ReviewNeurodegenerative Disease

Mouse models of neurodegeneration: Know your question, know your mouse

See allHide authors and affiliations

Science Translational Medicine  22 May 2019:
Vol. 11, Issue 493, eaaq1818
DOI: 10.1126/scitranslmed.aaq1818


Many mutant mouse strains have been developed as models to investigate neurodegenerative disease in humans. However, variability in results among studies using these mouse strains has led to questions about the value of these models. Here, we appraise various mouse models for dissecting neurodegenerative disease mechanisms and emphasize the importance of asking appropriate research questions. In therapeutic studies, we suggest that understanding variability among and within mouse models is crucial for preventing translational failures in human patients.


Neurodegenerative diseases are common, largely untreatable, and certainly incurable and create a huge health and social burden worldwide. For example, in 2018, more than 50 million people worldwide had dementia, of which ~70% was caused by Alzheimer’s disease (AD) with an overall cost of $1 trillion (1). Currently, ~66% of those with dementia live in low- or middle-income countries (1). For neurodegenerative movement disorders, ~10 million people worldwide currently suffer from Parkinson’s disease (PD) (2). These diseases are not necessarily illnesses of older age, for example, type 1 spinal muscular atrophy (SMA) is the biggest single genetic killer of children under 5 years of age, affecting up to 1 in 10,000 newborns. A typical mid-life–onset disorder such as amyotrophic lateral sclerosis (ALS) has been described in children as young as 11 years of age (3). As with many other human diseases, each manifestation of neurodegeneration is likely to vary from person to person and will probably require a tailored combination of treatments for different patient subgroups.

For decades, mice have been the mammal of choice for modeling neurodegenerative diseases because of our ability to exquisitely manipulate their DNA. However, few genetically modified mouse models fully recapitulate the human condition or have yet been instrumental in delivering treatments for neurodegeneration. This is leading some researchers to doubt the utility of mouse models in the hunt for therapies and to turn instead to alternative approaches, such as induced pluripotent stem cells (iPSCs) derived from fibroblasts or other cells of patients with neurodegenerative diseases. However, much of the uncertainty regarding mouse models stems from expecting each model to completely mimic human disease. Yet, mice only live for 2 years or so and exhibit many differences from humans: from the genome to neuroanatomy and from the immune system to behavior. Uncertainty also arises from the apparent lack of reproducibility of results from mouse studies across laboratories. Confounding variability is evident not only across studies and across laboratories but also even within cohorts of the same mouse genotype being studied within one laboratory.

Clearly, a multiplicity of models, including genetically modified mice and three-dimensional cellular systems, is required to understand neurodegenerative diseases. For example, human iPSCs provide cellular models and robust in vitro readouts that could be used for high-throughput analysis such as drug screens, which are not feasible in mice. However, mouse models remain essential because they enable us to take a holistic approach over the life span of an animal, giving us access to in vivo systemic interactions, between cell types (for example, glia and neurons), tissues (for example, muscle and neurons), and whole animal systems (for example, the immune system and the nervous system). They also provide access to developmental, metabolic, and behavioral outcomes over the natural history of a disease. In particular, because many of the symptoms in human patients that we are trying to treat are behavioral (for example, memory impairments and deficits in cognition), it is essential to understand how molecular, cellular, and circuit changes manifest as functional changes in the behavior of the organism. Mice also allow us to model environmental effects and individual responses over a lifetime in a way that is impossible in cellular systems.

Here, we argue that we need to acknowledge the differences between mice and humans and, most importantly, to tailor each mouse model to the research question being asked to elucidate mechanisms of neurodegeneration. We ask why mice with the same genetic mutation have divergent phenotypes in pathological and behavioral sequelae and how we can turn this to our advantage. What we have hitherto seen as a problem and a lack of reproducibility we now suggest may be an important opportunity for understanding gene interactions and interactions of genes with the environment, which may also underlie the variability of disease phenotypes observed in humans. Last, we discuss briefly the failure to translate mechanistic findings in mice to therapies in humans and suggest that this reflects the large diversity in the human clinical population that is currently not understood nor captured in mouse studies. Embracing and understanding such variability in animal models may therefore be an essential step in the translational process.


Despite the remarkable similarities between us, mice are not mini humans and humans are not large mice. There are species-specific differences in every aspect of our biology, and these can be extremely helpful for highlighting key issues in human neurodegeneration.

