Large heteromeric protein complexes are implicated in all aspects of gene expression, and uncovering their assembly mechanism is particularly challenging. Several different subunits, synthesized by separate mRNA molecules, must productively interact with their direct partners in the crowded cellular environment and sequentially build larger assemblies while minimizing off-pathway interactions and aggregation.

Co-translational assembly (co-TA) can facilitate the formation of protein complexes, whereby the newly synthesized nascent protein chain establishes the interaction with the partner before it is released from the ribosome1,2. Co-TA can be sequential (also termed directional), which is when a fully translated protein interacts with the partner nascent chain, or simultaneous (also termed symmetrical), which is when both interactors are nascent chains. Coupling specific subunit–subunit assembly with translation would reduce the exposure of aggregation-prone domains, facilitate the formation of intricate protein–protein interfaces and allow a sequential order for the assembly of different subunits1,3. Co-TA participates in the heterodimerization of several yeast proteins4,5,6,7, including the assembly of subunits of the nuclear pore complex8,9.

Many of the molecular machines involved in transcription initiation are large heteromeric protein complexes. Among them, the ~1.3-MDa basal transcription factor TFIID makes contacts with core promoter DNA elements, promotes TATA-binding protein (TBP) loading on core promoters and works as a scaffold for the formation of RNA polymerase II preinitiation complex (PIC) on all protein-coding genes10,11,12.

In metazoans, TFIID comprises TBP and 13 TBP-associated factors (TAFs)13, and it can be subdivided into three structural lobes (Fig. 1a)14,15. Single-particle cryo-electron microscopy (cryo-EM) models of yeast and human TFIID have shed light on the position and atomic interactions among its subunits10,16,17,18.

Fig. 1: A systematic assay expands the network of co-translational interactions in TFIID and identifies nascent TAF1 polypeptide as a central hub in the assembly process.
figure 1

a, Schematic structure of TFIID. Half-circle subunits represent HFD partners. b, RNA immunoprecipitation (RIP) assays using an antibody against endogenous human TAF10 on HeLa cell polysome extracts. Potential target mRNAs were tested by RT–qPCR. Data points correspond to technical duplicates from n = 3 biological replicates. c, RIP-coupled RT–qPCR assays using an antibody against endogenous mouse TAF10, performed on mESCs. Data points represent technical duplicates from n = 2 biological replicates. d, Schematic representation of the GFP-RIP-coupled RT–qPCR assay using HeLa cell lines expressing doxycycline (Dox)-inducible GFP-TAFs to systematically probe co-translational assembly in TFIID. e, Matrix summarizing the results of the systematic GFP-RIP assay in d. GFP-tagged TFIID subunits were used as baits in a GFP-RIP assay from polysome extracts (rows), and enrichment for TFIID-subunit mRNAs was assessed by RT–qPCR (columns). The area of each circle is proportional to mRNA log2(fold enrichment (FE)) over mock IP. Combinations whose FE was less than fourfold of that of negative control target mRNA (PPIB) are not shown in the plot and are considered negative. Gray circles represent hits for bait mRNA. Black circles represent Co-TA hits. Red circles highlight the widespread enrichment for TAF1 mRNA from RIP of several TFIID subunits. Stars indicate subunits for which GFP fusion resulted in ambiguous protein functionality. Results represent the mean of n = 2 biological replicates. f, RIP-coupled RT–qPCR assays against endogenous TAF6 performed on HeLa cells. The C-terminal location of the epitope prevented the detection of the nascent TAF6 protein, along with its own mRNA, and the simultaneous co-TA with TAF9 (TAF6 HFD partner). Data points correspond to technical triplicates from n = 2 biological replicates. g, RIP-coupled RT–qPCR assays against endogenous TAF7. Data points correspond to technical triplicates from n = 2 biological replicates. Bar graphs in the figure show the mean of the data. Antigen regions for the antibodies used are indicated. HFD, histone-fold domain; HEAT, HEAT-repeat domain; TAF1iD, TAF1-interaction domain; FRT-TO, FLP Recombination Target-Tet-ON; Ptet, tetracycline-responsive promoter. Cycloheximide (CHX) prevents ribosome dissociation from the mRNA. By contrast, puromycin (Puro) induces premature nascent polypeptide chain termination and release from the ribosome or mRNA.

Source data

Nine TAFs contain a histone-fold domain (HFD) that dictates five defined dimerization interfaces within the complex, namely TAF4–TAF12, TAF6–TAF9, TAF3–TAF10, TAF8–TAF10 and TAF11–TAF13. A set of five TAFs (TAF4, TAF5, TAF6, TAF9 and TAF12)—named core-TFIID—is present in two copies, constituting a pseudo-symmetrical unit within the complex that occupies both A and B lobes. These two lobes differ in the dimerization partner of TAF10: TAF3 in lobe A, and TAF8 in lobe B. Moreover, lobe A is characterized by the additional TAF11–TAF13 HFD pair (Fig. 1a). Lobe C comprises the structured domains of the TAF1–TAF7 dimer, TAF2 and the central HEAT domains of the two copies of TAF6. The B and C lobes are connected through the carboxy-terminal portion of TAF8, which directly interacts with TAF2 (ref. 19), whereas the A lobe remains flexibly connected through the TAF6 linker region.

How and where TFIID assembles in cells is a longstanding question. Classically, the holocomplex is isolated from nuclear extracts, while attempts to isolate endogenous assemblies in the cytoplasm led to the identification of preformed TAF2/TAF8/TAF10 and TAF11/TAF13 modules20,21. Another hint on the formation of cytoplasmic TFIID submodules came from the isolation of a stable TAF5/TAF6/TAF9 subcomplex22. These observations led to a model whereby different TFIID modules would be formed in the cytoplasm and holocomplex formation would take place in the nucleus23.

We have demonstrated co-TA events between three pairs of TFIID subunits in the cytoplasm, either sequential (TAF10–TAF8, TBP–TAF1) or simultaneous (TAF6–TAF9)24. With the aim of searching for additional co-TA events in TFIID, we carried out a broad combination of complementary approaches, and we identified a series of pairwise co-TA events that shape the early steps of TFIID assembly. Unexpectedly, we uncovered a new role for the TAF1 nascent protein as a co-translational ‘landing platform’ for preassembled TFIID submodules in the cytoplasm.


Co-translational interactions in TFIID identify TAF1 as a central hub

When reanalyzing our previously published TAF10 RNA immunoprecipitation (RIP)-microarray data24, TAF1 mRNA scored as a positive hit (Extended Data Fig. 1a). TAF1 is devoid of HFDs and it is not known to directly interact with TAF10 within TFIID, raising the possibility that higher-order co-translational interactions might take place. This observation prompted us to perform TAF10 RIP–qPCR on HeLa cells polysome extracts (Fig. 1b). Indeed, we detected a strong enrichment of TAF1 mRNA in TAF10 RIP assays, along with the expected mRNAs of TAF10 itself and its HFD partner TAF8. TAF1 mRNA enrichment was reproducible and puromycin-sensitive, suggesting a co-translational association of TAF10 with the nascent TAF1 polypeptide. To rule out biases from the cellular system or the antibody used, we performed the same experiment on E14 mouse embryonic stem cells (mESCs) using a different monoclonal antibody than the one used in the human system. We found that the anti-TAF10 RIP enriched Taf1 mRNA in mouse cells as well (Fig. 1c), suggesting that the phenomenon is conserved.

To systematically assess all co-translational assembly events within the TFIID complex, we used a series of inducible HeLa cell lines engineered to express each TFIID subunit as a fusion protein with an amino-terminal GFP tag25. These GFP-tagged TAFs have been shown to be incorporated into TFIID purified from nuclear extracts26. We performed GFP-RIP assays on polysome extracts for each individual TFIID subunit and used RT–qPCR to systematically test for enrichment of mRNAs encoding all the TFIID subunits (Fig. 1d). The results of this systematic RIP–qPCR screening confirmed the previously published TFIID co-TA subunit pairs (TAF10–TAF8, TAF6–TAF9 and TBP–TAF1), validating the general reliability of the system (Fig. 1e and Extended Data Fig. 1b). Strikingly, our systematic assay revealed that TAF1 mRNA was enriched in RIP experiments of several distinct TFIID subunits (Fig. 1e). TAF10 RIP also scored positive for TAF1, confirming the observations from endogenous TAF10 RIP assays (Fig. 1b–c). Apart from TAF10, RIP assays of TAF2, TAF4, TAF5, TAF8, TAF12 and TBP retrieved TAF1 mRNA (Fig. 1e).

