Ultraviolet radiation shapes dendritic cell leukaemia transformation in the skin – Nature

Patient samples

Patients with BPDCN seen at the Dana-Farber Cancer Institute provided informed consent to an IRB-approved research protocol permitting tissue collection and sequencing analysis. The demographic characteristics of the patient cohort are provided in Supplementary Table 1a. Healthy control participants for single-cell sequencing consented to IRB-approved research protocols from Brigham and Women’s Hospital or Lonza Bioscience that cover all of the study procedures; demographics are provided in Supplementary Table 3a.

Histopathology

Histological processing and immunohistochemical staining of patient bone marrow and skin tumour biopsies was performed according to routine clinical procedures in the Department of Pathology at the Brigham and Women’s Hospital, as previously described^26,43,44. Results are included in Supplementary Table 1a.

Targeted DNA sequencing and analysis

Targeted sequencing of fresh bone marrow samples using a 95-gene leukaemia panel (Rapid Haem Panel, n = 23 samples) and formalin-fixed, paraffin-embedded archival skin tumour samples using a 282-gene pan-cancer panel (Oncopanel, n = 9 samples) was performed for genes recurrently altered in myeloid malignancies and BPDCN^45,46,47,48. Mutation calls were manually inspected and verified in sequence alignment files for each sample. Combined mutation calls and VAFs across all samples are provided in Supplementary Table 2a. For the patient 2 bone marrow sample, we observed a lower read coverage for amplicons covering the ASXL1 gene relative to other samples and controls (67%, corresponding to a VAF of 0.33, located on chromosome band 20q11) and, accordingly, included this copy-number alteration in Fig. 1b.

WES analysis

WES analysis of cryopreserved bone marrow, skin tumour and germline samples (uninvolved skin for patient 7, bone-marrow-derived fibroblasts for patients 9 and 10) was performed using the Illumina HiSeq 4000 (2 × 150 bp, patient 7) and the BGISEQ-500 (2 × 100 bp, patients 9 and 10) platforms, as previously described⁴⁹. An overview of all of the profiled samples is provided in Supplementary Table 1b. Sequencing data for a total of 12 samples were mapped to the human genome reference (hg19; https://www.ncbi.nlm.nih.gov/data-hub/genome/GCF_000001405.13/) using BWA (v.0.7.15)⁵⁰. The resulting BAM files were further analysed and recalibrated with Picard (v.2.5.0)⁵¹ and the GATK toolkit (v.4.0.0.0)⁵². Somatic mutations were identified using Mutect2⁵³ by comparing to patients’ germline variations. Initial calls were filtered by estimated cross-sample contamination and artifacts related to orientation bias. Calls were then merged for each patient and further filtered by removing calls that showed a VAF lower than 0.1 in all samples, with the exception of mutations that were also identified by targeted sequencing (that is, mutations in TET2 in patient 10). We also excluded calls that showed a germline VAF that was greater than one-fifth of the highest VAF of other samples from the same patient, and mutations that were detected across multiple patients (with the exception of hotspot mutations in ASXL1 in patients 7 and 10). Resulting high-confidence mutation calls for each sample are provided in Supplementary Table 2b–d. For the patient 10 relapse bone marrow sample, which was collected after the patient received a stem cell transplant, we were unable to define sample-specific mutations owing to the high proportion of donor DNA. However, we could quantify mutations in this sample identified in other samples from the same patient.

Combined mutation calls were the basis for the inference of tumour phylogenies (Figs. 1c and 4a) and the inference of putative clonal architectures in patient bone marrow samples (Fig. 1d). Mutation calls were further analysed for mutational signatures defined in the COSMIC database⁵⁴ by applying the R package MutationalPatterns⁵⁵. The relative contribution of 30 different mutational signatures was calculated and scores for UV-light-associated signature 7 are indicated (Fig. 3b). For copy-number analysis, SNVs were jointly identified for all samples from each patient using bcftools (v.1.10.2; commands mpileup and call). The B (minor) allele frequency and read coverage (relative to germline samples) for each SNV was used to infer copy-number alterations (Extended Data Figs. 2 and 3). Public datasets (Extended Data Fig. 8a) were analysed from provided mutation calls, with the exception of data from ref. ²⁸, which were processed from raw sequencing reads (Sequence Read Archive: SRP301976) and analysed using Mutect2.

