RNA-Binding Specificity of the Human Fragile X Mental Retardation Protein

Abstract

Fragile X syndrome (FXS) is the most common form of inherited intellectual disability and is caused by a deficiency of the fragile X mental retardation protein (FMRP) in neurons. FMRP regulates the translation of numerous mRNAs within dendritic synapses, but how FMRP recognizes these target mRNAs remains unknown. FMRP has KH0, KH1, KH2, and RGG domains, which are thought to bind to specific RNA recognition elements (RREs). Several studies used high-throughput methods to identify various RREs in mRNAs that FMRP may bind to in vivo. However, there is little overlap in the mRNA targets identified by each study, suggesting that the RNA-binding specificity of FMRP is still unknown. To determine the specificity of FMRP for the RREs, we performed quantitative in vitro RNA binding studies with various constructs of human FMRP. Unexpectedly, our studies show that the KH domains do not bind to the previously identified RREs. To further investigate the RNA-binding specificity of FMRP, we developed a new method called Motif Identification by Analysis of Simple sequences (MIDAS) to identify single-stranded RNA (ssRNA) sequences bound by KH domains. We find that the FMRP KH0, KH1, and KH2 domains bind weakly to the ssRNA sequences suggesting that they may have evolved to bind more complex RNA structures. Additionally, we find that the RGG motif of human FMRP binds with a high affinity to an RNA G-quadruplex (GQ) structure that lacks single-stranded loops, double-stranded stems, or junctions.

INTRODUCTION

Fragile X syndrome (FXS) is the most common form of inherited intellectual disability. FXS is primarily caused by the deficiency of the Fragile X Mental Retardation Protein (FMRP) in neuronal cells [1]. The expansion of CGG trinucleotide repeats in the 5′-untranslated region of the FMR1 gene leads to gene silencing and FMRP deficiency [2–5]. FMRP has been proposed to regulate local protein synthesis within dendritic synapses [6–8]. FMRP deficiency results in the formation of dendrites with long, thin, and immature spines possibly caused by the deregulation of translation of specific neuronal mRNAs [9–11]. However, how FMRP binds to specific neuronal mRNAs to regulate their translation is unknown.

FMRP contains several conserved domains. Amongst the conserved domains, four are putative RNA-binding domains. Three K-Homology (KH) domains — KH0, KH1, and KH2 — and an arginine-glycine-glycine (RGG) motif are hypothesized to mediate FMRP binding to mRNAs [12,13]. The KH domain was first identified as a nucleic acid-binding domain in heterogeneous nuclear ribonucleoprotein K (hnRNP K) and has since been found in many eukaryotic (Type I) and prokaryotic (Type II) proteins [14,15]. A GXXG loop in the KH1 and KH2 domains of FMRP is conserved in many RNA-binding KH domains, such as the KH-type splicing regulatory protein (KSRP) and neuro-oncological ventral antigen (Nova-1 and 2) proteins, further suggesting FMRP KH domains play a role in binding specific RNAs [16,17]. A third KH domain was recently discovered upstream of the KH1 domain, termed KH0 [18,19]. Unlike KH1 and KH2, KH0 does not contain the GXXG loop, which suggests KH0 may not bind RNA or bind a different RNA motif. Mutations in the KH1 and KH2 domains of FMRP also result in FXS, indicating that the RNA-binding function of FMRP is essential for normal health [20,21]. The KH2 domain was shown to bind to an in vitro-selected kissing complex RNA called KC2 [22]. Additionally, the RGG motif binds to an in vitro-selected G-quadruplex forming RNA called Sc1 [23]. An NMR structure of the human FMRP RGG peptide bound to Sc1 RNA G-quadruplex (GQ) reveals how the disordered RGG motif is stabilized upon binding to the GQ structure [24]. A subsequent crystal structure reveals in atomic detail the interaction of each amino acid within the RGG motif with the GQ nucleotides of Sc1 [25]. Interestingly, the RGG peptide interacts not with the GQ region, but binds to the junction between the GQ and a double-stranded stem in SC1. While Sc1 was selected as a high-affinity binding partner for the RGG motif, it suggested that the RGG motif mediates FMRP binding to GQ structures in mRNAs. Indeed, the FMRP RGG peptide has been shown to bind to various GQ structures found in mRNAs with a range of binding affinities [26–30]. Given the ubiquity of potential GQ structures throughout the transcriptome, the presence of the RGG motif within FMRP hints at FMRP′s potential to bind a vast pool of target mRNAs in vivo.

The identification of FMRP target mRNAs is crucial for understanding the molecular function of FMRP and has been the focus of numerous studies. Most endeavors to identify FMRP target mRNA motifs have employed high-throughput methods such as coimmunoprecipitation of mRNA-FMRP complexes (Co-IP), cross-linking and immunoprecipitation combined with high-throughput sequencing (HITS-CLIP), photoactivatable ribonucleoside-enhanced cross-linking and immunoprecipitation (PAR-CLIP), RNAcompete, and targets of RNA-binding proteins identified by editing (TRIBE) [31–35]. Unfortunately, the proposed catalog of RNA sequence and structural motifs, or RNA-recognition elements (RREs), overlap poorly between each study [36]. Analysis of the proposed RREs yielded two promising RREs: clustered WGGA (W = A/U) sequences and the GACR (R = A/G) sequence [36]. The WGGA clusters were originally proposed to bind specifically to the KH1 domain of FMRP [31]. A subsequent study proposed that WGGA clusters could form GQ structures and therefore bind the RGG motif of FMRP instead [36]. The GACR sequences were enriched in RNAcompete for an FMRP fragment composed solely of the KH1 and KH2 RNA binding domains, suggesting that the GACR motif binds a KH domain [35].

Here, we use fluorescence anisotropy to quantify binding affinities between FMRP and RNAs of a defined length and sequence. We modeled the RNAs after proposed FMRP RREs [31,35,36]. Unexpectedly, our studies show that the KH0, KH1, and KH2 domains do not bind to the WGGA, ACUK, and GACR motifs. Previous studies showed that many KH domains recognize only four to five nucleotides in ssRNAs [16,17,37–39]. This inspired us to develop a new method to identify RNA sequences that bind to KH domains, which we named Motif Identification by Analysis of Simple sequences (MIDAS). We synthesized a pool of RNAs having five randomized nucleotides flanked by fixed sequences of eight nucleotides at the 5′ and 3′ ends. We used the randomized RNA pool to identify RNA molecules that bind to the KH domains of FMRP (the pool will contain 4⁵ = 1024 unique sequences). Surprisingly, the FMRP KH domains bind weakly to the selected RNA sequences suggesting that the KH domains of FMRP may not interact with simple RREs as the canonical KH domains. Instead, the FMRP KH domains may have evolved to recognize more complex RNA motifs. Additionally, consistent with previous studies, we find that the RGG motif of human FMRP binds to RNA GQ structures with a 100-fold range in affinities showing that FMRP can discriminate between the different GQ structures [26–30,40,41].

