IFNB, whereas cells expressing recordanelle^V1 or recordanelle^V2 retained this information throughout time.

Under certain conditions we observed that, in recordanelle^V1-containing cells, ISG detection was not only maintained but increased from day 0 to 2 after IFNB removal (Fig. 5b and Extended Data Fig. 7a,b); this could occur under conditions where the recorded ISG mRNAs are translated. This increase appeared specific to IFNB signaling as it was absent for the YWHAZ housekeeping gene. As we did not observe this behavior for recordanelle^V2, it seems that this version is superior as an RNA recording tool and we, therefore, proceeded with this recordanelle architecture.

To probe the limits of recording with recordanelle^V2, we next characterized its behavior over an extended period of time (Extended Data Fig. 7c–f, Supplementary Fig. 2 and 3). In nonpassaged cells, the IFNB signal is lost in recordanelle^MCP controls after 2 days, whereas, for recordanelle^V2-expressing cells, we still efficiently recorded ISGs after 5 days (Extended Data Fig. 7c,e). In passaged cells, recordanelles capture ISGs 3–4 days after stimulus removal but the signal is lost from day 5 onward. This indicates that cell division and growth contribute to the dilution of the recorded information.

Encouraged by the ability of recordanelle^V2 to capture the transient expression of specific ISGs, we proceeded with a deeper characterization of the recorded IFNB response with RNA-seq. After 20 h of IFNB treatment, under all conditions, we observed a robust IFNB response, as indicated by the differential expression of numerous ISGs (for example, OAS3 and IFIT3) (Fig. 5c and Extended Data Fig. 7g). Under control conditions, the differential response was undetectable for all but two genes 2 days later (Fig. 5c, top left versus top right). The two genes, IFI6 or IRF9, either exhibited the highest initial induction or are known to have prolonged expression after type I and II IFN stimulation³⁹. In contrast, cells expressing recordanelle^V2 retained all DEGs (that is, 100% recall) throughout time (Fig. 5c, bottom left versus bottom right). From these data, we can also draw three other conclusions. First, from the recorded information, we can reconstruct a detailed picture of the (historical) IFNB response, as indicated by a Gene Ontology term analysis of DEGs retained by recordanelle^V2 (with IFNB compared to no IFNB) within the mitochondrial fraction (Fig. 5d). Second, the response can also be captured at the full cell level, which contains the recordanelle, as indicated by concordance between whole cell and recordanelle LFCs (Fig. 5e and Supplementary Fig. 5). Third, recordanelle expression did not substantially alter the cellular transcriptome in the absence of IFNB treatment as indicated by a lack of differentially enriched biological processes when comparing recordanelle versus controls on day 0 and day 2 (Extended Data Fig. 7h and Supplementary Fig. 6). We further observed that while specific mRNAs can be very efficiently recruited into the mitochondrial fraction without an assembler protein (∆AP control, Fig. 2c), general mRNA recording is diminished when TAF15N is removed from recordanelle^V2 (Extended Data Fig. 7i,j).

One potential concern of a cytoplasmic RNA recorder is that it might lead to a sustained cell response to IFNB stimulation, effectively extending the inflammatory state. To interrogate this, we performed quantitative proteomics for cells expressing recordanelle^MCP or recordanelle^V2, collected before IFNB removal (day 0) and 2 days later. On day 0, we observed robust induction of several ISGs at the protein level. Then, 2 days after IFNB removal, all but two proteins returned to baseline (except RIGI and ISG15) and recordanelle^MCP and recordanelle^V2 were indistinguishable (Fig. 5f and Extended Data Fig. 8a,b). This indicates that, while recordanelle^V2 extends the lifetime of recorded RNAs, it does not lead to excessive protein translation of those recorded RNAs.

Taken together, we show that recordanelle^V2 does not alter the cellular transcriptome or proteome and can record complex and dynamic biological signaling events that are otherwise lost with conventional snapshot methods.

Recordanelles do not affect viability or disrupt mitochondrial function

Beyond not impacting the transcriptome, an ideal recordanelle should have minimal disruption to cellular physiology. To probe this, we examined cellular viability and mitochondrial function upon recordanelle expression (Extended Data Fig. 9).

We first determined that recordanelle^V1 and recordanelle^V2 did not significantly alter ATP production (Extended Data Fig. 9a) and did not substantially increase apoptosis (Extended Data Fig. 9b,c). Moreover, recordanelle expression did not substantially alter mitochondrial membrane potential in living cells (Extended Data Fig. 9d) or significantly affect aerobic metabolism (Extended Data Fig. 9e–n and Supplementary Notes).

Next, we used quantitative proteomics to evaluate the expression of all proteins within the mitochondrial respiratory chain complexes⁴⁰ in recordanelle-expressing cells. Relative to control transfections, we observed no distinct alterations induced by any of the recordanelles (Extended Data Fig. 8c), confirming that recordanelles do not impact mitochondrial function. In addition, when we compared the overall proteome of recordanelle^V2 to a mock transfection, we did not see any biological process being significantly affected (Supplementary Figs. 7–13).

Taken together, these results demonstrate that recordanelles do not substantially impact cell viability, the cellular proteome or mitochondrial functionality measured at the bulk and single-cell level under the conditions tested in this study. Moreover, we cannot exclude that the recordanelles have an impact beyond what was measured in these assays.

Recordanelles capture immune signaling at the single-cell level

Encouraged by our findings that recordanelles capture transient immune signaling events at both mitochondrial and full cell levels in bulk assays, we next investigated whether this system could preserve transcriptional signatures at single-cell resolution. To this end, we repeated our transient IFNB stimulation experiment for cells transfected with recordanelle^MCP or recordanelle^V2 (20-h stimulation with 60 U per ml IFNB, followed by 2 days without IFNB). Dimensionality reduction revealed a distinct cell cluster (cluster 6) that emerged exclusively upon IFNB stimulation and was entirely absent under unstimulated conditions on day 0 (Fig. 5g and Extended Data Fig. 10a). The top marker genes driving this cluster included many of the core ISGs previously detected by bulk RNA-seq (Fig. 5c,h); therefore, we term cluster 6 the IFNB response cluster (IRC). Across conditions at day 2, cluster proportions showed that, apart from two minor construct-specific clusters (cluster 8 for recordanelle^V2 and cluster 9 for recordanelle^MCP), unstimulated cells were largely indistinguishable (Fig. 5i). In contrast, among IFNB-stimulated cells at day 2, only recordanelle^V2 retained a substantial proportion of cells within the IRC. This demonstrates that recordanelle^V2 efficiently records IFNB stimulation at the single-cell level.

To specifically study the temporal evolution of this signal, we constructed a gene activity module using the ISGs defined in our bulk RNA-seq data. After validating that expression of this module was especially prominent in the IRC (Extended Data Fig. 10b), we observed that IFN activity was absent in all unstimulated cells (Extended Data Fig. 10c). Stimulated cells expressing recordanelle^MCP exhibited a substantial ISG signature at day 0 that disappeared completely 2 days after IFNB removal (Fig. 5j). In contrast, cells expressing recordanelle^V2 exhibited a similar signature that was retained 2 days after IFNB removal (Fig. 5k).

In conclusion, this demonstrates that recordanelle^V2 allows for efficient recording of endogenous immune signaling at the single-cell level.

Here, we introduce the recordanelle, a transcriptome-wide RNA recording technology for mammalian cells that continuously samples the cellular transcriptome in single cells. The recordanelle operates by enriching and protecting endogenous RNAs in a synthetic organelle mounted on the surface of mitochondria. This is achieved with a single fusion protein containing a mitochondrial anchor, assembler peptide and RBD. Our optimized recordanelles capture a subset of all mRNAs—reflecting up to 90% of the transcriptome—over multiple days and cell divisions, enabling transient and complex transcriptional events to be archived within single cells. Recordanelles achieve this without substantially altering cellular viability or mitochondrial function, as well as the cellular transcriptome and proteome under the conditions tested in this study. The degree of the apparent recording seems to be influenced by the cellular state, such that, without passaging, we can record IFNB signaling for at least 5 days, whereas the signal decays more quickly with passaging. We interpret this as proliferation increasing the turnover of the transcriptional record, whereby previously recorded transcripts can be lost and newly transcribed mRNAs are recorded in their place.

