An atlas of endogenous DNA double-strand breaks arising during human neural cell fate determination

Endogenous DNA double-strand breaks (DSBs) occurring in neural cells have been implicated in the pathogenesis of neurodevelopmental disorders (NDDs). Currently, a genomic map of endogenous DSBs arising during human neurogenesis is missing. Here, we applied in-suspension Breaks Labeling In Situ and Sequencing (sBLISS), RNA-Seq, and Hi-C to chart the genomic landscape of DSBs and relate it to gene expression and genome architecture in 2D cultures of human neuroepithelial stem cells (NES), neural progenitor cells (NPC), and post-mitotic neural cells (NEU). Endogenous DSBs were enriched at the promoter and along the gene body of transcriptionally active genes, at the borders of topologically associating domains (TADs), and around chromatin loop anchors. NDD risk genes harbored significantly more DSBs in comparison to other protein-coding genes, especially in NEU cells. We provide sBLISS, RNA-Seq, and Hi-C datasets for each differentiation stage, and all the scripts needed to reproduce our analyses. Our datasets and tools represent a unique resource that can be harnessed to investigate the role of genome fragility in the pathogenesis of NDDs. Measurement(s) DNA Double Strand Break • Transcriptome • Whole Genome Sequencing • Chromosome conformation capture assay • Tumor Suppressor p53-Binding Protein 1 Technology Type(s) sBLISS - Genome-wide detection of DNA double-strand breaks by in-suspension breaks labeling in situ and sequencing • RNA-seq of total RNA • copy number variation profiling assay • Hi-C • Immunocytochemistry Factor Type(s) Human Neural Cell Fate Specification Sample Characteristic - Organism Homo sapiens Sample Characteristic - Environment cell culture Measurement(s) DNA Double Strand Break • Transcriptome • Whole Genome Sequencing • Chromosome conformation capture assay • Tumor Suppressor p53-Binding Protein 1 Technology Type(s) sBLISS - Genome-wide detection of DNA double-strand breaks by in-suspension breaks labeling in situ and sequencing • RNA-seq of total RNA • copy number variation profiling assay • Hi-C • Immunocytochemistry Factor Type(s) Human Neural Cell Fate Specification Sample Characteristic - Organism Homo sapiens Sample Characteristic - Environment cell culture


Background & Summary
Incorrectly repaired DNA double-strand breaks (DSBs) pose a major threat to genome stability, as they can potentially lead to the formation of mutations and genomic rearrangements that can ultimately cause cancer and other disorders driven by genome instability 1 . Emerging evidence suggests that both transcription and the three-dimensional (3D) genome structure are associated with the formation of endogenous DSBs 2,3 . Transcription-associated DSBs seem to preferentially form around the transcription start site (TSS) of transcriptionally active genes as well as at chromatin loop anchors in proximity of highly transcribed genes, presumably because of the accumulation of DNA torsional stress in these regions, which requires resolution by transient DNA breaks formed by topoisomerases 4,5 .
In mouse neural stem cells, hotspots of endogenous DSBs (so-called recurrent DSB clusters) were previously detected around the TSS of highly transcribed genes involved in general cellular processes and along the gene body of long, neural-specific genes whose human orthologues had been previously implicated in neurodevelopmental disorders (NDDs) such as schizophrenia (SCZ) and autism spectrum disorder (ASD) [6][7][8] . Interestingly, many of these RDCs were also found in neurotypical human neural stem cells 9 . Moreover, neural stem cells derived from patients with a form of ASD marked by increased susceptibility to replication stress were shown to contain recurrent DSB clusters localized specifically within transcribed genes associated with ASD risk 10 .
Despite the emerging evidence linking endogenous DSBs to NDDs, genome-wide maps of DSBs spontaneously arising at different stages of human neurogenesis are currently missing. Such maps would allow correlating the genomic DSB landscape of cells at various stages of neural differentiation with other genomic and epigenomic features, providing clues on how DSBs might form in these cells and how their incorrect repair might Cells were harvested at three timepoints and processed for sBLISS, RNA-Seq, and Hi-C (see Supplementary Table 1 for the list of datasets). Note that different batches of NES cells from different passages (max 5 passages apart) were used for performing multiple replicate   ) before centrifuging them at 150 × g for 3 min. We filtered single-cell NPC and NEU cell suspensions first through 100 μm and then through 40 μm cell strainers (Corning) to obtain debris-free cell suspensions. We fixed cells resuspended at a density of 10 6 cells/mL in the appropriate growth medium by adding methanol-free paraformaldehyde (Thermo Fisher Scientific) to a final concentration of 2% and incubating the samples for 10 min at room temperature on a rolling shaker. Lastly, we inactivated the unreacted paraformaldehyde by adding glycine in 1X PBS to reach a final concentration of 125 mM and incubating for 5 min before washing the cells twice with DPBS. We stored samples prepared in this way at 4 °C for up to one week before proceeding with the sBLISS protocol (see below).

In-suspension BLISS (sBLISS).