In addition, novel subunit pairs undergoing co-TA were detected, including well-established HFD partners: TAF10 interacts co-translationally with nascent TAF3; TAF12 with nascent TAF4; and TAF11 and TAF13 are reciprocally enriched, hinting at symmetrical co-TA (Supplementary Table 1). Our systematic RIP assay also revealed co-TA among direct partner subunits that do not interact through an HFD. For instance, TAF2 and TAF8, which are known to interact directly in TFIID (Fig. 1a and Extended Data Fig. 1c), reciprocally enriched the partner’s mRNA, suggesting simultaneous co-TA. TAF5 enriched TAF6 mRNA, one of its direct interactors within core-TFIID: TAF6 contributes with a β-strand to the last blade of the TAF5 WD40 β-propeller domain (Extended Data Fig. 1d). Finally, TAF1 enriched the mRNA of TAF7, its direct partner within TFIID (Fig. 1e).

We noted that, for three of the GFP-fusion-protein-expressing cell lines (TAF6, TAF13 and TBP), we could not retrieve the bait mRNA in our RIP assays, and that the anti-GFP-TAF7 RIP failed to retrieve TAF1. TBP–TAF1 co-TA has already been shown with endogenous TBP RIP assays in our previous report24. To complete our systematic screening, we performed RIP assays with antibodies recognizing endogenous TAF6 and TAF7 and observed a robust puromycin-sensitive enrichment of TAF1 mRNA in both TAF6 and TAF7 RIP experiments (Fig. 1f,g and Extended Data Fig. 1e). Overall, these observations expand the repertoire of TFIID subunits that follow the co-TA pathway with their partners, and importantly identify the nascent TAF1 protein as a potential hub for the recruitment and assembly of many TFIID subunits.

TAFs are localized in the proximity of TAF1 mRNA in the cytoplasm

Our observations made on the basis of RIP assays suggest that, during TAF1 mRNA translation, several TFIID subunits physically associate with TAF1 nascent polypeptide. To physically localize and quantify these events in an endogenous cellular context, we combined immunofluorescence (IF) against several TFIID subunits with single-molecule RNA fluorescence in situ hybridization (smFISH) using HeLa cells.

First, we used this strategy to detect TAF1 nascent protein and estimate the fraction of actively translated TAF1 mRNAs. To this end, we used an IF-validated TAF1 antibody recognizing an N-terminal antigen and combined it with TAF1 mRNA smFISH (Fig. 2a). We used CTNNB1 as a negative control mRNA in smFISH. The average number of cytoplasmic mRNAs per cell for TAF1 was about 16, and it was about 120 for CTNNB1 (Fig. 2b). Next, we combined TAF1 IF with TAF1 or CTNNB1 smFISH (Fig. 2c) and quantified the number of TAF1 mRNA molecules co-localizing with TAF1 protein spots. About ~55% of TAF1 cytoplasmic mRNAs co-localized with TAF1 IF spots (Fig. 2d). This fraction decreased by more than tenfold upon puromycin treatment, proving a dependence on mRNA, ribosome and nascent chain integrity. The very low fraction (~1%) of co-localization with CTNNB1 mRNA was puromycin-insensitive and represents random co-localization. We also validated the specificity of the TAF1 smFISH signal by short interfering RNA (siRNA)-mediated knockdown of TAF1 (Extended Data Fig. 2). Roughly half of TAF1 mRNAs were detected as being actively translated in HeLa cells. However, nascent protein detection on poorly translated mRNAs might still be missed.

Fig. 2: Endogenous nascent TAF1 protein detection.
figure 2

a, Schematic overview of the imaging strategy used to detect actively translated TAF1 endogenous mRNAs with a combination of smFISH and IF. The antigen region recognized by the TAF1 antibody used in the assay is indicated. TAND, TAF1 N-terminal domain; T7iD, TAF7-interaction domain; BD, bromodomain. b, Representative confocal maximum intensity projections (MIPs) of smFISH against the negative control CTNNB1 and TAF1 mRNAs in HeLa cells. The plot on the right shows the absolute number of cytoplasmic mRNAs per cell (mean values are shown in green boxes, and total number of cells in brackets). In the boxplots, the center line marks the median, the box bounds mark the 25th and 75th percentiles, and the whisker limits are 1.5 × interquartile range. c, Representative multicolor confocal images for the co-localization assay shown in a and quantified in d. Each image is a single confocal optical slice. TAF1 protein IF and TAF1 mRNA detection in the merged image are shown in green and magenta, respectively. Co-localizing spots are indicated with yellow arrows. Zoomed-in regions (white squares) are shown on the right. Inset scale bars, 1 μm. d, Quantification of the fraction of mRNAs co-localized with protein signal for each experimental condition. Each open circle corresponds to an independent field of view (n = 3, total number of cells is in brackets).

Source data

Next, we applied the same strategy to TFIID subunits to assess their spatial proximity to TAF1 mRNA (Fig. 3a). First, we assessed the combination with TBP (lobe A component) (Fig. 3b), as its co-translational association with TAF1 has already been dissected24. We found that ~6% of cytoplasmic TAF1 mRNAs co-localized with TBP, whereas less than 1% of CTNNB1 mRNAs did so. The fraction of co-localized TAF1 mRNAs robustly decreased upon puromycin treatment, confirming the co-TA between the two subunits.

Fig. 3: Endogenous TFIID subunits are localized in physical proximity to TAF1 mRNA in the cytoplasm of human cells.
figure 3

a, Schematic overview of the imaging strategy used to detect co-TA events of endogenous TFIID subunits on TAF1 mRNA with a combination of smFISH and IF. bf, Representative multicolor confocal images from the co-localization assay shown in a for each assessed subunit. Each image is a single multichannel confocal optical slice. Protein IF and TAF1 mRNA detection are shown in green and magenta, respectively. Co-localizing spots are indicated with yellow arrows. Zoomed-in regions (white squares) are shown below each image. Inset scale bars, 1 μm. The plots on the right report the fraction of target mRNAs co-localized with protein signal for each experimental condition. Each open circle corresponds to an independent field of view (n = 3, total number of cells is in brackets).

Source data

We then performed the IF experiment with TAF4 (part of core-TFIID), TAF7 (lobe C component) and TAF10 (lobes A and B component). All tested TAFs positively co-localized with TAF1 mRNA (Fig. 3c–e). TAF4 (Fig. 3c) and TAF10 (Fig. 3e) co-localization levels with TAF1 were comparable to those of TBP (Fig. 3b), whereas levels for TAF7 (Fig. 3d) were considerably higher: ~40% of TAF1 mRNAs co-localized with TAF7 spots. Puromycin treatment consistently reduced the fraction of co-localization for all TAFs, although the reduction for TAF4 was not as large. TAF1 co-localization with SUPT7L—a subunit of the SAGA complex27—was very low (<1%) and not affected by puromycin (Fig. 3f), confirming the specificity of the results. We then probed cells for TBP and TAF7 subunits simultaneously using dual-color IF (Extended Data Fig. 3a). We found that ~50% of TBP-positive TAF1 RNA spots were simultaneously co-localized with TAF7. Puromycin treatment drastically reduced the frequency of co-localization, nearly abolishing the double-positive events (Extended Data Fig. 3b). Overall, these observations support the data from our systematic RIP–qPCR studies, further suggesting that multiple TFIID subunits are recruited on the TAF1 nascent polypeptide during TAF1 protein synthesis.

The cytoplasm is populated by multisubunit TFIID ‘building blocks’

To better understand how the co-TA events that we described above may participate in TFIID assembly, we set out to analyze the composition of potential TFIID assemblies in the cytoplasm. To this end, we immunopurified endogenous TFIID subunits from HeLa cytoplasmic extracts and analyzed the immunoprecipitated endogenous complexes by label-free mass spectrometry (MS) (Fig. 4a–f). All of our immunoprecipitation assays (IPs) invariably retrieved holo-TFIID from nuclear extracts, confirming the effectiveness of the antibodies used (Extended Data Fig. 4a).

Fig. 4: The cytoplasm is populated by multisubunit TFIID ‘building blocks’.
figure 4

af, IP of endogenous TFIID subunits coupled to label-free MS, performed on cytoplasmic extracts from human HeLa cells. Bar plots represent the average NSAF (normalized spectral abundance factor) value for each detected subunit in technical triplicates. Error bars represent the s.e.m. The antigen position of the TAF1 antibody used for IP is shown in f. g, Visual summary of cytoplasmic TFIID submodules inferred from IP–MS data. IPs in which the given submodule was enriched are indicated.