While the data reported here suggest that recordanelles do not substantially perturb cellular function under the tested conditions, it is important to clarify that transcriptional recorders such as recordanelles by their very nature are perturbations to cells. These perturbations include the burden of transgene expression, occupancy of the mitochondrial membrane and its immediate surroundings, unnatural interactions with the RNAs they track, depletion of RNAs from the cellular pool and thermodynamic costs of information storage. As the field moves from building technologies to applying them to ask fundamental questions about biology, exactly how these perturbations manifest and potentially confound our interpretation of biology remains to be seen.

Molecular recording technologies have emerged to solve fundamental challenges in observing, tracking and measuring biological systems throughout time. Whereas prior strategies rely on encoding information by modifying DNA, RNA or protein, our approach differs in that it functions by continuously storing unmodified biological molecules within a protective intracellular compartment that is straightforward to extract and sequence independent. Recordanelles are scalable, unbiased, fast, continuous and tunable and can record the whole transcriptome with single-cell sensitivity. Additionally, they only require the expression of a single protein of approximately 91 kDa (2.5 kb), which is small enough to be efficiently transfected or delivered with common viral vectors such as lentiviruses or adeno-associated viruses. We anticipate future efforts directed at improving or extending the recordanelle concept will focus on incorporating additional enzymes to additionally encode lineage or temporal order or recording biological molecules beyond RNA.

While engineering and optimizing recordanelles for RNA recording, we noted several alternative uses of the tool that offer additional opportunities. For instance, we demonstrated that the recordanelles could be used to profile the specificity of an RBD in living cells. This represents a live-cell alternative to other RBD profiling methods such as immunoprecipitation, RIP-seq, CLIP-seq and others^41,42. Furthermore, we showed that recordanelle^V1 may facilitate translation of the recorded RNAs, which led to amplified downstream signaling in the IFNB stimulation experiments, presumably occurring through the intrinsic feedback loop for this cytokine. Thus, recordanelle^V1 may be leveraged for artificially sustaining and amplifying intracellular signals of interest, which would be particularly advantageous in applications such as directed differentiation.

During the preparation of our manuscript, another transcriptome recording tool called TimeVault was introduced⁶. Whereas TimeVaults use free-floating vault particles that have been predominately tested for capturing 24-h expression snapshots; recordanelles use membrane-anchored biomolecular condensates to continuously record expression. While the tools are complementary, one main advantage of recordanelles is that they can continuously sample the transcriptome over arbitrary time windows and are, therefore, uniquely suited to biological questions that extend beyond 24 h and where the precise time of interest is unknown. Furthermore, recordanelles are sensitive enough to capture cellular signaling events in single cells.

In sum, recordanelles can continuously and unbiasedly record the transcriptome throughout time in single living mammalian cells, which offers numerous opportunities for understanding cellular function in the context of health and disease. Archiving transcriptional histories in the same living cells throughout time preserves transient information that is otherwise lost, thus providing a more detailed understanding of biology and how complex biological processes unfold.

Cell culture

All mammalian cell experiments were performed in authenticated HEK293T cells (American Type Culture Collection, CRL-3216) and tested for Mycoplasma. Cells were grown at 37 °C and 5% CO₂ in DMEM with 4.5 g l⁻¹ D-glucose, sodium pyruvate and GlutaMAX (Gibco) supplemented with 1% penicillin–streptomycin (Thermo Fisher Scientific), 10% FBS (Thermo Fisher Scientific) and 10 mM HEPES (GIBCO). Cells were passaged every 2 or 3 days. For culture of genetically engineered cell lines, medium was supplemented with puromycin (Thermo Fisher Scientific) at a final concentration of 3 µg ml⁻¹.

For transient transfections, cells were seeded at a cell density of 110,000 cells per well in 0.5 ml of growth medium for 24-well plates (Greiner), 500,000 cells per well in 2 ml of growth medium for six-well plates (Greiner) 1 day before transfection.

When indicated, in live-imaging experiments, transfected cells were transferred 1 day after transfection to Nunclon Delta Surface 48-well plates (Thermo Fisher Scientific).

In viability measurements using CellTiter-Glo, cells were seeded at a density of 15,000 cells per well in 100 µl of growth medium in 96-well white Nunclon Delta Surface plates (Thermo Fisher Scientific).

For IF experiments, cells were seeded on four-well Lab-Teks no. 1.0 borosilicate coverglass or 24-well glass bottom plates previously coated for 4 h at 37 °C, 5% CO₂ with 500 µl of 7.2 µg ml⁻¹ polyethyleneimine (branched, 25,000 Da; Sigma-Aldrich, 408727) in PBS and rinsed three times with 500 µl of PBS.

Constructs and cloning

Plasmids were generated by restriction enzyme-based cloning or Gibson assembly on pcDNA™3.1/Zeo⁽⁺⁾-derived backbones. Inserts were obtained from synthetized gene fragments or available Addgene plasmids. A list of all used plasmids is available in Supplementary Table 1. The sequence source for the different components was obtained as follows. Templates containing TOM20::EWSR1::MCP, TOM20::EWSR1::PCP, FUS, mCherry and EGFP were kind gifts from the E. A. Lemke lab. UBQL2 was ordered as a gene from GenScript; watermarked sequences for TAF15N, PABPC4 and EWSR1, the FUS IDR variants (FUS^6D, FUS^GA, FUS^QG, FUS^QG6D, FUS^SA, FUS^6D), G3BP1, G3BP1^∆ and G3BP1^{∆, F33W} and the RBDs PABPN1, PARN^28A,30A, CspE, Hfq, LARP1, LARP4A, Nfx1 and Nab2 were ordered as gene fragments from TWIST. The following sequences were derived from Addgene plasmids: DDX4 (101225), hnRNPA1C (101226), TDP43C (122669), EIF4E (178474), PABPC1 (155805), PABPC4 (19877) and Pab1 (37234).

The bidirectional promoter for the dual reporter (pEF1α-mCherry::ms2, pCMV-EGFP::pp7) was based on Addgene (plasmid 186980).

The pT2 donor plasmid with Dox-inducible mCherry was generated by cloning the mCherry::ms2 sequence into the pT2 backbone (Addgene, 62541) together with the tetOne system (Addgene, 138527), mTagBFP2 and the puromycin resistance gene (Addgene, 17433).

The mock plasmid consists of an empty pcDNA™3.1/Zeo⁽⁺⁾.

Transient transfection

Cells were transfected (at a confluency of 70–80%) using jetPRIME transfection reagent (Polyplus) following the manufacturer’s instructions (50 µl buffer, 250–500 ng of DNA and 0.5–1 µl of reagent for 24-well plates; 200 µl of buffer, 1–2 µg of DNA and 2–4 µl of reagent for 6-well plates; 10 µl of buffer, 50 ng of DNA and 0.1 µl of reagent for 96-well plates). Medium was exchanged 4 h after transfection.

For mitochondrial mRNA recruitment experiments, cells were transfected at a ratio of 1:1 with the recordanelle construct and the dual reporter (pEF1α-mCherry::ms2, pCMV-EGFP::pp7). For IF and FISH experiments, cells were transfected at a ratio of 1:1 with recordanelle plasmid and ms2-reporter plasmid (pCMV-mCherry::ms2) or a pp7-reporter plasmid (pCMV-mCherry::pp7).

For Dox recording experiments, cells were transfected with recordanelle plasmids and Dox was added upon medium exchange at a final concentration of 50 ng ml⁻¹. Then, 20 h after Dox treatment, cells were either harvested or rinsed with medium before adding fresh medium. Additional time points were then collected after 1 or 2 days.

For IFNB recording experiments, cells were transfected with recordanelle plasmids and human IFNB1a (Thermo Fisher, PHC4244) was added at a final concentration of 60 U per ml, 4 h after transfection. Then, 20 h after IFNB treatment, cells were either harvested or rinsed with medium (for at least 15 s) before adding fresh medium. Additional time points were then collected after 1, 2, 3, 4 or 5 days; optionally, cells were split after 2 days to collect additional time points (3, 4 and 5 days); optionally, cells were split again to collect samples after 6 and 7 days.