We performed sBLISS following the same protocol that we previously described in detail 11 . Briefly, we started by lysing the cell membrane in a lysis buffer containing 10 mM Tris-HCl/10 mM NaCl/1 mM EDTA/0.2% Triton X-100 pH 8 and incubated the samples for 60 min on ice. We then pelleted the cells at 150 × g for 5 min at room temperature, removed the supernatant, resuspended them in a permeabilization buffer consisting of 10 mM Tris-HCl/150 mM NaCl/1 mM EDTA/0.3% SDS pH 8 pre-warmed at 37 °C, and incubated the samples for 60 min at 37 °C. Next, we pelleted the cells at 150 × g for 5 min at room temperature, washed them twice with CutSmart Buffer (New England Biolabs)/0.1% Triton X-100 pre-warmed at 37 °C. To convert DSBs containing overhangs into a form that can be ligated to blunt adapters, we used the Quick Blunting Kit (New England Biolabs) and resuspended the cells in a final blunting reaction volume of 100 μL, followed by 60 min incubation at room temperature. We then washed the cells with CutSmart Buffer/0.1% Triton X-100 at room temperature and proceeded to in situ ligation of sBLISS adapters to the blunted DSB ends. For ligation we used 25 Weiss units (U) of 5 U/μL T4 DNA Ligase (ThermoFisher Scientific) and incubated the cells at 16 °C for 20-24 hours in a 100 μL ligation reaction mix containing 3 μL of 50 mg/mL BSA (ThermoFisher Scientific) and 12 μL of 10 mM ATP (ThermoFisher Scientific). For a sample containing ~10 6 cells, we used 4 μL of sBLISS adapter at 10 μM prepared as we described before 11 . The complete list of the sBLISS adapters used to generate the datasets described here is available in Supplementary www.nature.com/scientificdata www.nature.com/scientificdata/ 7.5 and adding 10 μL per sample of 800 U/mL Proteinase K (New England Biolabs). We incubated the samples for 14-18 h at 55 °C in a thermal mixer set at 800 shakes-per-minute, followed by addition of 10 μL of fresh Proteinase K the next morning and continuing the incubation for 1 hour. We then heat-inactivated the Proteinase K for 10 min at 95 °C and extracted gDNA using a 25/24/1 (vol./vol./vol.) solution of Phenol-Chloroform-Isoamyl Alcohol with 10 mM Tris pH 8.0, 1 mM EDTA (Sigma-Aldrich) and Chloroform (Merck), followed by ethanol precipitation. After drying, we dissolved the DNA pellets in 102 μL of Tris-EDTA by placing the tubes in a thermal mixer set to 50 °C for 15 min while shaking at 1,100 shakes-per-minute. After cooling the samples to 4 °C, we sonicated 100 μL of each sample using BioRuptor Plus (Diagenode) with the following settings: 30 sec on, 60 sec off, high intensity, 30 cycles. We concentrated the sonicated gDNA samples using AMPure XP beads (Beckman Coulter) and assessed the fragment size using a BioAnalyzer 2100 (Agilent Technologies). We aimed at obtaining gDNA fragments with sizes ranging from 300 to 800 bp, with a peak around 400-600 bp. We stored the sonicated and purified gDNA samples at -20 °C until we performed in vitro transcription (IVT) and library preparation. For IVT, we used 90-300 ng of each sonicated gDNA sample (see Supplementary Table 1 for the exact amount used for each sample) and the MEGAscript T7 Transcription Kit (Thermo Fisher Scientific) following the manufacturer's instructions, with the exception that we carried out all reactions for 14 hours at 37 °C after adding RiboSafe RNAse Inhibitor (Bioline) to each sample. Upon IVT completion, we degraded the gDNA by adding 2 U of RNase-free DNAse I (Thermo Fisher Scientific) to each sample and purified the amplified RNA (aRNA) with RNAClean XP beads (Beckman Coulter). Next, we ligated each aRNA sample to the Illumina RA3 adapter (purchased from Integrated DNA Technologies) using T4 RNA Ligase 2 (New England Biolabs) and incubated the samples for 2 hours at 25 °C. Thereafter, we reverse transcribed the ligated aRNA using the Illumina RTP primer (purchased from Integrated DNA Technologies) and SuperScript IV Reverse Transcriptase (Thermo Fisher Scientific), by following the manufacturer's instructions except for the incubation time, which we extended to 50 min at 50 °C followed by 10 min heat-inactivation at 80 °C. To prevent RNAse activity, we added RNaseOUT (Thermo Fisher Scientific) during the RA3 adapter ligation and during the reverse transcription. We amplified the resulting libraries using NEBNext Ultra II Q5 Master Mix (New England Biolabs), the Illumina RP1 primer (purchased from Integrated DNA Technologies), and one of the Illumina RPIX index primers (purchased from Integrated DNA Technologies), by performing 8 PCR cycles in 400 μL split into 8 PCR tubes to increase library complexity. The sequences of the RT, RP1 and RPIX primers used to generate the datasets described here are available in Supplementary Table 2. After PCR, we purified and size-selected the libraries using two-sided AMPure XP bead (Beckman Coulter) purification, aiming at obtaining libraries with a fragment size ranging between 300 and 900 bp. We assessed the final libraries on a BioAnalyzer High Sensitivity DNA chip (Agilent Technologies) and quantified them using the Qubit dsDNA HS Assay kit (Thermo Fisher Scientific). We sequenced all the libraries on a NextSeq 500 machine (Illumina) using the NextSeq 500/550 High Output Kit v2 (Illumina), performing 75 sequencing cycles and additional 6 cycles for index sequencing. We sequenced multiple indexed sBLISS libraries together aiming at obtaining ~25 million reads per library.
