Main

Previous work suggested that MSL2 may tether the MSL-DCC to DNA and that an intact CXC domain is required for X-chromosome discrimination5,6. To assess the DNA-binding specificity intrinsic to MSL2 comprehensively, we surveyed the Drosophila genome for MSL2-binding sites in vitro by DNA immunoprecipitation (DIP)7,8. Recombinant MSL2 was incubated with sheared genomic DNA (gDNA) purified from male Drosophila S2 cells. MSL2-bound DNA was recovered and sequenced.

Considering the lack of X-chromosome binding selectivity seen in previous in vitro studies, we did not expect to find that MSL2 preferentially retrieved DNA from distinct genomic loci, with a notable enrichment of sequences from the X chromosome (Fig. 1a). On the X chromosome, the MSL2 binding pattern was remarkably similar to the in vivo pattern that marks the positions of high-affinity binding sites (HAS; or chromatin entry sites) of the MSL-DCC (Fig. 1b). A total of 57 DIP sites coincided with in vivo HAS, although they show different signal intensities (Extended Data Fig. 1a, b). The results were similar if DIP followed by sequencing (DIP–seq) was performed with gDNA extracted from female cells or synthesized in vitro by whole-genome amplification (excluding the contribution of male-specific RNA contaminants or DNA modifications) (Extended Data Fig. 1c, d). It therefore appears that recombinant MSL2 has an intrinsic ability to enrich X-chromosomal sequences from complex genomic DNA.

Figure 1: Genome-wide MSL2 in vitro binding partially recapitulates the in vivo pattern.
figure 1

a, Chromosomal distribution of robust in vivo and in vitro MSL2 binding peaks, each determined by two independent experiments. The DIP–seq profile of heat shock factor (HSF)8 and the relative size of the chromosomes (genome) serve as references for uniform distribution. b, Representative profiles of MSL2 chromatin immunoprecipitation with sequencing (ChIP–seq) and DIP–seq experiments in a 1.3-Mb window on the X chromosome. Red and blue bars indicate the positions of robust peaks. Gene models are depicted in grey at the bottom.

PowerPoint slide

Full size image

We next assessed the contribution of the three known MSL2 domains to DNA binding (Fig. 2a and Extended Data Fig. 2a). Deletion of the RING finger domain that mediates MSL2 interaction with MSL1 (ref. 9) and contains E3 ligase activity10 had no obvious effect (Fig. 2b, c). Unexpectedly, however, deletion of a region rich in proline and basic amino acid residues (the Pro/Bas domain) that may bind RNA11 resulted in the complete loss of DNA binding (Fig. 2b).

Figure 2: The CXC domain of MSL2 increases X-chromosomal specificity.
figure 2

a, Linear representation of MSL2 domain organization and mutant proteins assayed in DIP. Point mutants in the CXC domain included a single (R543A) and double (R526A/R543A) mutant version. b, Representative DIP–seq profiles of wild-type and mutated MSL2 proteins in a region around the hiw gene. Red bars indicate HAS. c, Clustered heat map of DIP signals in all MSL2 in vitro peak regions. The red bar indicates a group of sites that show a prominent loss in signal upon CXC mutation. For some proteins, two independent replicates are shown. d, X-chromosomal enrichment over autosomes of robust wild-type and mutant MSL2 DIP–seq peaks. CXC indicates the peaks that significantly (false discovery rate (FDR) < 0.05) lose binding upon CXC depletion or mutation. The x-axis labels indicate the total number of peaks for each target in brackets.

PowerPoint slide

Full size image

Upon deletion or mutation of the CXC domain, binding to a subset of sites was much reduced (Fig. 2b, c). Statistical analyses revealed 56 regions that specifically required a functional CXC domain for binding. Notably, these ‘CXC-dependent’ sites displayed a higher enrichment on the X chromosome (Fig. 2d and Extended Data Fig. 2b). A total of 37 sites mapped to MSL2 in vivo peaks (HAS) on the X chromosome, and 2 sites corresponded to rare cases of autosomal sites that show MSL2 enrichment in vivo (Extended Data Fig. 1e, Extended Data Table 1 and Supplementary Table 1). Our data suggest that MSL2 interacts with DNA via two domains, CXC and Pro/Bas, and that the CXC domain is the major determinant of the selectivity for the X chromosome. While binding-site specificity can be achieved by cooperation between different transcription factors12, our finding suggests that cooperation between two different DNA-binding surfaces within this one protein may also refine its overall binding specificity.