Stable cell line generation

Genetically engineered cells expressing mCherry::ms2 under a Dox-inducible promoter were generated by transfecting cells in six-well plates as indicated previously with the pT2 donor plasmid and the SB100X Sleeping Beauty transposase plasmid (Addgene, 34879) at a 9:1 ratio. The following day, cells were split into T75 flasks. Selection was started the following day using puromycin (3 µg ml⁻¹) and maintained for a minimum of 7 days, with passaging according to cell density.

IF labeling, FISH and confocal microscopy

For IF, 1 day after transfection, cells were rinsed with PBS and fixed in 2% paraformaldehyde in PBS for 10 min at room temperature (RT), rinsed with PBS and permeabilized with 0.5% Triton X-100 in PBS (15 min, RT). Subsequently, they were rinsed twice with PBS and blocked with 3% BSA in PBS (90 min, RT). Then, cells were incubated overnight at 4 °C with primary antibody in blocking solution: Ab_HA (1:2,000; Sigma-Aldrich, h9658) used to stain epitope tagged recordanelles, Ab_AIF (1:500; Abcam, ab32516) as a mitochondrial marker and/or Ab_mCherry (1:2,000; Abcam, ab205402) used to boost mCherry fluorescence signal. The following day, cells were rinsed with PBS and incubated for 1 h at RT with secondary antibody in blocking solution (Thermo Fisher, A32728, A21206 and/or A-11041; 1:1,000). Cells were then washed with PBS and nuclei were stained with 1 mg ml⁻¹ Hoechst 33342 (Sigma-Aldrich) in PBS for 10 min at RT. Next, they were rinsed twice with PBS and fresh PBS was added before imaging.

For FISH coupled to IF imaging, FISH experiments were performed 1 day after transfection. Custom Quasar 670-dyed Stellaris FISH probes were designed against mCherry using the Stellaris RNA FISH Probe Designer (Biosearch Technologies) available online (www.biosearchtech.com/stellarisdesigner; version 4.2). Probe sequences are available in Supplementary Table 2.

For the staining shown in Fig. 2b and Extended Data Fig. 1c, the hybridization protocol was adapted from a previous study⁴³. Cells were rinsed with PBS and fixed with 4% freshly prepared formaldehyde in PBS (1 h, RT). Then, they were rinsed with PBS and permeabilized with 0.5% Triton X-100 solution in PBS (15 min, RT). Next, cells were rinsed with PBS and 2× SSC (0.3 M NaCl, 30 mM sodium citrate pH 7.4 and 50% formamide) and incubated with the hybridization buffer (0.6 M NaCl, 60 mM sodium citrate pH 7.4, 100 mg ml⁻¹ dextran sulfate (Sigma-Aldrich, D8906), 198 ng ml⁻¹ BSA, 2 mM VRC (ribonucleoside vanadyl complexes; Sigma-Aldrich, 94742), 80 ng ml⁻¹ Saccharomyces cerevisiae tRNA (Sigma-Aldrich, R8759), 50% formamide and the custom pool of FISH probes (250 nM) at 37 °C overnight in a sealed, humidified chamber. Subsequently, cells were washed with 4× SSC (0.6 M NaCl, 60 mM sodium citrate pH 7.4 and 50% formamide), 2× SSC (0.3 M NaCl, 30 mM sodium citrate pH 7.4 and 50% formamide), 1× SSC (0.15 M NaCl, 15 mM sodium citrate pH 7.4 and 50% formamide), 0.5× SSC (75 mM NaCl, 7.5 mM sodium citrate pH 7.4 and 50% formamide) and Tris-HCl NaCl buffer (0.1 M Tris-HCl and 150 mM NaCl, pH 7.5), each wash was incubated for 5 min at RT. Cells were then blocked with 3% BSA for 90 min at RT before they were incubated with primary antibodies overnight: Ab_HA (1:2,000; Sigma-Aldrich, h9658) to stain HA–epitope-tagged recordanelle constructs, Ab_AIF (1:500; Abcam, ab32516) and/or Ab_mCherry (1:2,000; Abcam ab205402). The following day cells were rinsed with PBS and incubated for 2 h at RT with secondary antibody in blocking solution (Thermo Fisher, A11001, A-11041 and/or A48254; 1:1,000). Next, they were rinsed twice with PBS and fresh PBS was added before imaging.

For the staining shown in Supplementary Fig. 4, the following protocol was used. First, 1 day after transfection, cells were rinsed with PBS and fixed in 2% paraformaldehyde in PBS for 10 min at RT, rinsed with PBS and permeabilized with 0.5% Triton X-100 in PBS (15 min, RT, 0.2 U per ml Protector RNase inhibitor (RNAINH-RO)). Cells were then blocked with 3% BSA in PBS (0.2 U per ml RNAINH-RO) for 90 min at RT before they were incubated with primary antibodies overnight: Ab_HA (1:2,000; Sigma-Aldrich h9658) to stain HA–epitope-tagged recordanelle constructs, Ab_AIF (1:500; Abcam, ab32516) and/or Ab_mCherry (1:2,000; Abcam, ab205402), with 3% BSA in PBS and 0.2 U per ml RNAINH-RO. The following day, cells were rinsed with PBS and incubated for 2 h at RT with secondary antibody in blocking solution (Thermo Fisher, A11001, A-11041 and/or A48254; 1:1,000, 0.2 U per ml RNAINH-RO). Next, cells were rinsed with PBS and 2× SSC (0.3 M NaCl, 30 mM sodium citrate pH 7.4 and 10% formamide) and incubated with the hybridization buffer (0.6 M NaCl, 60 mM sodium citrate pH 7.4, 100 mg ml⁻¹ dextran sulfate (Sigma-Aldrich, D8906), 198 ng ml⁻¹ BSA, 0.2 U per ml RNAINH-RO (Sigma-Aldrich, 94742), 80 ng ml⁻¹ S. cerevisiae tRNA (Sigma-Aldrich R8759), 10% formamide and the custom pool of FISH probes (250 nM) at 37 °C overnight in a sealed, humidified chamber. Subsequently, cells were washed with 4× SSC (0.6 M NaCl, 60 mM sodium citrate pH 7.4 and 10% formamide), 2× SSC (0.3 M NaCl, 30 mM sodium citrate pH 7.4 and 10% formamide),1× SSC (0.15 M NaCl, 15 mM sodium citrate pH 7.4 and 50% formamide), 0.5× SSC (75 mM NaCl, 7.5 mM sodium citrate pH 7.4 and 10% formamide) and Tris-HCl NaCl buffer (0.1 M Tris-HCl and 150 mM NaCl, pH 7.5). Each wash was incubated for 5 min at RT. Next, they were rinsed with PBS and fresh PBS was added before imaging.

Confocal microscopy images were acquired using a Leica SP8-FALCON microscope with an HC PL APO CS2 ×63/1.40 oil-immersion objective using 405-nm (for Hoechst and Alexa Fluor Plus 405), 488-nm (for Alexa Fluor 488), 577-nm or 578-nm (for Alexa Fluor 568), 650-nm (for Quasar 670) and 653-nm (for Alexa Fluor Plus 647) laser lines for fluorophore excitation. Images were processed using FIJI⁴⁴. Three-dimensional (3D) reconstructions were assembled using Imaris.

Mitochondria isolation

First, 24 h after transfection in six-well plates, cells were washed with PBS, detached and resuspended in 1 ml of growth medium. Two wells of the same conditions were pooled, pelleted at 500g for 5 min at 4 °C and washed with 1 ml of PBS. For experiments involving full cell RNA analysis, 10–20% of the sample was set aside and stored in RNA shield. The remaining sample was processed for mitochondrial isolation using the mitochondria isolation kit for cultured cells (Thermo Fisher, 89874) in 2-ml protein LoBind tubes. Buffers A and C were supplemented with 1× HALT protease inhibitor cocktail (Thermo Fisher) and buffer B was diluted 1:11 in buffer A (supplemented with HALT protease inhibitor). Cells were resuspended in 350 µl of buffer A by vortexing for 5 s at medium speed and incubated on ice for exactly 2 min. Then, 55 µl of the A + B mixture was added and cells were vortexed over the course of 5 min in 5-s intervals at maximum speed, kept on ice in between. Next 400 µl of buffer C was added and the tubes were gently inverted to mix. Cell debris was pelleted at 700g for 10 min at 4 °C and the supernatant, containing mitochondria, was transferred to a new 2-ml protein LoBind tube. Mitochondria were then pelleted at 3,000g for 15 min at 4 °C. The supernatant was discarded; the pellet was washed with 250 µl of buffer C and centrifuged at 12,000g for 5 min at 4 °C. After removing the supernatant, the purified mitochondrial pellet was brought to RT before RNA extraction.