WGS analysis

WGS analysis of germline, bone marrow and skin tumour samples at diagnosis and relapse (13 samples from patients 1, 3 and 12) was performed using the BGI DNBSEQ platform (2 × 100 bp). Bone marrows were profiled from cryopreserved samples, skin tumours were profiled from formalin-fixed paraffin-embedded (FFPE) samples and germline samples were profiled from both sample types (Supplementary Table 1b). Sequencing data were mapped to the human genome (hg19) reference using BWA (v.0.7.17)⁵⁰. The resulting BAM files for all of the samples from each patient were jointly analysed using Mutect2⁵³ by comparing to the matched germline sample, supplying both a germline resource (somatic-b37_af-only-gnomad.raw.sites.vcf) and a panel of healthy individuals (somatic-b37_Mutect2-exome-panel.vcf). Variant calls were filtered using GATK⁵² (v.4.2.3.0; commands LearnReadOrientationModel and FilterMutectCalls) and annotated using the Funcotator command (funcotator_dataSources.v1.6.20190124s). The resulting calls were further filtered by retaining only variants that had a coverage greater than 20 for cryopreserved samples and greater than 10 for FFPE samples in all samples per patient. Final mutation calls were defined at the latest timepoint per patient: in patient 1, we considered variants detected in both skin relapse 2 and bone marrow relapse (collected after the patient received a stem cell transplant) with a VAF of greater than 0.25. In patient 3, we considered variants detected in the diagnostic bone marrow with a VAF of greater than 0.1. In patient 12, we considered variants detected in bone marrow relapse with a VAF of greater than 0.25. These filtering steps were deemed to be appropriate owing to the lower quality of FFPE-derived samples, and challenges due to high proportions of donor DNA in patients who received a stem cell transplant. Variants were attributed to the first sample in which they were detected with a VAF greater than 0.1, and all subsequent samples.

The resulting mutation calls were visualized in tumour phylogenies (Extended Data Fig. 1d) and were analysed for mutational signatures similar to as described for the WES dataset above (Fig. 3b). Summaries of mutation calls for each sample are provided in Supplementary Table 2e. For copy-number analysis, single-nucleotide variants were jointly identified for all of the samples from each patient using bcftools (v.1.10.2; commands mpileup and call). The B allele frequency and read coverage (relative to germline samples, not shown) for each SNV was used to infer copy-number alterations (Extended Data Fig. 4).

scRNA-seq

scRNA-seq was performed on cryopreserved bone marrow aspirates. Cells were stored in liquid nitrogen, thawed using standard procedures and viable (propidium iodide negative) cells were sorted on the Sony SH800 flow cytometer. Next, 10,000–15,000 cells were loaded onto a Seq-Well array or 10x Genomics chip. Further processing was performed using the recommended procedures for the Seq-Well S³ (http://shaleklab.com/resource/seq-well/)⁵⁶ or the 10x Genomics 3′ v3/v3.1 chemistry. Seq-Well S³ libraries were sequenced on the NextSeq system (20 + 8 + 8 + 57 cycles) and 10x libraries were sequenced on the NovaSeq system (28 + 8 + 91 cycles for single-index or 28 + 10 + 10 + 75 cycles for dual index). Some of the data were previously reported^57,58 (Supplementary Table 3a). Serial samples from the same patient were loaded onto separate sequencing runs to avoid erroneous assignment of reads by index swapping between diagnosis/remission/relapse samples (this is particularly relevant for the identification of rare malignant cells).

