Markus Elsner

A modENCODE snapshot

Since 2007 a consortium of research groups has been studying the genomes of two model organisms, the fruitfly Drosophila melanogaster and the nematode worm Caenorhabditis elegans, in a project called model organism encyclopedia of DNA elements (modENCODE)1. The latest results from this project were described recently in two papers in Science2,3 and a suite of companion papers in Nature and Genome Research (http://blog. modencode.org/papers). The studies report both massive genome-scale data sets and analytic strategies for data integration. They substantially increase the annotated fractions of the fly and worm genomes and provide a wealth of data for understanding these model organisms and for developing new bioinformatic methods. Here we provide an overview of the data and some perspective from scientists on challenges for the field. The goal of modENCODE is to catalog sequence-based functional DNA elements in the fly and worm genomes. Such a catalog may be used to study regulatory networks and other emergent properties of the genomes, and, perhaps, to better understand the human genome. The project also seeks to generate experimental reagents for use by the research community. A summary of the new data sets is presented in Tables 1 and 2. To increase the number of functional genomic regions discovered, the studies analyzed organisms at different developmental stages. For the fly, ~700 data sets were generated from whole embryos, larvae and adult female and male insects as well as from a few cell lines and tissues. For the worm, ~240 data sets covered all major developmental stages along with some mutants, isolated tissues and animals exposed to pathogens. For both organisms, microarrays and sequencing were used to characterize gene expression, the binding sites of transcription factors and other proteins associated with DNA, origins of DNA replication, nucleosome turnover rates, salt-fractionated chromatin, the genomic locations of nucleosomes and the sites of different histone modifications. Looking forward, projects similar to modENCODE now seem feasible for studying other organisms and a broad range of biological problems. What will be the major challenges of such projects? Not sequencing, says Jun Wang, executive director of BGI in Shenzhen, China. He estimates that generating the equivalent of the fly modENCODE data set using today’s technology could take less than 2 months, including less than a month for library construction and a month for sequencing (although in practice more time may be required if several replicates are necessary). The main technical barrier, he says, will be preparing large numbers of samples from different tissues, developmental stages and conditions. A major challenge will be data integration, says Tom Gingeras of Cold Spring Harbor. For instance, robust integrative approaches are needed that combine genomic, transcriptional, regulatory and epigenomic signals, according to Olga Troyanskaya of Princeton University. Roded Sharan, a computational biologist at Tel Aviv University, agrees, adding that integrative analysis is required to identify an organism’s signaling and regulatory pathways and to elucidate how they vary over time and across cell types. He notes that current algorithmic work is focused on analyzing at most a few networks at a time and will have to be significantly scaled up to understand the complex developmental programs of fly or worm. Overall, says Troyanskaya, existing bioinformatic methods are not adequate, suggesting that novel ways of conceptualizing problems may be needed. New methods are needed to deal with the heterogeneity of the data types, to correct for technical and experimental biases and to detect biological signals hidden in experimental noise. Moreover, the sheer volume of data will require new approximation algorithms, computational infrastructure and strategies for disseminating the results. As a result of modENCODE, the catalog of DNA elements has grown larger, but many questions remain unanswered. The translation of annotated genomes into systems-level descriptions of the fly and the worm is a long-term goal. In worm, the number of candidate noncoding RNAs has increased severalfold, up from 1,061 at the start of the project, but the biological roles of these RNAs are not yet clear. Moreover, “pervasive post-transcriptional regulation of gene expression emerges as a theme from the modENCODE data,” says Thomas Sandmann, a fly geneticist at the German Cancer Research Center. In worm, ~22,000 genes were found to generate ~65,000 different transcripts; in fly, 74% of the ~17,000 genes showed at least one transcript isoform that differed from previous annotations. “Frankly,” says Sandmann, “we don’t have a good idea what this complexity is good for or how it works.” Other open questions involve the evolutionary conservation of DNA functional elements. Manolis Kellis, a member of the modENCODE consortium, explains that a “next step is tackling the comparative analysis of fly and worm to each other and to human, to understand the conservation of the regulatory principles learned, and the relevance of our results to the study of human biology and disease.” Finally, also in progress is ENCODE—a sister project analyzing functional elements in humans—which published results from a pilot project several years ago and is now progressing into its next stage of analysis across the entire human genome.