RESULTS

Isolating the RNA-binding domains of FMRP

To elucidate how FMRP specifically binds its target RNAs, we isolated specific regions of FMRP and quantified their binding affinity to a catalog of RNAs modeled after reported FMRP RREs. We sought to both test the validity of the proposed RREs within RNAs and clarify the roles of the KH0, KH1, KH2, and RGG domains in binding the RREs. We expressed and purified nine human FMRP (hFMRP) variants from E. coli for our studies (Figure 1). The N-terminus truncated human NT-FMRP construct was cloned by deleting the amino acids immediately upstream of the KH1 domain from the full-length sequences (NT-hFMRP encodes Arg218 to Pro632). This truncation eliminates the FMRP N-terminal domain (NTD), which contains the tandem Agenet 1 and Agenet 2 domains as well as the KH0 domain; the NTD is also responsible for FMRP dimerization [18]. The NT-hFMRP construct conserves the three proposed RNA-binding domains [42–44]. We also generated a NT-hFMRPΔKH1 construct encoding Phe281 to Pro632 to determine the binding specificity of the KH2 domain. Additionally, we made individual KH0, KH1, KH2, KH1-KH2, and KH0-KH1-KH2 fused downstream of the glutathione S-transferase (GST) tag. The fusion proteins are large enough to generate clear shifts in fluorescence anisotropy should the KH motifs bind the fluorescein-labeled RNAs.

Figure 1.

Design and purification of human FMRP constructs. (a) hFMRP is the full-length human FMRP isoform 1, spanning E2-P632 as indicated. The different truncation constructs of hFMRP that were made are illustrated. GST-hFMRP SGG is a fusion between the glutathione S-transferase and the RGG motif-containing sequence of hFMRP spanning G531-P632 where the 16 arginines spanning the RGG motif to the C-terminus were mutated to serines, illustrated using a different color for the C-terminus. (b) SDS-polyacrylamide gel showing the purified proteins illustrated above. Lanes are: molecular weight standards (MW), NT-hFMRP (47 kDa), NT-hFMRP ΔKH1 (40 kDa), GST-hFMRP RGG (40 kDa), GST-hFMRP SGG (39 kDa), GST-hFMRP KH0 (38 kDa), GST-hFMRP KH1 (36 kDa), GST-hFMRP KH2 (45 kDa), GST-hFMRP KH1-KH2 (52 kDa), GST-hFMRP KH0-KH1-KH2 (63 kDa), and GST-Nova-2 KH3 (38 kDa).

To evaluate the importance of the RGG motif in binding RNA GQ structures, we generated an hFMRP construct encoding Gly531 of the RGG domain to the C-terminal Pro632 and fused it downstream of the GST tag to generate a GST-hFMRP RGG fusion protein. Finally, we mutated all of the arginines to serines in the hRGG-encoding region, generating a GST-hFMRP SGG fusion protein. Together these FMRP constructs should delineate the contributions of the KH domains and the RGG motif towards binding to RNAs.

Design of model RNAs

We designed a set of RNAs modeled after three promising FMRP RREs — GQ structure, GACR clusters, and WGGA clusters — to determine if and how FMRP specifically binds these sequences (Table 1). Apart from the UG₄U 6-mer RNA, we standardized our model RNAs to 18 nucleotides in length to control for length-dependent binding affinities; M1 served as an 8-mer RNA size marker for UG₄U. We used PolyG₁₇ and UG₄U as two different model GQ RNAs to quantify and compare FMRP binding affinities for different RNA GQ structures [45–47]. UG₄U is noteworthy because the x-ray crystal structure showed that it forms a highly stable, four stacked G-tetrad structure [46], and we wanted to test whether the RGG domain could bind to a GQ structure that does not contain single-stranded regions, junctions or double-stranded stems.

Table 1.

Nomenclature and sequences of the RNAs

RNA	Sequence (5’ to 3’)
PolyG17	GGGGGGGGGGGGGGGGGU
UG4U	UGGGGU
(UGGA)4	UUGGAUGGAUGGAUGGAU
(GACG)4	UGACGGACGGACGGACGU
UGGm6ACU	UGGACUUGGm6ACUUGGACU
PolyC17	CCCCCCCCCCCCCCCCCU
ACUU	CCCCCCCACUUCCCCCCU
UGGA	CCCCCCCUGGACCCCCCU
CR1	GCUAUCCAGAUUCUGAUU

We generated (GACG)₄ and (UGGA)₄ 18-mer RNAs containing four tandem repeats of GACG and UGGA to test FMRP binding to the GACR cluster and WGGA cluster, respectively. The GACR array was reported to bind the KH1 and KH2 domains of FMRP [35]. The WGGA array was originally reported to bind the KH1 domain but was later hypothesized to assemble an intramolecular GQ structure that may bind the FMRP RGG motif [31,36]. We also tested FMRP binding to ACUU and UGGA RNAs to determine if a single ACUK (K = G/U) or WGGA motif is sufficient for binding to FMRP, respectively [31]. Finally, we used two 18-mer RNAs with different base compositions that cannot assemble any secondary structures as control RNAs, denoted PolyC₁₇, and CR1. All RNAs had a 3′-terminal uridine to prevent quenching of the fluorescein dye attached to the 3′ end [48]. The fluorescein dye was used to visualize the RNAs on gels and for performing fluorescence anisotropy binding studies.

Folding of model RNAs

We first tested whether our model RNAs assemble into folded structures by denaturing and native PAGE analyses (Figure 2). We tested whether UG₄U assembles the reported 24-mer intermolecular GQ composed of 4 stacked G-tetrads using two methods: (1) a denaturing urea PAGE in the presence of KCl, NaCl or LiCl as described previously [49], and (2) a native PAGE comparing UG₄U migration to the 8-mer M1, 18-mer PolyC₁₇ and 18-mer CR1 RNAs, which remain unfolded. The denaturing PAGE showed UG₄U migrating as two different species, with the presence of the slow-migrating species being K⁺-dependent (Figure 2a). The UG₄U GQ assembly was remarkably stable as it endured urea denaturing conditions even with Li⁺. The native PAGE showed UG₄U migrating much slower than the 8-nucleotide M1 RNA and slightly slower than the 18-mer RNAs (Figure 2c). Together, both PAGE analyses indicate that UG₄U indeed assembles the tetrameric GQ structure. All 18-nucleotide RNAs migrated as expected on a denaturing PAGE (Figure 2b). However, on a native PAGE, PolyG₁₇ migrated much slower than the control 18-nucleotide RNAs as a heterogeneous array (Figure 2c). This suggests the PolyG₁₇ array may comprise a mix of 36-nucleotide intermolecular GQ structures in different registers. This is consistent with a recent report that showed that PolyG₄₀ forms stable GQ structures [47].