XV-seq

We developed an improved method for targeted enrichment of genetic variants from scRNA-seq libraries that is compatible with Seq-Well S³ and 10x 3′ gene expression platforms. Compared to previous methods by us and others^26,59, we incorporated a number of computational and experimental steps for increased sensitivity and specificity: (1) we considered all mutations detected by WES, including synonymous mutations and mutations affecting untranslated regions (UTR). These mutations do not result in changes in the protein sequence, but can be used to infer clonal relationships. (2) We quantified detection of these mutations in the regular scRNA-seq data before enrichment. For example, of the 186 mutations detected across samples for patient 10, only 16 (8.6%) were detected in at least one transcript (Supplementary Table 2d). We found that detection in the regular scRNA-seq data is a good predictor of enrichment efficiency (Extended Data Fig. 6c). (3) We specifically considered loci of which only a single allele is present in the genomes of healthy and/or malignant cells. For these mutations, detection of the wild-type allele is as informative as the presence of the mutant allele (that is, if the wild-type is detected, the mutant must be absent; for heterozygous mutations, the mutant could remain undetected). In our dataset, this included a mutation in the X-chromosomal gene RAB9A (in a male patient), a focal deletion of CDKN2A/B, which occurred in cells already carrying loss of chromosome 9, and 3′ UTR mutations in SETX and SMARCC1, which also occurred in cells with loss of the other allele on chromosome 9 and chromosome 3. (4) Finally, we incorporated technical optimizations such as inclusion of dual indices, as outlined below.

XV-seq for Seq-Well S³

Compared with single-cell genotyping of Seq-Well S³ libraries that we previously reported²⁶, we adjusted primer designs to generate dual-indexed libraries. This increases the confidence that reads are assigned to the correct library, particularly when using Illumina instruments with patterned flow cells. We first designed biotinylated mutation-specific primers to detect each of the known mutations in a given sample (Supplementary Table 4a). As a starting material, we used amplified cDNA from the Seq-Well S³ protocol (also known as whole-transcriptome-amplified material). We then set up a biotin-PCR reaction to add a biotin tag and Nextera adapter to our gene of interest while retaining the unique molecular identifier (UMI) and cell barcode, as follows. We created a mixture containing a standard reverse primer at 3 µM (SMART-AC), and mutation-specific primers at a combined concentration of 3 µM. To prepare the template for the biotin-PCR reaction, we pooled and diluted whole-transcriptome-amplified products from the same sample and timepoint to 10 ng in a total volume of 10 µl. We next added 2.5 µl of primer mix (final concentration, 0.3 µM) and 12.5 µl of 2× KAPA HiFi Hotstart Readymix (Roche, NC0465187) to the template. We performed PCR using the following conditions: initial denaturation at 95 °C for 3 min; followed by 12 cycles of 90 °C for 20 s, 65 °C for 15 s and 72 °C for 3 min; and final extension at 72 °C for 5 min. After amplification, we purified the PCR product with 0.7× AMPure XP beads and captured biotinylated fragments using Streptavidin beads.

To add Illumina adapters, dual-indexed barcodes and a custom read primer binding sequence to the fragments, we performed a second PCR using the Streptavidin-bound product as a template (23 µl), with 2 µl of a 5 µM primer mix (N70D primers, Supplementary Table 4b) and 25 µl PFU Ultra II HS 2× Master Mix (Agilent, 600850). The parameters used for the second biotin-PCR were as follows: initial denaturation at 95 °C for 2 min; then 4 cycles of 95 °C for 20 s, 65 °C for 20 s and 72 °C for 2 min; followed by 10 cycles of 95 °C for 20 s and 72 °C for 2 min and 20 s; and then final extension at 72 °C for 5 min. After the second PCR, we magnetized the Streptavidin beads and saved/purified DNA from the supernatant with 0.7× AMPure XP beads. After eluting in 20 μl of TE, we magnetized the beads and saved the supernatant for sequencing on the Illumina NextSeq system.

XV-seq for 10x

We adjusted the Genotyping of Transcriptomes⁵⁹ protocol by (1) omitting staggered handles on gene-specific primers and (2) incorporating dual 10 bp library indices, which minimizes the chance of barcode swapping and ensures compatibility with 10x Genomics scRNA-seq v3.1 libraries and Illumina v1.0 and v1.5 chemistry. The starting material for transcript genotyping were the full-length cDNA libraries generated according to the 10x Genomics 3′ v3 or v3.1 scRNA-seq protocol. If cDNA quantities were limited, we performed a full-length cDNA PCR amplification using generic primers that bind to all transcripts (primers: PartialRead1 and PartialTSO; Supplementary Table 4c). The pre-enrichment PCR was set up by mixing 10 ng of cDNA template, forward and reverse primers at 0.3 μM each, 2× Kapa HiFi HotStart ReadyMix and H₂O up to 50 μl. The PCR was performed under the following conditions: initial denaturation at 95 °C for 3 min; followed by 6 cycles of 98 °C for 20 s, 67 °C for 15 s, 72 °C for 3 min; and a final extension of 72 °C for 3 min. After amplification, we purified the PCR product with 0.6× AMPure XP beads (Beckman Coulter Life Sciences, A63881).

The enrichment for loci of interest consists of two PCR reactions. For PCR1, to enrich for loci of interest (determined by targeted or exome sequencing), we designed primers to amplify specific regions (Supplementary Table 4a). We downloaded the transcript sequence in Geneious Prime 2020, and annotated the mutation of interest. We designed mutation-specific primers within 50 bases upstream of the mutation site (so that the mutation site would be captured in read 2 of the sequencing data). To add a read 2 sequence to this mutation-specific primer, we appended CACCCGAGAATTCCA at the 5′ end. PCR1 was performed using these mutation-specific primers and a generic forward primer (PartialRead1; Supplementary Table 4c). We mixed up to six mutation-specific primers per PCR1 reaction, as long as they targeted different transcripts. We prepared PCR1 reactions as follows: 100 ng cDNA was added to 0.25 μM forward primer and 0.25 μM mutation-specific primer(s), 20 μl 2× Kapa HiFi HotStart ReadyMix and H₂O up to 40 μl. The PCR was performed under the following conditions: a denaturation step at 95 °C for 3 min; followed by 10 cycles of 98 °C for 20 s, 67 °C for 15 s, 72 °C for 3 min; and a final extension 72 °C for 3 min. After amplification, we purified the PCR product with 1× AMPure XP beads. We next performed PCR2 to generate indexed libraries compatible with the Illumina NextSeq and NovaSeq machines.

For PCR2, we used a P5 sequence followed by a 10 bp index barcode and a read 1 sequence as a forward primer (XV-P5-i5-BCXX) and a P7 sequence followed by a 10 bp index barcode and a read 2 sequence as a reverse primer (XV-P7-i7-BCXX; Supplementary Table 4c). The PCR was set up as follows: 18 μl of the PCR1 product was added to 2 μl primers (0.25 μM each) and 20 μl Kapa HiFi HotStart ReadyMix. The PCR was performed under the following conditions: 95 °C for 3 min; followed by 6 cycles of 98 °C for 20 s, 67 °C for 15 s, 72 °C for 3 min; and a final extension 72 °C for 3 min. After amplification the PCR product was purified with 1× AMPure XP beads. Elution in 20 μl buffer TE typically yielded 5–50 ng μl⁻¹ with an average size of 300–1,500 bp, which was pooled for sequencing on the Illumina NextSeq or NovaSeq instruments with the goal of generating 10 million reads per library.

scRNA-seq computation analysis

Data from the Seq-Well protocol (healthy donor 6 and patient 9) were processed as described previously²⁶. In brief, demultiplexed fastq files were processed to maintain only cell barcodes with 100 reads and to append the cell barcode and UMI, derived from read 1, to the read identifier of read 2. The hg38 reference genome and annotations were downloaded from Ensembl (release 99), extended with RNA18S and RNA28S genes from UCSC, and finalized using Cell Ranger mkgtf with the recommended settings and the additional gene biotypes Mt_rRNA and rRNA. We then used STAR (v.2.6.0c) to align processed fastqs to hg38 and created a count matrix. Data from the 10x Genomics 3′ v3 and v3.1 platform (the remaining 15 samples) were processed using Cell Ranger (v.7.0.0) using the default settings and the same hg38 reference. Count matrices from both the Seq-Well and the Cell Ranger pipelines were processed to retain only cells with >2,000 UMIs, >1,000 genes and <20% mitochondrial alignments. From the count matrix, we removed mitochondrial genes (^MT-*), genes of the biotype rRNA (defined in the reference gtf file) and RNA18S/RNA28S. We maintained X- and Y-chromosomal genes, including ZRSR2 and IL3RA.

XV-seq computational analysis

To quantify detection of mutations in regular scRNA-seq libraries, we assessed every mutation detected by exome sequencing in the genome alignments for the respective sample. Mutations were quantified using samtools mpileup. For each base, information for cell barcode and UMI was obtained by setting the –output-extra option, and subsequently collapsed using R and the data.table package. Mutations that were most efficiently detected or that were of special interest were selected for XV-seq enrichment.

For analysis of XV-seq data, fastq files were processed using IronThrone-GoT (v.2.1) using the recommended set-up (https://github.com/landau-lab/IronThrone-GoT)⁵⁹. For patient 9, we used –bclen 12, –umilen 8 and a whitelist of cell barcodes that passed RNA-seq quality controls. For all of the other samples, we used –bclen 16, –umilen 12 and the whitelist 3M-february-2018.txt. For every mutation, we generated custom configuration files to distinguish between wild-type and mutant transcripts by one or several differing bases. If the mutation site was directly adjacent to the primer, the 3′ end of the primer was used as a shared sequence and additional bases were added to the wild-type/mutant sequences, taking into account that IronThrone-GoT allows for 20% of the bases in the analysed reads to be mismatched from the provided sequences. For MTAP, five configuration files were used, one for each of the potential splicing products indicating the CDKN2A/B deletion (Extended Data Fig. 6b). IronThrone-GoT jobs were submitted in Linux using the Sun Grid Engine with the options -pe smp 4 -binding linear:4 -l h_vmem=32g -l h_rt=96:00:00. After completion of the IronThrone-GoT run, we processed the generated information (summTable) by plotting the number of wild-type and mutant calls for different sequencing reads of each transcript (UMI). We used only transcripts that were supported by ≥3 reads and with at least threefold more wild-type than mutant calls or vice versa. For MALAT1.n.G3541A in Fig. 2e,f and Extended Data Fig. 8, we reduced the read threshold to 1. In the case of heterozygous mutations, cells in which a wild-type transcript is detected are not necessarily wild-type cells, as the mutated allele may have been missed. In the case of multiple mutations within the same gene (as is observed for TET2 in BPDCN), transcripts may show a wild-type result at one site while still harbouring a mutation in cis at a different position in the same transcript/allele. We added the genotyping information as metadata to Seurat objects with scRNA-seq expression data by joining based on cell barcodes.

To check the accuracy of our single-cell mutation calls, we validated the ASXL1.G642fs mutation in Pt10Dx using two different enrichment primers. This known oncogenic guanine insertion, resulting in ATCGGAGGGGGGGGT>ATCGGAGGGGGGGGGT, can be challenging to call. We enriched the mutation site from 10,106 high-quality single-cell transcriptomes using two different primers: ASXL1-1886 (CACCCGAGAATTCCAGTCACCACTGCCATAGAGAGG) and ASXL1-1898 (CACCCGAGAATTCCAATAGAGAGGCGGCCACCA; the first one is included in Supplementary Table 4a) (transcript-binding sequences are in bold). In the first experiment, we detected mutated ASXL1 transcripts in nine cells. In the second experiment, we detected mutated transcripts in eight cells. Seven of the cells overlapped between the two attempts, indicating striking concordance. We also detected wild-type ASXL1 transcripts in 33 and 32 cells in the two experiments, respectively. There was perfect overlap in 32 wild-type cells that were called between the two experiments with different ASXL1 enrichment primers. The agreement between these experiments, together with the orthogonal targeted DNA sequencing, which identified the same mutation, attests to the reliability of our mutation calls.