Source data

Cytoplasmic TAF2 was found associated with TAF8, in accordance with their co-TA (Fig. 1e), and TAF2 was partially integrated in the 8TAF complex (composed of core-TFIID, TAF2, TAF8 and TAF10; Fig. 4)21,28. The majority of immunopurified cytoplasmic TAF4 was in complex with TAF8, TAF6, TAF9/9B, TAF10 and TAF5 (Fig. 4b), which we interpreted as the 7TAF complex (core-TFIID, TAF8 and TAF10; the missing detection of TAF12 could be owing to the documented post-translational modifications of this subunit). The presence of small amounts of TAF11 co-purified with TAF4 hinted at the incorporation of the latter in a partial A lobe. In the cytoplasmic anti-TAF4 IPs, we found only TAF4, while in the nuclear anti-TAF4 IP we found peptides from TAF4 and its paralog TAF4B (Fig. 4b and Extended Data Fig. 4a). These findings suggest that the isolated cytoplasmic TAF4-containing building block is either lobe A or lobe B (as indicated in Fig. 4g), containing only one copy of TAF4. By contrast, the detection of both TAF4 and TAF4B in nuclear TAF4 IP suggests the isolation of holo-TFIID, as the holo-complex contains two copies of TAF4 family members (TAF4 and TAF4B).

Endogenous cytoplasmic TAF10 IP retrieved similar amounts of TAF10’s HFD partner TAF8, with a relevant portion of the heterodimer associated with core-TFIID and TAF2 in the 8TAF complex (Fig. 4c). In this case, the small amounts of TAF11 also hint that a fraction of TAF10 is incorporated in a partially assembled A lobe. On the other hand, we found some TAF11 associated with its HFD-partner TAF13 in the TAF11 IP20, with no detectable amounts of other lobe A subunits (Fig. 4d). Most immunopurified TAF7, the partner of TAF1 in TFIID, was not in the complex (Fig. 4e). Yet, cytoplasmic TAF7 co-purified with trace amounts of TFIID subunits—including TAF1 (Extended Data Fig. 4b). In contrast to nuclear extracts (Extended Data Fig. 4a), cytoplasmic TAF1 IP did not enrich any TFIID component, with the bait itself being barely detectable (Fig. 4f). Given that the IP-grade TAF1 antibody recognizes a C-terminal epitope along the protein, we conclude that the abundance of TAF1 mature protein in the cytoplasm is below the detection limit in this analysis.

These results demonstrate that the cytoplasm of HeLa cells is populated by different multisubunit TFIID submodules, likely representing stable intermediates along the assembly pathway of the complex (Fig. 4g). None of the cytoplasmic IPs, except for TAF7, co-purified TAF1, suggesting that it is present in small amounts in the cytoplasm and is the limiting factor in TFIID assembly. These findings further point to a co-translational recruitment mechanism whereby the preassembled TFIID ‘building blocks’ associate with nascent TAF1 polypeptide, in agreement with our RIP and imaging experiments.

TAF1 crosslinking hotspots are anchor points for TFIID building blocks

TAF1 is the largest subunit of TFIID (1,872 amino acids), and only ~47% of the protein structure has been solved. To rationalize how the nascent TAF1 polypeptide could work as a hub for TFIID assembly, we analyzed all available crosslinking-MS experiments performed on highly purified TFIID or PIC-incorporated TFIID10,18,19. The intercrosslinks between TAF1 and other TFIID subunits detected in at least two independent datasets indicate three main proximity and crosslinking ‘hotspots’ along TAF1 (Fig. 5a and Supplementary Table 2): (1) a loose region crosslinked with TBP and its interacting partners TAF11 and TAF13; (2) a well-defined hotspot rich in crosslinks with TAF6 along with single positions associated with TAF5, TAF8 and TAF9; and (3) a large central region that was extensively crosslinked to TAF7 and, to a lesser extent, to TAF2. The combination of the crosslinking hotspots (Fig. 5a) with TAF1 sequence features (conservation and structural disorder; Fig. 5b), annotated functional domains (Fig. 5c) and structural observations (Fig. 5d) shows that TAF1 is a flexible scaffold protein that connects all TFIID submodules by three main anchor points.

Fig. 5: Three crosslinking hotspots identified on TAF1 correspond to distinct anchor points for specific TFIID building blocks.
figure 5

a, Summary of TAF1-centred crosslinking-MS metanalysis derived from three independent studies10,18,19. Interprotein crosslinks between TAF1 and other TFIID subunits (TAFs are indicated with their corresponding numbers) are displayed along the protein as dotted lines or shaded ranges. Numbers above each crosslinked subunit indicate TAF1 positions or the range involved in the crosslinks, and the values in brackets indicate the number of distinct TAF1crosslinked positions. Only crosslinks reported in at least two independent datasets are shown. The known structural domains of TAF1 are depicted. b, Heatmaps of TAF1 conservation (ConSurf), structural disorder (Metapredict) and pLDDT (AlphaFold structural prediction confidence) scores are shown. Structures with pLDDT > 90 are expected to be modeled to high accuracy. A pLDDT between 70 and 90 means that the structure is modeled well, and structures with a pLDDT between 50 and 70 should be treated with caution. The scales of the different prediction scores are shown on the right. c, Full-length human TAF1 AlphaFold model. The initial structure was extended to better appreciate the different domains indicated along the protein. df, Structural models of the three main TAF1 anchor points (labeled A, B and C) in TFIID are shown in the insets. The model in d is the result of AlphaFold prediction of the TAF1–TAF11–TAF13–TBP subcomplex. Distinct TAF1 domains are colored as in a. Partner subunits are shown in shades of gray. TAF1 segments part of each anchor point are indicated. g, Schematic summary of the distinct interaction hotspots along the protein and the length of the intervening linker regions. TFIID lobes A and B are shown as circles. aa, amino acids; HMG: HMG-box domain; ZnK: zinc-knuckle domain.

TAF1 modular organization is shown on the AlphaFold model of the full-length protein (Fig. 5c). A substantial fraction of the protein (~48%) is predicted to be intrinsically disordered, including interdomain linker regions and the long acidic C-terminal tail (Fig. 5b). TAF1 contains two main well-structured regions: the TAF7 interaction domain (TAF7iD), which occupies the central portion of the protein, and two histone-reader bromodomains (BD) localized in tandem along the C-terminal tail (Fig. 5a). The TAF7iD, composed of the DUF3591 domain in concert with the RAP74 interaction domain (RAPiD), tightly associates with TAF7 and binds downstream of core promoter DNA18,29. Accordingly, these regions were modeled with high confidence by AlphaFold (Fig. 5b). The scarcity of TAF1 intraprotein crosslinks outside the TAF7iD and the tandem BDs is in accordance with the absence of other major structured domains along the protein (Extended Data Fig. 5a).

The three described crosslinking hotspots correspond to distinct anchor points (named here A, B and C) for specific TFIID submodules (Fig. 5d–f). The first TAF1 anchor point (A) would interact with TBP and the TAF11–TAF13 heterodimer, as the flexible TAF N-terminal domain (TAND) has been shown to directly interact with TBP and inhibit TBP DNA binding30,31,32. Additionally, removal of human TAND has been found to abolish the co-translational recruitment of TBP to TAF1 (ref. 24). The crosslinks of the TAF11–TAF13 heterodimer with TAF1 are consistently found in all datasets. They map on TAF1 Lys249, which lies within a conserved motif predicted with higher confidence and lower disorder scores than those of the flanking regions (Fig. 5a–c). Modeling TAF1 with TBP and TAF11–TAF13 with AlphaFold resulted in a ternary complex with the expected positioning of TAF1 TAND into the concave surface of TBP. The putative TAF11–TAF13 interaction motif of TAF1 was folded laterally in a pocket formed by the HFD subunits (Fig. 5d). The TAF1 Lys249 position in the model is compatible with all the experimental crosslinks with TAF11–TAF13 (Extended Data Fig. 5b).