At the genome level, ~90% of mouse and human DNA falls into regions of conserved synteny, distributed between the 46 human chromosome pairs and the 40 mouse chromosome pairs. Both species have similar numbers of protein coding genes, but roughly 1% of mouse genes appear to have no human equivalent and vice versa (4). These genomic differences can be an issue for mouse models. For example, approximately 1 in 800 babies are born with trisomy of human chromosome 21 (Hsa21), resulting in Down’s syndrome, which gives a greatly increased risk for early-onset AD (5). The long arm of Hsa21 has three regions of conserved homology with mouse chromosomes 10, 16, and 17 (Mmu10, 16, and 17), and partial trisomies of these chromosomes have been used to model Down’s syndrome (Fig. 1) (68). However, four Hsa21 protein-coding genes appear to be absent in the mouse genome, and at least four mouse genes found in the regions of homology seem absent in humans (9), so mouse partial trisomies cannot fully model human trisomy 21. The Tc1 mouse model, which carries a partial copy of Hsa21, partly overcomes this issue, but human gene regulation and protein interactions may not be the same in a mouse cellular environment, and this chromosome is lost stochastically from mouse cells (only ~70% of Tc1 mouse brain nuclei contain Hsa21) (7). No mouse can faithfully model Down’s syndrome, but analysis of phenotypes across a variety of mouse models of this disease has been valuable for teasing apart the complex biology of this human chromosomal disorder, providing that we take into account the characteristics of each model and ask specific questions about refined aspects of the syndrome that can be addressed in these models (6, 10).

Fig. 1 Differences in human and mouse genetic loci that may affect mouse models of neurodegeneration.

Humans and mice are separated by only 75 million years of evolution, yet there are many genetic differences between these species. This figure shows examples of genetic changes, not found in mice, that give rise to neurodegenerative disease in humans. Examples include human-specific repeat sequences that may mutate a gene such as the SVA (SINE-VNTR-Alu; red box) repeats that become inserted into the noncoding DNA of the FUKUTIN gene or the TAF1 gene, resulting in the neurodegenerative diseases Fukuyama muscular dystrophy or X-linked dystonia parkinsonism, respectively (exons, gray boxes) (24, 25). Gene regulation may be different between humans and mice as exemplified by OSTEOCRIN, which is expressed in the human but not in the mouse nervous system because primates have acquired an enhancer (red box) leading to neuronal expression (21). Splicing patterns may be different between human and mouse as occurs in the normal expression of the TAU gene (20): Mice have two to four isoforms per Tau gene (blue boxes), whereas humans have three to four isoforms per TAU gene (red boxes). Cryptic or skiptic DNA elements may be associated with disease in humans but do not exist in mice, such as the cryptic polyadenylation site in STATHMIN2 that is found in ALS (26). Gene copy number can be different between the two species, as with the single Smn gene found in mouse and its paralogous SMN1 and SMN2 genes in humans; mutation of SMN1 gives rise to SMA (11). Mouse genes are arranged over 19 autosomes and two sex chromosomes, whereas human genes lie within 22 autosomes and two sex chromosomes. The most common genetic form of cognitive disability, Down’s syndrome, arises from trisomy 21 and is modeled in mice either by creating an animal with the human chromosome (7) or by engineering duplications into the three mouse chromosomes 10, 16, and 17 that have regions of homology with human chromosome 21 (8). At the population level, we usually study inbred mouse strains that are genetically homogenous compared to human populations that are genetically diverse.

Credit: A. Kitterman/Science Translational Medicine

Gene copy number may differ between humans and mice. For example, humans have one copy of the SMN1 (survival motor neuron 1) gene and up to four copies of SMN2, which are paralogs that lie within a duplicated region on chromosome 5 (Fig. 1). SMN2 has a single base pair difference from SMN1 that results in alternative splicing such that only up to 20% of SMN2 transcripts encode functional protein. Mutations in SMN1 result in SMA, a neurodegenerative disease in which the lower motor neurons gradually die, and the rate of death is inversely related to the amount of functional SMN2. In the most severe form of the disease, children die before the age of four. Mice have been essential for studying SMA, but the mouse has only one Smn gene. Heterozygous Smn mice carrying one deleted allele and one wild-type allele are fully viable, whereas homozygous null mice die early in embryogenesis (11). Importantly, SMA has been modeled successfully by placing human SMN2 into Smn null mice (11). Now, a range of genetically different SMA mouse models exists, each carrying different patient mutations; some SMA mouse models carry an Smn gene that is mutated to resemble human SMN2. Each model has different advantages and disadvantages, such as reduced disease severity but a longer time in which to study disease processes (11). These models, including conditional mutant mice, have helped to dissect the timing of neuronal loss and to identify which neuronal populations are at risk. Furthermore, these mouse models have been critical for the development of therapies both conventional and genetic (11, 12). These include an adeno-associated virus gene therapy approach, now with U.S. Food and Drug Administration (FDA) approval for a phase 2 clinical trial, and the antisense oligonucleotide drug Spinraza, which in 2016 was approved by the FDA as the first therapy for SMA (13).