Dimensionality reduction and cell type annotation

Count matrices from healthy donors were imported into R (v.4.2.1) using Seurat (v.4.1.1) on a MacBook Pro with an M1 Max chip. Normalization, variable feature identification and data scaling were performed using the Seurat defaults⁶⁰. After principal component analysis, Harmony was used to integrate data from Seq-Well S³ and 10x v3 3′ scRNA-seq⁶¹. We then used Harmony reduction to determine clusters and UMAP coordinates. Integration from Seq-Well and 10x platforms generated clusters that were driven by biological (rather than technical) differences between cells (Fig. 2a). Clusters were annotated by expression of canonical marker genes such as CD34 (progenitors), CD14 (monocytes), haemoglobin (erythroid), IRF8 and TCF4 (pDCs; Supplementary Table 3b). This yielded 21 healthy reference cell types. One cluster (1.04% of healthy donor cells) was classified as doublets on the basis of co-expression of marker genes.

To annotate cell types from the samples of patients with BPDCN, we used the random-forest algorithm²⁶. Specifically, we used the R package randomForest (v.4.7-1.1) to generate a classification forest using marker genes of healthy donors (determined by Seurat’s FindAllMarkers function); we previously showed that this approach performs similarly to other reference-based classification algorithms⁵⁷. The confusion matrix and fivefold cross-validation both indicated 89.7% accuracy (Extended Data Fig. 5a). We next used the classification forest to assign each cell from the patient samples with prediction/probability scores for each reference cell type (function predict() with randomForest object and type = “prob”; see 3_RandomForest.R at https://github.com/petervangalen/Single-cell_BPDCN/). The reference cell type with the maximum prediction score was used for the patient cell classification. Cells that were classified as doublets (up to 2.34%) were excluded from further analysis. Projection of patient cells onto the UMAP of healthy donor cells was performed by plotting each patient cell at the coordinates of the normal cell with the highest prediction score correlation (Fig. 2b).

Annotation of host and donor single cells

To annotate single cells from the patient 10 relapse bone marrow sample for their origin (this patient received an allogeneic stem cell transplantation prior to relapse), we quantified SNVs specific to the host or donor genome in each single cell. We first identified all SNVs in the relapse bone marrow exome sequencing dataset, which represents a mixture of both genomes (n = 127,916; Extended Data Fig. 3b; see also the copy-number analysis above). For each SNV, we then quantified its B allele frequency in both the germline and relapse bone marrow sample. By applying thresholds on both frequencies, we identified variants that are informative for each genome (Extended Data Fig. 5e). We further removed variants that were located within broad copy-number alterations on chromosomes 3, 6 and 9, as well as on chromosomes X and Y. A total of 56,155 SNVs were identified in this manner, with 5,989 (10.7%) being homozygous in both genomes (that is, host A/A and donor B/B, or host B/B and donor A/A) and therefore informative for both alleles. These SNVs were quantified in the single-cell transcriptome data of the diagnostic and relapse sample using samtools mpileup. For each base, information for cell barcode and UMI was obtained by setting the –output-extra option. We then aggregated coverage for all host- and donor-specific alleles across the genome for each single cell. Cell annotations for the patient 10 relapse bone marrow sample were obtained for cells with a coverage of at least 10 and a donor-specific allele coverage of less than 10% (host cells, n = 4,453) or greater than 90% (donor cells, n = 2,664; illustrated in Extended Data Fig. 5f). A small fraction of cells with donor-specific allele coverage between 10% and 90% potentially reflect cell multiplets and were removed from further analysis. In total, 94.9% of patient 10 relapse cells were annotated for their host/donor origin. As a control, none of the single cells from the patient 10 diagnostic bone marrow sample were classified as donor cells.