The second hotspot (B) is the anchor point of both lobes A and B with TAF1. It is composed of three TAF1 stretches of conserved amino acids interspersed by loops of lower conservation, named TAF6-binding motifs (T6BMs, Fig. 5a–c). These motifs, which have recently been resolved by cryo-EM10, bridge the two copies of TAF6 HEAT domains, which in turn are connected to lobes A and B (Fig. 5e, see also Fig. 1a). The T6BMs occupy defined grooves and pockets across the pair of TAF6 HEAT domains at the center of TFIID (Fig. 5e). Modeling the entire TAF1 region containing the T6BMs allowed us to map all crosslinking sites otherwise positioned in unresolved flexible loops (Extended Data Fig. 5c), in perfect agreement with the experimental structure (Extended Data Fig. 5d–f). Overall, the T6BMs constitute most of the interface anchoring the two copies of TAF6 HEAT domains together (Extended Data Fig. 5g,h). The absence of crosslinked positions along the third T6BM (Fig. 5a) is due to the lack of lysine residues. Apart from TAF6, the crosslinks to other TAFs within this hotspot are likely driven by proximity rather than direct interactions.

The third hotspot (C) coincides with the TAF7iD (Fig. 5f). Besides the intricate fold adopted with TAF7, the DUF3591 loosely anchors the resulting TAF1–TAF7 globular domain to TAF2 (Extended Data Fig. 5i). Overall, structural and biochemical data support a scaffolding function of TAF1 within TFIID, thanks to its modular organization (Fig. 5g). TAF1 represents a flexible three-way anchor point that physically connects the three TFIID lobes through direct interactions with the two copies of TAF6, which in turn emanate into lobes A and B (Fig. 5g). Strikingly, all mapped crosslinks reside in the N-terminal half of TAF1 (Fig. 5a), leaving the ~700-aa region downstream of RAPiD free from crosslinks. This would allow TFIID assembly on the N-terminal half of TAF1 before the protein is released from the ribosome.

TAF1 depletion leads to an accumulation of TFIID building blocks in the cytoplasm

To investigate the role of TAF1 in the dynamics of cytoplasmic TFIID assembly, we perturbed the TAF1-dependent assembly by siRNA-mediated TAF1 knockdown (KD). Subcellular fractionation experiments revealed an enrichment of the protein levels of several TFIID subunits in the cytoplasmic fraction upon TAF1 KD. Specifically, the cytoplasmic extract was substantially enriched for core-TFIID subunits (TAF4, TAF5, TAF6 and TAF12) (Fig. 6a). This cytoplasmic increase in protein levels of TAF4/5/6/12 was not visible in the nuclear fraction, suggesting a specific cytoplasmic accumulation of those subunits. Also, TAF13, and to a lesser extent TBP, followed the same pattern. Instead, the levels of TAF8 and its partner TAF10 remained mostly unchanged. Notably, although the levels of TAF7 stayed constant in the cytoplasm, they were drastically reduced in the nuclear fraction, closely matching the depletion of TAF1. These observations show that TFIID subunits are differentially affected by TAF1 depletion. On the contrary, TAF4 and TAF7 KD under the same conditions did not reproduce the effect elicited by TAF1 silencing, suggesting that the observed phenomenon is TAF1-specific (Extended Data Fig. 6a).

Fig. 6: TAF1 depletion leads to an accumulation of TFIID building blocks in the cytoplasm.
figure 6

a, Subcellular fractionation of siRNA-transfected HeLa cells followed by western blot analysis of endogenous TFIID subunits distribution. GAPDH, lamin A/C and histone H3 were used as loading controls. The amount of loaded cytoplasmic (cyto.) extract is three times the amount of the nuclear (nucl.) extract counterpart. The positions of the molecular weight markers in kDa are indicated on the right. b, IP of endogenous TFIID subunits coupled to label-free mass-spectrometry (MS) performed on cytoplasmic extracts of siRNA-transfected HeLa cells. The circle area represents the average NSAF value for each detected subunit in technical triplicates. The NSAF scales are indicated on the right. Distinct TFIID subcomplexes are depicted on the top. Subunits are color-coded and arranged according to the subcomplexes indicated on the top. Black dots in the circles identify the protein used as bait in each IP. c, Same as in b, but the IPs were performed on nuclear extracts. CTR, non-targeting control siRNA. Co., complex.

Source data

To address whether the cytoplasmic increase of a subset of TFIID subunits would correspond to an accumulation of specific TFIID building blocks in the cytoplasm, we analyzed endogenous cytoplasmic subcomplexes composition by IP–MS upon TAF1 KD (Fig. 6b and Extended Data Fig. 6b). We selected IP-grade antibodies raised against a lobe C subunit (TAF2), a core TFIID subunit (TAF4) and a non-core TFIID subunit (TAF10). In good agreement with our findings, upon TAF1 KD, TFIID building blocks accumulated in the cytoplasm. The enrichment of cytoplasmic core-TFIID was evidenced in all IPs. Notably, the lobe A-specific subunits, TAF3, TAF11 and TBP, co-purified with TAF4 and TAF10 only following TAF1 depletion. We interpret these results as the cytoplasmic accumulation of different TFIID building blocks, including the 8TAF complex (B lobe+TAF2) and the A lobe complex, provoked by the impairment of the last, TAF1-dependent, assembly step before nuclear import.

IP–MS analyses on the nuclear fraction showed less dramatic rearrangements in subunit distribution, with an overall decrease in the abundance of all immunopurified TFIID subunits in TAF1 KD samples (Fig. 6c and Extended Data Fig. 6c). Overall, our observations together show that TAF1 is a major hub for the co-translational assembly of TFIID complex, from preassembled building blocks to subsequent nuclear translocation.


The self-assembly of large heterotypic multiprotein complexes in living cells poses major challenges to our understanding of cellular homeostasis. Here, we tackled the longstanding question of where and how the basal transcription factor TFIID assembles, and we comprehensively explored the landscape of co-TA events within TFIID. We uncovered TAF1 as the central hub in the assembly process and several previously undiscovered pairs of subunits that undergo co-TA (see Supplementary Table 1).

A hierarchical co-translational model for TFIID assembly

All our findings can be rationalized in a hierarchical model for TFIID assembly, which is stratified in three tiers of assembly events (Fig. 7). The first tier includes early events along the pathway: these are the formation of protein pairs, mostly through the dimerization of HFD-containing subunits. We find it remarkable that all the HFD pairs in TFIID assemble co-translationally, either directionally or symmetrically. The fact that several subunits used as bait in our cytoplasmic IP–MS data were not found as free proteins (Fig. 4) points either at a fast and efficient co-TA with their partners or to a degradation-driven removal of orphan subunits, although a combination of the two processes is also likely. Tier 1 also harbors interactions of non-HFD subunits, such as TAF2 and TAF5, which interact co-translationally with TAF8 and TAF6, respectively. All these directly interacting pairs are structurally well characterized10,18. The products of tier 1 assembly are free early multisubunit intermediates, likely stabilized by interactions with their partner. They are likely characterized by heterogeneous half-lives as free molecular species, since some of them can be isolated in our steady-state experiments (for example, TAF11/TAF13 HF pair), whereas others can be detected only as part of larger assemblies (for example, TAF4/TAF12 HF pair), yet some others are not detected at all (for example, TAF3/TAF10 HF pair) (Fig. 4). The products of tier 1 in turn access the second level of the assembly pathway by combining with each other in a few structurally constrained steps. Assembly in tier 2 occurs post-translationally and leads to the buildup of larger assemblies that were recurrently found in our IP–MS experiments, such as the 8TAF complex and a partially assembled lobe A.

Fig. 7: A co-translational hierarchical model for TFIID assembly.
figure 7

Scheme of the proposed cytoplasmic assembly model for TFIID that reconciles the experimental observations of the present work with previous structural and biochemical data. The assembly pathway can be subdivided into three tiers (colored and numbered horizontal stripes). Tier 3 represents the co-translational assembly of several TFIID building blocks on nascent TAF1 protein through three distinct interaction hotspots (labeled A, B and C), resulting in TFIID. Blue arrows with bases indicate directional co-TA events, and double-headed blue arrows specify reciprocal co-TA, as assessed by RIPs. For further details, see the Discussion section.

In tier 3, the products of tier 2 finally converge and engage co-translationally with the nascent TAF1 polypeptide (Fig. 7). An appealing idea is a sequential N- to C-terminal order of assembly, whereby different TFIID building blocks are recruited by the distinct assembly domains of nascent TAF1 as they emerge from the ribosome channel. The first N-terminal anchor point (A) would interact with TBP, which engages with nascent TAF1 by binding the TAND domain24. Our systematic survey confirmed this co-TA pair. TAF11–TAF13 dimer could also engage with TAF1 at anchor point (A), forming a ternary complex along with TBP (Fig. 5d). Biochemically, a recombinant complex formed by TAF1–TBP–TAF11–TAF13 and TAF7 can be readily purified33, and the direct interaction between TAF1 and TAF11–TAF13 is supported by crosslinking experiments and structural modeling (Fig. 5c,d and Extended Data Fig. 5b). However, TAF11–TAF13 did not score positive for TAF1 mRNA in our systematic RIP approach, opening the possibility of post-translational engagement, or weaker interactions.