With respect to gene expression, the ENCODE (Encyclopedia of DNA elements) project has examined regulatory regions of mouse and human genomes in multiple tissues and cell types and has demonstrated transcript profiles of a large number of human and mouse genes. The ENCODE project has revealed that much of the cis-regulatory landscape, including deoxyribonuclease 1 (DNase 1)–hypersensitive sites, are different between the two species (1417). Mouse and human also have different splice variants, and humans produce a greater number of splice isoforms (average of 3.4 isoforms per gene) compared to mice (2.4 isoforms per gene) (18). Splice isoform differences are particularly important in diseases such as the most common genetic form of frontotemporal dementia (FTD) (19), which is caused by mutations in the MAPT (microtubule-associated protein tau) gene that encodes the protein TAU (Fig. 1). Transgenic mice have been made using human MAPT complementary DNAs (cDNAs), and these have been helpful for pinning down mechanism but cannot address the abnormalities found in FTD that are caused by aberrant TAU splice isoform ratios. Now, however, mice with the complete human MAPT genomic region exist, and these animals carry the various human TAU splice variants. Importantly, for developing therapies, such mice respond to antisense oligonucleotide treatment by switching isoforms, thus giving insight into tauopathies and new models for developing therapies for FTD (20).

Non–protein-coding DNA may, or may not, be well-conserved between human and mouse. For example, osteocrin is an activity-dependent secreted factor that is induced by membrane depolarization of human, but not of mouse, neurons because of a primate-specific enhancer that has repurposed the expression of the gene encoding this protein (21). Mouse noncoding RNAs largely diverge in sequence from those of human (22). Mutations in non–protein-coding gene regions can be causative for neurodegeneration. For example, in mice, transfer RNA (tRNA) mutations can lead to neurodegeneration of the cortex, cerebellum, and hippocampus (23). In humans, mutation of the 3′ untranslated region (3′UTR) of the FUKUTIN gene causes Fukuyama muscular dystrophy; mutation within an intron of the human TAF1 (TATA-box binding protein associated factor 1) gene results in X-linked parkinsonism dystonia (24, 25). In both of these human disorders, the mutation is an insertion of a human-specific SVA repeat that is not found in mice. Cryptic or skiptic DNA elements may be associated with disease in humans but may not exist in mice, such as the cryptic polyadenylation site in STATHMIN2 (a neuronal growth associated factor) that is present in sporadic and some genetic forms of ALS (Fig. 1) (26). Noncoding changes also modulate disease outcomes: for example, by 2017, more than 3000 different genome-wide association studies (GWAS) reported >30,000 single-nucleotide polymorphism (SNP)–disease associations, the vast majority of which lie in noncoding regions (27), indicating the importance of gene regulation in disease including neurodegenerative disease (23, 28).

Unexpectedly, small differences in human and mouse amino acid sequences can have large effects on phenotype. Humans with three copies of the wild-type APP gene encoding amyloid precursor protein succumb to early-onset AD, whereas three copies of the wild-type mouse App gene do not lead to similar amyloid deposition in mouse brain (5). Human and mouse APP are highly homologous, but the 17–amino acid differences between them include three key residues that affect how APP protein forms amyloid deposits, and this may, in part, explain why human and mouse respond differently to having three “doses” of the wild-type APP gene (29). Studying these differences may help to shed light on how the primary amino acid sequence of APP affects amyloid deposition. Similarly, human and mouse superoxide dismutase 1 (SOD1) proteins have a few key amino acid differences, likely making the human protein more prone to aggregation than the mouse equivalent (30). The SOD1D83G mutation causes autosomal dominant ALS in humans, but in mice, the picture is very different. A mouse with the identical nucleotide and amino acid mutation (Sod1D83G) only exhibits progressive motor neuron loss when homozygous for the mutation. This mouse shows only mild upper motor neuron loss (~20% by 29 weeks of age) and a similar loss (23%) of lower motor neurons by 15 weeks of age. Remarkably, after these time points, motor neuron loss stops and motor neuron numbers remain stable for at least a year, although accompanied by a severe peripheral neuropathy (31). Dissecting the phenotype of the Sod1D83G mouse has revealed that motor neuron axonal degeneration and cell body degeneration are separate phenomena (at least initially). Potentially, this provides insights into the human disorder SOD1-ALS and the separate loss-of-function and gain-of-function effects arising from a single mutation, although the mouse does not fully model human ALS (31).