Figure 2.

Analysis of the structures formed by the model RNAs. (a) Urea PAGE of UG₄U denatured and then refolded in the presence of 100 mM LiCl, 100 mM NaCl or 100 mM KCl. (b) Urea PAGE of the denatured 18-mer model RNAs. (c) Native PAGE of the model RNAs. G-quadruplex species are marked with red arrows and a blue bar; (GACG)₄ hairpin monomer and dimer species are marked with green and magenta arrows, respectively; (UGGA)₄ structure is marked with a cyan arrow. (d) Fold increase in fluorescence anisotropy of 5 nM fluorescein-labeled RNAs upon the addition of 430 nM anti-GQ BG4 scFV antibody. All RNAs are labeled at 3′ end with fluorescein. The standard deviations from two experiments are shown. (e) Increase in fluorescence intensity of NMM with increasing concentrations of unlabeled RNA. The inset shows the data for the (UGGA)₄ RNA performed at lower concentrations of RNA. The closed and open symbols show RNAs in KCl and LiCl, respectively. The standard deviations from three experiments are shown.

The (GACG)₄ and (UGGA)₄ RNAs migrated differently from the other RNAs. The (GACG)₄ RNA migrated as two species, one migrating slowly at a similar rate as PolyG₁₇ and the other migrating faster than the unstructured 18-nucleotide control RNAs (Figure 2c). Unlike PolyG₁₇, (GACG)₄ does not contain four repeats of at least two consecutive guanines required to assemble an intramolecular GQ structure [50,51]. The slower (GACG)₄ species may be a stable dimer structure formed by intermolecular base pairs (Supplemental Figure 1). The faster (GACG)₄ species could then be the monomeric 18-mer stem-loop structure, which would be expected to migrate faster than the unfolded control RNAs.

The (UGGA)₄ RNA is predicted to assemble an intramolecular GQ structure composed of 2 stacked G-tetrads [52,53]. It is indeed migrating at a similar rate as one of the PolyG₁₇ species within the PolyG₁₇ array (Figure 2c). To verify whether (UGGA)₄ is forming a GQ structure, we used the BG4 antibody that binds specifically to GQ structures [54]. The binding of the BG4 antibody to the fluorescein-labeled RNAs was monitored by fluorescence anisotropy, which should increase because the RNA-BG4 antibody complex will tumble slower than the free RNAs. We observed a substantial increase in anisotropy with UG₄U and PolyG₁₇, which is consistent with the fact that these RNAs form GQ structures (Figure 2d). The (UGGA)₄ RNA showed a modest increase in anisotropy, suggesting that the BG4 antibody binds poorly to the GQ structure formed by (UGGA)₄. Alternatively, the change in anisotropy is small because the fluorescein attached to the RNA is dynamic in the (UGGA)₄ RNA-BG4 antibody complex and does not correlate with the tumbling rate of the complex. No change in anisotropy was observed with (GACG)₄, indicating that this sequence does not form a GQ structure.

To further validate the formation of GQ structures, we used the N-methyl mesoporphyrin IX (NMM) dye, which shows enhanced fluorescence when it intercalates into GQ structures [55,56]. We examined the presence of GQ structures in PolyG₁₇, UG₄U, (UGGA)₄, (GACG)₄, and tRNA reconstituted in buffer containing either KCl or LiCl with the NMM dye. The tRNA served as a negative control. Consistent with the native gel analysis and the BG4 antibody assay, we observed increased fluorescence intensity with PolyG₁₇, UG₄U, and (UGGA)₄ but not with (GACG)₄ and tRNA (Figure 2e). The increase in the fluorescence intensity was higher when the RNAs were in KCl than in LiCl, demonstrating that PolyG₁₇, UG₄U, and (UGGA)₄ form GQ structures.

FMRP binds to different G-quadruplex RNAs

Little is known regarding how FMRP specifically binds its target mRNAs. A previous study suggested that the KH1 and KH2 domains of Drosophila FMRP bind to the ribosome while the RGG motif is available to bind the target mRNA [57]. Previous reports also showed that the FMRP RGG motif binds to the in vitro selected Sc1 RNA GQ structure and to a variety of GQ structures from potential mRNA targets of FMRP [23,25–30,40,41,58]. However, it is unknown whether FMRP can bind with a high affinity to a GQ structure that does not have any single-stranded regions, double-stranded stems, or junctions. We employed fluorescence anisotropy to determine the equilibrium dissociation constants (K_D) of several human FMRP variants binding to RNAs that assemble different GQ structures or contain RREs that have been reported to bind FMRP (Table 1).

We observed that NT-hFMRP binds to PolyG₁₇ with a K_D of 14 ± 2 nM and UG₄U with a K_D of 110 ± 16 nM (Figure 3a, Table 2). Additionally, NT-hFMRP binds to (UGGA)₄ with a K_D of 710 ± 50 nM. Since all three RNAs form GQ structures, our results show that NT-hFMRP binds to the different GQ structures with a 100-fold range in binding affinities. Interestingly, NT-hFMRP failed to bind to the (GACG)₄ RNA, which was modeled after the GACR array reported to bind to the KH domains of FMRP (Figure 3a). The four tandem GACG repeats may be insufficient for binding to FMRP. The (GACG)₄ RNA could assemble a mixed population of GC stem-loop monomers and double-stranded dimers. FMRP′s inability to bind either the GC stem-loop monomer or dimer would suggest FMRP does not bind double-stranded RNA assemblies even if they are G-rich. NT-hFMRP did not bind to RNAs having a single UGGA or ACUU, which cannot form GQ structures (Figure 3a). Finally, despite lacking the KH1 domain, NT-hFMRPΔKH1 showed a binding pattern similar to NT-hFMRP (Figure 3b, Table 2).

Figure 3.

FMRP binding to model RNAs. Fold increase in fluorescence anisotropy of 5 nM fluorescein-labeled RNAs upon the addition of 0–800 nM of (a) NT-hFMRP, (b) NT-hFMRP ΔKH1, (c) GST-hFMRP RGG, and (d) GST-hFMRP SGG. Bovine serum albumin (BSA) binding to the GQ-assembling RNAs was also measured to test for non-specific protein binding by the two RNAs. Binding trials were performed at least three times, with error bars depicting standard deviation. The symbols indicate the identity of the different RNAs.

Table 2.