The second interaction anchor point (B)—the T6BMs—would interact with two copies of TAF6 HEAT domains, bringing together lobe A and lobe B (Figs. 7 and 5g). Interestingly, TAF1 evolved distinct binding motifs to recognize corresponding identical surfaces from the two TAF6 copies within TFIID (Extended Data Fig. 5h). The third anchor point (C) recruits TAF7, which interacts with the TAF1 central domain (DUF3591 and RAPiD) (Fig. 5g). Notably, in our RIP experiments, TAF7 enriched TAF1 mRNA and vice versa, opening the possibility of a simultaneous co-translational interaction between the two. Such an ordered addition would entail a remarkable degree of coordination, potentially reinforced by binding cooperativity among the modules as they join the growing assembly. Yet, in our imaging data, we detected TAF7-positive TAF1 RNA spots lacking TBP signal and vice versa, hinting at a potential independent binding mode (Extended Data Fig. 3).

Upon completion of TAF1 protein synthesis, the assembled TFIID is released and readily translocated in the nucleus. A subset of subunits scored negative for TAF1 mRNA in our RIP assays: these include TAF3, TAF9, TAF11 and TAF13. Therefore, it is possible that they join the complex post-translationally or through their interaction partners. The benefits of a hierarchical co-translational assembly have been recently theorized in the framework of yeast nuclear pore assembly9. The proposed model may also apply to our findings, in which co-TA is pervasively exploited for the hierarchical assembly of TFIID in the cytoplasm of mammalian cells.

Our data are in agreement with the published TAF interactions, the cryo-EM TFIID structures10,23 and the previous descriptions of partial TFIID assemblies28,34. According to recent bioinformatic analyses, during evolution, proteins that assemble co-translationally have sustained large N-terminal interfaces in order to promote co-translational subunit recruitment35. In agreement, out of the eight larger subunits of TFIID that participate in co-TA as nascent polypeptides (TAF1, TAF2, TAF3, TAF4, TAF6, TAF7, TAF8 and TAF9), all except TAF4 have their interaction domains in the N terminus.

TAF1: a ‘driver’ and limiting factor along the assembly line

TAF1 mRNA was enriched in the majority of our RIPs (Fig. 1 and Supplementary Table 1), and it was found in physical proximity of several TFIID subunits in the cytoplasm (Fig. 3). In this compartment, the levels of TAF1 protein seem to be limiting with respect to TFIID building blocks (Fig. 4). TAF7 IP enriched the whole spectrum of TFIID subunits, including TAF1, albeit at very low levels (Fig. 4e and Extended Data Fig. 4b). TAF7 also showed the highest levels of co-localization with TAF1 mRNA (Fig. 3d), pointing at a remarkable co-TA efficiency between TAF1 and TAF7. The higher assembly efficiency is consistent with the detection of a fully assembled complex in TAF7 cytoplasmic IP. The interaction interface between TAF1 and TAF7 is remarkably intricate, with deeply intertwined β-strands from each protein contributing to a common β-barrel29. It would be conceivable that such an interface would form only concomitantly with folding during protein synthesis, imposing a structural constraint solved by co-TA. Curiously, TAF7 levels decreased proportionally with TAF1 depletion in the nucleus, hinting at a partner stabilization effect (Fig. 6a), similar to the one observed between TAF10 and TAF8 (ref. 24).

TAF1 depletion led to the accumulation of several TFIID building blocks in the cytoplasm, revealing a key role of this subunit in driving complex assembly and consequent relocation in the nucleus (Fig. 6). We propose that nascent TAF1 nucleates the late steps of TFIID assembly in the cytoplasm by tethering different submodules of the complex, and, once released from the ribosomes, the whole assembly efficiently shuttles in the nucleus (Fig. 7). This process may act as a quality checkpoint before nuclear import. In agreement, both in yeast and in metazoans TAF1 is an essential gene36,37,38.

A central role of a single nascent subunit for the co-translational assembly of protein complexes has been demonstrated for the COMPASS histone methyltransferase in yeast, in which a specific subcomplex is directly assembled on nascent Set1 protein, stabilizing the latter from degradation39. Set1 behaved as a co-translational ‘driver’ subunit, simultaneously promoting complex assembly and limiting its abundance. Other examples of subunits potentially working as co-translational drivers for complex assemblies have been uncovered in fission yeast4 and are supported by structural analyses40. We argue that an equivalent process in mammalian cells is led by TAF1 as the driver subunit for TFIID co-translational assembly, which culminates with the tethering of distinct building blocks on TAF1 nascent polypeptide. In this regard, TAF1 mRNA offers the longest coding sequence (CDS) among TFIID components, implying there is a prolonged timeframe in which co-translational binding events can occur. By taking into account an estimate of average translation speed of ~5.6 codons s–1 in mammalian cells41, translating TAF1 CDS would take ~5.6 min. The last assembly domain along TAF1 completely emerges from the ribosome around position 1240, granting an additional window of time of about 1.9 min to ultimate co-TA before ribosome release. In this regard, the analysis of ribosome profiling (Ribo-seq) merged datasets showed a wide region of sparse ribosome-protected fragments, encompassing all three T6BMs and extending inside the DUF3591 domain-encoding region (Extended Data Fig. 7). The low signal in the T6BMs region hints at fast elongation rates, which would rapidly expose all three T6BMs for the co-translational recruitment of the respective TFIID building blocks. Downstream of this region, translation slows down, as suggested by the higher ribosome occupancy. This would also buy time to establish productive co-translational interactions with upstream anchor points.

Open questions

Our findings reveal an unprecedented mechanism for TFIID biogenesis, answering longstanding questions and opening new ones. One of remarkable importance is how efficiently co-TA occurs. A prerequisite for co-translational interactions is an actively translated mRNA. Our imaging approach detected nascent TAF1 protein on roughly half of the correspondent cytosolic messengers (Fig. 2). By using this observation as a proxy for the proportion of actively translated TAF1 mRNAs, the observed frequency of Co-TA events for the other probed subunits (Fig. 3) would be underestimated. In the future, the observation of co-translational binding events in living cells might offer a quantitative dimension to this field.

A second point is whether co-TA is an efficient option for complex assembly or an obligate path. Co-TA might be the sole opportunity for assembly domains characterized by structural constraints, such as the TAF1–TAF7 interface. Instead, interactions mediated by classical binding pockets, extended surfaces or short linear motifs can rely also on post-translational assembly. However, co-translational interactions have the advantage of abolishing partially unfolded/unstable intermediates by kinetically anticipating their complexed state. Although it seems reasonable to hypothesize that natural selection promoted molecular features favoring co-translational interactions, it has proven hard to disentangle co-translational from post-translational assembly experimentally, since both mechanisms ultimately depend on protein synthesis. Note also that other chaperone-mediated assembly processes play a role in multisubunit complex assembly pathways42.

Third, our data open new questions on the nuclear import mechanism adopted by TFIID or its building blocks. The observation that a defined set of subcomplexes accumulates in the cytoplasm upon TAF1 depletion opens the possibility of distinct entry routes to the nucleus. A fully assembled TAF1-containing complex could be the most efficiently translocated molecular species, with several subcomplexes relying on TAF1 for nuclear import. Conversely, other building blocks might access the nucleus autonomously, as shown for the TAF2–TAF8–TAF10 module21,43. The interrogation of a systematic interactome survey of the major nuclear transport receptors on human cells by BioID44 showed that all the detected TFIID subunits shared the same import systems, mainly the α-importins IMA1 and IMA5. This is consistent with the idea that TFIID is transported across the nuclear pore as a pre-assembled entity. Intriguingly, in this study, TAF1 was one of the main biotinylated TFIID subunits, suggesting that TAF1 can directly interact with the nuclear transport receptors and drive nuclear import, as recently found in yeast45.

Our study provides an understanding of a series of steps underlying the assembly mechanism of the general transcription factor TFIID. We envision that the principles of hierarchical co-translational assembly could apply to the biogenesis of most large heteromeric multiprotein complexes in living cells.