RNA extraction and reverse transcription

Extraction of RNA from the isolated mitochondria or full cell analysis was extracted using the Quick-RNA Microprep Kit (Zymo) following the manufacturer’s instructions. Optionally, for storage in between cells or mitochondrial, pellets were stored in DNA/RNA shield (Zymo). RNA concentrations were measured using a Nanodrop (Thermo Fisher Scientific) and, for each sample, 100–250 ng of RNA was used as input for reverse transcription using qScript Ultra SuperMix (Quanta) following the manufacturer’s instruction.

Cell-to-CT RNA extraction

For Dox recording and IFNB recording experiments in 24-well plates, RNA was extracted using a Homebrew cell-to-CT protocol adapted from a previous study⁴⁵. Cells were washed with PBS and lysed in 200 µl of lysis buffer (9.6 mM Tris-HCl pH 7.8, 0.5 mM MgCl₂, 0.44 mM CaCl₂, 10 μM DTT, 0.1% (w/v) Triton X-114 and 3 U per ml proteinase K in UltraPure water, pH ~7.8). Before lysis, buffer was supplemented with additional DNAse I for a final concentration of 300 U per ml (Sigma-Aldrich, D2821-50KU) and proteinase K for a final concentration of 6 U per ml (Sigma-Aldrich, P4850-5ML). Cells were lysed by carefully pipetting up and down avoiding generation of bubbles. The lysis reaction was incubated at RT for exactly 8 min. Next, lysed solution was transferred to stop solution (1 mM proteinase K inhibitor (EMD Millipore, 539470-10MG), 90 mM EGTA and 113 μM DTT in UltraPure water) on ice at a ratio of 11:1. Stopped solution was incubated for 5 min on ice and 1.5 µl was used as input for reverse transcription using qScript Ultra SuperMix (Quanta) and a final reaction volume of 10 µl.

qPCR

cDNA transcripts were quantified with at least two technical replicates using KAPA SYBR FAST qPCR kits (Roche) and a 384-well QuantStudio5 (Thermo Fisher) according to the manufacturer’s instructions, with 2 μl of prediluted template (1:10 for cell-to-CT samples, 1:200 for Quick-RNA-Microprep samples) and 400 nM primer pair for a final total reaction volume of 10 μl. A water negative control was included to detect possible reagent contamination or nonspecific amplification. Gene expression was analyzed normalizing to MT-ATP6 (housekeeping mitochondrial gene) or YWHAZ (housekeeping cytosolic gene) using the 2^−ΔΔCt method⁴⁶. A complete list of used qPCR primers is available in Supplementary Table 3.

Spike-in addition for quantification of generalized enrichment

For bulk RNA-seq experiments assessing global RNA enrichment or stabilization (including the capture of specific RNAs through MCP/PCP and endogenous transcripts through poly(A)-binding proteins), a fixed amount of ERCC spike-in mix no. 1 (Thermo Fisher, 4456740) was added to the RNA elution buffer. Samples were processed for either whole-cell or mitochondrial RNA extraction as previously described. For each condition, two wells of a six-well plate were pooled (total volume: 2 ml) and the pooled material was split into an 1.8-ml fraction for mitochondrial isolation followed by RNA extraction and a 100-µL fraction for full cell RNA extraction. RNA was eluted in 46 µl of elution buffer consisting of 40 µl of nuclease-free water and 6 µl of diluted ERCC spike-in mix no. 1 (diluted 1:100 in nuclease-free water).

RNA-seq library preparation

Bulk RNA-seq libraries were prepared using the Illumina stranded mRNA prep kit (Illumina) according to the manufacturer’s instructions (document 1000000124518 v02) starting from 500 ng (for transcript enrichment experiments) or 250 ng (for IFNB recording experiments) of purified RNA. Input RNA was quantified using the Qubit RNA BR assay kit on a Qubit Flex fluorometer (Thermo Fisher Scientific). Library fragment-size distributions were individually assessed using the 5300 Fragment Analyzer system (Agilent) following a 1:10 dilution. Libraries were pooled and sequenced on an AVITI system (Element Biosciences) using the AVITI 150-cycle sequencing kit (Cloudbreak FS High Output) in single-end mode across multiple runs to achieve the desired sequencing depth. All experiments were performed using three independent biological replicates per condition.

To characterize mitochondrial fraction enrichment and full cell global RNA perturbation of recordanelle^V2, purified RNA with ERCC spike-in mix no. 1 was obtained as indicated previously. RNA was quantified using the Qubit RNA BR assay kit on a Qubit Flex fluorometer (Thermo Fisher Scientific) and, where applicable, diluted below 100 ng µl⁻¹ in nuclease-free water. Purified RNA was then processed using the bulk RNA-seq service from Plasmidsaurus. The experiment was performed using five biological replicates per condition.

RNA-seq alignment and quantification

Reference genomes were constructed individually for each condition by appending plasmid-derived exogenous transcripts (to avoid read multimapping between identical sections in different recordanelles) and, where applicable, ERCC spike-in sequences to the human reference genome (GRCh38, Ensembl release 115). Raw reads were processed using the nf-core/rnaseq pipeline (version 3.21.0)⁴⁷ with the STAR-Salmon aligner.

For libraries prepared using the Illumina Stranded mRNA prep kit, preprocessing parameters included 5′ clipping (1 bp) and a minimum length of 50 bp (–clip_R1 1 –length 50). Alignment was refined using STAR splice junction overhang constraints (–alignSJDBoverhangMin 5 and –alignSJoverhangMin 10) and Salmon quantification included correction for sequence and G+C bias (–seqBias –gcBias).

For libraries prepared by Plasmidsaurus, preprocessing parameters included read deduplication using UMI-tools and alignment refinement using STAR, with parameters adapted from Plasmidsaurus recommendations (–outFilterIntronMotifs RemoveNoncanonical –alignIntronMax 1000000 –alignIntronMin 20 –alignSJoverhangMin 8 –alignSJDBoverhangMin 1 –outFilterType BySJout –outFilterMismatchNoverLmax 0.1). Transcript quantification was performed using Salmon without length correction (–noLengthCorrection) to account for the 3’-biased nature of the libraries.

Bulk RNA-seq analysis in R

Gene-level quantification matrices from the nf-core output were analyzed in R (version 4.5.2); the G+C-corrected and length-corrected matrix (salmon.merged.gene_counts_length_scaled) for the Illumina stranded libraries or the uncorrected matrix (salmon.merged.gene_counts) for the 3’-biased Plasmidsaurus libraries were initially merged and empty rows (0 counts across all samples within the experiment) were removed. Counts per million (CPM) were computed using edgeR (version 4.8.2) and genes with low expression were filtered out if they failed to meet the criteria of maximum CPM ≥ 7 in at least one condition and mean CPM ≥ 5 across samples. Differential expression was modeled using DESeq2 (version 1.50.2) using a no-intercept design, where the grouping variable encodes the full experimental condition of each sample, to facilitate arbitrary contrast between conditions. Library size normalization was experiment dependent. For the generalized transcript enrichment experiments, where the recordanelle is expected to enrich a large fraction of the transcriptome, which violates the assumption that most genes are unchanged, size factors were estimated with DESeq2 ‘estimateSizeFactors’ using the ERCC spike-ins and with mitochondrially encoded transcripts as control features. For the specific transcript enrichment and IFNB recording experiments, default DESeq2 size factors (median of ratios across all genes) were used. Pairwise comparisons were performed using the Wald test and effect sizes were reported as FCs with Benjamini–Hochberg FDR correction. For the generalized transcript enrichment experiments characterizing recordanelle^V2, independent filtering was disabled because the near-identical P-value distributions otherwise assigned NA adjusted P values to a large fraction of genes. DEGs were defined on the basis of experiment-specific thresholds: |log₂FC| > 1.5 and FDR < 0.01 for specific transcript enrichment experiments, |FC| > 1.5 and FDR < 0.01 for generalized transcript enrichment experiments and |FC| > 2 and FDR < 0.01 for IFNB recording experiments. Genes with extreme effect sizes (|log₂FC| > 10 in any contrast) were excluded to prevent compression of the volcano plot.