Sequence analyses within CXC-dependent and CXC-independent binding sites for MSL2 yielded two distinct motifs. Whereas the CXC-independent binding sites shared low-complexity GA repeats (Extended Data Fig. 3a), the CXC-dependent peaks centre around a more complex variation of the MSL response element (MRE), with a notable 5′ extension (Figs 3a, 4c and Extended Data Fig. 3c). Remarkably, this novel motif can predict in vivo MSL2 binding (HAS) better than the MRE, as its position weight matrix (PWM) is superior in classifying whether MRE hit regions overlap HAS (Fig. 3b and Extended Data Fig. 3b, d). Applying low thresholds (q ≤ 0.2) we found 2,667 instances of this motif throughout the genome (Supplementary Table 1), with an approximately twofold enrichment on the X chromosome. Higher-scoring matches to the consensus sequence tend to be more strongly enriched on the X chromosome. For example, the 34 best matches are 9.8-fold enriched on the X chromosome (Extended Data Fig. 3e, f). However, 18 of those instances were not bound in vitro by MSL2 in a CXC-dependent manner, indicating that the recognition sequence represented by a PWM cannot fully explain this binding mode.

Figure 3: The CXC domain reads out nucleotide sequence and additional features.
figure 3

a, Motif discovered by the MEME motif-discovery tool in CXC-dependent binding regions (E-value = 3.9 × 10−158). b, Receiver operating characteristic (ROC) curves representing the performance of MRE and the new motif PWMs in predicting whether genomic MRE instances (35,659) overlap with a HAS (266). Areas under the curves (AUCs) are provided in brackets. c, A list of oligonucleotides used in DIP experiments. Nucleotides highlighted in red are mutations introduced based on the predictions of our classification model. d, DIP experiments using synthetic DNA representing wild-type or mutated binding sites in the genomic context. Results from qPCR amplification were normalized for their input and shown as enrichment over an unbound fragment. Data are mean ± s.e.m. for 3 biological replicates. e, Individual feature importance evaluated on sets of CXC-dependent motif instances defined by increasing score thresholds. For each feature a ROC analysis on CXC-dependent binding was performed. The AUCs of all features were scaled from 0 to 100 at each threshold level. Only features which ranked at least twice among the top five are reported. Numbers of instances (CXC-bound, non-CXC-bound) are provided underneath the x-axis tick labels. f, DIP experiments using the wild-type CG1492 sequence and a mutant in which the DNA roll at position +1 was reduced by mutating positions −1 and +2. Results from qPCR amplification were normalized for their input and shown as enrichment over an unbound fragment. Data are mean ± s.e.m. for 4 biological replicates.

PowerPoint slide

Full size image
Figure 4: The CXC-dependent sites are pioneer HAS.
figure 4

a, Representative profiles of MSL2 ChIP–seq from S2 cells treated with RNAi against GST (control) or mle. Red bars indicate binding sites that are maintained in the absence of MLE. b, MSL2 signal on 37 HAS matching CXC-dependent in vitro binding sites (PionX) or 272 non-matching ones (non-PionX) during SXL knockdown in Kc cells. Signals were averaged across 4 biological replicates and normalized to the mean signal at time point 0. Curves depict mean and s.e.m. across all sites within one class. c, Comparison of motifs found in MSL2-bound regions using different experimental approaches. See main text for details. Shown are the top scoring motifs except for ‘all (in vitro)’ which places second after a low-complexity GA-repeat similar to Extended Data Fig. 3a. d, Schematic representation of the 10-bp deletion that generated PionX motifs on the D. miranda neo-X chromosome21.