Single-stranded siRNAs for in vivo gene silencing

1063 The discovery of RNA interference (RNAi) in the late 1990s was almost immediately recognized as providing a general mechanism for specifically targeting thousands of disease-associated genes. But the difficulty of delivering the necessary RNA oligomers to cells in vivo has vexed the field ever since. Two recent papers in Cell1,2 now suggest that single-stranded RNAs (ssRNAs) could be a valuable alternative to the more commonly used double-stranded molecules. Double-stranded, small interfering RNAs (ds-siRNAs) are the most efficient way of inducing RNAi, but they do not readily cross cellular membranes and require special delivery vehicles to gain access to the cytoplasm of cells. The need to formulate ds-siRNAs not only increases the complexity of designing siRNA-based therapies but has also been the source of many of the adverse effects observed in clinical trials of RNA drugs. In addition, both strands of the double-stranded molecule can, in principle, be incorporated into the RNA-induced silencing complexes that mediate RNAi. Although methods to guide the strand selection toward the desired RNA strand have been developed over the years, researchers remain worried about off-target effects and associated toxicities. In contrast, ssRNAs do not need to be formulated to enter cells and can be delivered in a simple saline solution. By definition, they also eliminate the possibility of off-target effects due to erroneous strand incorporation. Despite these potential advantages, ssRNAs have not been widely used because they are substantially less potent inhibitors of gene expression than their double-stranded counterparts and, even more importantly, are very unstable in serum and in cells. Lima et al.1 and Yu et al.2 now show how chemical modifications can be used to increase both the stability and the potency of ss-siRNAs for in vivo applications. The introduction of up to 14 phosphorothioates, alternating 2′-fluororibose and 2′-methoxyribose and additional 2′-methoxyethylribose modifications increases the half-life of the ssRNA in cell homogenates from less than 2 minutes to more than 8 hours. Both papers show that the modified ss-siRNAs engage the cellular RNAi machinery and that the interaction with the key AGO2 protein depends on the presence of a 5′ phosphate. To overcome the rapid dephosphorylation that RNA undergoes in vivo, the authors1,2 introduce a metabolically stable 5′-(E)-vinylphosphonate modification. Similarly to siRNAs, ss-siRNAs will lead to degradation of their target mRNAs when they are perfectly complementary but will only inhibit translation when they contain a central mismatch. In mice, efficient silencing of the target gene was observed in the liver after either intravenous or subcutaneous delivery1 and in the brain after direct intraventricular infusion of the ss-siRNA2. In experiments in a Huntington’s disease model, Yu et al.2 showed that they can design ss-siRNAs that are >30-fold more active against the diseasecausing allele of mouse Htt (huntingtin) than against the normal allele, reducing the abundance of the disease-causing protein by up to ~80%. Lima et al.1 targeted three different mRNAs (PTEN, Factor VII and ApoC) in wild-type mice, reducing the liver transcript abundance by up to ~70%, especially when an additional lipid modification was introduced that increased targeting to the liver. Although these papers introduce ss-siRNAs as a potentially useful therapeutic option, it remains to be seen how advantageous these agents will prove to be in practice. Even though the chemically modified ssRNAs are much more effective than the unmodified versions, the doses required for efficient gene silencing are still relatively high. For systemic treatment, doses between 100 mg and 300 mg per kilogram of body weight were needed to achieve a meaningful knockdown of genes in the liver, and only very modest knockdown was observed in other tissues at these doses. Even with direct infusion to the brain, 300 μg a day was required. “Toxicity at these doses might be a worry, especially as very little is known about the metabolism of the modified bases. More work is certainly needed,” says Phil Zamore of the University of Massachusetts Medical School in Worcester. “Although the high doses are not necessarily a problem, more follow-up work that carefully characterizes toxicities and especially off-target effects of the ss-siRNA will be needed,” says Mark Kay of Stanford University in Palo Alto, California, adding that “one also needs to be careful when extrapolating from mice to man. With the lipid-based delivery vehicles, for example, many problems only became apparent in humans.” But as these are the first reports of successful use of ss-siRNAs in vivo, there may be plenty of room for improvement. How rapidly technology can develop has been seen with the ds-siRNA delivery vehicles, in which doses needed for gene knockdown in the liver of mice were brought down from >10 mg to <0.01 mg RNA per kilogram of body weight within a few years3. “It is still early days for ss-siRNAs,” says Kay, “and we don’t know what will work best in the end.”