After sequencing, we demultiplexed the raw data based on index sequences using Illumina's BaseSpace and retrieved the corresponding FASTQ files. We processed the raw sequencing reads using the sBLISS pipeline that we previously described in detail 11 . Briefly, the pipeline first selects reads containing the expected sBLISS prefix comprised of the unique molecular identifier (UMI; 8 nucleotides, nt) and the sample barcode (BC; 8 nt) using SAMtools 20 (v1.10) and Scan for Matches 21 , allowing at most one mismatch in the barcode. The pipeline then clips the prefixes off and stores them, while the trimmed reads are aligned to the human reference genome (GRCh37/hg19) using BWA-MEM 22 (v0.7.17-r1188) with default options. Afterwards, the pipeline discards the reads with mapping quality scores lower than 30 and removes PCR duplicates by searching for adjacently mapped reads (at most 10 bp apart along the reference genome) with at most one mismatch in their UMI sequence. Finally, the pipeline returns a list of unique DSB locations and the corresponding number of unique UMIs identified at each position, which we used for all downstream analyses described below. To visualize the distribution of DSBs along selected genes, we uploaded the corresponding BED files to the UCSC Genome Browser (https://genome.ucsc.edu/cgi-bin/hgTracks) and selected the squish display mode in the Track Settings.

Quantification of DSBs by immunofluorescence.
We performed immunofluorescence staining for the DSB marker 53 binding protein 1 (53BP1) following the same procedure described above for stem and neural differentiation markers, except for the imaging part (see Table 1 for a list of antibodies and dilutions used). For imaging, we mounted the coverslips in non-hardening Vectashield medium (Vector labs). We imaged the samples on the same Nikon Ti-E microscope equipped with a Sona 4.2B-6 sCMOS camera (Andor Technology) using a using a CFI Plan Apochromat Lambda 100X Oil objective (Nikon). We acquired 10 image stacks per sample, each consisting of at least 48 focal planes spaced 0.2 µm apart, aiming at imaging at least 150 cells for each differentiation stage (NES, NPC, and NEU). To segment the cells, we first used a custom script written in MATLAB (see Code Availability) to generate 2D nuclei segmentation masks using the Otsu's method to find a global threshold of the fluorescence intensity in the DNA channel for each field of view. We visually inspected and, if needed, manually corrected all the segmentation masks. We then measured the fluorescence intensity in the channel corresponding to 53BP1 by integrating the images axially over the segmentation masks, using the mid 48 slices in each z-stack. When the images contained more than 48 slices, we selected a range of z-values where the corresponding gradient magnitude was as high as possible (this corresponds to performing an axial centering around the middle section of the nuclei).
Total RNA-Seq. We cultured NES, NPC, and NEU cells as described above and rinsed them with DPBS Mg+/Ca+ (Sigma-Aldrich). We added 1 mL of TRI Reagent Solution (Thermo Fisher Scientific) per million cells directly to the cells attached to the flasks. After incubating for 5 min at room temperature, we detached the cells by pipetting the TRI Reagent Solution on them multiple times and pipetting the detached cells up-and-down www.nature.com/scientificdata www.nature.com/scientificdata/ multiple times until the solution became clear. We then transferred the solution to a 1.5 mL tube and stored the samples at −20 °C overnight. The next day, we added 20% (vol./vol.) chloroform (Sigma-Aldrich) to each sample, mixed well and incubated for 3 min at room temperature, followed by centrifugation at 12,500 × g for 15 min at 4 °C. We then transferred the upper phase and thoroughly mixed this with an equal volume of ice-cold 70% ethanol. Next, we pipetted the mixture onto a PureLink RNA Mini Spin Cartridge (Thermo Fisher Scientific) and proceed to RNA purification following to the manufacturer's instructions. RNA-Seq libraries were prepared by the National Genomics Infrastructure at the Science for Life Laboratory (SciLifeLab) in Stockholm, Sweden, using the TruSeq Stranded Total RNA Library Prep kit with RiboZero from Illumina. We assessed and quantified the final libraries on a BioAnalyzer 2100 (Agilent Technologies) and Qubit using a High Sensitivity RNA chip (Thermo Fisher Scientific). We sequenced the libraries on a NextSeq 500 machine (Illumina) using the NextSeq 500/550 High Output Kit v2 (Illumina) and performing 75 sequencing cycles and 6 additional cycles for index sequencing. We pooled three replicates for each of the three cell types, aiming at obtaining at least 50 million reads per library.