When indicated, Gene Ontology term analysis was performed in R using enrichR (version 3.4) and the database ‘GO_Biological_Process_2025’.

scRNA-seq library preparation and analysis

Cells were transfected in six-well plates with recordanelle^MCP or recordanelle^V2, stimulated with IFNB as previously described and harvested after 20 h of IFNB stimulation or 48 h of recovery after IFNB removal. At the moment of harvest, cells were washed with PBS, detached with TrypLE and resuspended in 1 ml of growth medium. Cells were then pelleted at 500g for 5 min at 4 °C and washed with 1 ml of 0.04% BSA in PBS. Cells were filtered using 70-μm Flowmi cell strainers (Bel-Art), counted and diluted in 0.04% BSA in PBS at a concentration of 1.5 million cells per ml.

Cell numbers and viability were subsequently assessed on a Cellometer K2 image cytometer (Nexcelom Bioscience) using ViaStain AOPI staining solution (Nexcelom Bioscience, CS2-0106-5mL) and PD100 cell counting slides (Nexcelom Bioscience, CHT4-PD100-003).

Single-cell emulsion, barcoding and gene expression library generation were performed on the 10x Genomics Chromium platform using the GEM-X Universal 3′ gene expression v4 assay with 4-plex on-chip multiplexing (OCM), according to the manufacturer’s instructions (10x Genomics, CG000768 Rev B). Four samples per time point were multiplexed per OCM reaction, loading cell suspensions to recover approximately 5,000 cells per sample following the manufacturer’s recommendation. For library generation, cDNA was initially amplified with 11 PCR cycles and sequencing libraries were initially prepared with five cycles of sample index PCR. Because the first two sequencing runs gave low output, libraries were subsequently reprepared from 500 ng of amplified cDNA using 11 cycles of index PCR. Amplified cDNA and final libraries were analyzed on a Fragment Analyzer 12-capillary system (Agilent, 5200 Fragment Analyzer, M5310AA) using the next-generation sequencing HS reagents (Agilent, HS NGS Fragment Kit (1–6,000 bp), DNF-474-1000) and the Qubit 4 fluorometer (Thermo Fisher Scientific, Q33238) with the Qubit 1× dsDNA HS assay kit (Thermo Fisher Scientific, Q33231). Libraries were pooled in equimolar amounts based on Fragment Analyzer molarity and repooled after an initial sequencing run to equalize read depth across samples. Sequencing was performed on an AVITI24 system (Element Biosciences) using the Cloudbreak Freestyle high-output 150-cycle chemistry at a loading concentration of 25 pM with a 1% PhiX spike-in and a run configuration of 29/10/10/91 cycles (read 1/index i7/index i5/read 2) across multiple runs to achieve the desired sequencing depth.

Raw reads were aligned and quantified using Cell Ranger version 10.0.0 (10x Genomics). A custom reference was built with ‘cellranger mkref’ from the human reference genome (GRCh38 Ensembl release 115) restricted to protein-coding genes and extended with two ectopic sequences composed by the full-length recordanelle^V2 and the MCP sequence surrounded with corresponding 50 bp junctions for recordanelle^MCP to avoid read multimapping. Each 4-plex OCM pool was processed with ‘cellranger multi’, with samples demultiplexed by their barcodes (OB1–OB4), and the eight per-sample outputs were combined with ‘cellranger aggr’ using mapped-read depth normalization. For downstream analysis, transgene unique molecular identifiers (UMIs; summed across the two appended features) were assigned per cell only to the recorder construct expressed in the corresponding sample.

Cell Ranger output was subsequently analyzed in R (version 4.5.2) using Seurat (version 5.4.0). Potential doublets and cells with reduced complexity were filtered out applying thresholds on the basis of the number of genes detected and the mitochondrial fraction per cell (<1,500 or>7,000 genes detected,>8% MT reads). Gene counts were log-normalized and cell cycle S and G2/M phase scores were computed using established gene sets⁴⁸. The 2,000 most variable genes were selected by the variance-stabilizing transformation, excluding the recorder transgenes, and the data were scaled regressing out the S and G2/M scores and the total UMI count per cell. Principal component analysis was performed on the scaled expression matrix (50 principal components) and, to correct for the effect introduced by day, principal components were integrated with Harmony (version 1.2.4). The first 20 Harmony dimensions were used to compute a uniform manifold approximation and projection (UMAP) embedding and a shared-nearest-neighbor graph; cells were clustered with the Louvain algorithm at a resolution of 0.5. Residual low-complexity cells were removed at the cluster level; one first-pass cluster with a median gene count below 2,200 was discarded and the surviving cells were then reembedded and reclustered as described above. Cluster marker genes were identified with Seurat’s ‘FindAllMarkers’ (detected in at least 25% of cells in either group with an absolute log₂FC of at least 0.25) downsampling each cluster to a maximum of 4,000 cells. A per-cell IFNB response score was computed with Seurat’s ‘AddModuleScore’ over the ISG signature defined by bulk RNA-seq (IFI6, ISG15, IRF9, IFI44L, OAS3, BST2, PARP9, IFIT1, IFITM1, DTX3L, STAT1, DDX60, IFIT3, IFITM3 and LGALS3BP). The cluster with the highest mean ISG activity was further defined as the IRC.

Quantification of apoptosis through annexin V staining

Toxicity of various recordanelle plasmids was assessed using the Incucyte SX5 (version 20231.3.3.0 Rev2) equipped with G/O/NIR modules (Sartorius) for continuous monitoring of apoptosis. Cells were seeded in 24-well plates and transfected as described above. Then, 4 h after transfection, the growth medium was replaced and supplemented with Incucyte annexin V NIR dye (Sartorius, 4768) at a 1:400 dilution. As a positive control for apoptosis, a mock-transfected condition was supplemented with puromycin (Thermo Fisher Scientific) at a final concentration of 3 µg ml⁻¹. Cells were maintained at 37 °C in 5% CO₂ and imaged every 90 min for 48 h using the AI Scan with a ×10 objective (phase contrast, orange 400 ms and NIR 400 ms; four images per well). For each condition three replicates were analyzed.

To facilitate extended imaging periods without confounding effects arising around 100% confluency, an independent set of cells was transfected and reseeded at a lower density in a parallel experiment. Cells were seeded in 24-well plates and transfected as previously indicated. Then, 24 h after transfection, cells were washed with PBS, detached with TrypLE and resuspended in 1 ml of growth medium. Next, 60 µl of the cell suspension was reseeded in 48-well plates in a final volume of 290 µl to achieve suitable density. After 3 h, growth medium was replaced and supplemented with Incucyte annexin V NIR dye at a 1:400 dilution. As a positive control for apoptosis, cells were treated with puromycin at a final concentration of 3 µg ml⁻¹. Cells were maintained at 37 °C and 5% CO₂ and imaged every 90 min for 66 h using the AI Scan with a ×10 objective (phase contrast, orange 400 ms and NIR 400 ms; four images per well). For each condition four replicates were analyzed.

Data were analyzed using the AI Cell Health module (Sartorius, BA-04871) with the following parameters: phase segmentation sensitivity, 0.8; dead score threshold, 0.15; minimum area, 90 µm²; maximum eccentricity, 0.98. For NIR channel segmentation, ‘surface fit no mask’ was applied. Cells were classified with the cell-by-cell classification using gate values of 140.52 µm² for HEK area and 0.2686 for NIR mean intensity. Toxicity was reported as the fraction of total cells exhibiting high NIR intensity.

Quantification of transfection efficiency using live-cell imaging

Transfection efficiency was assessed using the Incucyte SX5 (version 20231.3.3.0 Rev2) equipped with G/O/NIR modules (Sartorius) for continuous monitoring of fluorescence intensity in the orange channel. Cells were seeded in 24-well plates and transfected as described above with mCherry plasmid or mock plasmid. Cells were maintained at 37 °C and 5% CO₂ and imaged every 90 min for 48 h using the AI Scan with a ×10 objective (phase contrast, orange 400 ms and NIR 400 ms; four images per well). For each condition, three replicates were analyzed.