Genome organization by the slice

volume 35 number 5 may 2017 nature biotechnology The human genome—46 long molecules of DNA bound to proteins and RNA—is not stored in the nucleus in a random fashion, like spaghetti in a bowl. Yet technologies to investigate its three-dimensional organization are few, and the ability to crosscheck experimental findings using independent methods has been very limited. Writing in Nature, Pombo and colleagues1 now report an entirely new way of mapping chromatin topology. The approach, called genome architecture mapping (GAM), works with very small numbers of cells and reveals additional levels of genomic organization beyond what is easily accessible with current techniques. That chromatin is not evenly distributed in the nucleus was apparent from early microcopy studies on stained nuclei, which showed distinct patterns of euchromatin and heterochromatin. Later, fluorescence in situ hybridization was applied to image the spatial relationships between a few specific genomic loci. In the early 2000s, the field was revitalized with the development of chromosome conformation capture (3C) methods, ranging from the original 3C itself, which detects interactions between two selected loci, to Hi-C, which investigates pairwise-interactions between all genomic loci2,3. These approaches harness the power of proximity ligation4 and high-throughput sequencing to generate twodimensional genome-wide maps of loci that are in physical proximity, which can then be computationally reconstructed into the threedimensional architectures most likely to fit the observed contacts. 3C methods have been dogged by concerns that the sample processing steps—cell fixation, DNA digestion, and proximity ligation to link adjacent DNA regions—introduce biases in the data. Indeed, in some instances, validation experiments with in situ hybridization produced data that were difficult to reconcile with 3C results5. Another drawback of 3C methods is that they cannot directly measure organizational levels higher than point-to-point contacts. The new GAM method also begins with a fixation step, but here the similarity to 3C ends. In the next step, ultramicrotome cutting and laser microdissection are used to collect 220-nm thin, randomly oriented slices of individual nuclei. The genetic material in each slice is sequenced, and three-dimensional information is extracted from the sequencing data by analyzing how often specific sequences co-occur in the same slice. “The approach is quite orthogonal to 3C methods, which is really important for the field,” says Job Dekker, a researcher at the University of Massachusetts Medical School and one of the inventors of 3C. “You do get stuck when you only have one way of doing an experiment,” notes Erez Lieberman Aiden, professor at Baylor College of Medicine and Rice University, in Houston. Pombo and colleagues1 apply GAM to human embryonic stem cells, as these have been extensively studied with 3C methods. Analyzing nuclear profiles from 400 cells, they find that GAM achieves a resolution of 30 kb, and a cosegregation analysis is consistent with previous Hi-C results, showing, for example, the existence of two nuclear compartments and of topologyassociated domains (defined by frequent interactions within, and few outside, the domain). “It is pleasing to see that the initial data are largely consistent with Hi-C data,” says Dekker. To identify specific interactions between individual genomic loci, the authors develop a more sophisticated model that distinguishes random contacts from specific interactions. The model, called SLICE, calculates how often random contacts between loci are expected to occur as a function of genomic distance based on the experimental data. It then identifies loci that are detected in the same nuclear slice significantly more often than would be expected if their contacts were random. SLICE takes into account various confounding factors, such as the detection efficiency of a genomic locus and genomic resolution. At the global level, Pombo and colleagues1 find a significant enrichment of connections between enhancers and active genes (defined previously by RNA-seq), which is especially pronounced at transcription start and end sites. In general, it is more difficult to detect interactions between three or more loci. In this respect, it is interesting that GAM detects threeway interactions between genomic regions at a resolution of hundreds of kilobases, roughly the size of topology-associated domains. A notable advantage of GAM is that it requires only small numbers of cells. And in principle it can be applied not only to dissociated cells, as shown by Pombo and colleagues1, but also to cells in fixed or frozen tissues, without the need to isolate the cells or culture them. “This might be the killer app for the technology,” says Aiden. Genome structure in cells in their native environment has been largely unexplored. “Especially for rare or difficult-toculture cell types, the method is of high interest,” explains Dekker. Importantly, the authors also show that GAM uncovers higher levels of genome association, such as radial distribution of chromosomes, sub-chromosomal regions, and areas of chromosomal compaction, which could only be deduced indirectly from 3C data. Although GAM is relatively labor-intensive and requires specialized skills and equipment, its potential to probe the hidden recesses of the nucleus, either independently or in combination with 3C and optical methods, appears vast. As Aiden notes, “One has to keep in mind that this is only the first report. Future improvements, such as cutting thinner slices, more sophisticated slicing strategies, or improved statistical analysis methods will certainly make the method even more powerful.”

