A novel human fetal liver-derived model reveals that MLL-AF4 drives a distinct fetal gene expression program in infant ALL

Although 90% of children with acute lymphoblastic leukemia (ALL) are now cured1, the prognosis of infant-ALL (diagnosis within the first year of life) remains dismal2. Infant-ALL is usually caused by a single genetic hit that arises in utero: rearrangement of the MLL/KMT2A gene (MLL-r). This is sufficient to give rise to a uniquely aggressive and treatment-refractory leukemia compared to older children with the same MLL-r3–5. The reasons for disparate outcomes in patients of different ages with identical driver mutations are unknown. This paper addresses the hypothesis that fetal-specific gene expression programs co-operate with MLL-AF4 to initiate and maintain infant-ALL. Using direct comparison of fetal and adult HSC and progenitor transcriptomes we identify fetal-specific gene expression programs in primary human cells. We show that MLL-AF4-driven infant-ALL, but not MLL-AF4 childhood-ALL, displays expression of fetal-specific genes. In a direct test of this observation, we find that CRISPR-Cas9 gene editing of primary human fetal liver cells to produce a t(4;11)/MLL-AF4 translocation replicates the clinical features of infant-ALL and drives infant-ALL-specific and fetal-specific gene expression programs. These data strongly support the hypothesis that fetal-specific gene expression programs co-operate with MLL-AF4 to initiate and maintain the distinct biology of infant-ALL.