PowerPoint slide

Full size image

PWMs model the base readout of DNA sequences with the implicit assumption that each nucleotide at a given position contributes to binding independently of other positions. Physical interactions of neighbouring base pairs, however, alter the structural conformation of the DNA double helix (often referred to as the DNA shape), which may manifest as variations in the minor groove width, roll, helix twist, and propeller twist. Many proteins depend on both base identity and localized helix shape to recognize their binding site13,14,15. Using a pentamer-based model built from all-atom Monte Carlo simulations of DNA structures16, we calculated DNA shape parameters at each base position of the low-stringency motif hits, with 20-base-pair (bp) extensions on either side. To complement these position-centred features we also calculated regional mono- and dinucleodide frequencies (k-mers) in 4-bp windows along the hit sequences. Principal component analysis (PCA) revealed that a combination of DNA shape and k-mer features was able to separate the two classes of sequences: those that were bound in a CXC-dependent manner (CXC-bound) and those that were not bound in a CXC-dependent manner (non-CXC-bound). Sequences in the latter group were either not bound at all (2,502) or were bound independently of the CXC domain (111) (Extended Data Fig. 4a and Supplementary Table 2). This suggested that at least some of the DNA features might improve binding prediction. Indeed, classification models constructed with our additional feature sets performed much better than a PWM-score model in predicting CXC-dependent binding sites on all motif hits (Extended Data Fig. 4b).

Guided by the good performance of our classification model using both the PWM-hit-score and k-mer features, we predicted mutations that would convert robust CXC-bound sites to non-CXC-bound sites. The model suggested that the best discriminating residues would localize to the 5′ part of the motif and not to the GA-rich region (Fig. 3c). To test these predictions, we modified the DIP experiment by mixing appropriately diluted DNA oligonucleotides, representing either a native site or its mutated version, into the genomic DNA. The efficiency of DNA retrieval of experimental oligonucleotides and control genomic loci was quantified by quantitative PCR (qPCR). The results confirmed our predictions (Fig. 3d), leading us to conclude that the main determinants for CXC-dependent binding reside within the first eight bases at the 5′ end of the consensus motif. Notably, this is the part of the motif that diverges most from the MRE.

To achieve a switch in the predicted class from CXC-bound to non-bound in the context of the unbalanced data set of low-threshold instances (54 bound sites, 2,613 non-bound sites) required at least three mutations. This inevitably affected the motif score, making it difficult to distinguish the effects of base and shape readout. To reduce class imbalance and to evaluate the contribution of shape features to CXC-dependent binding of sequences with high similarity to the motif consensus, we limited the analysis to fewer sites through the stepwise increase of motif score thresholds. Figure 3e reveals the relative success of the PWM score compared with a selection of additional features in predicting CXC-dependent binding. In a more balanced data set of motifs consisting of the 95 best motif hits (29 sequences CXC-bound, 66 non-bound) the PWM score was no longer a good predictor and DNA shape features became increasingly relevant. In particular, ‘roll at position +1’ (that is, the roll between the first two base pairs of the motif) turned into the best-performing predictor when sequences with PWM scores higher than 21 were considered. We therefore focused on the 34 highest-scoring motif sequences (q < 0.05), which are highly enriched on the X chromosome; however, only 16 of them are bound in a CXC-dependent manner by MSL2. We systematically scanned these high-scoring sequences for statistically significant shape differences between the CXC-bound and non-bound classes at any nucleotide position (Extended Data Fig. 4d). The results confirmed that ‘roll at position +1’ was appreciably different between the two classes. To test experimentally the importance of this feature we changed the degree of roll at position +1 from >4° to < −2° by either replacing cytosine at position +1 or the two adenines at positions −1 and +2. These mutations led to a clear reduction in MSL2 binding (Fig. 3c, d, f). We were also able to convert a sequence that was not efficiently bound by MSL2 into one that was by changing the roll at position +1 from < −1.9° to >4° (Extended Data Fig. 4c). Adding the DNA shape feature ‘roll at position +1’ to our PWM-hit-score classification model resulted in substantially improved performance when applied to the complete list of 2,667 motif hits (Extended Data Fig. 4b). We therefore conclude that the ability of MSL2 to distinguish true binding sites from a large collection of irrelevant elements with highly related sequences also relies on structural features.