After sequencing, we demultiplexed the raw data based on the index sequences using Illumina's BaseSpace and downloaded the corresponding FASTQ files. We performed sequencing quality checks for all the experiments using FastQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/, v0.11.9). We removed the adapter sequences using TrimGalore 23 (v0.6.4_dev) with default parameters. We filtered the reads against human rRNA and tRNA sequences obtained from NCBI using Bowtie2 24 (v2.4.1) with the option --sensitive-local. We used the reads that failed to align in the previous step as input for STAR 25 (v2.7.0e) and mapped them to the human reference genome (GRCh37/hg19) using the GENCODE 26 (v19) gene reference annotation with the following parameters: --twopassMode Basic --alignSJoverhangMin 8 --alignSJDBoverhangMin We removed PCR duplicates using the Picard MarkDuplicates tool (http://broadinstitute.github.io/picard/, v2.21.7) with default parameters. We carried out transcript expression quantification using Salmon 27 (v1.2.0) in pseudo-alignment mode with library type (--l ISR) and correction of sequence-specific (--seqBias) and position-specific (-posBias) biases, and the GENCODE (v19) annotation and corresponding transcriptome as references. We imported the transcripts per million (TPM) tables obtained with Salmon into R (https://www.r-project.org/, v4.0.3) using the tximport package 28 with lengthScaledTPM scaling method and no scaling for differential gene expression analysis. We performed gene expression levels normalization across samples and differential gene expression calculations using the DESeq 2 R package and the IHW R package for Independent Hypothesis Weighting. We adjusted the P values for multiple hypotheses testing with the Benjamini and Hochberg method 29 and used a false discovery rate (FDR) of 0.001 and absolute Log2 fold-change of 1 as thresholds to extract genes significantly changing their expression between conditions. We applied a variance stabilizing transformation (VST) using the vst function from the DESeq 2 package. We then extracted the expression values of differentially expressed genes, calculated their z-score across conditions and replicates and the Euclidean distance between gene vectors using the dist function from the stats R package, and performed hierarchical clustering using the hclust function from the stats R package. The number (5) of final clusters was obtained through the cutree function from the R Stats package with h parameter set to 5. We performed all Gene Ontology (GO) enrichment analyses using the Biological Processes (BP) ontology and the compareCluster function from the clusterProfiler R package. We carried out the individual GO terms (e.g., Neurogenesis, Axonogenesis, etc.) enrichments across differentially expressed gene clusters by generating the corresponding contingency tables and applying the Fisher's exact test to assess their statistical significance using the fisher .test function from the Stats R package, with alternative hypothesis set to greater. We retrieved the individual GO terms by performing manual queries across the AmiGO 2 database (http://amigo.geneontology.org/).

Hi-C.
We prepared Hi-C libraries using the Arima-HiC kit (Arima Genomics) following the manufacturer's instructions. Briefly, we crosslinked the cells with 2% formaldehyde (following the same procedure described above for preparing cells for sBLISS) and then used approximately 1 million fixed cells as input for each cell type. Subsequently, we used 2.5 μg (for NES) or 1.5 μg (for NPC and NEU) of Hi-C template for biotin pull-down and library preparation according to the Arima Genomics User Guide for Library Preparation using KAPA Hyper Prep Kit. Specifically, we used 6 (for NES) or 8 (for NPC and NEU) PCR cycles for library amplification. We then pooled corresponding Arima-HiC replicate libraries and sequenced each pool on one SP flowcell on the NovaSeq 6000 system (Illumina) in paired-end sequencing mode aiming at obtaining ~800 million reads per library. Sequencing was carried out at the National Genomics Infrastructure at the Science for Life Laboratory (SciLifeLab) in Stockholm, Sweden.
After sequencing, we processed the raw sequencing data using HiCUP 30 (v0.7.4) with default parameters. Briefly, the pipeline employs Bowtie2 24 (v2.4.1) to align the reads to the human reference genome (GRCh37/ hg19) and filters out experimental artefacts (i.e., circularized, re-ligated, and duplicate reads). We generated 'digest' files using the hicup_digester command with the option --arima. We used the HiCUP output files (BAM format) containing only valid, non-redundant read pairs, as input for pairtools (https://pairtools.readthedocs. io/en/latest/, v0.3.0). First, we converted and sorted the BAM files into pairsam format using the parse and sort modules in pairtools. In addition to individual replicates, we generated pooled samples for each cell type (NES, NPC, and NEU) by merging matched replicates using the merge module in pairtools. We marked read duplicates using the pairtools dedup module with option -mark-dups and filtered the results selecting only specific pair types (i.e., 'UU' , 'UR' and 'RU') via the pairtools select module, which produced an output in pairs format. Finally, we added the fragment information using the fragment_4dnpairs.pl convenience script provided alongside the Juicer pipeline 31 , and converted the pairs files into hic format using the Pre module from Juicer-Tools www.nature.com/scientificdata www.nature.com/scientificdata/ (v1.22.01). Unless explicitly required by the software/package used, we normalized all contact matrices with Knight-Ruiz Matrix Balancing (KR) using Juicer-Tools. For each chromosome, we extracted from the hic files the intra-chromosomal interactions matrix at 1 megabase (Mb) and 100 kb resolution using Straw 31 without normalization. We assessed the reproducibility of Hi-C data using the GenomeDISCO 32 (v1.0.0) run_all module.
A/B compartments calling. For each cell type (NES, NPC, and NEU), we extracted the first and second eigenvectors at 1 Mb resolution from the hic files, using the compartments module from the FAN-C 33 toolkit (v0.9.17). For each chromosome, we determined the correctness of the eigenvectors by correlating the first and second eigenvector of each developmental stage with the first eigenvector of the others and switched the first with the second eigenvector if the Pearson's correlation of the latter was higher than the former. In addition, although FAN-C orients the eigenvectors by using the average GC-content information, we re-assigned the orientation of each eigenvector by using the average RNA-Seq signal across positive and negative regions. We defined regions with high RNA-Seq signal as ' A' compartments and regions with low RNA-Seq signal as 'B' compartments.
TAD boundaries calling and classification. To call TAD boundaries, we used Straw 34 without normalization to extract the intra-chromosomal interaction matrix for each chromosome at 100 kb resolution. We then imported the hic files into R and used them as input for the TimeCompare function in TADCompare 35 . We classified TAD boundaries into six different categories: (1) Early Appearing (i.e., boundaries detected in NPC and NEU but not in NES); (2) Early Disappearing (i.e., boundaries detected in NES but not in NPC and NEU); (3) Late Appearing (i.e., boundaries detected in NEU but not in NES and NPC); (4) Late Disappearing (i.e., boundaries detected in NES and NPC but not in NEU); (5) Dynamic (i.e., boundaries detected only in NES and NEU or only in NPC); and (6) Common (i.e., boundaries detected in all three cell types). In addition, for each cell line we extracted the insulation scores and plotted their average across a 200 kb region centered on the TAD boundary, using the insulation and boundaries modules from the FAN-C toolkit.