Cell culture

Human HeLa cells (CCL-2; ATCC) were obtained from the IGBMC cell culture facility and cultured in DMEM (4.5 g L–1 glucose) supplemented with 10% fetal calf serum (Dutscher, S1810), 100 U ml penicillin and 100 μg ml–1 streptomycin (Invitrogen, 15140-130). E14 mESCs (ES Parental cell line E14Tg2a.4, Mutant Mouse Resource and Research Center) were obtained from the IGBMC cell culture facility and cultured on gelatinized plates in feeder-free conditions in KnockOut DMEM (Gibco) supplemented with 20 mM l-glutamine, penicillin–streptomycin, 100 µM non-essential amino acids, 100 µM β-mercaptoethanol, N-2 supplement, B-27 supplement, 1000 U ml–1 LIF (Millipore), 15% ESQ FBS (Gibco) and 2i (3 µM CHIR99021, 1 µM PD0325901, Axon MedChem). Cells were grown at 37 °C in a humidified, 5% CO2 incubator.

GFP-fusion cell lines generation

GFP-TAFs fusion cell lines used in this study were described in ref. 26. Briefly, the coding sequences for the human TFIID subunits (TAF1, TAF2, TAF3, TAF4, TAF5, TAF6, TAF7, TAF8, TAF9, TAF10, TAF11, TAF12, TAF13 and TBP) were obtained by PCR using the appropriate cDNA clone and gene-specific primers flanked by attB sites followed by BP-mediated GATEWAY recombination into pDONR221, according to the manufacturer’s instructions (Invitrogen). The cloned sequence was verified by sequencing and it was transferred to the pcDNA5-FRT-TO-N-GFP Gateway destination vector by LR recombination, according to the manufacturer’s protocol (Invitrogen). HeLa Flp-In/T-REx cells, which contain a single FRT site and express the Tet repressor25, were grown in DMEM, 4.5 g L–1 glucose (Gibco), supplemented with 10% vol/vol fetal calf serum (Gibco). All the GFP-fusion destination vectors were co-transfected with a pOG44 plasmid that encodes the Flp recombinase into HeLa Flp-In/T-REx cells using polyethyleneimine (PEI) to generate stable doxycycline-inducible expression cell lines. Recombined cells were selected with 5 μg ml–1 blasticidin S (InvivoGen) and 250 μg ml–1 hygromycin B (Roche Diagnostics) 48 h after PEI transfection. Cells were maintained in DMEM supplemented with 10% Tet-free fetal calf serum (Pan Biotech, P30-3602), blasticidin S, hygromycin B and penicillin–streptomycin.

RNA immunoprecipitation against endogenous TFIID subunits

Polysome extract preparation and RIPs wexre performed essentially as described in ref. 24. HeLa cells grown on 15-cm plates (~90% confluent) were treated either with 100 μg ml–1 cycloheximide (CHX, Merck, C1988) for 15 min or with 50 μg ml–1 puromycin (Puro, Invivogen, ant-pr-1) for 30 min in the incubator at 37 °C. Plates were placed on ice and cells were washed twice with ice-cold PBS and scraped in 2 ml lysis buffer (20 mM HEPES-KOH pH 7.5, 150 mM KCl, 10 mM MgCl2, 0.1% NP-40, 1 × PIC (complete EDTA-free protease inhibitor cocktail, Roche, 11873580001), 0.5 mM DTT (ThermoScientific, R0862), 40 U ml–1 RNasin Ribonuclease Inhibitor (Promega, N2511)) supplemented either with CHX or Puro. Cell suspension was homogenized with 10 Dounce strokes using a B-type pestle on ice. Lysates were incubated 15 min on ice and cleared by centrifugation at 17,000g. The supernatant represents the polysome extract.

For each IP, 1.2 mg protein G Dynabeads (Invitrogen, 10004D) was used. The antibodies used are listed in Supplementary Table 3. Dynabeads were washed twice in buffer IP100 (25 mM Tris-HCl pH 8.0, 100 mM KCl, 5 mM MgCl2, 10% glycerol, 0.1% NP-40). Each antibody (5–10 μg per IP) was coupled to Dynabeads in 100 μL buffer IP100 for 1 h at room temperature (RT) while being agitated. Mock IPs were performed using mouse or rabbit IgG. Antibody-coupled Dynabeads were washed twice in buffer IP500 (25 mM Tris-HCl pH 8.0, 500 mM KCl, 5 mM MgCl2, 10% glycerol, 0.1% NP-40) and three times in buffer IP100. For each IP, 1 ml of polysome extract (equivalent to ~107 cells) was used as input. A 10% equivalent volume of the input was kept at 4 °C for input normalization. IP reactions were incubated with rotation at 4 °C overnight. The next day, Dynabeads were washed four times with 0.5 ml high-salt was buffer (25 mM HEPES-KOH pH 7.5, 350 mM KCl, 10 mM MgCl2, 0.02% NP-40, 1 × PIC, 0.5 mM DTT, 40 U ml–1 RNasin Ribonuclease Inhibitor) supplemented either with CHX or Puro. RNA from the resulting immunopurified material was extracted using NucleoSpin RNA XS kit (Macherey-Nagel, 740902) in 100 μL RA1 lysis buffer and purified according to the manufacturer’s protocol. The input sample was extracted and processed in parallel with the IPs.

GFP-fusion RIP

GFP-RIPs using inducible cell lines were performed as described for endogenous RIPs, with the following modifications. The day of the experiment, the expression of the GFP-tagged TFIID subunit was induced by addition of 1 μg ml–1 doxycycline (Dox) for 2 h. For GFP-TAF3 and GFP-TAF10 cell lines Dox treatment was omitted due to their leaky basal expression. Cells were treated with CHX, lysed and polysome extracts prepared as described in the previous section. GFP-IPs were carried out using 40 μL GFP-Trap Agarose beads (ChromoTek, gta-20). Mock IPs were carried out using an equivalent volume of protein G Sepharose beads. Beads were incubated with polysome extracts for 4 h at 4 °C, washed and RNA purified as described in the previous section.


Reverse transcription reaction was performed using SuperScript IV First-Strand Synthesis System (Invitrogen, 18091050) and random hexamers according to manufacturer instructions. The resulting cDNA was diluted 1:10. Two or three technical replicates of qPCR using 2 μL cDNA, primers listed in Supplementary Table 4 and LightCycler 480 SYBR Green I Master (Roche, 04887352001) were performed in a LightCycler 480 instrument (Roche). Input (%) normalization for RIP samples was performed by applying the formula \(100 \times 2^{[({\mathrm{Ct}}_{\mathrm{input}}-6.644)-{\mathrm{Ct}}_{\mathrm{RIP}}]}\). Fold-enrichment normalization was performed by dividing RIP input (%) by mock input (%).