To investigate further the role of MSL2-binding sites in X chromosome dosage compensation, we first attempted to monitor the interactions of MSL2 with HAS in vivo, with minimal contributions from other DCC subunits. Genetic studies had shown that the assembly of a mature MSL-DCC bound to the non-coding roX RNA in male flies is compromised by inactivating the RNA helicase maleless (MLE). Under those circumstances, the remaining MSL2–MSL1 sub-complex is bound to a small subset of HAS17. We recreated this scenario in S2 cells by using RNA interference (RNAi) against mle expression, and found that MSL2 binding was preferentially retained at HAS, corresponding to CXC-dependent binding sites (Extended Data Fig. 5b). The 25 HAS that were most resistant to MLE depletion (Fig. 4a) revealed a shared sequence, bearing a strong resemblance to the CXC-dependent motif (Fig. 4c). By contrast, the 25 sites most sensitive to MLE depletion (bound only by the complete DCC) shared a GA motif similar to the one found in CXC-independent in vitro binding sites (Extended Data Fig. 5a). This suggests that under physiological conditions, the MSL2–MSL1 sub-complex directly contacts a subset of HAS in a CXC-dependent manner in the absence of associated protein and RNA subunits.

It is possible that these chromosomal interactions represent an intermediate of MSL-DCC assembly. To test this hypothesis, we initiated de novo MSL-DCC assembly in female Kc cells by reducing the expression of the sex-lethal gene Sxl. The SXL protein prevents MSL2 expression and thus the dosage compensation program in female cells. Upon depletion of SXL (Extended Data Fig. 5c, d), binding of newly expressed MSL2 to CXC-dependent HAS was stronger and occurred earlier when compared to CXC-independent ones (Fig. 4b). Consistent with this finding, hierarchical clustering of MSL2 signals from the common set of Kc and S2 peak regions revealed 30 sites that acquire strong MSL2 binding ability 3 days after SXL depletion (Extended Data Fig. 5d). De novo motif discovery on these sites revealed a consensus sequence that resembles the one in the CXC-dependent sites (Fig. 4c). Our data strongly suggest that those sites identified in vitro as CXC-dependent are pioneering binding sites for MSL2 in vivo. We therefore refer to them as PionX sites, and to their defining motif as the PionX motif.

The notion that PionX sites are important for dosage compensation is further supported by evolutionary considerations. Drosophila miranda represents a unique system to study how newly evolving X chromosomes acquired dosage compensation. The D. miranda neo-X chromosome is a sex chromosome that began to evolve just 1 million–2 million years ago18. Owing to the relatively short evolutionary time span, the neo-X chromosome still retains many autosomal features, but has already acquired partial dosage compensation. Recent work has identified the MSL-DCC-binding sites on all D. miranda X chromosomes19. De novo motif analysis yielded the typical GA-rich MREs for the older, fully compensated X-chromosomal arms XL and XR. Notably, though, the consensus sequence derived from the neo-X chromosome clearly resembled the PionX signature19 (Fig. 4c).

The neo-Y chromosome originated from the fusion of one Müller-C chromosome to the Y chromosome, resulting in evolutionary pressure on the second Müller-C chromosome to become the neo-X chromosome. We found the PionX motif (but not the MRE) to be particularly enriched on the D. miranda neo-X chromosome but not on the related Drosophlia pseodoobscura Müller C autosome, supporting the idea that this motif represents a new X-chromosome-specific feature (Extended Data Fig. 5e). Careful comparison of neo-X-chromosome sequences with the homologous regions in D. pseodoscura revealed that the novel MSL-DCC-binding sites were acquired by diverse molecular mechanisms, including point mutations and short insertions/deletions of precursor sequences20. About half of them originated from precursor sequences contained in a D. miranda-specific helitron transposon21. The homologous neo-Y helitron does not contain PionX motifs—only precursor sequences in which the 5′ CAC motif and the 3′ GA-rich element are separated. On the neo-X chromosome these two parts are fused by a 10-bp deletion to form PionX consensus motifs19,21 (Fig. 4d). The insertion of a PionX consensus motif derived from the D. miranda neo-X chromosome into an autosome of D. melanogaster led to strong, ectopic binding of the MSL-DCC. By contrast, the corresponding homologous neo-Y-chromosome sequence, in which the 5′ and GA sequences are split by a 10-bp insertion, did not recruit the complex19,21. A similar experiment used the strongest DCC-binding site from the neo-X chromosome and the corresponding neo-Y-chromosomal fragment, showing MSL-DCC recruitment to the former, but not to the latter19. While the neo-Y-chromosomal fragment does not contain a PionX motif, the evolved neo-X chromosome contains nine of them (Extended Data Fig. 5f). Collectively, these observations suggest that PionX motifs play an important role in de novo acquisition of dosage compensation.