Data were analyzed using the AI Cell Health module (Sartorius, BA-04871) with the following parameters: phase segmentation sensitivity, 0.8; dead score threshold, 0.15; minimum area, 90 µm²; maximum eccentricity, 0.98. For orange channel segmentation, ‘surface fit no mask’ was applied. Cells were classified with the cell-by-cell classification using gate values of 140.52 µm² for HEK Area and 0.0572 for orange mean intensity. Transfection efficiency was reported as the fraction of total cells exhibiting high orange intensity.

Live quantification of mitochondrial membrane potential

Alteration of mitochondrial function by different recordanelle plasmids was assessed using the Incucyte SX5 (version 20231.3.3.0 Rev2) with G/O/NIR modules (Sartorius) for continuous monitoring of mitochondrial membrane potential. Cells were seeded in 24-well plates and transfected as previously described. As additional controls for mitochondrial perturbation, cells were transfected with a mock plasmid and the growth medium was supplemented with 2 or 8 µg ml⁻¹ Dox 4 h after transfection.

Then, 24 h after transfection, cells were washed with PBS, detached and resuspended in 1 ml of growth medium. Next, 60 µl of the cell suspension was reseeded in 48-well plates in a final volume of 290 µl to achieve suitable seeding density. Then, 3 h later, all wells were supplemented with 145 µl of growth medium containing 80 nM Incucyte MMP orange reagent (Sartorius, 4775) and incubated for 1 h to allow the cells to take up the dye.

Subsequently, controls for mitochondrial depolarization and hyperpolarization were prepared following the manufacturer’s instructions; mock-transfected wells were additionally supplemented with 145 µl of medium containing either FCCP or oligomycin A, achieving final concentrations of 20 µM FCCP and 5 µg ml⁻¹ oligomycin A. Remaining wells received 145 µl of growth medium without additional compounds to reach a final concentration of 20 nM MMP orange reagent across all samples. Cells were maintained at 37 °C and 5% CO₂ and imaged every 120 min for 66 h using the AI Scan with a ×10 objective (phase contrast, green 250 ms, orange 400 ms; four images per well). For each condition, four replicates were analyzed.

Data were analyzed using the AI Cell Health module (Sartorius, BA-04871) with the following parameters: phase segmentation sensitivity, 0.8; dead score threshold, 0.15; minimum area, 90 µm²; maximum eccentricity, 0.98. For the orange channel segmentation, ‘surface fit no mask’ was applied. Cells were classified using the cell-by-cell classification using gate values of 140.52 µm² for HEK area and 0.2686 for orange mean intensity. Mitochondrial membrane potential per cell was calculated as the total mean intensity object average for all classified HEK cells.

Viability measurements

Cell viability for recordanelle-transfected cells in 96-well plates, assessed through ATP production, was quantified using the CellTiter-Glo luminescent cell viability assay (Promega, G7570) according to the manufacturer’s instructions. As a positive control for cell death, an untransfected condition was supplemented with puromycin at a final concentration of 3 µg ml⁻¹ 4 h after transfection. Luminescence was measured 24 and 48 h after transfection on a Tecan SPARK plate reader. For each condition, five replicates were analyzed.

Seahorse assay

Cells were seeded in six-well plates and transfected as described above. As additional controls for mitochondrial perturbation, cells were transfected with a mock plasmid and the growth medium was supplemented with 2 or 8 µg ml⁻¹ Dox 4 h after transfection. Then, 24 h after transfection, cells were washed with PBS, detached with TrypLE, resuspended in 1 ml of growth medium and subsequently counted.

Mitochondrial respiration was measured with a Seahorse XF96 cellular flux analyzer (Agilent Technologies) using a MitoStress test assay (Agilent Technologies, 103015-100). Seahorse XF96 cell culture microplates (Agilent Technologies, 103794-100) were prepared in advance, coating for 4 h at 37 °C, 5% CO₂ with 100 µl of 7.2 µg ml⁻¹ polyethyleneimine (branched, 25,000 Da, PEI-25K; Sigma-Aldrich, 408727) in PBS and subsequently rinsed three times with PBS. For each condition, 12 wells of the XF96 PEI-25K-coated cell culture microplate were seeded at a density of 20,000 cells per well and incubated for 24 h at 37 °C, 5% CO₂.

After 24 h, the Seahorse measurement was performed following the manufacturer’s instructions. Wells were washed with warmed Seahorse serum-free XF medium (XF base medium supplemented with 25 mM glucose, 2 mM glutamine and 1 mM sodium pyruvate, pH 7.4) and then filled with 180 μl of fresh Seahorse assay medium. Cells were then incubated in a CO₂-free incubator at 37 °C for 45 min before running the assay.

After calibration, the cartridge plate was replaced with the culture plate, and the assay was continued according to the manufacturer’s recommendation. Using Seahorse Wave controller 2.4 software (Agilent), oxygen consumption rate (OCR) and extracellular acidification rate were monitored under basal conditions. This was followed by the sequential addition of oligomycin (1.5 µM), FCCP (0.5 µM) and rotenone + antimycin A (0.5 µM) (final concentration in the well) to estimate the contribution of individual parameters for basal respiration, ATP production, proton leak, maximal respiration and spare respiratory capacity. After the assay, the cell number in each well was measured by Hoechst staining (10 mg ml⁻¹ Hoechst 33342 for 30 min, followed by two washes with PBS). Fluorescence intensity was measured using a Tecan SPARK plate reader and data were normalized to the fluorescence intensity per well.

The Mito Stress Test Report Generator (version 4.0) by the XF96 analyzer automatically calculated the XF cell mito stress test parameters from Wave after cell count normalization and exported to Excel. For OCR curves and individual parameters, the values are presented as the mean ± s.e.m. of three biological repeats in all experiments performed, with at least five replicate wells in the Seahorse XF96 analyzer.

Proteomics

Cells were seeded in six-well plates and transfected with mCherry::pp7, mock, recordanelle^MCP, recordanelle^V1 or recordanelle^V2 constructs as described above. They were subsequently subjected to transient IFNB stimulation as described above. Then, 0 or 2 days after IFNB removal, cells were harvested. To this end, they were rinsed with PBS, detached with TrypLE and subsequently resuspended in medium. Cells were pelleted by centrifugation (500g, 5 min, 4 °C) and washed three times with ice cold PBS. Cells were counted at the last wash step and 250,000 cells were pelleted. Dried pellets were frozen on dry ice and stored at −70 °C until lysis.

Cells were brought to RT and lysed with lysis buffer containing (100 mM TEAB pH 8.5, 5% SDS and 10 mM TCEP) for 10 min at 95 °C followed by 10 min of sonication (30 s on, 30 s off per cycle) on a Pixul system (Active Motif) using the following settings: pulse, 50 N; PRF, 1 kHz; burst rate, 20 Hz. Protein extracts were alkylated using 15 mM iodoacetamide at 25 °C in the dark for 30 min. For each sample, 20 µg of protein lysate was captured, digested and desalted using the SP3 approach⁴⁹ with the fluent liquid handling platform (Tecan Group). In brief, Speed Beads (45152105050250 and 65152105050250, GE Healthcare) were mixed 1:1, rinsed with water and diluted to the 8 μg µl⁻¹ stock solution. Samples were adjusted to a final volume of 45 µl and 10 µl of the bead stock solution was added. Proteins were bound to the beads by addition of 125 µl of 100% acetonitrile to the samples, which were then incubated for 8 min at RT with a gentle agitation (200 rpm). Next, samples were placed on a magnetic rack and incubated for 5 min. Supernatants were removed and discarded. The beads were washed twice with 160 µl of 70% (v/v) ethanol and once with 160 µl of 100% acetonitrile. Samples were placed off the magnetic rack and 50 µl of digestion mix (10 ng µl⁻¹ of trypsin in 50 mM triethylammonium bicarbonate) was added to them. Digestion was allowed to proceed for 12 h at 37 °C. After digestion, samples were placed back on the magnetic rack and incubated for 5 min. Supernatants containing peptides were collected and dried under vacuum.