Immunofluorescence–single-molecule inexpensive RNA FISH

RNA detection was performed using smiFISH46. Primary probe sets (24 single oligonucleotides) against target coding sequences were designed using Oligostan in R, as described in the software documentation46. Probes sequences are reported in Supplementary Table 4. Primary probes were synthesized by Integrated DNA Technologies (IDT) in plate format, and were dissolved in TE buffer at 100 μM. 5′ and 3′ Cy3-labeled secondary probe (FLAP) was synthesized by IDT and purified by high-performance liquid chromatography. Primary probes were mixed in an equimolar solution in TE at 0.83 μM per probe. To prepare a 50× smiFISH composite probes mix, 4 μL primary probe mix was mixed with 2 μL 100 μM secondary probe solution in 20 μL final reaction volume in 100 mM NaCl, 50 mM Tris-HCl pH 8.0 and 10 mM MgCl2. The annealing reaction was performed in a thermocycler with the following conditions: 3 min at 85 °C, 3 min at 65 °C, 5 min at 25 °C. 50× smiFISH probes mix was stored at –20 °C. The day before the experiment, HeLa cells were seeded on coverslips (no. 1.5H, Marienfeld, 630–2000) in a 12-well plate (0.2 × 106 cells per well). The day after, cells were treated either with 100 μg ml–1 CHX for 15 min or with 50 μg ml–1 Puro for 30 min in the incubator at 37 °C. Then, cells were directly processed for immunofluorescence. All buffer solutions were filtered (0.22-μm filter). Cells were washed twice with PBS (containing CHX for cells treated with it) and fixed with 4% paraformaldehyde (Electron Microscopy Sciences, 15710) in PBS for 10 min at RT. Cells were washed twice with PBS and incubated for 10 min at RT in blocking/permeabilization solution (BPS) (1× PBS, 1% BSA (MP, 160069), 0.1% Triton-X100 (Merck, T8787), 2 mM vanadyl ribonucleoside complexes (VRC, Merck, R3380)). Cells were incubated for 2 h at RT with the following primary antibodies diluted in BPS: TAF1 (1:1000, rabbit pAb, Abcam, ab188427), TAF4 (3 μg ml–1, mouse mAb, 32TA 2B9), TAF7 (1:250, rabbit pAb, no. 3475), TAF10 (3 μg ml–1, mouse mAb, 6TA 2B11), TBP (2 μg ml–1, mouse mAb, 3TF1 3G3) or SUPT7L (rabbit pAb, Bethyl, A302-803A). A secondary-only control sample was incubated with BPS devoid of primary antibody. After three 5-min PBS washes, cells were incubated for 1 h at RT (light-protected) with AF488-conjugated secondary antibodies diluted 1:3,000 in BPS (goat anti-mouse IgG, A11001 or goat anti-rabbit IgG, A11008, Life Technologies). For dual-color IF, we also used Alexa Fluor Plus 647-conjugated secondary antibody (goat anti-mouse IgG, A32728). After three 5-min PBS washes, a second fixation step was performed with 4% paraformaldehyde in PBS for 10 min at RT. Cells were washed twice with PBS and equilibrated in hybridization buffer (2× SSC buffer, 10% formamide (Merck, F9037)) for at least 10 min at RT. An equivalent volume of the following mixes was prepared: Mix1 (2× smiFISH probes mix, 2× SSC buffer, 30% formamide, 0.68 mg ml–1 E. coli tRNA (Roche, 10109541001)) and Mix2 (0.4 mg ml–1 BSA (NEB, B9000S), 4 mM VRC, 21.6% dextran sulfate (Merck, D8906)). Mix1 and Mix2 were combined 1:1 and thoroughly mixed by vortexing. Then, 45 μL of the resulting solution were applied on the surface of a 10-cm plastic dish that served as hybridization chamber. Each coverslip was applied upside-down on the smiFISH mix drop. A hydration chamber (a 3.5-cm plate filled with hybridization buffer) was included. The hybridization chamber was sealed with parafilm and incubated overnight at 37 °C, in the dark. The day after, each coverslip was washed twice at 37 °C for 30 min in 2 ml hybridization buffer. In the second wash, 0.5 μg ml–1 DAPI (Merck, MBD0015) was included for nuclear counterstain. After two PBS washes, coverslips were mounted with 5 μL Vectashield (Vector Laboratories, H-1000) and sealed with nail polish. For Alexa Fluor Plus 647 imaging, mounting was performed with Aqua-Poly/Mount47 (Polysciences, 18606).

Confocal microscopy and image processing

Cells processed for immunofluorescence/smFISH were imaged using spinning disk confocal microscopy on an inverted Leica DMi8 equipped with a CSU-W1 confocal scanner unit (Yokogawa), with a 1.4-NA ×63 oil-objective (HCX PL APO lambda blue) and an ORCA-Flash4.0 camera (Hamamatsu). DAPI, AF488 (IF) and Cy3 (smFISH) were excited using a 405 nm (20% laser power), 488 nm (70%) or 561 nm (70%) laser line, respectively. For dual-color IF experiments, Alexa Fluor Plus 647 was excited using the 642 nm laser line. Three-dimensional image acquisition was managed using MetaMorph software (Molecular Devices). Images of 2,048 × 2,048 pixels (16-bit) were acquired with a xy pixel size of 0.103 μm and a z step size of 0.3 μm (~30–40 optical slices). Multichannel acquisition was performed at each z-plane. Multicolor fluorescent beads (TetraSpeck Fluorescent Microspheres, Invitrogen, T14792) were imaged alongside the samples. Chromatic shift registration was performed with Chromagnon48 using the fluorescent beads hyperstack as reference. Image channels were split, and maximum intensity projections (MIPs) were generated in Fiji49 using a macro. smFISH RNA spots were detected and counted using the RS-FISH Fiji plugin50 on MIPs. Briefly, anisotropy coefficient calculation was performed on a smFISH z-stack image, and spot detection on MIPs was performed in ‘advanced mode’ (no RANSAC, compute min/max intensity from image, use anisotropy coefficient for DoG, add detections to ROI-Manager, mean background subtraction, Sigma = 1.25, DoG and intensity thresholds were manually adjusted). All detected RNA spots were saved as region of interest (ROI) selections and used to create an RNA spots label map image (each spot is identified as a pixel with a distinct value) using a custom Fiji macro. A CellProfiler51 pipeline was used to segment cells and allocate and count cytoplasmic RNA spots. Briefly, DAPI images were used to identify nuclei as primary objects using a minimum cross-entropy thresholding method, smFISH background fluorescence was used to identify cell boundaries as secondary objects and cytoplasmic regions were derived by subtracting nuclei from cells. The ‘RelateObjects’ function was used to assign each RNA spot to the mother object cytoplasm. The total number of cytoplasmic RNA spots per image was computed. To count the number of cytoplasmic RNA spots per cell, cells touching the image border were excluded. The detection of cytoplasmic RNA spots (smFISH) co-localizing with protein spots (IF) was performed manually on chromatic-shift-corrected multichannel z-stack images. To avoid operator bias in image annotation, image files were randomized using a custom Fiji macro script before the analysis. The position of cytoplasmic RNA spots was used as reference to check for the presence of resolution-limited particles in the IF channel, distinct from the background and overlapping in xyz with the RNA spots. The position of each positive co-localization event was recorded in ROI manager. To account for RNA abundance, the number of RNA spots that co-localized with protein spots was normalized to the total number of cytoplasmic RNA spots per image and expressed as a fraction. If not specified otherwise, images shown in the main figures correspond to representative subsets of single optical planes from chromatic-shift-corrected confocal images. Brightness and contrast adjustments were applied on the entire image in Fiji to facilitate the visualization, without background clipping.

siRNA transfection

Control (siCTR) and TAF1 siRNAs were purchased from Horizon (ON-TARGETplus Non-targeting Control Pool D-001810-10-05, ON-TARGETplus Human TAF1 siRNA SMARTpool L-005041-00-0010) and resuspended in nuclease-free H2O. For large-scale transfections, 2.5 × 106 HeLa cells were seeded in 10-cm plates. The next day, cells were transfected using Lipofectamine 2000 (Invitrogen, 11668019) using a low-volume transfection protocol. In brief, after medium removal, cells were treated with 17.5 μL lipofectamine 2000 diluted in 2.8 ml Opti-MEM (Gibco, 31985062) for 15 min at 37 °C. Then, 56 pmol of siRNA diluted in 0.7 ml Opti-MEM was added dropwise to the cells and gently mixed, achieving a 16 nM final siRNA concentration. After ~5 h of incubation at 37 °C, the transfection mix was replaced with prewarmed complete DMEM. Cells were collected 48 h post-transfection.

Western blot

Samples were loaded on SDS–PAGE gels with 0.5% 2,2,2-trichloroethanol (TCE, Sigma-Aldrich) added for stain-free protein detection52. The gel was activated for one minute with UV and the proteins were transferred to a nitrocellulose membrane following standard procedures. Specific proteins were probed with the primary antibodies listed in Supplementary Table 3 and HRP-conjugated secondary antibodies. To reprobe the membrane with an antibody raised in a different species, the previous secondary antibody was inactivated with 10% acetic acid according to53. Detection was performed using a ChemiDoc Touch system (BioRad) and images were visualized in ImageLab v6.0 software (BioRad).

Subcellular fractionation

Adherent cells were washed with cold PBS twice and harvested by scraping on ice. Cell suspension was centrifuged at 400 × g for 5 min at 4 °C and the pellet was resuspended in 4 packed cell volumes (PCV) of hypotonic buffer (50 mM Tris-HCl pH 8.0, 1 mM EDTA, 1 mM DTT, 1× PIC). After 30 min incubation on ice, cells were lysed with 10 hits of Dounce homogenizer and centrifuged at 2,300g for 10 min at 4 °C. The supernatant was saved as cytoplasmic extract. Nuclei were washed once in hypotonic buffer and resuspended in 3.5 PCV hypertonic buffer (50 mM Tris-HCl pH 8.0, 0.5 mM EDTA, 500 mM NaCl, 25% glycerol, 1 mM DTT, 1× PIC). Nuclei were lysed with 20 hits of Dounce homogenizer, incubated with agitation for 30 min at 4 °C and centrifuged at 19,000g for 30 min at 4 °C. The supernatant was saved as nuclear extract. Cytoplasmic and nuclear extracts were dialyzed against 25 mM Tris-HCl pH 8.0, 5 mM MgCl2, 100 mM KCl, 10% glycerol, 0.5 mM DTT, 1× PIC at 4 °C using DiaEasy dialyzers (BioVision K1013-10), and protein concentration was measured using the Bradford assay (BioRad, 5000006).