In summary, we provide three lines of argument suggesting that PionX sites are X-chromosome-specific determinants that function early in the series of events that lead to exclusive targeting of the X chromosome and correct dosage compensation. First, PionX sites are bound by an MSL2–MSL1 sub-complex in the absence of all other subunits, a state that may reflect an early intermediate of MSL-DCC assembly at HAS. Second, PionX sites are the first to be occupied during de novo establishment of dosage compensation. Finally, PionX motifs arose during the early phase of neo-X-chromosome evolution in D. miranda.

A pertinent conceptual advance from our study is the understanding that not all HAS contain the same amount of information. The subset of PionX sites are not necessarily sites of highest MSL2 occupancy in vivo (Extended Data Fig. 1b), but contribute an important qualitative element of X-chromosomal discrimination. This discrimination is not wholly apparent from the consensus motif as it also relies on the shape of the DNA at the MSL2-binding site.

The initial recruitment of MSL2 to PionX sites on the X chromosome may trigger the distribution of the complex to nearby non-PionX HAS within the chromosomal territory, thereby further amplifying the difference in MSL2 occupancy between the X chromosome and the autosomes. It is likely that other factors contribute to the stability of the targeting system in vivo, such as the cooperativity of MSL2 domains within what is presumed to be a dimeric complex22; the assembly of functional complexes within the X-chromosomal territory owing to transcription of roX RNA from the X chromosome23; synergistic interactions between different MSL-DCC complexes and with the CLAMP protein at clustered MREs24; and a supportive organization of the conformation of the X chromosome25.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment.

Protein purification

MSL2 proteins were expressed in Sf21 cells and purified by Flag affinity chromatography as described5.

Genomic DNA preparation

The pellet from 6 × 107 S2 or Kc cells was suspended in 1.2 ml of lysis buffer (10 mM Tris pH 8, 100 mM NaCl, 25 mM EDTA pH 8, 0.5% SDS, 0.15 mg ml−1 of proteinase K) and incubated at 56 °C overnight. After addition of sodium acetate to a final concentration of 0.3 M, the nucleic acids were extracted with phenol–chloroform and precipitated with an equal volume of isopropanol at −20 °C for 1 h. Precipitated nucleic acids were centrifuged and washed with 70% ethanol. Dried pellets were resuspended in TE buffer and sonicated with Covaris AFA S220 (microTUBEs, peak incident power 175 W, duty factor 10%, cycles per burst 200, 430 s) to generate 200-bp fragments. After RNase digestion (0.1 mg ml−1, 1 h at 37 °C), DNA was purified with the GenElute kit (Sigma). Synthetic DNA was produced using the Repli-g kit (Qiagen) with 20 ng of gDNA as starting material.

DIP–seq

DIP–seq experiments were performed as in ref. 7 with few modifications. In brief, 400 ng of gDNA was incubated with either 80 nM of MSL2–Flag or mutated recombinant protein at 26 °C for 30 min in 100 μl of binding buffer (100 mM KCl, 2 mM MgCl2, 2 mM Tris-HCl pH 7.5, 10% glycerol, 10 μM ZnCl2). For DIP experiments in the presence of synthetic DNA, 10 pM of the specified synthetic DNA was added to the reaction. 10% of the reactions was taken as input material and subjected to quantitative PCR and/or deep sequencing. DNA–protein complexes were immunoprecipitated using 15 μl of Flag bead slurry (M2, Sigma) for 15 min at room temperature and washed twice with 100 μl of binding buffer to eliminate unbound DNA. After digestion with proteinase K (0.5 mg ml−1, 1 h at 56 °C), DNA was purified with the GenElute kit (Sigma) and subjected to qPCR and/or deep sequencing. The DIP experiments in presence of synthetic DNA were performed using the deltaRING construct (three different protein preps).

Cells, RNAi, ChIP–seq

All cells used in this study were authenticated performing karyotyping and staining for the MSL-DCC and regularly tested for mycoplasma contamination.

Double-stranded RNAi fragments were generated from PCR products obtained using the following oligonucleotides: mle RNAi: 5′-TTAATACGACTCACTATAGGGAGAATGGATATAAAATCTTTTTTGTACCAATTTTG-3′; 5′-TTAATACGACTCACTATAGGGAGAACAGGGCGCATGACTTGCT-3′. Sxl RNAi-1: 5′-TAATACGACTCACTATAGGGAGAGATCACAGCCGCTGTCC-3′; 5′-TAATACGACTCACTATAGGGAGATACCGAATTAAGAGCAAATAATAA-3′. Sxl RNAi-2: 5′-TAATACGACTCACTATAGGGAGACCCTATTCAGAGCCATTGGA-3′; 5′-TAATACGACTCACTATAGGGAGAGTTATGGTACGCGGCAGATT-3′.