Chromatin loops calling and classification. To call chromatin loops at 5 kb resolution, we applied Mustache 36 (v0.1.4) to the hic files using the --pThreshold 0.1 parameter. Then, we imported the list of chromatin loops into R and classified them in six categories, as described above for TAD boundaries: (1) Early Appearing (i.e., loops absent in NES and present in NPC and NEU); (2) Early Disappearing (i.e., loops detected in NES but not in NPC and NEU); (3) Late Appearing (i.e., loops detected in NEU but not in NES and NPC); (4) Late Disappearing (i.e., loops detected in NES and NPC but not in NEU); (5) Dynamic (i.e., loops detected in NES and NEU or only in NPC); and (6) Common (i.e., loops detected in all three cell types).

Analysis of DSBs at promoter regions and along gene bodies.
We first selected all expressed protein-coding genes (average TPM ≥ 1 across all RNA-Seq samples) and computed the number of DSBs in their promoter (defined as the region from 2 kb upstream to 1 kb downstream of the TSS) or in their gene body (i.e., from the TSS to the transcription end site or TES). The only exception was when we calculated the ratio of DSBs in the promoter over the gene body, in which case we computed the number of DSBs from 1 kb downstream of the TSS to the TES. We discarded genes shorter than 2 kb to avoid overestimating the signal along the gene body for regions shorter than 1 kb. To compare the DSB burden across promoter regions and gene bodies within the same sample, we normalized the DSB counts per million mapped reads (CPM) or per kilobase gene length per million mapped reads (CPKM), respectively. To compare the DSB burden across different samples (e.g., to compare DSB metaprofiles around the TSS in NES, NPC, and NEU cells), we downsampled the DSB counts in each sample to the size of the smallest dataset (~15 million DSBs). We added a pseudo-count for plotting or to avoid null denominators in ratios whenever we applied a logarithmic transformation.

Analysis of DSBs at TAD boundaries and around chromatin loop anchors.
To compute the DSB burden around TAD boundaries, we used a window of 200 kb centered on each TAD boundary to calculate how many of the downsampled DSBs (~15 millions) were mapped in these regions. To assess DSBs around chromatin loop anchors, we generated a list of significant chromatin loops at 5 kb resolution for each differentiation stage using HiCCUPS from Juicer tools (v1.22.01). We used the central positions of the upstream and downstream loop anchors as loop start and end reference points, respectively. We then used a window of 40 kb centered on these reference points to calculate how many of the downsampled DSBs (~15 millions) were mapped in these regions. Lastly, we normalized the generated metadata profiles by dividing the DSB counts by the lowest value of the density function, which was then considered as signal baseline.
Analysis of DSBs in NDD risk genes. We manually compiled a list of genes associated with increased risk for either SCZ or ASD by parsing the Supplementary Data in ref. 37 and Table S2 in ref. 38 , respectively, and retaining only unique gene names (see Supplementary Table 3). We then mapped the gene names to their corresponding Entrez identifier using the AnnotationDbi 39 R package. We computed the number of DSBs in the promoter region and along the gene body of each gene as described above and used the distribution of all protein-coding genes (after filtering out those included in the SCZ and ASD risk genes list) as background reference.

Technical validation
Validation of the 2D cell culture model system of neural differentiation. We first aimed at validating our in vitro model of neural cell fate specification by performing immunofluorescence staining for various neural lineage markers (Methods). As expected, NES cells expressed neuroepithelial stem cell protein (Nestin) and sex determining region Y-box 2 (SOX2), two markers of actively self-renewing neuroepithelial stem cells, and showed neural rosette formation (Fig. 1b, NEU). After 6 days of in vitro differentiation, the cells started to divide less frequently and short projections expressing the neuronal marker microtubule-associated protein 2 (MAP2) began to form (Fig. 1b, NPC). After 35 days of differentiation, the cultures consisted mainly of neuronal cells with highly entangled MAP2-positive fibers and, amidst them, few cells expressing the glial fibrillary acidic protein 2 (GFAP) marker of differentiation towards astrocytes (Fig. 1b, NEU).
It has been reported that reprogramming of somatic cells and prolonged passaging of human iPSCs can result in the formation of karyotypic abnormalities such as chromosomal amplifications or deletions 41 . The iPSC-derived NES cell line that we used to generate the datasets described here (see Methods) was previously shown to be karyotypically normal and stable over multiple passages 15,42 . However, to exclude that chromosomal alterations formed during the course of our experiments, we performed low-pass (1X sequencing depth) whole-genome sequencing followed by copy number calling using genomic DNA extracted from NES cells and NEU cells after 35 days of differentiation (Methods). For both differentiation stages, the genome-wide copy number profiles were flat (Fig. 1c,d), indicating that the cells remained karyotypically stable throughout our experiments.