Membrane channels built from DNA

125 DNA ‘origami’ is a technique for shaping long strands of DNA into nearly any threedimensional structure1. with their recent publication in Science, Langecker et al.2 reveal the potential of this approach for a new set of applications centered on membrane pores. The paper describes a transmembrane channel made entirely from DNA and DNA-bound cholesterol moieties. As DNA nanostructures can be readily engineered because of the relatively simple rules that govern their folding, membrane channels with customized properties could find uses as biosensors and nucleic acid or protein sequencers. In DNA origami, a single-stranded DNA, often obtained from the M13 phage genome, is folded by hundreds of ‘staples’. The staples are short, single-stranded oligonucleotides that are complementary to two or more sequences in the scaffold and can therefore crosslink distant sites. with the help of computer programs, the sequences of the staples needed to create even complex structures can be easily determined. The structures form spontaneously in a simple one-pot thermal annealing reaction when scaffold and staple DNAs are mixed. Previously, it was thought DNA nanostructures could only be used in aqueous solutions or surfaces. “I think everybody was surprised that a highly charged molecule like DNA can be made to penetrate a lipid membrane,” says Ulrich Keyser of the University of Cambridge (UK), whose group has developed a DNA-based channel to functionalize solid-state nanopores3. The inspiration for the design of the DNA transmembrane channel came from the bacterial protein pore α-hemolysin. Similar to α-hemolysin, the structure designed by Langecker et al.2 consists of a channel that spans the lipid bilayer and a barrel-shaped cap that binds to the surface of the membrane. The channel comprises five DNA helices that surround a central opening with a diameter of 2 nm. Membrane adhesion is mediated by 26 cholesterol moieties attached to the membrane-facing side of the cap. The authors demonstrate that the channel conducts an electrical current proportional to the potential that is placed across the membrane. Interestingly, the channel shows stochastic gating—fluctuations between an open and a closed conformation—similar to what is observed in many naturally occurring protein pores. The gating dynamics are amenable to engineering, as the insertion of a small heptanucleotide at a defined position within the channel substantially increases the time spent in the off state. Many different kinds of nanopores are currently under investigation for molecular detection applications, such as DNA sequencing. The data presented by Langecker et al.2 hint that DNA pores might be suitable for such applications by showing that the kinetics of the unfolding of DNA hairpins passing through the pore can be monitored and that different lengths of DNA can be distinguished. For practical uses, the properties of DNA pores will have to be fine-tuned for each specific application, but DNA nanostructures are especially easy to modify. “A key advantage of DNA-based nanostructures is the ease with which they can be engineered to improve on your basic design,” says Björn Högberg of the Karolinska Institute in Stockholm. “In addition to making defined changes to the DNA structure itself, one can place basically any functional group anywhere in the structure with high spatial precision.” Various heteroelements, ranging from small molecules (e.g., cholesterol2) and peptides4 to enzymes5 and carbon nanotubes6, have been positioned at precise locations within DNA nanostructures. In addition to in vitro applications, future uses might also be found in medical therapies. “One can imagine quite sophisticated drug delivery applications of DNA nanostructures. For the DNA channel presented here, one could think of coupling gating to the absence or presence of cellular proteins or RNAs in order to only target specific cell types,” suggests Keyser. Although DNA nanotechnology holds promise to provide molecular machines in the future, more development is needed before it can reach its potential. “A major problem is the availability of cheap methods to synthesize sufficient amounts of customized long-scaffold DNA,” says Keyser. Engineering of heteroelements also requires optimization. As Högberg notes, “we need to improve the precision of the placement of the heteroelements; compared to what is needed to build an active site of an enzyme, for example, we are still quite imprecise.”