The culture of Drosophila male S2 (subclone L2-4, provided by P. Heun), female Kc cells and RNAi against mle and Sxl were performed as previously described3 with modifications. At days 3, 6 and 9 after the initial treatment with dsRNA, Sxl RNAi cells were split and either collected for ChIP experiments and western blot analyses or treated again with Sxl dsRNA. For S2 cells, ChIP experiments were performed using a Covaris AFA S220 (PIP 100 W, DF 20%, 200 CB for 30 min) to generate chromatin fragments of sizes averaging 180 bp. For Kc cells, ChIP experiments were performed as before26 with modifications. In brief, about 4 × 107 cells were suspended in ice-cold homogenization buffer and fixed with 1% formaldehyde for 10 min at room temperature. After quenching with 125 mM glycine, the cells were collected and washed three times with ice-cold RIPA buffer (1% Triton X-100, 0,1% Na deoxycholate, 0,1% SDS, 140 mM NaCl, 10 mM Tris-HCl ph 8, 1 mM EDTA). Fixed nuclei were sonicated in RIPA buffer with a Covaris sonifier (PIP: 140, DF 20%, CB: 200) for 30 min.

Antibodies

MSL2, MLE and Lamin antibodies were previously described10. The SXL antibody was obtained from F. Gebauer.

Library preparation and sequencing

The Diagenode MicroPlex library kit was used to prepare libraries from 1–2 ng of input, DIP or ChIP DNA quantified using the Qubit dsDNA HS Assay kit (Life Technologies Q32851). The libraries were sequenced on a HighSeq 1500 (Illumina) instrument to yield roughly 15 million–25 million reads of 50-bp single-end sequences per sample.

Oligonucleotides

Double-stranded synthetic DNA fragments were obtained by annealing equimolar concentrations (10 μM) of complementary oligonucleotides. All oligonucleotides used in the DIP studies are listed in Extended Data Table 2.

Data analysis

If not indicated otherwise, data were processed using R (http://www.r-project.org) or Bioconductor (http://www.bioconductor.org) and function calls with default parameters. For hierarchical clustering of binding sites based on MSL2 signals, we applied the complete method on Euclidean distances.

Read processing, coverage and normalized coverage

Sequence reads were aligned to the D. melanogaster release 6 reference genome using Bowtie27 version 1.1.1 allowing only for single matches to the reference (parameter –m 1). We extended the matched reads to a total of 200 bp and calculated for each sample a per-base genomic coverage vector by cumulating the total spans of all sequenced fragments.

We defined target signal enrichment as the standardized difference between normalized immunoprecipitate and corresponding normalized input coverage using:

in which i denotes genomic position, and coverage denotes number of fragments covering i.

Peak calling, definition of robust peak sets and chromosomal enrichment

Peaks were called using Homer28 findPeaks version 4.7.2 with the parameters: style = factor; size = 200; fragLength = 200; inputFragLength = 200 and C = 0. All peaks were called using the corresponding input samples as controls. We defined peaks as robust if the region was called in at least two biologically replicated samples. X-chromosomal enrichment describes the ratio of X-chromosomal peak density to autosomal peak density, with density being the number of peaks divided by length of chromosome(s).

Definition of CXC-dependent sites (PionX)

We first defined a robust set of in vitro MSL2-binding regions by combining the peaks called in the two MSL2 DIP–seq experiments performed on S2 gDNA and the one performed on Kc gDNA. We calculated the average signal enrichment over the input for the profiles described in Fig. 2c. We then tested for signal differences in samples with intact CXC domain against the ones with a deleted or mutated CXC domain using a linear model (R package, limma) including the cell type (origin of gDNA) as random effect. CXC-dependent sites were defined with an FDR threshold of < 0.05 and a fold change < 0.