We next sought to determine the extent to which normal fetal gene expression programs contribute to the distinct molecular profile of MLL-AF4 infant-ALL. We compared bulk RNAseq for sorted human fetal liver (FL) hematopoietic stem and progenitor cell (HSPC) subpopulations previously generated in our lab 19 to a human adult bone marrow (ABM) HSPC RNA-seq dataset 20 . We carried out differential gene expression analysis between comparable subpopulations of FL and ABM HSPCs along the B lineage differentiation pathway (Fig. 1e).
The hematopoietic stem cell (HSC) subpopulation shows the greatest number of differentially expressed genes between FL and ABM (3,787 genes), reducing at each subsequent stage of B lineage differentiation (1,509 genes differentially expressed between FL committed B progenitors (CBP) and ABM common lymphoid progenitors (CLP)) ( Fig. 1e). A total of 5,709 genes were differentially expressed between FL and ABM in a least one HSPC subpopulation when we combined all differentially expressed gene lists (Fig. 1e, Supplementary Table 2).
We repeated the clustering analysis of the patient dataset based on these 5,709 genes and found that they were capable of separating MLL-AF4 infant-ALL from MLL-AF4 childhood-ALL ( Fig. 1f, Supplementary Fig. 2a). Comparing differentially expressed genes in both the normal and leukemic setting, we found 72 genes that were significantly upregulated in both normal  Table 2), including IGF2BP1, a member of the fetal-specific LIN28B gene expression pathway 21 , which has previously been reported to positively regulate HOXB4 expression 22 (Supplementary Fig.   2c). Together, these data suggest that the molecular profile of the human fetal HSPCs that form the target cells for leukemic transformation plays a role in determining the distinct gene expression profile of MLL-AF4 infant-ALL.
To test the hypothesis that to accurately model MLL-AF4 infant-ALL, the MLL-AF4 translocation should be expressed in human fetal HSPCs, we directly induced the most common t(4;11)/MLL-AF4 translocation in infant-ALL (with the MLL breakpoint in intron 11 13 ) in 13-15 post-conception week (pcw) human FL CD34+ cells by CRISPR-Cas9 genome editing. Edited samples (n=3) and biologically-matched mock-edited controls (n=3) were transferred to MS-5 co-cultures to facilitate expansion of successfully edited cells along the B lineage (designated CRISPR MLL-AF4+).
Finally, we wanted to ask whether inducing an MLL-AF4 translocation in human FL gave rise to a model that specifically recapitulated the molecular profile of MLL-AF4 infant-ALL. The only humanized mouse model of MLL-AF4 ALL that has previously been published introduced a chimeric MLL-Af4 fusion gene into human neonatal (cord blood (CB)) HSPC (hereafter referred to as CB MLL-Af4+ ALL) 27 . We hypothesized that it may recapitulate MLL-AF4 childhood-ALL, and could be used as a comparison to CRISPR MLL-AF4+ ALL.
To examine the fetal and post-natal gene expression programs that are key to determining the age-related differences between MLL-AF4 ALLs, we used the 139 genes up-or downregulated in both FL (compared to ABM) and MLL-AF4 infant-ALL (compared to MLL-AF4 childhood-ALL) (Supplementary Table 2). Clustering analysis based on this core gene list showed that, while CRISPR MLL-AF4+ ALL was similar to MLL-AF4 infant-ALL patients, CB MLL-Af4+ ALL clustered away from MLL-AF4 infant-ALL and closer to MLL-AF4 childhood-ALL patients (Fig.   5a). To explore this in more detail, we carried out differential gene expression analysis between CRISPR MLL-AF4+ ALL and CB MLL-Af4+ ALL, followed by Gene Set Enrichment Analysis (GSEA). We found that CRISPR MLL-AF4+ ALL was significantly enriched for genes upregulated in both FL HSPCs and MLL-AF4 infant-ALL compared to CB MLL-Af4+ ALL (Fig.   5b).
Comparing MLL-AF4 binding at promoters genome-wide in both models, we found that MLL-AF4 in CRISPR MLL-AF4+ ALL showed greater enrichment (normalized ChIP-seq reads/bp) at the promoters of infant-ALL-and FL-specific genes compared to MLL-Af4 in CB MLL-Af4+ ALL (Fig. 5c). However, at all other genes, MLL-AF4/MLL-Af4 enrichment was comparable ( Fig. 5c). At iALL-and FL-specific genes IGF2BP1 (Fig. 5d) and HOXB4 (Fig. 5e), we observed an MLL-AF4 peak in CRISPR MLL-AF4+ ALL but not in CB MLL-Af4+ ALL. These data suggest that MLL-AF4 may play an active role in maintaining fetal gene expression programs in infant-ALL. Increased levels of H3K79me2 are a commonly used marker of MLL-AF4 activity 23,28 . Therefore, using one of the unique features of our model, we carried out H3K79me2 ChIP-seq for the first time in identical primary human FL HSPC before and after leukemic transformation. We observed increased levels of H3K79me2 at MLL-AF4 peaks in FL-and iALL-specific genes such as IGF2BP1 (Fig. 5d) and HOXB4 (Fig. 5e) in CRISPR MLL-AF4+ ALL, further suggesting that MLL-AF4 actively maintains the expression of these fetalspecific genes in MLL-AF4 infant-ALL.
The mechanisms by which the same MLL-r driver mutation could cause more aggressive disease and worse outcomes in infant-ALL compared to childhood-ALL have always been unclear. We hypothesized that there must be intrinsic biological differences between infant-ALL and childhood-ALL blasts, unrelated to the driver mutation, that underlie these age-related differences. Here, we identify the unique molecular profile of MLL-AF4 infant-ALL using primary patient data. Reasoning that this profile drives the distinct phenotype of infant-ALL, we set out to identify factors that could explain it. We find that maintenance of fetal-specific gene expression programs account for a large proportion (~40%) of the unique molecular profile of MLL-AF4 infant-ALL, suggesting that it is the specific fetal target cell(s) in which it arises that provide the permissive cellular context for aggressive infant-ALL.
Human fetal HSPCs are more proliferative than ABM HSPCs 32,33 , and they differentiate down distinct developmental pathways 34,35 , some of which are virtually absent in adult life. Therefore, maintenance of fetal HSPC characteristics provides a possible explanation for the highlyproliferative, therapy-resistant nature of infant-ALL. One of the biggest challenges to understanding the biology of infant-ALL and developing novel, more effective therapies has been the lack of pre-clinical models 36 that capture the unique characteristics and aggressive nature of the disease. By targeting a t(4;11)/MLL-AF4 translocation to primary human FL HSPCs, we have created the first bona fide MLL-AF4 infant-ALL model. Our results finally confirm that a human fetal cell context is permissive, and indeed probably required; to give rise to an ALL that recapitulates key phenotypic and molecular features of poor prognosis MLL-AF4 infant-ALL. CRISPR MLL-AF4+ mice represent a previously lacking model in which the function of MLL-AF4 can be investigated in the appropriate human fetal cell context. Moreover, because CRISPR MLL-AF4+ cells were generated by CRISPR-Cas9 genome editing, they express both MLL-AF4 and the reciprocal AF4-MLL at physiological levels. Therefore, CRISPR MLL-AF4+ ALL also provides an opportunity to explore the contribution of the reciprocal fusion protein during leukemogenesis, which has been a topic of debate in the MLL-r ALL field 37,38 . Finally, the infant-ALL-like features of CRISPR MLL-AF4+ ALL make this an important model for future preclinical testing of novel therapies. To our knowledge, we are the first to report CNS disease in an MLL-AF4 mouse model, which is a common clinical feature of infant-ALL that can lead to CNS relapse 4 . Therefore, the ability of novel treatments to eradicate blasts from the CNS is an important consideration, and this can now be tested in CRISPR MLL-AF4+ ALL.