To further characterize our model system, we performed three total RNA-Seq experiments on NES, NPC, and NEU cells (from different NES passages), followed by differential gene expression and gene ontology (GO) analysis (Methods). Experimental replicates corresponding to the same differentiation stage clustered together, whereas the largest difference was observed between NES and NEU samples, as expected (Fig. 1e). We identified 4,583 differentially expressed genes forming two larger and three smaller clusters marked by different expression patterns (Fig. 1f). Cluster 1 was significantly enriched in genes linked to DNA replication, which became downregulated during the transition from NES to NEU, as expected given that the latter cells stop cycling (Fig. 1f,g). On the other hand, cluster 2 and, to a lesser extent, clusters 3-5 were significantly enriched in GO terms related to differentiation towards neural lineages and neuronal functions, such as 'Neurogenesis' , 'Neuron differentiation' , and ' Axonogenesis' (Fig. 1h-n). Altogether, these results demonstrate that our in vitro differentiation system is a good proxy of the morphological and transcriptional changes that occur during neural cell lineage specification. sBLISS reproducibly detects endogenous DSBs. Next, we performed two sBLISS experiments on NES, NPC, and NEU cells (from different NES passages), following the same procedure that we previously described in detail 11 (Fig. 2a and Methods). We first compared DSB counts across replicates at decreasing genomic bin sizes (100, 25, and 10 kb). DSB counts were highly correlated between replicates (Pearson's correlation coefficient (PCC) between replicates: 0.95 ± 0.02, 0.88 ± 0.03, 0.82 ± 0.03 at 100, 25, and 10 kb resolution, respectively, mean ± s.d.), highlighting the reproducibility of sBLISS (Fig. 2b-d). To compare the DSB counts between differentiation stages, we normalized the DSB counts based on the genomic DNA input to the IVT step in sBLISS (Methods). The normalized DSB burden progressively increased as cells differentiated, with NEU samples yielding the highest number of DSB ends sequenced (Fig. 2e). To validate this finding using an orthogonal approach, we performed immunofluorescence staining for the DSB marker, 53BP1 (Methods). Indeed, NPC and especially NEU cells displayed substantially higher 53BP1 levels compared to NES cells (Fig. 2f,g). Altogether, these results demonstrate that sBLISS is a valid approach to reproducibly capture changes in the burden of endogenous DSBs.
DSBs are enriched at the promoter and along the gene body of highly transcribed genes. Using sBLISS, we previously showed that endogenous DSBs are enriched around the TSS of transcribed protein-coding genes in immortalized chronic myeloid leukemia TK6 cells, as well as in primary mouse enterocytes 11 . Other studies using alternative DSB detection approaches, such as DSBCapture 43 , END-seq 44 . and High-Throughput Genome-wide Translocation Sequencing (HTGTS) 45 , have also reported similar enrichment patterns of both endogenous and treatment-induced DSBs around the TSS of actively transcribed genes in different cell types, including mouse neural stem cells 4,[6][7][8]11,43,46,47 . Hence, to further validate our sBLISS datasets, we assessed whether the endogenous DSBs detected by sBLISS in NES, NPC, and NEU cells are also enriched around the TSS of protein-coding genes in a transcription-associated manner (Methods). The normalized DSB counts around the TSS were significantly higher (P < 0.0001, Wilcoxon test, two-tailed) in the top expression quartile gene group, in all three differentiation stages analyzed (Fig. 3a-c). The DSB burden around the TSS of highly expressed genes was significantly higher (P < 0.0001, Wilcoxon test, two-tailed) in NEU compared to NES cells, suggesting that differentiated, post-mitotic neuronal cells might experience higher levels of transcription-related torsional stress in promoter regions. Next, we investigated whether the gene body of highly expressed genes also harbors more DSBs compared to lower expressed genes. Indeed, in all three differentiation stages, protein-coding genes in the top expression quartile displayed significantly higher (P < 0.0001, Wilcoxon test, two-tailed) normalized DSB counts along the gene body compared to genes in the bottom quartile (Fig. 3d-f). Visual inspection of the DSB tracks along selected genes confirmed that DSBs accumulate around the TSS and along the gene body of genes upregulated during the differentiation of NES cells towards NEU (Fig. 3g,h). Altogether, these results demonstrate that highly expressed genes at different stages of neural cell fate specification are hotspots of endogenous DSB formation, in line with previous observations by our and other groups 4,6-8,11,43,46,47 , further corroborating the validity of our datasets. www.nature.com/scientificdata www.nature.com/scientificdata/ CpG content at promoters correlates with fragility. To further investigate endogenous DSBs at promoter regions, we assessed whether the CpG content of the sequence surrounding the TSS is associated with the frequency of DSBs detected by sBLISS. 72% of human promoter sequences contain CpG islands that are considered important for regulation of development and that have been associated with constitutively nucleosome-depleted regions, altered modifications and positioning of nucleosomes, and distinct patterns of transcription initiation 48,49 . We therefore examined the DSB distribution at protein-coding gene promoters classified as CpG High or CpG Low based on their CpG content (Methods). In all three stages of differentiation, CpG High promoters carried a significantly higher (P < 0.0001, Wilcoxon test, two-tailed) burden of DSBs compared to CpG Low ones (Fig. 4a-c). Notably, CpG High promoters in NEU cells showed the highest density of endogenous DSBs in the region upstream of the TSS (Fig. 4d,e). Comparison with RNA-Seq data showed that genes with a CpG High promoter were expressed at significantly higher levels in all three stages of differentiation, suggesting that the CpG content might indirectly affect the fragility of promoters by mediating high transcriptional activity in these regions (Fig. 4f-h).