De novo motif discovery and genome-wide motif searches

We searched for enriched motifs in peak regions using MEME29, with the zero or one occurrence per sequence (zoops) model, except for searches in D. Miranda. Here we applied the any number of repetitions (anr) model, given the extreme amplification of motifs in some of the peak regions. Genome-wide searches were performed with FIMO29 version 4.10.0 and an initial P-value cutoff of 1 × 10−5.

Definition of HAS and MRE

MSL2 in vivo peaks were called on two published high-quality profiles (GEO accession codes GSM929148 and GSM929149). A total of 309 overlapping peak regions (304 on the X chromosome, 5 on the autosomes) were defined as HAS. MEME-based de novo motif discovery using the zoops model yielded a consensus sequence that we refer to as MRE. This MRE closely matches the original definition2,3 (Extended Data Fig. 3c).

Performance comparison of MRE and PionX PWM

On each of the genome-wide 36,410 MRE hits we calculated the score of the PionX PWM after extending the hit region by 2 bp 5′ to the MRE consensus. We then determined the overlap of the hit regions with the 309 HAS. If more than one MRE hit matched to the same HAS, we kept the hit with the highest PionX score. We determined the ROC of each PWM by continuous thresholding of the respective scores using the match to a HAS as response. The analysis comprises 35,659 instances mapping to 266 HAS.

Definition of mle RNAi-resistant sites

Average MSL2 enrichment was calculated on all 309 robust MSL2 in vivo binding sites (HAS) in control and MLE RNAi ChIP–seq samples (3 biological replicates each). We then tested for difference in signals between the two experimental groups using limma (R). We defined the most resistant as the 25 sites with the highest moderated t values. Accordingly, the 25 sites with the lowest t values were defined as most sensitive sites.

Definition of strong MSL2-binding sites upon Sxl RNAi in Kc cells

Average MSL2 enrichment was calculated for all 309 HAS including the 90 robust peak regions arising in Kc cells (4 biological replicates for the Sxl RNAi at each time point, 2 for the controls). The signals were clustered hierarchically. Two clusters with the strongest gain were combined to constitute a set of 30 sites.

DNA shape calculation and extended feature description

The initial set of regions subjected to extended feature analysis was defined by applying low thresholds (q ≤ 0.2) on FIMO motif searches for the PionX motif. We obtained 2,667 hits (Supplementary Table 1), 54 of which were bound in a CXC-dependent manner by MSL2, 111 bound in a CXC-independent manner by MSL2 and 2,502 were unbound in our in vitro DIP–seq experiments. We refer to unbound instances as well as CXC-independently bound ones as non-CXC-bound.

DNA shape parameters were calculated with the DNAshape program16. While minor groove width and propeller twist refer to specific nucleotide positions, roll and helix twist specify structural parameters between adjacent bases. For the sake of simplicity, we assigned the values of roll and helical twist to the preceding base (‘roll at position 1’ actually specifies the roll between the bases of nucleotides at position 1 and 2).

The total set of features considered in this study (with the number of variables in brackets) comprise: the PWM-hit-score (1), nucleotide composition at each position from −20 bp to +20 bp around the motif (244), minor groove width from −18 bp to +18 bp around the motif (57), roll from −19 bp to +18 bp around the motif (58), helix twist from −19 bp to +18 bp around the motif (58), propeller twist from −18 bp to +18 bp around the motif (57), nucleotide frequencies in six consecutive 4-bp windows starting at position 1 of the motif (24), dinucleotide frequencies in six consecutive 4-bp windows starting at position 1 of the motif (96). Minor groove width, roll, twist and propeller twist constitute the shape features (230). Mono- and dinucleotide frequencies constitute the k-mer features (120). The total number of features was 595.

Machine learning

Classification models for feature evaluation were built using simple logistic classifier30. ROC curves of the classifiers were based on tenfold cross-validation.

The importance of all features were ranked by measuring their correlation (Pearson’s) with the class label on the whole set of PionX PWM hits with a q ≤ 0.2. Features selected as relevant for and present at CXC-dependent binding of the hiw, CG8097 and CG1492 genes were: ‘CA in window 1’, ‘TT at window 2’ and ‘T at window 2’. Mutations were proposed based on the results of feature selection and the presence of respective k-mers in the sites selected for mutation. The proposed mutations were evaluated by simple logistic classifier trained using the PWM score and k-mers on the full data set. Modified sites were designed to result in the switch of the predicted class from CXC-bound to not CXC-bound.