CRISPR-Cas9 MLL-AF4 translocation
CRISPR-Cas9 genome editing was carried out using a previously described protocol 40 . MLL and AF4 sgRNAs (Synthego) were first tested for editing efficiency individually in FL CD34+ cells. Cryopreserved CD34+ cells from a single primary human FL sample were thawed and placed into suspension culture at a density of 2.5x10 5 cells/ml in StemLine II (Sigma) supplemented with SCF (100ng/ml), FLT3L (100ng/ml) and TPO (100ng/ml) (Peprotech) for 12 hours. Cells were harvested and electroporated with either (i) Cas9 protein (IDT) only or (ii) a Cas9/sgRNA RNP using a Neon TM Transfection System (Thermo Fisher). Electroporated cells were placed into fresh suspension culture media to recover overnight. Cells were harvested and bulk genomic DNA was extracted using a DNeasy Blood and Tissue Kit (Qiagen). A ~1kb region of DNA around the target cut site was amplified by PCR and Sanger sequenced (Eurofins). Sanger sequencing traces from samples edited with RNPs were compared to traces from Cas9 only controls using the ICE Analysis online tool (Synthego, https://ice.synthego.com). Editing efficiency is reported as the percentage of indels detected ( Supplementary Fig. 3a).
Electroporated cells were placed into fresh suspension culture media to recover overnight before subsequent in vitro culture and in vivo transplantation experiments.
Xenograft transplantation [8][9][10][11][12] week old female NSG mice were sub-lethally irradiated with two doses of 1.25Gy six hours apart (2.5Gy total) and injected via the tail vein with 25,000-35,000 edited FL cells ( CRISPR MLL-AF4+, n=3; Cas9 control, n=5; or Cas9 plus MLL-sgRNA control, n=1) plus 30,000 wild-type, unedited, sex-mismatched FL CD34+ carrier cells. Engraftment was monitored by peripheral blood sampling every 3 weeks. Human CD45+ cells were sorted from peripheral blood samples to carry out MLL-AF4 and AF4-MLL RT-qPCR for the detection of successfully edited cells. Animals were monitored regularly using a standardized physical scoring system, and any mouse found to be in distress was humanely killed. Mice were considered leukemic if they met at least 3 of the following criteria: (i) overt signs of disease (hunching, lack of movement, weight loss, paralysis), (ii) splenomegaly, (iii) PB blast count over 50%, (iv) peripheral organ infiltration, (v) detection of the MLL-AF4 translocation in both BM and spleen.

Flow cytometry
Cells were stained with fluorophore-conjugated monoclonal antibodies in PBS with 2% FBS and 1mM EDTA for 30 minutes and analyzed using BD LSR II or Fortessa X50 instruments.
Antibodies used are detailed in Supplementary Table 5. Analysis was performed using FlowJo software where gates were set using unstained and fluorescence minus one controls.

Histopathology
On termination, samples of ~0.5-1cm 2 were taken from the spleen and liver of CRISPR MLL-AF4 + and Cas9 control mice and fixed in 10% formaldehyde. After fixation, tissues were processed and paraffin embedded. 4µm paraffin sections we cut onto Superfrost Plus adhesive slides,

RT-qPCR
Total RNA was extracted from cells using an RNeasy Micro Kit (Qiagen). cDNA was generated from polyA mRNA using a SuperScript III kit (Invitrogen). qPCR was carried out on cDNA using SYBRGreen master mix (Thermo Fisher) and a QuantStudio3 Real-Time PCR System (Thermo Fisher). For list of qPCR primers used see Supplementary Table 5.