Immunoprecipitation coupled to LC–MS/MS analysis

Specific and mock (anti-GST) antibodies were coupled either with 200 μL Protein G Sepharose (large-scale IPs, Fig. 4 and Extended Data Fig. 4) or with 2.7 mg Protein G Dynabeads (medium scale IPs, Fig. 6) in IP100 buffer (25 mM Tris-HCl pH 8.0, 5 mM MgCl2, 100 mM KCl, 10% glycerol, 0.1% NP-40, 0.5 mM DTT, 1× PIC) in agitation for 1 h at RT. Antibody-coupled beads were washed twice in IP500 buffer (25 mM Tris-HCl pH 8.0, 500 mM KCl, 5 mM MgCl2, 10% glycerol, 0.1% NP-40, 0.5 mM DTT, 1× PIC) and three times in buffer IP100. Antibody-coupled beads were incubated with cytoplasmic (3–30 mg, medium–large-scale IPs) or nuclear (1–10 mg) extracts overnight at 4 °C. The day after, beads were washed twice with IP500 for 5 min at 4 °C and three times with IP100. Immunopurified proteins were eluted in 0.1 M glycine pH 2.7 and immediately buffered with 1 M Tris-HCl pH 8.0. Eluates were precipitated with TCA (Merck, T0699) overnight at 4 °C and centrifuged at 14,000 g for 30 min at 4 °C. Protein pellets were washed twice with cold acetone and centrifuged at 14,000 g for 10 min at 4 °C. Pellets were denatured with 8 M urea (Merck, U0631) in 0.1 M Tris-HCl, reduced with 5 mM TCEP for 30 min and alkylated with 10 mM iodoacetamide (Merck, I1149) for 30 min, with protection against light. Both reduction and alkylation were performed at RT and in agitation. Double digestion was performed with endoproteinase Lys-C (Wako, 125-05061) at a 1:100 ratio (enzyme:protein) in 8 M urea for 4 h, followed by an overnight modified trypsin digestion (Promega, V5113) at a 1:100 ratio in 2 M urea for 12 h.

Samples were analyzed using an Ultimate 3000 nano-RSLC coupled in line, via a nano-electrospray ionization source, with the LTQ-Orbitrap ELITE mass spectrometer (Thermo Fisher Scientific) or with the Orbitrap Exploris 480 mass-spectrometer (Thermo Fisher Scientific) equipped with a FAIMS (high Field Asymmetric Ion Mobility Spectrometry) module. Peptide mixtures were injected in 0.1% TFA on a C18 Acclaim PepMap100 trap-column (75 µm ID × 2 cm, 3 µm, 100 Å, Thermo Fisher Scientific) for 3 min at 5 µL min–1 with 2% ACN and 0.1% FA in H2O and then separated on a C18 Acclaim PepMap100 nano-column (75 µm ID × 50 cm, 2.6 µm, 150 Å, Thermo Fisher Scientific) at 300 nL min–1, at 40 °C with a 90 min linear gradient from 5% to 30% buffer B (A: 0.1% FA in H2O/B: 80% ACN, 0.1% FA in H2O), with regeneration at 5% B. Spray voltage was set to 2.1 kV, and the heated capillary temperature was set at 280 °C. For Orbitrap Elite, the mass spectrometer was operated in positive ionization mode, in data-dependent mode with survey scans from m/z 350–1,500 acquired in the Orbitrap at a resolution of 120,000 at m/z 400. The 20 most intense peaks from survey scans were selected for further fragmentation in the Linear Ion Trap with an isolation window of 2.0 Da and were fragmented by CID with normalized collision energy of 35% (TOP20CID method). Unassigned and single charged states were excluded from fragmentation. The ion target value for the survey scans (in the Orbitrap) and the MS2 mode (in the linear ion trap) were set to 1E6 and 5E3, respectively, and the maximum injection time was set to 100 ms for both scan modes. Dynamic exclusion was set to 20 s after one repeat count, with mass width at ± 10 ppm. For Orbitrap Exploris 480 MS associated with the FAIMS module, a combination of two compensation voltages, −40 V and −55 V, was chosen, with a cycle time of 1 s for each. For the full MS1 in DDA mode, the resolution was set to 60,000 at m/z 200 and with a mass range set to 350–1400. The full MS AGC target was 300%, with an IT set to Auto mode. For the fragment spectra in MS2, the AGC target value was 100% (Standard) with a resolution of 30,000 and the maximum Injection Time set to Auto mode. Intensity threshold was set at 1E4. Isolation width was set at 2 m/z and normalized collision energy was set at 30%. All spectra were acquired in centroid mode using positive polarity. Default settings were used for FAIMS with voltages applied as described previously, and with a total carrier gas flow set to 4.2 L min–1.

Mass spectrometry data analysis

Proteins were identified by database searching using SequestHT (Thermo Fisher Scientific) with Proteome Discoverer 2.4 software (PD2.4, Thermo Fisher Scientific) on human FASTA database downloaded from UniProt (reviewed, release 2021_06_03, 20380 entries, Precursor and fragment mass tolerances were set at 7 ppm and 0.6 Da, respectively, and up to 2 missed cleavages were allowed. For the data acquired on the Orbitrap Exploris 480, the software Proteome Discoverer 2.5 version was used with a human fasta database from UniProt (reviewed, release 2022_02_21, 20291 entries). Precursor and fragment mass tolerances were set at 10 ppm and 0.02 Da respectively, and up to 2 missed cleavages were allowed. For all the data, oxidation (M, +15.995 Da) was set as a variable modification, and carbamidomethylation (C, + 57.021 Da) as a fixed modification. Peptides and proteins were filtered with a false discovery rate (FDR) at 1%. Label-free quantification was based on the extracted ion chromatography intensity of the peptides. All samples were measured in technical triplicates. The measured extracted ion chromatogram (XIC) intensities were normalized on the basis of median intensities of the entire dataset to correct minor loading differences. For statistical tests and enrichment calculations, not detectable intensity values were treated with an imputation method, where the missing values were replaced by random values similar to the 10% of the lowest intensity values present in the entire dataset. Unpaired two-tailed t-tests, assuming equal variance, were performed on obtained log2 XIC intensities. Normalized spectral abundance factors (NSAF) were calculated for each protein, as previously described54. To obtain spectral abundance factors (SAF), spectral counts identifying a protein were divided by the protein length. To calculate NSAF values, the SAF values of each protein were divided by the sum of SAF values of all detected proteins in each run. All raw LC–MS/MS data have been deposited to the ProteomeXchange via the PRIDE repository with identifier PXD036358.

Crosslinking-MS metanalysis, protein sequence analysis and modeling

For the metanalysis on the available crosslinking-MS experiments performed on human TFIID, we retrieved and combined the curated datasets from ref. 18 (one dataset, apo-TFIID), ref. 19 (one dataset, apo-TFIID) and ref. 10 (five datasets of TFIID incorporated in preinitiation complex variants: cPICscp, cPICpuma, mPICscp, hPICscp, p53hPIChdm2), for a total of seven datasets. We included only intra and interprotein crosslinks involving TAF1 and found in at least two different datasets. If a crosslink was present only among the Chen et al.10 datasets, it was considered only if it scored as significant in more than one dataset (probability score < 0.05). The resulting subset of common TAF1 crosslinks is reported in Supplementary Table 2.

TAF1 conservation and structural disorder prediction were computed using ConSurf55 and Metapredict56, respectively. The TAF1 full-length model, corresponding to the UniProt entry P21675, was downloaded from the AlphaFold Protein Structure Database ( For visual clarity in Figure 5c, the model backbone was manually extended at low-confidence coil regions in UCSF ChimeraX57. The TAF1–TBP–TAF11–TAF13 subcomplex was modeled using AlphaFold2_advanced ColabFold implementation with standard settings ( and using the following protein fragments as input: TAF1 (1–300 aa), TBP (150–339 aa), TAF11 (50–211 aa), TAF13 (full length). The TAF1–TAF6HEAT–TAF6HEAT–TAF8 subcomplex was modeled using AlphaFold2 Multimer extension on COSMIC2 server with standard settings45,59 and using the following protein fragments as input: TAF1 (300–550 aa), TAF6 (215–482 aa) and TAF8 (130–220 aa). All structural models were visualized, analyzed and rendered in UCSF ChimeraX57.

Reporting summary

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