High-speed imaging in a flow

841 sion engineering circuits in human cells. Khalil et al. use artificial zinc finger domains, which can be readily engineered to recognize various target DNA sequences, as the basis for a modular, customizable set of synthetic transcription factors (sTFs) and demonstrate how to build and tune these sTFs for higher-order synthetic gene circuits in yeast, a useful testing ground for circuits that might help bridge to studies in human cells. A second group, Wei et al., exploit a family of bacterial pathogen proteins known as effectors that can modulate host cell signaling and dampen immune responses. They use effectors to selectively modulate mitogen-activated, protein kinase–regulated signaling pathways, including those responding to osmolarity in yeast, and show that these effectors can be used to reshape signaling input and output behavior. They use the same effectors as a switch to toggle on and off the activity of human CD4+ T cells in response to a simple drug signal. Control switches like these might find application in adoptive immunotherapy. (Cell 150, 647–658, 2012; Nature 488, 384–388, 2012) SJ

Ribosomes reveal proteome complexity

Structural variation in the human genome can be discovered by comparing paired-end next-generation sequencing reads from an individual to the human reference genome. Hormozdiari et al. improve the accuracy of discovering variants in repetitive regions of the genome through simultaneous analysis of several genomes, rather than the one-by-one approach that has been the norm. The key to discovering structural variants lies in the identification of ‘discordantly mapped’ paired-end reads: pairs that match regions in the reference that are further apart or closer together than expected are indicative of an insertion or deletion, respectively. Existing approaches fall short in repetitive genomic regions when there are several possible discordant mappings for a read pair, indicative of several possible variants. Hormozdiari et al. describe algorithms, called CommonLAW, that harness discordant read pairs from multiple genomes to determine which of the discordant mappings is most likely. Applying the new algorithms to genomes sequenced in the 1000 Genomes Project reduced the number of false-positive calls of mobile element insertions by >20-fold and provided moderate improvements to calling deletions. (Genome Res. published online, doi:10.1101/gr.120501.111, 2 November 2011) CM ribosomes reveal proteome complexity