RNA-sequencing
Approximately 3x10 5 CD45+CD19+ cells were sorted from the bone marrow of 3 primary CRISPR MLL-AF4 + recipient mice and 3 control primary recipient mice (Cas9 control, n=2; Cas9 plus MLL-sgRNA, n=1). Total RNA was extracted using an RNeasy Mini Kit (Qiagen). Poly(A) purification was conducted using the NEB Poly(A) mRNA magnetic isolation module as per the manufacturer's protocol. Library preparation was carried out using the Ultra II Directional RNA Library Prep Kit (NEB, E7765). RNA libraries were sequenced by paired-end sequencing using a 150 cycle high output kit on a Nextseq 500 (Illumina). RNA-seq protocols for sorted subpopulations of FL HSPC have been previously described in 19 .

IgH rearrangement analysis
Samples were screened for IgH complete (VH-DH-JH) and IgH incomplete (DH-JH) rearrangements using BIOMED-2 protocols to detect clonality. DNA was extracted from cells from the bone marrow of 3 primary CRISPR MLL-AF4 + recipient mice. IgH rearrangements were analyzed as described in 35 .

ChIP-sequencing
The full protocol is described in 31 . In short, up to 5x10 7 cells were sonicated (Covaris) following the manufacturer's protocol and incubated with antibody overnight. Magnetic protein A and G beads (ThermoFisher Scientific) were used to isolate antibody-chromatin complexes.
Antibodies used are detailed in Supplementary Table 5. Beads were washed three times using a solution of 50mM HEPES-KOH (pH7.6), 500mM LiCl, 1mM EDTA, 1% NP40 and 0.7% sodium deoxycholate and once with Tris-EDTA. Samples were eluted and Proteinase K/RNase A-treated. Samples were purified using a ChIP Clean and Concentrator kit (Zymo).
DNA libraries were generated using the NEBnext Ultra DNA library preparation kit for Illumina (NEB). Libraries were sequenced by paired-end sequencing using a 75 cycle high output kit on a Nextseq 500 (Illumina).

NGS analysis
For RNA-seq, following sequencing, QC analysis was conducted using the fastQC package (http://www.bioinformatics.babraham.ac.uk/projects/fastqc). Reads were mapped to the human genome assembly using STAR. The featureCounts function from the Subread package was used to quantify gene expression levels using standard parameters. This was used to identify differential gene expression globally using the edgeR package. Differential gene expression was defined by an adjusted p-value (FDR) of less than 0.05. Infant ALL RNA-seq datasets were analyzed as described previously 35 .
To derive a FL vs ABM gene signature, bulk RNA-seq for sorted subpopulations of FL HSPC 19 were compared to matched sorted subpopulations of ABM HSPC 20 Table 2).
For ChIP-seq, quality control of FASTQ reads, alignment, PCR duplicate filtering, blacklisted region filtering and UCSC data hub generation was performed using an in-house pipeline (https://www.biorxiv.org/content/10.1101/393413v1) as described. The HOMER tool makeBigWig.pl command was used to generate bigwig files for visualization in UCSC, normalizing tag counts to tags per 1x10 7 . ChIP-seq peaks were called using the HOMER tool findPeaks.pl with ChIP input sample used to estimate background signal. Gene profiles were generated using the HOMER tool annotatePeaks.pl.

Statistics
Two-tailed Mann-Whitney, Log-rank (Mantel-Cox) tests and ANOVA followed by multiple comparisons testing were used to compare experimental groups as indicated in the figure legends. Statistical analyses were performed using GraphPad Prism v7.00 or R v4.0.1. Data are expressed as mean ± SEM unless otherwise indicated.

Data availability
Further information and requests for resources and reagents may be directed to and will be fulfilled by the corresponding authors, Dr Anindita Roy (anindita.roy@paediatrics.ox.ac.uk) and Dr Thomas A Milne (thomas.milne@imm.ox.ac.uk).
The accession number for the RNA-seq and ChIP-seq data generated during this study is NCBI GEO: XXXXX