Next, we wondered how DSBs are distributed along the promoter and gene body of the same gene. To this end, we computed the ratio between the normalized DSB counts in the promoter region and along the gene body for all human protein-coding genes and classified the genes into four different groups (A, B, C, D) based on the calculated ratio (Fig. 4i). Interestingly, considerably more genes were assigned to group D (i.e., genes with DSBs predominantly concentrated in the promoter region) in NEU compared to NES and NPC cells, whereas more      . The violin plots extend from minimum to maximum, and the boxplots inside the violins extend from the 25th to the 75th percentile, with the horizontal bar representing the median, and whiskers extending from -1.5 × IQR to + 1.5 × IQR from the closest quartile, where IQR is the inter-quartile range. (j) Number of genes in each of the four groups shown in (i). (k) Distributions of gene expression levels measured by RNA-Seq in the four gene groups shown in (i), for genes classified as CpG High or CpG Low based on the frequency of CpG dinucleotides in their promoter region. The number of genes (n) in each group is shown below each boxplot. (l) Same as in (k) but for gene length in kiloSbases (kb). Gene length was calculated as the distance from the TSS to the transcription end site of each gene. In all the boxplots shown in the figure, each boxplot extends from the 25th to the 75th percentile, the horizontal bar represents the median, and whiskers extend from -1.5 × IQR to + 1.5 × IQR from the closest quartile, where IQR is the inter-quartile range. Black dots, outliers. The asterisks in (a-c), (f-h) and (k,m) indicate a P value less than 0.01 (**), 0.001 (***) or 0.0001 (****) (Wilcoxon test, two-tailed). ns, not significant. (2022) 9:400 | https://doi.org/10.1038/s41597-022-01508-x www.nature.com/scientificdata www.nature.com/scientificdata/ genes were assigned to group B (i.e., genes with DSBs more abundant in the gene body compared to the promoter region) in NES and NPC compared to NEU cells (Fig. 4j). Expression levels progressively increased from group A to group D and were consistently higher for genes with CpG High promoters (Fig. 4k). The latter genes were, on average, also shorter compared to genes harboring CpG Low promoters (Fig. 4l). Altogether, these results indicate that gene fragility is generally higher in NEU compared to NES cells, and that the frequency and pattern of DSBs along protein-coding genes correlate with multiple factors, including CpG abundance in the promoter region, gene length, and expression levels.

DSBs are enriched in the A chromatin compartment.
To further validate our datasets, we turned our attention to the 3D genome, which, in concert with transcription, has been previously implicated in the formation of endogenous DSBs 5,50 . To this end, we performed two Hi-C experiments on NES, NPC, and NEU cells (from different NES cell passages), and called A/B compartments, TADs, and chromatin loops following well-established procedures (Methods). As replicate Hi-C datasets were highly correlated (Fig. 5a), we pooled them for all subsequent analyses. We first examined the distribution of DSBs across A/B compartments 12 (Methods). The fraction of the genome classified as belonging to the A compartment was ~50% in NES cells and progressively decreased during the differentiation towards neurons (Fig. 5b). A small fraction (~9%) of the genome switched compartment type during differentiation, with A → B transitions being more frequent (~6%) compared to B → A transitions (~3%) (Fig. 5c,d). This reflects the progressive reduction of the A compartment from NES to NEU (Fig. 5b) and is consistent with the previous finding that the interaction strength within the B compartment increases during the differentiation of mouse NPCs to neurons 51 . We then assessed the normalized DSB counts across compartments and found them to be significantly higher (P < 0.0001, Wilcoxon test, two-tailed) in the A compartment, in all three stages of differentiation ( Fig. 5e-g). This is consistent with the notion that the A compartment corresponds to active chromatin 12 and with our finding that DSBs are enriched in highly expressed genes in all three differentiation stages (see Fig. 3). Interestingly, the A compartment contained a higher burden of DSBs in NES compared to NEU cells, whereas the B compartment was more fragile in NEU compared to NES cells ( Fig. 5e-g). To assess whether these differences might be related to A/B compartment switches during the differentiation process, we compared the normalized DSB counts between A/B compartments that either remained stable or changed during the transition from NES to NEU (Methods). The DSB burden was significantly higher (P < 0.0001, Wilcoxon test, two-tailed) in A/B compartments that switched during differentiation, with the largest difference observed in NES cells (Fig. 5h-j). The biggest increase in the DSB burden was observed within genomic regions that switched from B to A compartment during the transition from NES to NEU (Fig. 5k-m). Interestingly, A-compartment regions switching to the B compartment showed a slightly but significantly lower (P < 0.01, Wilcoxon test, two-tailed) DSB burden in NES cells compared to stable A regions, but this difference leveled out in NPC and NEU cells (Fig. 5k-m). The lowest burden of DSBs was observed in genomic regions classified as stably belonging to the B compartment, in all three stages of differentiation (Fig. 5k-m). These results suggest that the DSB burden of a given genomic region depends on its global transcriptional activity and whether the region changes compartment during differentiation.
DSBs are enriched at TAD boundaries and around chromatin loop anchor sites. Next, we examined how the endogenous DSBs detected by sBLISS in NES, NPC, and NEU cells are distributed with respect to TADs and chromatin loop anchors (Methods). The average TAD size was significantly higher (P < 0.01, Wilcoxon test, two-tailed) in NEU compared to NES cells, whereas the TAD insulation score remained constant (Fig. 6a,b). This is reminiscent of the larger TAD structures previously observed in mature neurons [52][53][54] . Consistent with the progressive expansion of the B compartment, the proportion of B-type TADs increased during the transition from NES to NEU, while the fraction of TADs overlapping different compartments slightly decreased (Fig. 6c,d). DSBs were locally enriched around TAD boundaries and chromatin loop anchors, as well around CCCTC-binding factor (CTCF) motifs in all three differentiation stages ( Fig. 6e-h). These findings are in agreement with prior reports according to which chromatin loop anchors represent fragile sites where DSBs accumulate most likely as a consequence of topological stress associated with loop extrusion and active transcription of nearby genes 4,50 . We then classified TADs and chromatin loops based on their dynamics throughout the differentiation process ( Fig. 6i-k and Methods). Unlike for A/B compartments, the DSB burden around TAD boundaries and chromatin loop anchors remained substantially unchanged regardless of the dynamics of these regions during differentiation (Fig. 6l-q). Collectively, these results suggest that, while endogenous DSBs form non-randomly with respect to TADs and chromatin loops, structural changes occurring in these regions during neural cell fate determination do not seem to reshape their fragility landscape.
Endogenous DSBs form preferentially within expressed NDD risk genes. In mouse neural stem cells, endogenous DSBs indirectly detected by HTGTS 45 are enriched in genes whose human orthologues are associated with increased risk for schizophrenia (SCZ) and autism spectrum disorder (ASD) [6][7][8] . Furthermore, in human neural progenitor cells, recurrent DSB clusters identified by HTGTS map along the gene body of multiple genes involved in SCZ and ASD 9 . Therefore, to further validate our datasets, we explored whether the endogenous DSBs detected by sBLISS in human NES, NPC, and NEU cells are also enriched in genes associated with these disorders. To this end, we retrieved a list of 1,169 genes previously associated with increased SCZ risk 37 as well as 100 genes previously associated with ASD 38 (Supplementary Table 3). The normalized DSB counts inside the promoter region and along the gene body of these genes were significantly higher compared to the counts in the corresponding regions of all other human protein-coding genes (background), in all three cell differentiation stages analyzed (Fig. 7a-l and Methods). Furthermore, among SCZ and ASD risk genes, the DSB burden around promoters as well as along gene bodies was significantly higher in NEU compared to NES cells (Fig. 7a-l). The ten most fragile SCZ and ASD genes included genes with well-known neuronal functions as well as genes that www.nature.com/scientificdata www.nature.com/scientificdata/ www.nature.com/scientificdata www.nature.com/scientificdata/ have been previously implicated in neuro-psychiatric disorders, including MAPK10, CDC42BPA, CLIP1 and PCDH9 among SCZ risk genes, and TCF7L2, CHD2, MAP1A, FOXP1, RFX3, and TBL1XR1 among ASD risk genes (Fig. 8a,b). In all cases, the burden of DSBs in the promoter region of these top fragile genes was consistently higher in NEU compared to NES cells (Fig. 8a-d). This suggests that, in the later stages of neural cell fate specification, these genes might experience a higher amount of transcription-related damage, possibly because of higher expression levels. Indeed, comparison with RNA-Seq data revealed that SCZ and ASD risk genes were significantly more expressed (P < 0.0001, Wilcoxon test, two-tailed) in NEU compared to NES cells, as well as compared to background genes (Fig. 8e,f). Altogether, these results demonstrate that, in agreement with previous observations, endogenous DSBs accumulate at gene loci associated with increased risk for NDDs, with differentiated neural cells exhibiting the highest amount of DSBs in these genes. Although most of the endogenous DSBs forming inside NDD risk genes during neurogenesis are most likely correctly repaired, it is tempting to speculate that repeated DSB repair errors occurring during neurodevelopment might change the promoter sequence of these genes and instigate pathogenic effects by affecting their expression levels and/or correct time of expression. We anticipate that the approach, tools, and datasets presented here will facilitate future studies aimed at testing this fascinating hypothesis. (d-f) Same as in (a-c) but for normalized DSB counts along the gene body (from the first TSS of the gene to the last transcription end site). (g-i) Same as in (a-c) but for genes associated with ASD risk (see Table S2 in ref. 38 ). (j-l) Same as in (h-j) but for normalized DSB counts along the gene body. The asterisks in (a-l) indicate a P value less than 0.01 (**), 0.001 (***) or 0.0001 (****) (Wilcoxon test, two-tailed).

Usage Notes
We have generated what is, to our knowledge, the first atlas of endogenous DSBs that form spontaneously in an in vitro model mimicking the process of human neural cell fate specification. The sBLISS BED files that we provide (see Data Records) can be readily used to visualize the pattern of DSBs along any gene of interest using freely available tools such as the UCSC Genome Browser (https://genome.ucsc.edu/) or the Integrative Genomics Viewer (https://software.broadinstitute.org/software/igv/). Moreover, the accompanying RNA-Seq and Hi-C datasets represent a valuable resource that can be further mined to investigate the interplay between 3D genome organization and transcriptional dynamics during neural cell fate specification. We provide all the custom scripts and relative documentation needed to reproduce the pre-processing of sBLISS, RNA-Seq, and Hi-C datasets as well as all the analyses described here. One limitation of our study is that it was conducted on an in vitro model system that most likely only partially recapitulates the landscape of genome fragility in the nervous system. Hence, future studies applying sBLISS to more complex in vitro systems, such as brain organoids, as well as to nuclei extracted directly from brain tissue are needed to fully characterize the landscape of genome fragility in the developing and adult brain.