Equilibrium dissociation constant, K_D (nM) of FMRP binding to RNAs

RNA	NT-hFMRP	NT-hFMRP ΔKH1	RGG	SGG
UG4U	110 ± 16	80 ± 7	210 ± 18	No binding
PolyG17	14 ± 2	17 ± 3	8.6 ± 1.2	No binding
(UGGA)4	710 ± 50	840 ± 39	> 4000	No binding
(GACG)4	No binding	Not tested	No binding	Not tested
UGGm6ACU	1250 ± 610	Not tested	Not tested	Not tested

FMRP RGG motif specifically binds to RNA G-quadruplex structures

To determine whether the RGG motif is responsible for binding to the RNA GQ structures, we tested the binding of our model RNAs to GST-hFMRP RGG. Consistent with previous studies [23,25–30,40,41,58], the RGG motif was sufficient for binding PolyG₁₇ and UG₄U GQ structures (Figure 3c, Table 2). In fact, GST-hFMRP RGG bound PolyG₁₇ with a K_D of 8.6 ± 1.2 nM and UG₄U with a K_D of 210 ± 18 nM, similar affinities as NT-hFMRP. The nanomolar affinity of NT-hFMRP for UG₄U shows that the RGG domain can bind to GQ structures that do not contain single-stranded loops, double-stranded stems, or junctions. Interestingly, GST-hFMRP RGG bound (UGGA)₄ with a K_D of > 4000 nM, a significantly lower affinity than NT-hFMRP. This suggests that the region from the KH2 to the RGG domain may be necessary for binding to (UGGA)₄. To confirm that the RGG motif is required for binding to GQ RNAs, we made a GST-hFMRP SGG construct with all sixteen arginines in the RGG motif and C-terminal domain changed to serines. The GST-hFMRP SGG protein did not bind to PolyG₁₇ and UG₄U GQ structures, demonstrating that the RGG motif is essential for binding to GQ RNAs (Figure 3d).

Finally, we tested whether the FMRP RGG motif is binding to the GQ assemblies of PolyG₁₇ and UG₄U, or unfolded forms of the RNAs. To assess specific binding to GQ assemblies, we compared the binding efficiency of GST-hFMRP RGG to PolyG₁₇ and UG₄U in the presence of 75 mM KCl versus 75 mM LiCl. If FMRP could bind unfolded forms of PolyG₁₇ and UG₄U, then we would observe increased binding efficiency in LiCl, where both RNAs favor unfolded states. Conversely, if FMRP specifically binds the folded form, then we would observe decreased binding efficiency due to a drop in the folded population of RNA. Indeed, substituting Li⁺ in place of K⁺ decreased the net anisotropy change for PolyG₁₇, reflecting a smaller population of the RNA binding to FMRP (Supplemental Figure 2a). Additionally, GST-hFMRP RGG bound to PolyG₁₇ with a 50-fold lower affinity in the presence of Li⁺ (K_D = 480 ± 62 nM), consistent with a recent report [40]. In the case of UG₄U, we observed only a slight decrease in anisotropy with Li⁺ (Supplemental Figure 2b). This may be because Li⁺ does not significantly destabilize the UG₄U GQ assembly. While we see some destabilization by Li⁺ in our urea-PAGE, the urea may be catalyzing it. In both RNA cases, binding was not completely depleted in Li⁺. This suggests the cation substitution did not destabilize the protein. Instead, the protein was tightly binding to the remaining population of stable PolyG₁₇ and UG₄U GQ assemblies. The concomitant drop in the extent of the anisotropy shift with the drop in the GQ population supports the conclusion that the RGG motif specifically binds to RNA GQ structures.

The KH domains of FMRP do not bind to the WGGA and GACR motifs

Our studies showed that NT-hFMRP binds with low affinity to (UGGA)₄ RNA and does not bind to (GACG)₄ RNA. Furthermore, we showed that (UGGA)₄ forms a GQ structure suggesting that the RGG motif mediates the interaction of NT-hFMRP with (UGGA)₄. To confirm that the KH domains of FMRP do not bind to (UGGA)₄ and (GACG)₄ RNAs, we performed binding studies with the KH domains of FMRP fused to the GST tag (Figure 1). GST-hFMRP KH0, GST-hFMRP KH1, GST-hFMRP KH2, GST-hFMRP KH1-KH2, and GST-hFMRP KH0-KH1-KH2 did not bind to (UGGA)₄ and (GACG)₄ RNAs demonstrating that the KH domains of FMRP do not bind to the WGGA and GACR motifs (Supplemental Figure 3).

FMRP does not bind with high affinity to m⁶A RNA

Recent studies indicated that FMRP preferentially binds to mRNAs with the N⁶-methyladenosine (m⁶A) modification [59–61]. We, therefore, tested the binding of NT-hFMRP to the UGGm⁶ACU model RNA having a single m⁶A modification in the middle of the sequence. UGGACU is the consensus sequence for m⁶A modification in mRNAs and is sufficient for binding to YTH-domain proteins [62–64]. Surprisingly, our results show that NT-hFMRP binds poorly to the m⁶A modified RNA suggesting that at best, FMRP can only form a low-affinity complex with m⁶A-containing mRNAs (Supplemental Figure 4).

The KH domains of FMRP do not bind to short ssRNA motifs

The above results indicated that the KH domains of FMRP do not bind to the previously identified RREs. Biochemical and structural studies have shown that KH domains generally recognize 4 to 5 nucleotides in ssRNAs [16,17,37–39]. For example, the KH3 domain of Nova-2 protein binds to the UCACC sequence [37,38]. Structural studies showed that the UCACC binds to a cleft formed by the invariant and variable loops of Nova-2 KH3. Similarly, the KH3 and KH4 domains of ZBP1 bind to ACAC and CGGAC sequences, respectively [39]. Thus, KH domains typically recognize short, ssRNA sequences.

The realization that KH domains recognize 4 to 5 nucleotides motivated us to develop a method called MIDAS (Motif Identification by Analysis of Simple sequences) to determine the RNA-binding specificity of the FMRP KH domains (Figure 4a). We synthesized a pool of RNA molecules having 5 nucleotides randomized in the middle and flanked by 8 nucleotides of fixed sequences in the 5′ and 3′ ends (the pool will have 4⁵ = 1024 unique sequences). The sequence of the fixed nucleotides was explicitly designed to avoid any of the known KH domain binding RREs. The RNA pool was labeled at the 5′-end with ³²P and incubated with 10 μM protein having one or more KH domain. The RNA•protein complex was separated from free RNA using a native gel (Figure 4b). The band corresponding to the RNA•protein complex was excised, the RNA purified, and the sequences of the purified RNA molecules were identified by next-generation sequencing.

Figure 4.