Barcodes galore for developmental biology

volume 36 number 10 oCTober 2018 nature biotechnology The emerging field of regenerative medicine has much to gain from basic research into the mechanisms of development, tissue regeneration and wound healing, all of which depend on the faithful execution of differentiation programs. Although these processes have been intensively studied for decades, a new class of methods based on Cas9 genome editing provides a welcome boost. In a recent paper in Science (Kalhor, R. et al. Science 361, eaat9804; 2018), Kalhor et al. show how genomic ‘scarring’ with Cas9 can be adapted to trace cell lineages during mouse embryo development. The use of Cas9 for lineage tracing was pioneered by Shendure and colleagues (McKenna, A. et al. Science 353, aaf7907; 2016) in zebrafish. Their work, as well as subsequent improved methods, relied on errorprone repair of the double-strand breaks induced by Cas9. The small insertions and deletions introduced during repair act as barcodes; if several sites are cut at different time points during animal development, the accumulated errors can be used to reconstruct lineage trees. The approach had previously been applied only in lower vertebrates and in cell culture. Because mouse development is more complex and more prolonged than that of the previously studied organisms, Kalhor et al. designed a system that increases the diversity of barcodes over longer time periods. To do so, they used so-called homing guide RNAs (hgRNAs), which direct Cas9 to a genomic locus encoding the hgRNA (Kalhor, R., Mali, P. & Church, G.M. Nat. Methods 14, 195–200; 2017). This system enables continuous barcode evolution, because errors introduced in the targeting sequence also change the expressed sgRNA, which retains the ability to target its genomic locus. Using a library of hgRNA constructs, Kalhor et al. generated a chimeric founder mouse, called mouse for actively recording cells 1 (MARC1), that carries 60 different hgRNA constructs in its genome. “Moving up the evolutionary ladder is a very important step and will allow addressing even more interesting questions,” says Jan-Philipp Junker of the Max Delbrück Center in Berlin, Germany, “and the use of hgRNAs is a creative way to obtain the necessary barcode complexity.” To study early development, the authors crossed MARC1 with mice constitutively expressing Cas9. The different hgRNA sites were modified at different rates. ‘Fast’ hgRNA sites were modified in almost all cells by E8.5, whereas ‘slow’ hgRNA sites were edited in only a minority of cells, even in adult animals. A group of intermediate hgRNAs that accumulated mutations throughout embryonic development were also observed. To validate the accuracy of the method, the authors reconstructed the first lineages in mouse development at embryonic day (E) 12.5. By sampling two tissues derived from each of the three basic developmental lineages, they accurately traced the differentiation of blastomeres into the embryonic inner cell mass and trophectoderm, as well as the subsequent differentiation of the inner cell mass into the primitive endoderm and epiblast, and of the trophectoderm into the placenta. Notably, a correct lineage tree could be constructed with as few as three hgRNAs in 50% of cases. If only fast and intermediate hgRNAs were considered, three hgRNAs were sufficient to construct a correct tree in 90% of cases. Kalhor et al. applied the method to address a longstanding question in mouse development. Both the anterior–posterior and the lateral axes are established in the embryonic neural tube by E8.5, although the order in which the axes emerge had been unclear. Dissecting the left and right sides of the cortex, cerebellum and tectum, the authors found that the barcodes of cells from the left and right side of each brain region were more similar to each other than to cells from any other brain regions, suggesting that the anterior–posterior axis is established before the lateral axes. In principle, the diversity of repair outcomes (theoretically, >1074 combinations of barcodes are possible for the 41 active hgRNAs in MARC1) is sufficient to uniquely label every single cell in an adult mouse as well as all the nodes of the developmental tree. Although the extent to which this diversity can be observed in practice remains to be seen, many open questions can clearly be addressed with this technique. “The logical next steps for the technology are to move to single-cell analysis to obtain lineage information of individual cell types and to couple it with inducible or tissue-specific Cas9 mice to investigate processes later in development or after experimental interventions,” says Junker. Beyond development, applications in tumor development and metastasis may also be possible. Eventually, cell-type-specific barcodes might even reduce determination of the mouse brain connectome to a sequencing problem.