Dried peptides were resuspended in 0.1% aqueous formic acid and 100 ng of peptides were subjected to liquid chromatography–tandem mass spectrometry analysis using an Astral Zoom mass spectrometer fitted with a Vanquish Neo (both Thermo Fisher Scientific) and a custom-made column heater set to 60 °C. Peptides were resolved using a reverse-phase high-performance liquid chromatography column (75 μm × 30 cm) packed in-house with C18 resin (ReproSil Saphir 100 C18, 1.5 µm of resin; Dr. Maisch) at a flow rate of 0.2 µl min⁻¹. The following gradient was used for peptide separation: from 2% B to 5% B over 1 min to 25% B over 22 min to 35% B over 4.5 min to 95% B over 0.5 min, followed by 4.5 min at 95% B to 2% B over 0.5 min. Buffer A was 0.1% formic acid in water and buffer B was 80% acetonitrile and 0.1% formic acid in water.

The mass spectrometer was operated in data-independent acquisition mode. MS1 scans were acquired in the Orbitrap in profile mode at a resolution of 240,000 full width at half maximum, a scan range from 380 to 980 m/z, normalized automatic gain control (AGC) target set to 500% and a maximum ion injection time of 30 ms. MS2 scans were acquired in the Astral analyzer over a precursor mass range of 380 to 980, an isolation window of 4 m/z without window overlap, a maximum injection time of 8 ms, a scan range from 150 to 2,000 m/z and a normalized AGC target of 800% with window placement optimization enabled. Peptides were fragmented by higher-energy collisional dissociation with collision energy set to 25%, preaccumulation was enabled and one microscan was acquired for each spectrum.

The acquired files were searched using the Spectronaut (Biognosys version 20.3) directDIA workflow against a human database (downloaded from UniProt on June 25, 2025; P00001–P0005 were added manually; Supplementary Table 4) and 392 commonly observed contaminants. Quantitative fragment group data (FG.MS2RawQuantity) were exported from Spectronaut. Protein-level quantification, normalization, missing-value imputation and differential-abundance testing were performed using ProteoFlux version 1.8.8. (https://doi.org/10.5281/zenodo.18640999). Downstream filtering, significance calling and visualization were performed in R. Within each contrast, proteins were retained only if (1) they were quantified (nonimputed) in at least two of three replicates in at least one of the two conditions and (2) they were supported by at least two peptides consistently detected across all replicates of at least one condition. Lastly, proteins whose FC direction was discordant with the underlying MS2 fragment-count evidence (normalized spectral-count misalignment > 1.5, as computed by ProteoFlux) were excluded. Proteins quantified in only one condition of the contrast (present in ≥2/3 replicates of one condition and undetected in the other) were treated as qualitative ‘on/off’ changes. Proteins were considered significantly differentially abundant at an adjusted P value < 0.01 and an absolute log₂FC > 1.

When indicated, GO-term analysis was performed in R using enrichR (version 3.4) and the database ‘GO_Biological_Process_2025’.

Statistics and reproducibility

Statistical analysis of qPCR data, as well as CellTiter-Glo and Seahorse assays, was performed using GraphPad Prism software. The used statistical tests are indicated in the figure legends and significant differences are displayed using asterisks in the figures (*P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001 and ****P ≤ 0.0001).

Regarding FISH experiments, for each condition, multiple fields of view were taken (typically at least ten). Analysis of RNA recruitment by FUS^QG variants was performed once; analysis of RNA recruitment by TAF15N variants was performed at least three times with similar results.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

The code for this publication is available from Zenodo (https://doi.org/10.5281/zenodo.22096128)⁵⁰.

1. Stark, R., Grzelak, M. & Hadfield, J. RNA sequencing: the teenage years. Nat. Rev. Genet. 20, 631–656 (2019).

   Article
   CAS 
   PubMed
   
   Google Scholar

2. Manno, G. L. et al. RNA velocity of single cells. Nature 560, 494–498 (2018).

   Article
   PubMed
   PubMed Central
   
   Google Scholar

3. Herzog, V. A. et al. Thiol-linked alkylation of RNA to assess expression dynamics. Nat. Methods 14, 1198–1204 (2017).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

4. Rodriques, S. G. et al. RNA timestamps identify the age of single molecules in RNA sequencing. Nat. Biotechnol. 39, 320–325 (2021).

   Article
   CAS 
   PubMed
   
   Google Scholar

5. Armbrust, N. et al. Non-destructive transcriptomics via vesicular export. Nat. Commun. 17, 3812 (2026).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

6. Chao, Y.-K., Wu, M., Gong, Q. & Chen, F. A. A genetically encoded device for transcriptome storage in mammalian cells. Science 391, eadz9353 (2026).

   Article
   CAS 
   PubMed
   
   Google Scholar

7. Horns, F. et al. Engineering RNA export for measurement and manipulation of living cells. Cell 186, 3642–3658 (2023).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

8. Choi, J. et al. A time-resolved, multi-symbol molecular recorder via sequential genome editing. Nature 608, 98–107 (2022).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

9. Schmidt, F., Cherepkova, M. Y. & Platt, R. J. Transcriptional recording by CRISPR spacer acquisition from RNA. Nature 562, 380–385 (2018).

   Article
   CAS 
   PubMed
   
   Google Scholar

10. Reinkemeier, C. D. & Lemke, E. A. Synthetic biomolecular condensates to engineer eukaryotic cells. Curr. Opin. Chem. Biol. 64, 174–181 (2021).

    Article
    CAS 
    PubMed
    
    Google Scholar

11. Boeynaems, S. et al. Phase separation in biology and disease; current perspectives and open questions. J. Mol. Biol. 435, 167971 (2023).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

12. Bracha, D., Walls, M. T. & Brangwynne, C. P. Probing and engineering liquid-phase organelles. Nat. Biotechnol. 37, 1435–1445 (2019).

    Article
    CAS 
    PubMed
    
    Google Scholar

13. Lim, F. & Peabody, D. S. RNA recognition site of PP7 coat protein. Nucleic Acids Res. 30, 4138–4144 (2002).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

14. Bertrand, E. et al. Localization of ASH1 mRNA particles in living yeast. Mol Cell 2, 437–445 (1998).

    Article
    CAS 
    PubMed
    
    Google Scholar

15. Stavru, F., Nautrup-Pedersen, G., Cordes, V. C. & Görlich, D. Nuclear pore complex assembly and maintenance in POM121- and gp210-deficient cells. J. Cell Biol. 173, 477–483 (2006).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

16. Reinkemeier, C. D. & Lemke, E. A. Dual film-like organelles enable spatial separation of orthogonal eukaryotic translation. Cell 184, 4886–4903 (2021).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

17. Shin, Y. et al. Spatiotemporal control of intracellular phase transitions using light-activated optoDroplets. Cell 168, 159–171 (2016).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

18. Bracha, D. et al. Mapping local and global liquid phase behavior in living cells using photo-oligomerizable seeds. Cell 175, 1467–1480 (2018).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

19. Wei, M.-T. et al. Nucleated transcriptional condensates amplify gene expression. Nat. Cell Biol. 22, 1187–1196 (2020).

    Article
    CAS 
    PubMed
    
    Google Scholar

20. Reinkemeier, C. D., Girona, G. E. & Lemke, E. A. Designer membraneless organelles enable codon reassignment of selected mRNAs in eukaryotes. Science 363, eaaw2644 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

21. Patel, A. et al. A liquid-to-solid phase transition of the ALS protein FUS accelerated by disease mutation. Cell 162, 1066–1077 (2015).