Illustration of MIDAS method. (a) Cartoon diagram of the MIDAS process. Purified GST-Nova-2 KH3, GST-hFMRP KH0, and GST-hFMRP KH1-KH2 are individually incubated with the RNA library containing a central N5 sequence. The RNAs bound by each protein are separated by an electrophoretic mobility shift assay (EMSA) and purified from the gel. The input RNA sequences and the selected RNA sequences are then reverse transcribed, amplified, and Illumina sequenced. (b) Autoradiograph of EMSA. Contrast- enhanced image showing the free RNA library migrates as an array (black bar) and the RNAs bound by Nova-2 KH 3, FMRP KH0, and FMRP KH1-KH2 migrate similar distances as distinct bands. Control lanes are the MIDAS N5 RNA library incubated without protein.

To validate the MIDAS method, we used the well-studied Nova-2 KH3 domain [37,38]. As shown in Figure 4b, Nova-2 KH3 binds to some of the RNA pool molecules resulting in the slower migrating RNA•protein complex band in the native gel. Surprisingly, we observed very faint RNA•protein complex bands with both FMRP KH0 and FMRP KH1-KH2, suggesting very weak binding to the RNA pool molecules. Additionally, we observed no binding to the RNA pool with the individual FMRP KH1 and FMRP KH2 domains (data not shown). We analyzed the RNA sequences that bind to Nova-2 KH3. Interestingly, the highly enriched sequences are UUUCA, UUUAA, UUCAC, UUCAU, and UCAAA (Table 3 and Supplemental Table1). The previously studied UCACC sequence was ranked much lower than the highly enriched sequences [37,38]. To further validate the MIDAS method, we synthesized RNA molecules corresponding to the highly enriched sequences and performed equilibrium binding studies (Figure 5a). We find excellent agreement between the enrichment factor and the K_D for Nova-2 KH3 binding to the selected RNA molecules, which shows that the MIDAS method is capable of identifying authentic single-stranded RREs (Table 3).

Table 3.

MIDAS enrichment factor (EF) and binding affinity for RNAs

	RNA	EFa	Kd (nM)
Nova-2 KH3
	UUUCA	17.3	179 ± 23
	UUCAC	14.7	739 ± 82
	UUUAA	17.5	1038 ±120
	UCACC	2.8	2991 ± 450
	CCACC	1.1	6991 ± 1848
FMRP KH0
	UCACC	0.9	> 10,000b
	CCACC	0.6	> 10,000b
	GACUC	0.2	> 10,000b
	UUCAC	0.2	> 10,000b
	UUUCA	0.1	> 10,000b
FMRP KH1-KH2
	UCACC	0.5	> 10,000b
	CCACC	0.3	> 10,000b
	GACUC	0.3	> 10,000b
	UUCAC	0.3	> 10,000b
	UUUCA	0.1	> 10,000b

^aEnrichment factors were normalized to the relative abundance of selected RNAs from the EMSA gel

^bEstimated affinity based on EMSA selection using 10 μM FMRP KH domains

Figure 5.

KH domains binding to RNAs identified by MIDAS. Fold increase in fluorescence anisotropy of 5 nM fluorescein-labeled RNAs upon the addition of 0–8000 nM of (a) GST-Nova-2 KH3, (b) GST-hFMRP KH0, (c) GST-hFMRP KH1-KH2, and (d) GST-hFMRP KH1-KH2-KH3. Binding trials were performed at least three times, with error bars depicting standard deviation. The symbols indicate the identity of the different RNAs tested for each protein construct.

Analysis of the RNA molecules that bind to the FMRP KH0 and FMRP KH1-KH2 domains indicated that the enriched sequences are UCACC and CCACC. However, the overall enrichment factors for FMRP KH0 and FMRP KH1-KH2 were significantly lower compared to Nova-2 KH3 (Table 3). Binding studies showed that FMRP KH0 and KH1-KH2 does not bind to UCACC at the standard concentration range used in our anisotropy binding assay (Figure 5b and c). It was not feasible to test higher concentrations of proteins with this assay. Therefore, we used an electrophoretic mobility shift assay (EMSA) to estimate the binding affinity of FMRP KH0 and KH1-KH2 for UCACC. The EMSA indeed showed that hFMRP KH0 and KH1-KH2 bind to UCACC, but the affinity is weaker than that of Nova-2 KH3 binding to UCACC (K_D = 3 μM) (Supplemental Figure 5). This suggests that the MIDAS method can identify even low-affinity RNA ligands.

To test whether the weak binding may be the result of having a single RRE, we synthesized RNAs having two UCACC sequences separated by 5, 10, or 15 nucleotides. The distance between the UCACC sequence was varied because we do not know what the optimal spacing is for the multiple KH domains in FMRP to bind to the RNA. Additionally, we used the GST-hFMRP KH0-KH1-KH2 construct to evaluate potentially synergistic binding when all three KH domains of hFMRP are present. Binding studies showed that hFMRP KH0-KH1-KH2 does not bind to the RNAs having the two UCACC (Figure 5d). Our results indicate that the KH domains of FMRP do not bind to simple RREs, as proposed by previous studies. Instead, the KH domains of FMRP have likely evolved to bind to a more complex RNA or RNA-protein structure that is not represented in our MIDAS RNA pool. Indeed, FMRP is known to interact with the ribosome, and the KH domains may bind to a unique structure in the ribosome [33,43,57,65–70].

DISCUSSION

FMRP has been estimated to regulate the translation of nearly 4% of the mRNAs in the human fetal brain [12]. However, identifying the RRE of FMRP in these mRNAs has remained challenging. Early RNA in vitro selection studies showed that the KH2 motif binds to kissing complex RNAs, whereas the RGG motif binds to a GQ-forming RNA [22,23]. In addition, high-throughput approaches have been used to determine the RREs of FMRP. For example, the WGGA and the ACUK motifs were shown to be enriched in FMRP mRNA targets by the PAR-CLIP method [31]. The WGGA and ACUK motifs were proposed to bind to the KH1 and KH2 domains of FMRP, respectively. In contrast to the above study, the GACR motif was identified using the KH1 and KH2 domains of FMRP by the in vitro technique called RNAcompete [35]. Further in-depth analysis of the FMRP target datasets indicated that the GACR motif is highly enriched, the WGGA motif is modestly enriched, and the ACUK motif is not enriched [36]. Moreover, the WGGA motifs are clustered together in the mRNA targets suggesting that they are enriched in the FMRP dataset because they form GQ structures [36].

In the PAR-CLIP method, 4-thiouridine is incorporated into nascent mRNA transcripts to improve the efficiency of cross-linking proteins that bind to mRNAs. However, as noted previously [36], this may bias the selection of mRNA sequences having more uridine residues, which may explain the enrichment of the ACUK motif in the FMRP target dataset. Consistent with this interpretation, our studies show that FMRP does not bind to an RNA having a single ACUU sequence. Furthermore, the PAR-CLIP method used full-length FMRP to identify mRNA targets, which may bias the selection of mRNAs having GQ structures because of the presence of the RGG domain. Consistent with this explanation, our results show that an RNA with four repeats of the WGGA motif forms a GQ structure and binds to the RGG domain but not to the KH domains of hFMRP.

Surprisingly, our studies show that FMRP does not bind to (GACG)₄, which was identified by the RNAcompete method and appears to be highly enriched in the FMRP target datasets [35,36]. The discrepancy between our data and the previous RNAcompete data may be explained by the size of the FMRP protein or the length of the RNA used in both studies. We used NT-hFMRP and several deletion constructs to determine K_D. In contrast, the RNAcompete method used a KH1-KH2 fragment, which may not have the same structure and the RNA-binding specificity as the larger protein fragment. Alternatively, we used 18 nucleotide RNA for our binding studies, whereas the RNAcompete used 30 – 41 nucleotides RNAs, which may fold into more complex structures that could bind to the KH1-KH2 fragment.

Interestingly, we show that FMRP has the highest binding affinity for PolyG₁₇ and UG₄U RNAs, which form stable GQ structures. Furthermore, the RGG domain of FMRP is sufficient for binding to PolyG₁₇ and UG₄U. These results are consistent with previous studies that showed the RGG domain of FMRP binds specifically to GQ structures [23–30,40,58]. Taken together, our results show that the KH0, KH1, and KH2 domains do not bind to the WGGA, ACUK, and GACR motifs, whereas the RGG domain binds with high affinity to RNA GQ structures. We developed the MIDAS method to identify in an unbiased manner single-stranded RREs that could bind to the KH0, KH1, and KH2 domains of hFMRP. Although the MIDAS method was able to identify RREs that bind with high affinity to the Nova-2 KH3 domain, we were only able to identify RREs that bind with weak affinity to the KH0 and KH1-KH2 domains of FMRP. The inability to identify any RREs that bind to the KH domains of hFMRP using the MIDAS method is consistent with our data showing that they do not bind to model RNAs having the previously identified RREs. One possibility for our unexpected results is that the phosphorylation of hFMRP at Ser500 is essential for the KH domains to bind to single-stranded RNAs with high affinity [71,72]. Although Ser500 is not located within the KH domains, FMRP phosphorylation may induce some structural rearrangement in the protein that increases the binding affinity of the KH domains for single-stranded RNAs [71].

A second possibility is that the KH domains of hFMRP recognize more complex RNA motifs than represented in our randomized RNA pool. Indeed, Darnell and co-workers showed that an in vitro selection experiment performed with an RNA pool having 25 randomized nucleotides did not yield any RNA binders for the hFMRP KH1-KH2 domains [22]. They obtained RNA binders only when they used an RNA pool with 52 randomized nucleotides. The in vitro selected RNA binders formed a kissing complex RNA motif, which binds to the KH2 domain of FMRP, most likely in a different mode than simple RREs binding to the canonical KH domains. These results suggest that that the KH domains of FMRP are more specialized than the canonical KH domains that recognize 4–5 nucleotides in single-stranded RNA. The evolutionary changes in the sequence of the KH domains of FMRP may inhibit their binding to simple single-stranded RNA sequences so that FMRP may bind to a unique structure such as the ribosome [33,43,57,65–70].

MATERIAL AND METHODS

Protein expression and purification

The human FMR1 isoform 1 gene was assembled from E. coli codon-optimized gene blocks (IDT). The gene encoding the full-length isoform 1 human FMRP (hFMRP) spanning residues E2-P632 was subcloned into the LIC expression vector pMCSG7 (DNASU plasmid repository) with the 5′ TEV cleavage site deleted, conferring a N-terminal hexahistidine tag. The N-terminus truncated human FMRP (NT-hFMRP) spanning residues R218-P632 was subcloned into the LIC expression vector pMCSG7, conferring a N-terminal hexahistidine tag with the 5′ TEV cleavage site deleted. The NT-hFMRP ΔKH1 construct spanning residues F281-P362 was subsequently generated from NT-hFMRP via PCR deletion of the KH1 domain. The human FMRP RGG domain (GST-hFMRP RGG) spanning residues G531-P632 was subcloned into the LIC expression vector pMCSG10 (DNASU) which confers a N-terminal hexahistidine-GST fusion tag that is cleavable using TEV protease. The mutant SGG domain (GST-hFMRP SGG) was assembled from an E. coli codon-optimized gene block where all 16 arginine residues of the RGG region were mutated to serines. The resulting SGG fragment was subcloned into the LIC expression vector pMCSG10. The human Nova-2 KH3 domain (GST-Nova-2 KH3) spanning residues K406-G492 was subcloned into the LIC expression vector pMCSG10. The human FMRP KH0 domain (GST-hFMRP KH0) spanning residues F126-R201, human FMRP KH1 domain (GST-hFMRP KH1) spanning residues A216-E280, human FMRP KH2 domain (GST-hFMRP KH2) spanning residues F281-K425, human FMRP tandem KH1-KH2 domains (GST-hFMRP KH1-KH2) spanning residues A216-K425, and human FMRP tandem KH0-KH1-KH2 domains (GST-hFMRP KH0-KH1-KH2) spanning residues F126-K425 were also subcloned into the LIC expression vector pMCSG10.

The NT-hFMRP, NT-hFMRP ΔKH1, GST-hFMRP RGG, GST-hFMRP SGG, GST-Nova-2 KH3, and the five GST-hFMRP KH construct expression plasmids were all transformed in E.coli BL21(DE3) cells (Novagen). NT-hFMRP and NT-hFMRP ΔKH1 were purified using Ni-NTA affinity chromatography (Qiagen). GST-hFMRP RGG, GST-hFMRP SGG, and the five GST-hFMRP KH constructs were purified using glutathione affinity chromatography (GE Healthcare). All proteins were further purified using Superdex 75 16/60 or Superdex 200 16/60 (GE Healthcare) gel filtration chromatography and stored in their respective running buffers. Gel filtration buffer composition varied between FMRP constructs. NT-hFMRP and NT-hFMRP ΔKH1 gel filtration running buffer contained 25 mM Tris pH 7.4, 150 mM KCl and 1 mM DTT. GST-hFMRP RGG and GST-hFMRP SGG gel filtration running buffer contained 50 mM Tris pH 7.5 and 1 mM DTT. GST-Nova-2 KH3 gel filtration running buffer contained 10 mM HEPES pH 7.5, 100 mM KCl, 5 mM MgCl₂ and 1 mM DTT. All five GST-hFMRP KH domain constructs were purified in gel filtration running buffer containing 50 mM Tris pH 7.4, 150 mM NaCl and 1 mM DTT.

Purification of RNAs

The PolyG₁₇, UG₄U, CR1, PolyC₁₇, PolyA₁₇, ACUU, UGGA, (GACG)₄, (UGGA)₄ and UGGm⁶ACU RNAs were purchased from Thermo Scientific/Dharmacon with a fluorescein tag at the 3′ end. For the NMM assay, PolyG₁₇ and UG₄U were also purchased without a fluorescein tag, and (UGGA)₄ was in vitro transcribed from annealed oligonucleotide template using T7 RNA polymerase. Each RNA was purified under standard denaturing conditions by urea polyacrylamide gel electrophoresis (PAGE) and then extracted from the gel.

The RNA in solution was purified by chloroform extractions followed by ethanol precipitation. The purified RNAs were dissolved in RNase-free water. The RNA concentration was measured by absorbance at 260 nm using a spectrophotometer.

PAGE analysis of labeled RNAs

Denaturing urea PAGE analysis of the UG₄U RNA was performed as described previously [49]. 20 pmol UG₄U was reconstituted in 50 mM Tris pH 7.8 and either 100 mM KCl, 100 mM NaCl or 100 mM LiCl. The UG₄U samples were heated at 37°C for 90 minutes, then incubated at 25°C for 5 minutes in a thermal cycler. Then formamide loading solution was added to each RNA sample to a final 8 μL volume. The samples were heated at 60°C for 5 minutes and then incubated at 25°C for 5 minutes in a thermal cycler. Finally, the samples were run on a 16% urea polyacrylamide gel at room temperature for 4 hours at 600 V.

For denaturing urea PAGE analysis of the 18-mer model RNAs, 10 pmol PolyG_17, 3 pmol PolyC₁₇, 2 pmol ACUU, 2 pmol UGGA, 6 pmol CR1, 4 pmol PolyA₁₇, 6 pmol (GACG)₄, and 10 pmol (UGGA)₄ in 20 mM Tris-HCl pH 7.45 and 100 mM KCl were run on a 15% urea polyacrylamide gel at room temperature for 2 hours at 25 W.

For native PAGE analysis of all the model RNAs, 10 pmol M1 5′ UTR, 50 pmol UG₄U, 10 pmol PolyG_17, 3 pmol PolyC₁₇, 2 pmol ACUU, 2 pmol UGGA, 6 pmol CR1, 4 pmol PolyA₁₇, 6 pmol (GACG)₄, and 10 pmol (UGGA)₄ in 20 mM Tris-HCl pH 7.45, 100 mM KCl and 5% (v/v) glycerol were run on a 15% native polyacrylamide gel containing 1X TBE. The gel was run at 4°C for 4 hours at 11 Watts.

BG4 antibody assay

Antibody BG4 scFv was purchased from Millipore Sigma. 100 nM BG4 was mixed with 5 nM fluorescein-labeled RNA by diluting BG4 into fluorescence anisotropy buffer (20 mM Tris pH 7.7, 75 mM KCl, 5 mM MgCl₂, and 100 ng/μL total tRNA from E. coli) followed by mixing with fluorescence anisotropy buffer supplemented with 5 nM fluorescein-labeled RNA directly in the well of a non-binding 96-well black flat bottom plate (Greiner) in a final 200 μL volume. The plate was incubated at room temperature for 30 minutes. Samples were excited at 470 nm and emission was measured at 520 nm with a 20 nm bandwidth; optimal signal gain was determined per read. Fluorescence anisotropy was measured on a Tecan Safire 2 plate reader in polarization mode. All experiments were performed a minimum of 2 or 3 times to determine the standard deviations.

NMM assay

N-methyl mesoporphyrin IX (NMM) was purchased from Frontier Scientific. NMM stock solution was prepared by dissolving the stock in 0.2 N HCl to 5 mg/mL (equal to 8.61 μM). Fresh working solution was prepared by diluting the NMM stock solution to 400 μM in 10% (v/v) DMSO. Various concentrations of the RNAs were reconstituted in 158 μL solution containing 20 mM Tris pH 7.7, 5 mM MgCl₂, 1 mM DTT, and either 100 mM KCl or 100 mM LiCl. The samples were then heated at 90°C for 2 minutes and slowly cooled to room temperature over 1 hour. Samples were briefly spun down and collected before adding 2 μL of 400 μM NMM to each 158 μL sample, to a final 5 μM NMM per sample. The samples were incubated at room temperature for 10 minutes before taking fluorescence measurements (JASCO FP-8500). Samples were excited at 400 nm and emission was scanned from 560 nm to 650 nm with a 5 nm excitation and emission bandwidths; automatic signal gain was used. The peak NMM fluorescence (610 nm) was plotted as a function of RNA concentration. All experiments were performed a minimum of 2 or 3 times to determine the standard deviations.

Fluorescence anisotropy RNA binding assay

The various FMRP constructs were titrated against 5 nM fluorescein-labeled RNAs. Fluorescence anisotropy was measured on a non-binding 96-well black flat bottom plate (Greiner) using a Tecan Safire 2 plate reader in polarization mode. Some of the RNA fluorescence anisotropy measurements were also performed using a Tecan Spark plate reader.

FMRP was diluted into fluorescence anisotropy buffer supplemented with DTT (20 mM Tris pH 7.7, 75 mM KCl, 5 mM MgCl₂, 100 ng/μL total tRNA from E. coli, and 1 mM DTT) and mixed with fluorescence anisotropy buffer supplemented with fluorescein-labeled RNA within the plate wells in a final 200 μL volume. The plate was incubated at room temperature for 30 minutes. Samples were excited at 470 nm and emission was measured at 520 nm with a 20 nm bandwidth; optimal signal gain was determined per read.

To quantify the binding affinity between FMRP and RNA, the anisotropy data from each binding assay was normalized to the initial value without protein, plotted and fit to the quadratic equation below to calculate the equilibrium dissociation constant (K_D) as described previously [73]:

$$\frac{\left[P+RNA\right]}{\left[RNA\right]}=\frac{\left(\left[P\right]+\left[RNA\right]+K_D\right)-\sqrt{{\left(\left[P\right]+\left[RNA\right]+K_D\right)}^2-4\left[P\right]\left[RNA\right]}}{2\left[P\right]}$$

Where [P+RNA]/[RNA] is the anisotropy value, [RNA] is the RNA concentration, and [P] is the protein concentration. GraphPad Prism software (Graphpad Software Inc.) was used to perform the curve fits. All experiments were performed a minimum of 2 or 3 times to determine the standard deviations.

MIDAS: DNA template and primers

The MIDAS N5 DNA template sequence: 5′-mAmATGCGTANNNNNTGGATCCCTATAGTGAGTCGTATTA-3′ was purchased from IDT and has two 2′-O-methyl substitutions (mA) at the 5′ terminus to reduce the amount of n+1 product during in vitro transcription. The five randomized positions were produced by hand mixing the nucleotides to give equal ratio of all four nucleotides. 18T7T sequence: 5′-TAATACGACTCACTATAG-3′. First-strand primer: 5′-GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAG AATGCGTA-3′. Second-strand primer: 5′-TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGGGGATCCA-3′. Index 1 Read primer: 5′-CAAGCAGAAGACGGCATACGAGAT [i7] GTCTCGTGGGCTCGG-3′. Index 2 Read primer: 5′-AATGATACGGCGACCACCGAGATCTACAC [i5] TCGTCGGCAGCGTC-3′.

MIDAS: Separation of RNA•protein complex using a native gel

The MIDAS RNAs were generated by in vitro transcription of annealed 18T7T and MIDAS template DNAs using T7 RNA polymerase. The sequence of the 21-mer MIDAS N5 RNA is 5′-GGGAUCCANNNNNUACGCAUU-3′. The transcription products were purified under standard denaturing conditions by urea polyacrylamide gel electrophoresis (PAGE) and then extracted from the gel. Each RNA was purified by chloroform extractions followed by ethanol precipitation. The purified RNAs were dissolved in RNase-free water. The RNA concentration was measured by absorbance at 260 nm using a spectrophotometer.

The MIDAS N5 RNA pool was 5′ dephosphorylated with alkaline phosphatase (NEB) and labeled with [γ−³²P]-ATP and T4 polynucleotide kinase (NEB). An EMSA was used to detect and isolate protein-bound MIDAS RNAs from the pool. Briefly, in a 10 μL sample, 100,000 cpm of MIDAS N5 RNA was incubated with 10 μM protein (or protein storage buffer in control lane) in RNA binding buffer (50 mM Tris acetate pH 7.7, 50 mM potassium acetate and 5 mM magnesium acetate, 1 mM DTT, 0.1 mg/mL BSA, and 100 ng/μL total tRNA from E. coli) for 45 minutes at room temperature. Glycerol was added to 5% final concentration and the samples were run on an 8% (79:1 acrylamide:bisacrylamide) 0.5X TBE native polyacrylamide gel at 4°C for 2 hours and 30 minutes at a constant 300 V. The gel was exposed to a phosphor screen at −80°C and imaged on a Typhoon FLA 9500 phosphorimager (GE) after 18 hours. The gel was thawed at room temperature and the shifted band for each protein was extracted from the gel. Each selected RNA was purified by chloroform extractions followed by ethanol precipitation. The purified RNAs were dissolved in 35 μL RNase-free water.

MIDAS: Library preparation

The selected MIDAS RNAs were used to generate a cDNA library for next generation sequencing. This was accomplished in three simple steps: (1) synthesize the first-strand cDNA by reverse transcription, (2) synthesize the second-strand by primer extension, and (3) amplify the dsDNA by PCR. First, 8 μL of the selected RNA and 2 μL of 10 μM MIDAS reverse transcription primer were combined and annealed by heating at 65°C for 5 minutes and cooling to room temperature for 2 minutes before placing on ice. The 10 μL annealed RNA-primer was mixed with 10 μL 2X cDNA synthesis mix containing 2X FS buffer (100 mM Tris pH 8.3, 150 mM KCl, and 6 mM MgCl2), 1 mM dNTPs, 10 mM DTT, and 10 U Superscript III reverse transcriptase (ThermoFisher). The reverse transcription was carried out in a thermal cycler: (1) 10 minutes at 25°C, (2) 1 hour and 30 minutes at 30°C, (3) 15 minutes at 70°C, (4) 37°C hold for RNase H digestion. On step 4, the 20 μL first-strand samples were placed on ice for 1 minute and spun down. Then 1 μL of 5 U/μL RNase H (NEB) was added and the sample was incubated at 37°C for 20 minutes.

To the 21 μL first-strand cDNA sample, we added: 3.5 μL of 100 mM NaCl, 6.4 μL of 100 mM MgCl₂, 10 μL of 10 mM dNTPs, 10 μL of 10 μM MIDAS cDNA Klenow primer, 2 μL of 5 U/μL Klenow fragment 3′ → 5′ exo⁻ (NEB) and water to a final 100 μL volume. This made an ideal 100 μL primer extension reaction containing 10 mM Tris pH 8.2, 15 mM KCl, 35 mM NaCl, 7 mM MgCl₂, and 1 mM DTT. The reaction proceeded for 30 minutes at room temperature and was then deactivated for 20 minutes at 75°C. The second-strand product was then amplified by PCR with sample-specific Index 1 and Index 2 primers containing their unique i7 and i5 index sequences.

PCR was carried out using a KOD HS polymerase system (MilliporeSigma). 50 μL PCRs were carried according to manufacturer protocol using 5 μL of the 100 μL second-strand product. The DNA was amplified over 25 cycles (annealing temperature of 60°C and a 3 second extension time). The DNA library was finally cleaned up using an Amico Ultra-0.5 mL centrifugal filter with a 30,000 Da molecular weight cut-off. The 50 μL PCR product was diluted in 450 μL water and concentrated down to approximately 50 μL a total of three times. The DNA concentration was measured by absorbance at 260 nm using a spectrophotometer. The Nova-2 KH3, hFMRP KH0, hFMRP KH1-KH2, and total input DNA libraries were mixed in equimolar ratios and then sequenced with Illumina HiSeq 4000 at the UCSD IGM Genomics Center.

MIDAS: NGS data analysis

We obtained more than four million reads for each Illumina sequencing run. To identify and rank the favourite motifs of each protein we used a custom Python script, which first filtered the data by sequence length to show only the N5 region sequences. Doing this revealed exactly 1,024 unique 5-mer sequences. Counts from each of the 1,024 sequences were compared between the protein-selected libraries and the total input library to generate a list of enrichment factors. To determine if any of the top enriched sequences had a significant number of counts coming from mutated sequences, the data were also filtered by sequence length to show the N5 region plus 3 bases upstream and 3 bases downstream. This revealed how many counts for each N5 motif came from the wild type sequence and how many came from mutated sequences. More than 99% of the reads were the correct wild type sequence and only counts coming from the wild type sequences were used to calculate the enrichment factor. Identical sequencing data were obtained when the Illumina sequencing was repeated with another set of independent library preparations showing that the process is reproducible.