BPDCN signature generation and single-cell gene expression analysis

To generate a single-cell transcriptional signature specific for malignant BPDCN cells, we made two groups of cells: (1) cells classified as pDCs from healthy donors and (2) cells classified as pDCs from patients with marrow involvement. Cells classified as pDCs without progression mutations from patients without marrow involvement were also included in group 1 as they were similar to pDCs from healthy donors (Extended Data Fig. 7a). Cells classified as pDCs with progression mutations from patients without marrow involvement were not included in differential gene expression analysis because they were suspected circulating malignant BPDCN cells based on mutation and gene expression patterns. We randomly selected at most 50 cells per sample, so that the analysis would not be dominated by samples with a high number of pDCs. We then compared the two groups using the Seurat function FindMarkers and selected genes with twofold higher expression in the second group (log₂-transformed fold change > 1) and an adjusted P < 1 × 10⁻³⁰ (ref. ⁶²; Wilcoxon Rank Sum test, 45 genes; Extended Data Fig. 7b and Supplementary Table 3c). To identify malignant cells, we scored all 87,011 single-cell transcriptomes for this 45-gene signature using the Seurat function AddModuleScore (Extended Data Fig. 7c,d). We then defined malignant BPDCN cells as all cells classified as pDCs in patients with bone marrow involvement, as well as all cells (regardless of their initial classification) with a BPDCN signature score exceeding 0.5 (Extended Data Fig. 7f–h). To ensure that reclassification of a small proportion of cells as malignant BPDCN cells was justified, we checked marker gene expression, reasoning that the absence of canonical markers would support reclassification. Indeed, reclassified pro-B cells lacked CD19 and reclassified plasma cells lacked CD138 (Extended Data Fig. 7g). Using Seurat objects with scRNA-seq expression data and metadata (including cell type annotations and XV-seq mutation calls joined based on cell barcodes), we performed all downstream single-cell analyses in R with extensive use of the tidyverse⁶³.

In vitro differentiation of dendritic cells and UV exposure

HOXB8-FL cells were derived as described previously³³ from bone marrow cells of mice constitutively expressing Cas9 (Jackson Laboratory, 026179). HOXB8-FL cells were cultured in RPMI-1640 (Gibco, 11875093) supplemented with 10% FBS (Sigma-Aldrich, F2442), 1% penicillin–streptomycin (Corning, 30002CI), 50 ng µl⁻¹ mouse FLT3L (BioLegend, 550706) and 1 µM oestrogen (Sigma-Aldrich, E2758). Cells were resuspended in fresh medium every 2–3 days. For DC differentiation, HOXB8-FL cells were washed once in RPMI-1640, then resuspended in the same medium without oestrogen. For UV exposure on day 6 after-oestrogen withdrawal, cells were resuspended in PBS and exposed to the indicated doses of UV using an XL-1500 Spectrolinker. Cells were then resuspended in differentiation medium and analysed by flow cytometry on day 8 after oestrogen withdrawal. For CRISPR-mediated knockout of Tet2, Cas9-expressing HOXB8-FL cells underwent lentiviral transduction of sgRNA using the pLKO5.sgRNA.EFS.GFP vector (Addgene, 57822). Data were combined from two Tet2-targeting sgRNAs (GAATACTATCCTAGTTCCGAC and GAACAAGCTCTACATCCCGT). For controls, data were combined from non-transduced cells and cells transduced with sgRNA targeting a safe harbour region (ATGTACAACACAAACGAAGT). Tet2-sgRNA-induced indels were validated using PCR amplicon next-generation sequencing (Extended Data Fig. 10f). For flow cytometry analysis, differentiated HOXB8-FL cells were incubated for 10 min in Fc block (BD, 553141), then stained for CD11b Alexa Fluor 700 (BioLegend, 101222), CD11c PE/Cyanine7 (BioLegend, 117317), B220 APC/Cyanine7 (BioLegend, 103224), Siglec-H PE (BioLegend, 129605) and MHC-II PerCP/Cyanine5.5 (BioLegend, 107626). cDCs were defined as CD11c⁺CD11b⁺B220-MHC-II⁺ and pDCs were defined as CD11c⁺CD11b⁻B220⁺Siglec-H⁺ (Extended Data Fig. 10d,e). DAPI staining was used to exclude non-viable cells.

Reporting summary

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