Epigenome editing to the rescue

315 Epigenetic dysregulation of gene expression is a major contributor to many human diseases. Targeted alteration of the epigenome became possible with the advent of customizable DNA binding domains, and the approach was quickly tested in animal models1 and in clinical trials. In recent years, the development of easily programmable genome editors based on CRISPR–Cas9 has renewed interest in epigenome editing technologies and their therapeutic applications2. Now, a new study in Cell3 has demonstrated the power of using specific editing of epigenomic marks to reverse the effects of a genetic mutation. Epigenome editing makes use of the same customizable DNA binders (zinc finger proteins, TALEs or CRISPR–Cas9) that are used for genome editing or for general transcriptional activation or repression. But instead of being fused to a nuclease or to a transcriptional activator or repressor, the DNA binder carries an enzyme that puts in place or erases a specific epigenetic mark2. Although epigenome editors have been a boon to scientists investigating the mechanisms of epigenetic regulation, their utility for therapeutic purposes has not yet been tested. As a first step toward therapeutic epigenome editing, Liu et al.3 studied a Cas9based DNA demethylase in a model of fragile X syndrome. This condition affects about 1:3,600 males and is the most common cause of male intellectual disability. It is caused by a trinucleotide repeat expansion in the 5′ UTR of the FMR1 gene. In individuals with more than 200 of these repeats, the repeat region is hypermethylated, leading to formation of heterochromatin at the gene promoter and gene silencing. To reverse the hypermethylation, the authors designed single-guide RNAs that target a catalytically inactive Cas9, which has been fused to the catalytic domain of the DNA methylcytosine dioxygenase TET1, to the hypermethylated repeats. Testing the constructs in patient-derived induced pluripotent stem cells (iPSCs), they observed a 96% reduction in the methylation levels of the repeats and an almost complete restoration of FMR1 expression. “A surprising and important finding was that the reversal of repeat methylation was so closely linked to the removal of heterochromatin marks and the appearance of active chromatin marks at the promoter,” says Charles Gersbach, professor of biomedical engineering at Duke University in Durham, North Carolina. Off-target demethylation events were rare. Although ChIP-seq detected >1,000 sites that were at least transiently bound by the demethylase, only 29 of these loci showed substantial demethylation. Of the 28 off-target genes affected, most showed no change, and none showed more than a fourfold change, in expression levels. “One of the lessons that we learned from gene therapy is that we have to carefully assess safety of the treatments before going into a clinical trial,” says Angelo Lombardo of the San Raffaele Telethon Institute for Gene Therapy in Milan. “It is reassuring that the offtarget effects observed here are limited, but one needs to keep in mind that epigenetic off-target effects might be much more context-dependent than genomic off-target events.” Restoration of FMR1 expression was sufficient to rescue cellular phenotypes associated with fragile X syndrome. Neurons that differentiated from methylation-edited iPSCs retained close to normal FMR1 expression levels and showed none of the electrical hyperactivity of affected neurons. Similar results were obtained if iPSC-derived neurons were treated after differentiation, although in this case demethylation and expression restoration remained incomplete. The authors also tested whether reactivation of FMR1 is maintained in vivo by transplanting methylationedited neural precursor cells into the brains of mice. After three months, about half of the neurons derived from the implanted cells still expressed FMR1. Interestingly, at least in vitro, maintaining expression of FMR1 did not require sustained dCas9–Tet1 activity. When the authors expressed a Cas9 inhibitor, FMR1 expression was unchanged for at least two weeks. Further characterization of the therapeutic effects of dCas9–Tet1 was complicated by the limitations of the available mouse models. Inserting the extended repeats into mouse Fmr1 does not result in DNA hypermethylation or gene silencing. The limitations of the fragile X mouse models also make it difficult to assess whether post-natal reactivation of FMR1 will be sufficient to achieve a cure. “While this is an important first step, it remains to be seen how well this approach will translate into clinical applications. Delivery is an important issue as the dCas9– Tet1 fusion is too large for commonly used AAV [adeno-associated virus] vectors, although it should be possible to split the construct and deliver it in separate vectors,” says Lombardo. And Gersbach highlights the potential issue of continuous transgene expression: “Ideally, one would like to avoid long-term expression of the epigenome editing tool, but additional research is necessary to evaluate if a hit-and-run approach would work here.”

Mosaic mice ace functional genomics

volume 35 number 10 oCTober 2017 nature biotechnology an estimated 43 million people. The Next Generation Cassava Project is implementing genomic selection to systematically improve the yield and nutritional value of this African staple. As modern techniques are brought to bear on other crops, such as yams and millets, they will continue to enrich the world’s table. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests.

Erratum: Stabilizing prospects for a universal flu vaccine

In the version of this article initially published, on p.369, “European consumers” were said to account for “over €1.1 (