    Article
    CAS 
    PubMed
    
    Google Scholar

22. Altmeyer, M. et al. Liquid demixing of intrinsically disordered proteins is seeded by poly(ADP-ribose). Nat. Commun. 6, 8088 (2015).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

23. Wang, J. et al. A molecular grammar governing the driving forces for phase separation of prion-like RNA binding proteins. Cell 174, 688–699 (2018).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

24. Yang, P. et al. G3BP1 is a tunable switch that triggers phase separation to assemble stress granules. Cell 181, 325–345 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

25. Sanders, D. W. et al. Competing protein–RNA interaction networks control multiphase intracellular organization. Cell 181, 306–324 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

26. Guillén-Boixet, J. et al. RNA-induced conformational switching and clustering of G3BP drive stress granule assembly by condensation. Cell 181, 346–361 (2020).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

27. Gilat, A., Alexandrov, A. I., Dubreuil, B. & Levy, E. D. Mapping interactions between disordered regions reveals promiscuity in biomolecular condensate formation. Nat. Commun. 17, 7069 (2026).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

28. Borden, K. L. B. The eukaryotic translation initiation factor eIF4E wears a ‘cap’ for many occasions. Translation 4, e1220899 (2016).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

29. Webster, M. W. et al. mRNA deadenylation is coupled to translation rates by the differential activities of CCR4–NOT nucleases. Mol. Cell 70, 1089–1100.e8 (2018).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

30. Wigington, C. P., Williams, K. R., Meers, M. P., Bassell, G. J. & Corbett, A. H. Poly(A) RNA-binding proteins and polyadenosine RNA: new members and novel functions. Wiley Interdiscip. Rev. RNA 5, 601–622 (2014).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

31. Mattijssen, S., Kozlov, G., Fonseca, B. D., Gehring, K. & Maraia, R. J. LARP1 and LARP4: up close with PABP for mRNA 3′ poly(A) protection and stabilization. RNA Biol 18, 259–274 (2021).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

32. Mikulecky, P. J. et al. Escherichia coli Hfq has distinct interaction surfaces for DsrA, rpoS and poly(A) RNAs. Nat. Struct. Mol. Biol. 11, 1206–1214 (2004).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

33. Feng, Y., Huang, H., Liao, J. & Cohen, S. N. Escherichia coli poly(A)-binding proteins that interact with components of degradosomes or impede RNA decay mediated by polynucleotide phosphorylase and RNase E. J. Biol. Chem. 276, 31651–31656 (2001).

    Article
    CAS 
    PubMed
    
    Google Scholar

34. Soucek, S., Corbett, A. H. & Fasken, M. B. The long and the short of it: the role of the zinc finger polyadenosine RNA binding protein, Nab2, in control of poly(A) tail length. Biochim. Biophys. Acta Gene Regul. Mech. 1819, 546–554 (2012).

    Article
    CAS 
    
    Google Scholar

35. Virtanen, A., Henriksson, N., Nilsson, P. & Nissbeck, M. Poly(A)-specific ribonuclease (PARN): an allosterically regulated, processive and mRNA cap-interacting deadenylase. Crit. Rev. Biochem. Mol. Biol. 48, 192–209 (2013).

    Article
    CAS 
    PubMed
    
    Google Scholar

36. Viphakone, N. et al. TREX exposes the RNA-binding domain of Nxf1 to enable mRNA export. Nat. Commun. 3, 1006 (2012).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

37. Boeynaems, S. et al. Poly(A)-binding protein is an ataxin-2 chaperone that regulates biomolecular condensates. Mol. Cell 83, 2020–2034 (2023).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

38. Lovén, J. et al. Revisiting global gene expression analysis. Cell 151, 476–482 (2012).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

39. Cheon, H. et al. IFNβ-dependent increases in STAT1, STAT2, and IRF9 mediate resistance to viruses and DNA damage. EMBO J. 32, 2751–2763 (2013).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

40. Seal, R. L. et al. Genenames.org: the HGNC and PGNC resources in 2026. Nucleic Acids Res. 54, D1098–D1107 (2025).

    Article
    
    Google Scholar

41. Hafner, M. et al. CLIP and complementary methods. Nat. Rev. Methods Prim. 1, 20 (2021).

    Article
    CAS 
    
    Google Scholar

42. Zhao, J. et al. Genome-wide identification of Polycomb-associated RNAs by RIP-seq. Mol. Cell 40, 939–953 (2010).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

43. Pierce, J. B. et al. Strategies for investigating nuclear–cytoplasmic tRNA dynamics in yeast and mammalian cells. Methods Cell. Biol. 122, 415–436 (2014).

    Article
    PubMed
    
    Google Scholar

44. Schindelin, J. et al. Fiji: an open-source platform for biological-image analysis. Nat. Methods 9, 676–682 (2012).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

45. Joung, J. et al. Genome-scale CRISPR–Cas9 knockout and transcriptional activation screening. Nat. Protoc. 12, 828–863 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

46. Livak, K. J. & Schmittgen, T. D. Analysis of relative gene expression data using real-time quantitative PCR and the 2−^ΔΔCT method. Methods 25, 402–408 (2001).

    Article
    CAS 
    PubMed
    
    Google Scholar

47. Ewels, P. A. et al. The nf-core framework for community-curated bioinformatics pipelines. Nat. Biotechnol. 38, 276–278 (2020).

    Article
    CAS 
    PubMed
    
    Google Scholar

48. Tirosh, I. et al. Dissecting the multicellular ecosystem of metastatic melanoma by single-cell RNA-seq. Science 352, 189–196 (2016).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

49. Hughes, C. S. et al. Single-pot, solid-phase-enhanced sample preparation for proteomics experiments. Nat. Protoc. 14, 68–85 (2019).

    Article
    CAS 
    PubMed
    
    Google Scholar

50. Reinkemeier, C. D. et al. Recordanelles for transcriptome-wide single-cell RNA recording in human cells. Zenodo https://doi.org/10.5281/zenodo.22096128 (2026).

51. Michalska, A., Blaszczyk, K., Wesoly, J. & Bluyssen, H. A. R. A positive feedback amplifier circuit that regulates interferon (IFN)-stimulated gene expression and controls type I and type II IFN responses. Front. Immunol. 9, 1135 (2018).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

52. Wilder, C. L. et al. A stimulus-contingent positive feedback loop enables IFN-β dose-dependent activation of pro-inflammatory genes. Mol. Syst. Biol. 19, MSB202211294 (2023).

    Article
    
    Google Scholar

53. Karim, M. M. et al. A mechanism of translational repression by competition of Paip2 with eIF4G for poly(A) binding protein (PABP) binding. Proc. Natl Acad. Sci. USA 103, 9494–9499 (2006).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

54. Kahvejian, A., Svitkin, Y. V., Sukarieh, R., M’Boutchou, M.-N. & Sonenberg, N. Mammalian poly(A)-binding protein is a eukaryotic translation initiation factor, which acts via multiple mechanisms. Genes Dev. 19, 104–113 (2005).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

Download references

Author notes

1. These authors contributed equally: Christopher D. Reinkemeier, Marcos Manero-Carranza.

2. Requests for materials should be sent to Randall J. Platt.

Authors and Affiliations

1. Department of Biosystems Science and Engineering, ETH Zurich, Basel, Switzerland

   Christopher D. Reinkemeier, Marcos Manero-Carranza, Rick Farouni, Federica Filice, Alexander Klotz, Konstantin Schneider-Heieck, Tim L. Andermann & Randall J. Platt

2. Botnar Research Center for Child Health, Basel, Switzerland

   Randall J. Platt

3. Department of Chemistry, University of Basel, Basel, Switzerland

   Randall J. Platt

4. NCCR Molecular Systems Engineering, Basel, Switzerland

   Randall J. Platt

Contributions

Conceptualization, C.D.R., M.M.C. and R.J.P. Formal analysis, C.D.R., M.M.C. and R.F. RNA-seq analysis, M.M.C. and R.F. Methodology, C.D.R., M.M.C., F.F., A.K., K.S.H., T.L.A. and R.J.P. Investigation, C.D.R., M.M.C. and R.J.P. Visualization, C.D.R. and M.M.C. Funding acquisition, C.D.R., M.M.C. and R.J.P. Supervision, C.D.R. and R.J.P. Writing—original draft, C.D.R., M.M.C. and R.J.P. Writing—review and editing, C.D.R., M.M.C. and R.J.P.

Corresponding authors

Correspondence to
Christopher D. Reinkemeier or Randall J. Platt.

Competing interests

C.D.R., M.M.C. and R.J.P. have filed a patent application (Switzerland Patent application no. CH000281/2026) related to this work. The remaining authors declare no competing interests.

Peer review information

Nature Biotechnology thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.

Reporting Summary (download PDF )

Peer Review File (download PDF )

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions