In vitro selection with a site-specifically modified RNA library reveals the binding preferences of N⁶-methyladenosine (m⁶A) reader proteins

Abstract

Epitranscriptomic RNA modifications can serve as recognition elements for the recruitment of effector proteins (i.e., “readers”) to modified transcripts. While these interactions play an important role in mRNA regulation, there is a major gap in our understanding of the sequence determinants critical for binding of readers to modified sequence motifs. Here, we develop a high-throughput platform, relying upon in vitro selection with a site-specifically modified random sequence RNA library and next-generation sequencing, to profile the binding specificity of RNA modification reader proteins. We apply our approach to interrogate the effect of sequence context on the interactions of YTH-domain proteins with N⁶-methyladenosine (m⁶A)-modified RNA. We find that while the in vitro binding preferences of YTHDC1 strongly overlap with the well-characterized DR(m⁶A)CH motif, the related YTH-domain proteins YTHDF1 and YTHDF2 can bind tightly to non-canonical m⁶A-containing sequences. Our results reveal the principles underlying substrate selection by m⁶A reader proteins and provide a powerful approach for investigating protein-modified RNA interactions in an unbiased manner.

Graphical abstract

INTRODUCTION

RNA behavior in the cell is regulated by its interactions with a large complement of RNA-binding proteins¹. These proteins recognize specific RNA molecules and affect gene expression through the control of processes, including splicing, turnover, trafficking, and translation. Characterizing the molecular determinants underlying RNA-protein binding, including how RNA sequence and structure influence these interactions, is therefore an important component towards a unified understanding of gene expression regulation.

Recently, a growing number of chemical modifications on eukaryotic mRNA² have emerged as a new modality for post-transcriptional gene regulation and have been termed the “epitranscriptome.” The most abundant of these modifications, N⁶-methyladenosine (m⁶A), is found at ~10,000 sites in the human transcriptome^{3, 4}. The m⁶A modification affects mRNA stability⁵, translation^6–9, splicing¹⁰, and nuclear export^{11, 12}, and has been implicated in diverse biological processes including development^13–15, innate immunity^{16, 17}, DNA damage signaling¹⁸, and cellular proliferation¹⁹. Emerging evidence suggests that other epitranscriptomic marks such as 5-methylcytidine²⁰, N¹-methyladenosine²¹, pseudouridine^{22, 23}, N⁴-acetylcytidine²⁴, 2’-O-methylation²⁵ and N⁶,2ʹ-O-dimethyladenosine^{26, 27} can also affect mRNA behavior in the cell. New approaches are needed in order to study the molecular mechanisms underlying the interpretation of this RNA modification code.

How do post-transcriptional modifications affect mRNA properties in the cell? On the one hand, modifications may influence RNA structure by modulating inter- or intramolecular RNARNA interactions. Such a mechanism has been demonstrated for m⁶A²⁸ and has many precedents among tRNA modifications²⁹. Alternatively, modifications can also serve as a binding platform to recruit modification-specific RNA-binding proteins (or “readers”) to modified RNA transcripts³⁰. Indeed, numerous m⁶A readers and anti-readers have now been identified and functionally characterized^{3, 5, 31, 32}. Most prominent among these are the YTH-domain proteins which bind to m⁶A-modified sequences and are broadly conserved in eukaryotes³³. The human genome encodes 5 YTH-domain proteins with distinct biological functions^{5, 6, 9, 10, 34}. While we now have structural models for how these proteins bind to methylated RNA^35–37, we lack a comprehensive understanding of how m⁶A readers select their mRNA substrates, including the sequence determinants underlying these interactions.

RNA-protein interactions can be studied in cells using cross-linking and immunoprecipitation combined with high-throughput sequencing approaches (e.g., HITS-CLIP, CLIP-Seq, PAR-CLIP, CRAC)³⁸. While these methods are easily generalizable and report on native RNA-protein interactions in a highly parallel fashion, they have not been adapted to specifically interrogate RNA modification-dependent interactions. Moreover, sequence bias can be introduced through reliance upon base-specific photocrosslinking chemistry or non-canonical nucleotides, and they do not typically provide insight into interaction affinity.

An alternative approach to investigate RNA-protein interactions involves the use of the SELEX/in vitro selection strategy described by Gold³⁹ and Szostak⁴⁰. In this approach, an unbiased random sequence RNA library is subjected to affinity-based selection against a protein target of interest. Iterated cycles of in vitro transcription, selection, reverse transcription, and amplification enable the identification of tight-binding sequences and reveal the sequence-binding preferences of the target protein. Indeed, in vitro selection has been applied to query the substrate binding preferences of numerous RNA-binding proteins in an unbiased manner^{39, 41–43}. While modified bases have been incorporated into random sequence RNA libraries used for in vitro selection^{44, 45}, since these libraries are made by in vitro transcription of a randomized DNA template, modification location, and stoichiometry cannot be controlled, except in the case where one of the native NTPs is replaced entirely with a modified NTP.

In this manuscript, we interrogate the substrate preferences of three mammalian m⁶A reader proteins using in vitro selection with a random-sequence, site-specifically modified RNA library and high-throughput sequencing (Fig. 1). First, we develop conditions for the direct synthesis of random-sequence RNA containing a single, defined m⁶A site. Next, we perform affinity selection of an m⁶A-modified RNA library against readers YTHDF1, YTHDF2, and YTHDC1, and identify bound sequences using Illumina sequencing. Finally, bioinformatic analysis and sequence clustering, as well as biophysical validation of enriched k-mer motifs centered around the m⁶A residue, provides a fingerprint of the binding preferences for each m⁶A reader protein.

Figure 1.

In vitro selection strategy to reveal the sequence preferences of m⁶A reader proteins. A large pool of random, m⁶A-containing RNA sequences is interrogated with a bead-bound reader, and enriched sequences are selectively eluted and elucidated by high-throughput sequencing.

MATERIALS AND METHODS

Oligonucleotide synthesis

Solid-phase synthesis of random sequence libraries and fluorescein-labeled RNA probes was performed on an ABI 394 oligonucleotide synthesizer (Applied Biosystems) using standard conditions and commercial phosphoramidites and oligosynthesis reagents (Glen Research). For synthesis of random sequence RNA, a custom mix was created by combining each TBDMS-protected RNA phosphoramidite in the appropriate ratio. After cleavage and deprotection, RNA and RNA-DNA hybrid libraries were purified by denaturing PAGE. 5’-fluorescein-labeled RNA probes were purified by reverse-phase HPLC and validated by high-resolution ESI-MS (Table S1). The sequence for the hybrid DNA-RNA libraries 1 and 2 is as follows: 5’-AAGCTTCCCGGGCTGCAGGGATCC-NNNNNA*NNNN-GCCGCGGGAATTCTCCCT-3’, where the constant regions consist of DNA and the central region is randomized RNA with either m⁶A (library 1) or adenosine (library 2) at position 6. The sequence of RNA library 3 is 5’-NNNNNNN-m⁶A-NNNNNNNU-3’.

Reverse transcription and Sanger sequencing

Following the manufacturer’s protocol, each library (1 fmol) was reverse-transcribed in a 10 μL volume with SuperScript II reverse transcriptase (Invitrogen, 0.25 μL) using reverse primer 1 (Table S2). The incomplete first-strand cDNA was extended with the Klenow fragment of DNA polymerase I (NEB, 0.5 μL) at 37°C for 1 hour, and the original DNA-RNA template was digested with RNAse H (NEB, 0.5 μL) at 37°C for 20 minutes. Each enzyme was heat inactivated at 75°C for 15 minutes before adding the next. A small aliquot of the RT reaction (2–5 μL) was PCR-amplified under standard conditions for OneTaq DNA polymerase (NEB) using the aforementioned reverse primer 1, forward primer 2 (Table S2), a 60.5°C annealing temperature, and a 30 second elongation. The size of the amplicon was checked by gel electrophoresis on 3% agarose, and a small aliquot of the unpurified amplified mixture (5–8 μL), along with the forward and reverse primers, was submitted for Sanger sequencing (Genewiz) to assess base content across the random region. Results were analyzed using SnapGene software.

Illumina sequencing

Library 1 was processed for reverse transcription and amplification as described above, and the amplicon was purified by agarose gel (3%) electrophoresis. The amount of double-stranded template was then quantified by Quant-iT PicoGreen assay (Invitrogen) following the supplier’s directions, and 20 μL of amplicon (5 ng/μL) were submitted for Illumina sequencing. To assess percent incorporation of each base at the random position, reads were uploaded to the Galaxy workflow system (https://galaxy.princeton.edu) and processed with the FastQC tool. For post-selection sequencing of library 3, a small portion of the selection elutions (6 μL) was converted to cDNA using the NEBNext Ultra RNA Library Prep Kit for Illumina (NEB) following the supplier’s instructions. Given the small amount of material in each elution, the kit’s 3’ adapter, RT primer, and 5’ adapter were diluted 1:1 in RNAse-free water prior to use. Differently indexed primers (index 4, 5, and 6) were employed for each sample at the PCR stage. After amplification, barcoded PCR amplicons were gel purified, combined, and submitted for Illumina sequencing on a MiSeq Micro flowcell (Illumina) as paired-end 2 × 150 nt reads following the manufacturer’s protocol.

Bioinformatics

Sequences were uploaded to and demultiplexed on the Princeton HTSeq database system and transferred to the Princeton Galaxy instance, where read quality was assessed using FastQC. Adapters were trimmed using Cutadapt (Galaxy Version 1.16.4), and the 15-base randomized regions were excised using “Trim sequences” (Galaxy Version 1.0.2). Results were then downloaded and further analyzed on a SLURM compute cluster. Sequences were filtered for the methylated (or unmethylated control) adenosine in the template-designed position. Sequence case was changed to lowercase to identify the target adenosine. Positional base frequencies were calculated and overall flat transfac motif logos generated using WebLogo 3.6.0. Observed/expected ratios were calculated based on the wikiselev/bioinformatics-algorithms github page (https://github.com/wikiselev/bioinformatics-algorithms/wiki/Kmer-expected-number-of-occurrences-in-a-DNA-string). Case-sensitive k-mers were counted and sorted by descending abundance. They were then greedily clustered by 70% identity (without replacement). Logos were generated per cluster using ceqlogo from MEME suite 4.10.0_1. All steps, unless otherwise noted were performed using custom in-house perl scripts and executed on the cluster in batch per k-mer size.

Protein expression and purification

Plasmids encoding cDNA for YTH proteins were obtained from Addgene: YTHDC1 (NP_001026902.1) (#85167)⁴⁶, YTHDF1 (NP_060268.2) (# 70087)⁵, and YTHDF2 (NP_057342.2) (# 52300)⁵. The YTH domains of these proteins were cloned into pGEX-6P-1 (YTHDF2) or a pET28a vector (YTHDC1 and DF1) for protein expression in Escherichia coli. All sequence-verified constructs were transformed into E. coli strain BL21 (Rosetta). His₆-YTHDC1 (residues 345–509) and His₆-YTHDF1 (residues 389–523) were expressed overnight in at 18°C with 0.2 mM isopropyl-β-D-thiogalactopyranoside (IPTG). Cells were lysed by sonication in lysis buffer (20 mM HEPES pH 7.4, 200 mM NaCl, 5 mM β-mercaptoethanol, 5 mM imidazole, 0.5% Triton X-100, supplemented with 1 mM PMSF, EDTA-free protease inhibitor tablet (Roche), and benzonase) and purified by affinity chromatography with Ni-NTA resin (ThermoFisher) according to the manufacturer’s recommendations. GST-YTHDF2 (residues 383–553) was overexpressed overnight in at 18°C with 0.2 mM IPTG. Lysis by sonication was performed in buffer containing 1X TBS, 150 mM NaCl, 5 mM EDTA, 1 mM DTT, 0.2 mg/mL lysozyme, and 1% Triton X-100, supplemented with 1 mM PMSF, protease inhibitor tablet, and benzonase. The lysate was purified using Pierce glutathione agarose resin (ThermoFisher) following manufacturer’s instructions. Following affinity purification, all proteins were fractionated on a HiLoad 16/600 Superdex 200 pg preparative size exclusion column (GE Healthcare) with buffer containing 20 mM HEPES pH 7.4, 220 mM NaCl, 1 mM DTT. The most concentrated fractions were combined and further concentrated to 10–12 mg/mL.

In vitro selection protocol

For selections with libraries 1 and 2, His₆-tagged YTH domains (50 pmol) were immobilized to pre-equilibrated magnetic His-Tag Dynabeads (Invitrogen) in binding buffer (200 μL) containing 50 mM sodium phosphate pH 8.0, 300 mM NaCl, and 0.01% Tween 20 by incubating overnight at 4°C with end-to-end rotation (experimental binding capacity: YTHDC1 = 4 μg protein/μL bead slurry, YTHDF1 = 3.5 μg protein/μL bead slurry). Bead-bound proteins were washed with washing buffer (200 μL, 50 mM sodium phosphate pH 8.0, 300 mM NaCl, 0.05% Tween 20) three times, the last wash being supplemented with 100 μg/mL salmon sperm DNA as blocking agent. The RNA library (10 pmol) was then applied to the washed beads in selection buffer (100 μL, 50 mM sodium phosphate pH 8.0, 300 mM NaCl, 0.05% Tween 20, 100 μg/mL salmon sperm DNA), and the mix was incubated at room temperature for 1 hour with shaking. The unbound flow-through was discarded, and the beads were washed three times with washing buffer, using the last wash to transfer the beads to a clean tube. Protein-library adducts were eluted with elution buffer (50 μL, 50 mM sodium phosphate pH 8.0, 300 mM NaCl, 300 mM imidazole) at room temperature for 10 minutes. The elution was isolated and desalted using an Illustra MicroSpin G-25 column (GE Healthcare) following the manufacturer’s protocol. Selections with the GST-tagged YTHDF2 YTH domain were performed on Pierce glutathione magnetic agarose beads (Invitrogen, experimental binding capacity = 16 μg protein/μL settled bead) following the above protocol with the following buffers: binding buffer (1X TBS pH 7.4, 3 mM DTT, 0.01% Tween 20), washing buffer (1X TBS pH 7.4, 1 mM DTT, 0.05% Tween 20), selection buffer (1X TBS pH 7.4, 1 mM DTT, 0.05% Tween 20, 100 μg/mL salmon sperm DNA), and elution buffer balanced at pH 8 (1X TBS pH 7.4, 1 mM DTT, 50 mM reduced glutathione). The selection of library 3 against YTH proteins was carried out using the aforementioned selection protocol, but the library and protein-bead amounts were increased 10-fold. Washing, application, and elution volumes were kept the same. Prior to selection, library 3 was 5’-phosphorylated using T4 polynucleotide kinase (NEB) following the manufacturer’s recommendations.

Quantitative reverse-transcription PCR

All measurements were carried out in triplicate on a Viia 7 Real-Time PCR System (Applied Biosystems) using a MicroAmp Fast Optical 96-well plate (Applied Biosystems). Following selection, a small portion of the elutions containing the enriched sequences (1 μL) was converted to first-strand DNA with SuperScript II as indicated above (see Sanger sequencing) using primer 1 as the reverse primer (Table S2). The qPCR reactions were prepared using PowerUP SYBR Green Master Mix (Invitrogen) according to the manufacturer’s recommendations, including primer 2 as the forward primer (Table S2). Recoveries from each elution sample were determined by fitting the experimental C_t value to the pertinent standard curve (Figure S3).

Binding assays

All MST experiments were conducted in triplicate at 25°C on a Monolith NT.115 instrument (Nanotemper) using standard Monolith NT.115 capillaries. The following parameters were employed in Expert Mode: blue laser excitation, 60% excitation power, 30 seconds MST on, 5 seconds MST off. The buffer for the experiment and in which the 2X probe stocks and the 2X protein stocks were prepared was 20 mM Tris HCl pH 7.5, 150 mM NaCl, 5 mM MgCl₂, and 0.05% Tween 20. GGACU control and GGm⁶ACU control probes (50 nM working concentration) were titrated with decreasing concentrations of the purified YTH domains of YTHDC1 and YTHDF2 (12-point titration in 2-fold dilutions) starting with an initial protein concentration of 50 μM YTHDC1 or YTHDF2. Probes 1/2 (30 nM) and 3/4 (50 nM) were titrated against the YTH domain of YTHDC1 (16-point titration in 2-fold dilutions, 100 μM initial protein concentration for 1 and 3, 176 μM initial concentration for 2 and 4). Probes 5/6 (50 nM) were titrated against the YTH domain of YTHDF1 (16-point titration in 2-fold dilutions, 140 μM initial concentration for probe 5, 192 μM initial concentration for probe 6). Probes 7/8 (50 nM) were titrated against the YTH domain of YTHDF2 (16-point titration in 2-fold dilutions, 100 μM initial concentration for both probes). Mixtures were incubated for 15 minutes at room temperature before loading into the capillaries and MST measurements taken. Data were recorded in the MO.Control software (Nanotemper) and analyzed in the MO.Affinity Analysis software (Nanotemper) using TJump analysis. MST Values were normalized, plotted against protein concentration, and fit to a four-parameter dose-response equation (Hill model) to determine the dissociation constant (K_d). Graphs in the main text and SI were generated with GraphPad Prism.

RESULTS

Synthesis of RNA libraries

In order to profile the binding preferences of m⁶A reader proteins using in vitro selection (Fig. 1), we needed to generate a library of m⁶A-containing RNA sequences. Since RNA-binding proteins have been shown to recognize short sequence motifs¹ and library complexity increases rapidly with sequence length, we designed a short 10-mer oligonucleotide with a single m⁶A residue at position 6, resulting in a pool of 4⁹ (~260,000) unique, m⁶A-containing sequences (library 1, Fig. 2A). Additionally, we flanked the random RNA region with defined DNA primer binding sites to facilitate reverse transcription, PCR amplification, and sequencing analysis after the in vitro binding selection. While random sequence RNA libraries containing modified and unmodified nucleotides can be prepared by in vitro transcription of the corresponding random sequence DNA^{44, 45}, enzymatic synthesis cannot install modified bases site-specifically. Therefore, we prepared the library directly by solid-phase oligonucleotide synthesis using an m⁶A phosphoramidite, a custom mix of A, C, G, and U designed to generate equal proportions of each unmodified base at the randomized positions, and deoxynucleotide phosphoramidites for the flanking DNA primer binding sites.

Figure 2.

Design and characterization of random sequence libraries used in this study. (A) Structure of libraries 1 and 2. The libraries consist of a random RNA region with a central m⁶A (1) or adenosine (2) residue flanked by constant DNA sequences as primer binding sites. (B-C) Relative abundance of each ribonucleobase across the random regions of library 1 and 2. The purified libraries were reverse-transcribed, PCR-amplified, and analyzed by Sanger sequencing (see Fig. S2 for gel characterization). (D) Percent incorporation of each base across the random sequence region in library 1. The library was prepared as in panel B and subjected to by high-throughput sequencing (Illumina) to quantify base content at the random region.

First, we developed reaction conditions for direct chemical synthesis of random sequence RNA. We started by synthesizing a test library using a custom mix of TBDMS-protected RNA phosphoramidites (A:C:G:U = 0.26:0.25:0.29:0.20) based upon conditions reported by Bartel and co-workers for random sequence DNA synthesis⁴⁷. After synthesis and purification, we subjected the library to reverse transcription, PCR amplification, and Sanger sequencing to measure the relative abundance of each base at the randomized positions. Interestingly, our initial library contained over-incorporation of G at the expense of C (Fig. S1), suggesting that relative RNA phosphoramidite reactivity differs from that of the corresponding DNA monomers. Therefore, we prepared several additional libraries using phosphoramidite mixes with varying compositions (decreasing the concentration of G monomer while increasing C monomer concentration). Sanger sequencing analysis of these libraries enabled the selection of optimal reaction conditions (A:C:G:U = 0.28:0.35:0.15:0.22) that produced comparable amounts of the 4 bases across the random sequence region (Fig. S1). This custom mix was then used to synthesize library 1, containing the RNA motif NNNNN-m⁶A-NNNN (Fig. 2B, Fig. S2) and the corresponding unmodified library (2, Fig. 2C, Fig. S2) containing the motif NNNNN-A-NNNN. We subjected library 1 to high-throughput sequencing (Illumina) and observed comparable distribution of all bases at the random sites (between 22% and 27%) and exclusive presence of adenosine signal (m⁶A is read as A during reverse transcription) at the modified position, further validating our synthetic protocol (Fig. 2D).

Validation of selection platform with YTH-domain proteins

Next, we evaluated the ability of m⁶A reader proteins to bind and enrich bona fide substrate sequences by affinity selection (Fig. 3A). For this purpose, we chose YTH-domain proteins YTHDC1, YTHDF1, and YTHDF2, which have all been characterized as m⁶A readers, and tested binding of our m⁶A-modified and unmodified random sequence libraries. Briefly, library 1 or 2 was incubated with an excess of bead-immobilized YTH domain and bound sequences were eluted after stringent washing. Quantitative reverse-transcription PCR (RT-qPCR), using appropriate standard curves generated from library 1 and 2 (Fig. S3), was then used to measure the amount of library present in the elution. Analysis of selections against all 3 YTH-domain proteins indicated that only a minor fraction, ranging from 10⁻⁷ to 10⁻⁴, of the input library interacted with the bead-bound protein (Fig. 3B). Notably, for all 3 m⁶A readers that we tested, we observed ~100-fold increased recovery of the m⁶A-modified library compared to the unmodified library (Fig. 3B), indicating that our in vitro selection conditions can reliably distinguish sequences based upon their affinity, and validating the preference of YTH-domain proteins for m⁶A-modified RNA substrates.

Figure 3.

Recovery of sequences in libraries 1 and 2 after affinity selection with YTH-domain reader proteins. (A) Workflow for qPCR validation of selection. Libraries 1 and 2 are probed for binding to immobilized YTH readers, bound targets are enriched, and recovery is quantified by qPCR. (B) Recovery of libraries 1 and 2 upon binding to YTH-domain proteins as determined by qPCR (see Fig. S3 for qPCR standard curves). Libraries were incubated with excess immobilized YTH reader, and after strenuous washing bound sequences were selectively eluted. The eluted targets were then reverse-transcribed, and the resulting first-strand cDNA was amplified by qPCR. Amount of material in each elution was extrapolated from the experimental threshold cycles (C_t) and the respective qPCR standard curve. Values represent mean +/− s.d. (n=3).

High-throughput sequencing analysis of YTH-domain protein selections

Having validated our proposed selection strategy, we sought to elucidate the identities of each m⁶A-reader’s preferred sequences through high-throughput sequencing. To avoid any potential bias that could be introduced by the DNA priming regions of libraries 1 and 2 (Fig. 2A), we synthesized a new 15-mer RNA-only library, 3, consisting of a single m⁶A nucleotide surrounded on both sides by seven fully randomized positions (Fig. S4). We then performed in vitro selection against YTHDC1, YTHDF1, and YTHDF2, and prepared bound library sequences for high-throughput sequencing by adaptor ligation, reverse transcription, and PCR amplification. Illumina-based sequencing was then performed on these samples yielding between 1–2 million sequence reads per selection.

In order to analyze selection results, we developed a custom script to extract the abundance of k-mers centered around the m⁶A residue. Based on our sequencing coverage and the assumption that enrichments after 1 round of selection are likely to be modest, we considered k-mers up to length 11 (~10⁶ possible sequences). We first chose to focus on the 10 most abundant 5-mer motifs from each selection (Table 1), as the consensus sequence motif for m⁶A modification sites as mapped by antibody-based sequencing is of the form DR(m⁶A)CH (D = A/G/U; R = A/G; H = A/C/U)^{3, 4, 48}. Together, these top 5-mer motifs comprise ~8–10% of all sequence reads for their respective selection. Consistent with the structural homology between YTHDC1, YTHDF1, and YTHDF2³³, we observed considerable overlap in their enriched 5-mer motifs, with 4 out of the top 10 motifs in each selection (GGACG, GGACA, GUAGA, and GGACU) (Table 1, yellow) shared between all three proteins, and 8/10 top 5-mer motifs shared between YTHDF1 and YTHDF2 (which share 90% sequence identity in their YTH domains, as compared to only ~30% sequence identity with YTHDC1). Among the top 5-mer motifs enriched in all three selections, we found GG(m⁶A)CU, the most abundant m⁶A-containing 5-mer in the mammalian transcriptome and a validated substrate of YTH domain proteins^{3, 4, 35, 48}, suggesting that our in vitro selection approach can effectively capture known binding sequences. We also found GG(m⁶A)CA⁴⁸, another abundant DR(m⁶A)CH-matching motif (Table 1), as well as GG(m⁶A)CG, which has been reported as the most abundant non-DR(m⁶A)CH m⁶A-containing pentamer sequence in mammalian cells⁴⁸.

Table 1:

The 10 most abundant m⁶A-centered 5-mers and 11-mers enriched upon in vitro binding selections with YTH readers. 5-mers and 11-mers shared between selections are highlighted in yellow and purple, respectively.

Position	YTHDC1	YTHDF1	YTHDF2
5-mers
1	GGACG (1.35)a	CUAGA(1.30)	GUAGA (1.15)
2	GGACA (0.98)	GUAGA (1.23)	GGACG (0.94)
3	AGACG (0.93)	GGACG (1.08)	CUAGA(0.91)
4	GUAGA (0.88)	CGAUC(1.03)	GGACA (0.87)
5	GGACU (0.85)	CUAGU (0.94)	GUAGG (0.82)
6	GAACG (0.82)	GGACA (0.91)	CGAUC (0.74)
7	AGACU (0.72)	GCAGA(0.91)	GCAGA (0.73)
8	GUAGG (0.71)	GGACU (0.85)	GGACU (0.71)
9	GAACU (0.59)	GGAGA (0.82)	GGAGA (0.70)
10	GAACA (0.59)	CGACU (0.79)	AGACG (0.67)
11-mers
1	AAAGGACGUGG (0.0050)	CGGCUAUAGAA (0.0080)	GUGCUAUAGAA (0.0062)
2	UUUGGACGUGG (0.0039)	CGGCUAGAAUA (0.0079)	GUGCUAUAGAU (0.0046)
3	AAAGGACGUGA (0.0037)	GGCCGAUCUGA (0.0078)	CGGCUAGAAUA (0.0039)
4	GAUCUACUGAA (0.0036)	CGGCGAUCUUU (0.0076)	CGGCUAGAACA (0.0037)
5	AAUGGACGUGA (0.0035)	CGGCUAUUUGA (0.0075)	CGGCUAUGAAA (0.0036)
6	UGACGAUCUGA (0.0034)	CGGCUAGAAUU (0.0071)	GGCCGAUCUGA (0.035)
7	UGAAGACGUGG (0.0032)	CGGCUAUUGAA (0.0069)	CGGCUAGAAUU (0.0035)
8	AAAAGACUGGG (0.0032)	CGGCUAUGAAU (0.0068)	GGCGUAGAAAA (0.0035)
9	GGGCAAAAGAG (0.0032)	CGGCUAGAGAA (0.0067)	CGCCGAUCUGA (0.0035)
10	AAAAUAAAGGG (0.0031)	CGACGAUCUGA (0.0066)	CGGCUAGUAGA (0.003)

^apercent abundance for each sequence (sequence counts/total sequence counts for selection) is shown in parenthesis.

The enriched 5-mer motifs in our library selection against YTHDC1 show clear overlap with the canonical m⁶A-containing sites that have been identified in transcriptomic modification-sequencing studies. There is a strong preference for C after the m⁶A residue (8/10 of the top 10 motifs possess this feature) (Table 1) and clear selection of purines at the n-1 and n-2 positions, consistent with reported m⁶A-modifications sites^{3, 4, 48}, but seemingly less selection at the n+2 position. In contrast, our selections with YTHDF1 and YTHDF2 appear to enrich a greater variety of motifs including several non-canonical sequences (e.g., CUAGA, CUAGU, Table 1) bearing little resemblance to the DR(m⁶A)CH consensus motif.

In order to examine the wider sequence context of enriched m⁶A-containing motifs, we next analyzed 11-mers centered around m⁶A (Table 1). Among the top 10 most abundant 11-mers enriched in the YTHDC1, YTHDF1 and YTHDF2 selections, we observed sequence abundances ranging from 0.003–0.008% of all sequence counts, indicating enrichment factors of ~30–80-fold, assuming equal abundance of all sequences in the starting pool. Again, we saw enrichment of similar motifs in the selections with YTHDF1 and YTHDF2 (Table 1, 3/10 shared 11-mer motifs in purple), further demonstrating the biochemical similarities between these two proteins. Interestingly, none of the enriched 11-mers in these two selections contained the DR(m⁶A)CH sequence around the m⁶A residue, suggesting that this motif is not required for recognition by YTHDF1/2 proteins. In contrast, in the YTHDC1 selection, we found strong enrichment for 11-mers containing a central 5-mer GG(m⁶A)CG sequence and the DR(m⁶A)CH-matching AG(m⁶A)CU motif (Table 1). As with its enriched 5-mers, YTHDC1 maintains a preference for C at the +1 position (Table 1; 6/10 sequences). Taken together, the similarities in the YTHDF1/2 selection results compared with YTHDC1 support the notion that protein sequence (and presumably structural) similarity underlies binding preference for distinct m⁶A-containing sequences.

In vitro binding analysis of YTH-domain proteins with selection motifs

To validate and further expand upon our selection results, we characterized the affinity of enriched 11-mer sequences against their relevant protein targets. Rather than picking individual k-mers to evaluate, we performed clustering of k-mers based on sequence similarity to generate sequence logos representing the predominant 11-mer motif present in a particular family of related sequences (Fig. 4A, 4B, and 4C). Sequence logos were then ranked by the number of total sequence counts contained within. Next, we synthesized 4 m⁶A-containing 11-mer sequences (probes 1, 3, 5, and 7) encompassing highly enriched sequence logos from all 3 selections, as well as the corresponding unmethylated control sequences (probes 2, 4, 6, and 8) (Fig. 4A, 4B, and 4C; Table S1) and measured binding using microscale thermophoresis (MST)^{49, 50}. Gratifyingly, methylated probe 1, which represents the top sequence logo (Fig. 4A) and the most strongly enriched 11-mer motif from selection against YTHDC1 (Table 1), bound to the protein with a dissociation constant of 0.39 +/− 0.071 μM (Fig. 4D). Probe 2, which contains the same sequence but lacks methylation, bound with ~40-fold lower affinity (K_d = 16.9 +/− 0.8 μM) (Fig. 4D), demonstrating the importance of the methyl group for the interaction. Similarly, probe 3, which was designed based on the second most abundant sequence logo, bound to YTHDC1 in an m⁶A-dependent manner, exhibiting ~40-fold lower K_d than the corresponding unmethylated sequence 4 (Fig. 4E). As a comparison, we also measured binding between YTHDC1 and 10-mer oligonucleotides containing the methylated or unmethylated GGACU motif and observed similar affinity and m⁶A selectivity (Fig. S5A, Table S1) as to our selection sequences.

Figure 4:

Enriched families of 11-mer sequence motifs identified in YTH-domain selections. (A-C) The 5 most abundant clustered 11-mer logos elucidated from the selection of library 3 with YTH-domain m⁶A readers. m⁶A is centered at position 6 of the logo, and U is represented as T. Values above the logos represent the percentage of sequences in each selection encompassed by the logo. Sequence logos from which probes were synthesized are boxed. Briefly, library 3 was selected against excess immobilized YTH reader as in Figure 3B, and the elution was reverse transcribed and PCR-amplified with barcoded primers before being subjected to high-throughput sequencing. (D-G) Binding of methylated (m⁶A) and unmethylated (adenosine) probes derived from the logos (in A-C) to YTH-domain reader proteins. Binding assays were performed by microscale thermophoresis (MST) with 30 nM (1 and 2) or 50 nM (3-8) probe and increasing concentrations of protein (see Table S1 for probe sequences). Values represent mean +/− s.d. (n=3). K_d values were calculated by fitting the data points to a four-parameter dose-response curve.

We next measured the affinity of YTHDF1 and YTHDF2 for methylated sequences identified in their selections. For this purpose, we chose the third most abundant logo from the YTHDF1 selection (Fig. 4B) (used for probes 5/6) and the most abundant logo from the YTHDF2 selection (Fig. 4C) (used for probe 7/8), which differs by only 1 nucleotide from the most abundant logo selected by YTHDF1 (Fig. 4B). While these sequences do not resemble the DR(m⁶A)CH motif, both methylated probes bound tightly (YTHDF1: probe 5, K_d = 0.51 +/− 0.045 μM; YTHDF2: probe 7, K_d = 0.79 +/− 0.018 μM) and in an m⁶A-dependent fashion to their cognate protein, exhibiting 25–40-fold selectivity for methylated over unmethylated sequences (Fig. 4F and 4G). Indeed, we found that these non-canonical m⁶A-containing sequences bound to YTHDF1/2 with 2–3-fold higher affinity and greater m⁶A specificity than a GG(m⁶A)CU-containing 10-mer oligonucleotide (Fig. S5B).

Finally, we asked whether sequences exhibiting low enrichment in a selection would bind poorly towards that protein target. For this purpose, we tested methylated probes 1 and 3 (identified in the YTHDC1 selection) against YTHDF1/2, and methylated probe 7 (identified in the YTHDF1 selection) against YTHDC1; these sequences did not exhibit strong enrichment in all 3 protein selections. Consistent with our selection results, we indeed observed little binding of these non-selected sequences, characterized by either irregular dose-response traces, non-saturating binding curves, or weak thermophoretic changes even at high protein concentrations (Fig. S6).

Taken together, our results demonstrate that a single round of in vitro selection combined with high-throughput sequencing can identify tight-binding methylation-specific RNA-protein interactions from a random sequence site-specifically m⁶A-modified library and generate substrate binding profiles for distinct m⁶A reader proteins.

DISCUSSION

In this manuscript, we develop a strategy based on in vitro selection and high-throughput sequencing to profile the sequence-binding preferences of RNA modification reader proteins. We apply our approach using a site-specifically m⁶A-modified random sequence RNA library to investigate the effect of sequence context on the binding of YTH-domain reader proteins to m⁶A-containing RNA. Our results reveal distinct m⁶A-binding preferences among different families of YTH-domain proteins and provide a general strategy for characterizing modification-dependent RNA-protein interactions.

Interactions between m⁶A-modified mRNA and its corresponding reader proteins, most prominently the YTH-domain proteins, play an important role in the biological function of this epitranscriptomic modification. In mammals, the 5 YTH-domain proteins regulate distinct aspects of the mRNA lifecycle, including splicing¹⁰, nuclear export^{11, 12}, translation^6–9, and degradation⁵. How do these proteins find their relevant RNA substrates in the cell? In our study, we aimed to investigate in an unbiased fashion the sequence determinants underlying recognition of m⁶A residues by YTHDF1, YTHDF2, and YTHDC1, established m⁶A reader proteins. Our work reveals several insights into these modification-dependent interactions. YTHDC1 shows a preference for binding canonical DR(m⁶A)CH-like m⁶A-containing sequences, with a strong preference for G(m⁶A)C-containing sequences, consistent with prior biochemical and structural studies of this protein^{35, 37}. Interestingly, a preference for G at the +2 position is also strongly supported by our data. In contrast, YTHDF1 and YTHDF2, which exhibit very similar binding profiles to one another, do not show strong selection for DR(m⁶A)CH-like m⁶A sequences. Instead, our selection data identified 11-mer motifs containing pyrimidine bases at the −1 and −2 positions, and lacking C at the +1 position, which we demonstrated bind tighter to YTHDF1/2 than GG(m⁶A)CU-containing oligos of similar length. Since these proteins are known to bind canonical DR(m⁶A)CH sequences as well^{5, 6}, our data suggest that YTHDF1/2 can recognize a more diverse collection of m⁶A-modified RNA sequences than YTHDC1, and demonstrate the importance of sequences flanking the m⁶A residue for recognition. Interestingly, the strong similarity in sequence binding preference between YTHDF1 and YTHDF2 raises the question of whether these proteins compete for the same RNA binding sites in cells; since these two proteins have different effects on mRNA behavior, additional mechanisms regulating RNA-protein binding over space and time could function to ensure proper recruitment to m⁶A-modified RNAs.

Our method surveys the interactions of m⁶A reader proteins with all possible singly-modified m⁶A-containing RNA sequences. Of course, only a subset of these sequences may exist in vivo (presumably determined by the substrate preferences of m⁶A writer and eraser enzymes), and therefore our in vitro binding results must be interpreted in the context of validated m⁶A-modified sequences. Nevertheless, the finding that YTHDF1/2 can bind tightly to non-canonical m⁶A-containing motifs suggests that related m⁶A sequences may exist in the transcriptome. Indeed, single-nucleotide m⁶A sequencing approaches have revealed the presence of such sequences and have indicated that the m⁶A-modified transcriptome is more diverse than previously appreciated^{48, 51}.

Finally, we envision that site-specifically modified random RNA libraries can be applied to probe the effect of diverse epitranscriptomic marks and sequence context on fundamental nucleic acid-related processes including protein-RNA binding, catalysis by modification writer and eraser enzymes, and templated polymerization. Combined with nucleic acid indexing and massively parallel sequencing strategies, different library chemistries and experimental conditions can be interrogated in a single experiment. Such efforts are currently underway in our laboratory.

CONCLUSION

Herein, we develop an in vitro selection approach to interrogate modification-dependent RNA-protein interactions with a site-specifically modified random sequence RNA library. We apply our strategy to characterize the effects of sequence context on the binding of YTH-domain proteins, established m⁶A reader proteins, to m⁶A-modified RNA. Our results demonstrate that YTHDC1 and YTHDF1/2 possess distinct sequence-binding preferences, suggesting a mechanism for their recruitment to different m⁶A-modified mRNA substrates in the cell. Taken together, our study provides insight into m⁶A-dependent protein-RNA interactions and provides a general and unbiased approach for investigating the effect of RNA modifications on diverse biochemical processes.