Cellular and metabolic pathways may be common to both species but may have a different importance in each. Mice lacking Hexa (the gene encoding the hexosaminidase subunit alpha) were generated to model the lysosomal storage disorder Tay-Sachs disease, which in humans is caused by loss of HEXA function. In finding out why the null mouse does not overtly model this deadly childhood disorder, researchers teased out an alternative metabolic pathway that is present in mice but of less importance in humans, although it could nevertheless still be used as a possible therapeutic route for treating this disease (32). This highlights the important point that humans and mice have evolved to fill different niches, and so, major biochemical, metabolic, and physiological pathways may be different between us. For example, compared to humans, mice have a higher specific metabolic rate (metabolic rate per gram of tissue) and larger deposits of brown fat that are essential for thermoregulation (33). Other differences include cell membranes that have more polyunsaturated fatty acids in mouse than in human (34) and the different dietary requirements, for example, mice can synthesize vitamin C, whereas humans cannot (33). Fundamental differences in the mouse and human immune systems (33) must be taken into account as the connections between immunology and neurodegeneration continue to be elucidated.

Genes themselves may be more important to one species than the other. For example, although we mostly share essential genes that are lethal if deleted, SOD1 is essential in humans but not in mice (35); Sod1 null mice survive, whereas no SOD1 null humans have been reported. This is important when modeling SOD1-ALS, because the majority of mutations result in some loss of SOD1 activity, which may have a more severe outcome in human motor neurons than in mouse motor neurons (36).

Differences in gross morphology are also important for modeling human disease, for example, size is closely correlated with metabolic rate in mammals (33). With respect to the gut and the microbiome, the relative length of the small intestine to the colon is larger in mice than in humans, probably because the cecum is important for microbial fermentation of foods in mice (33, 37). Mice and humans have clear neuroanatomical differences in the brain and spinal cord, although less so in the peripheral nervous system. In the brain, correspondence between human and mouse regions is largely characterized by histological staining patterns and the connections between regions, but there are many key differences. For example, the human brain has substantially more white matter than the mouse brain: The amount of white matter increases as a cubic function of the size of an animal because more axons are needed to innervate the body of larger animals (38). Also, some cell types found in the human brain have not been detected in the mouse brain (39). Furthermore, we are largely visual animals, whereas mice rely more heavily on other sensory inputs, such as olfaction.

Life span is another important difference when considering mouse models of neurodegenerative diseases that occur in mid-life or old age. For example, the widely used mouse strain C57BL/6J typically lives for just more than 2 years compared to, say, people in Japan who have an average (male and female combined) life expectancy of greater than 83 years. A mouse succumbing to disease at 1 year of age may be in late middle age (depending on the strain), whereas at 3 months of age (often the age when mice are studied), that animal may be still a young adult. Moreover, aging in mice does not necessarily reflect the same processes as in humans (40).

For both humans and mice, we need to carefully tease apart the effects of normal aging from the effects of neurodegeneration. This is an issue when investigating AD in individuals with Down’s syndrome because premature aging is also part of the clinical picture of Down’s syndrome (41). Moreover, a related point for mouse models is that individual inbred mouse strains may have alleles that predispose them to progressive phenotypes as the mice age, such as retinal degeneration, hearing loss, and other defects, even before any mutation is placed onto these strains (42). There are many apocryphal stories of researchers testing mice with visual cues, unaware that both littermate controls and mutant mice have vision loss. Therefore, we need to know about the normal aging characteristics of both mice and humans when investigating neurodegenerative disease phenotypes.

Undoubtedly, one of the most difficult areas to tackle for neurodegeneration studies is that of behavior. Nevertheless, although behavioral changes are highly variable among people affected by neurodegenerative diseases, there are common patterns such as memory loss or disinhibition, depending on the disorder. This is such a difficult area to appraise but is so important.


Whereas most human neurodegenerative diseases are sporadic, meaning that we are all at risk (usually that risk increases with age), there are often rare familial monogenic forms of the disease that enable us to create animal models to study underlying mechanisms of pathogenesis. For example, up to 40% of FTD cases and 10% of ALS cases are due to mutations in single causative genes. Up to 5% of AD is inherited in an autosomal dominant manner due to mutations in one of three genes such as APP, presenilin 1 (PS1), and presenilin 2 (PS2). About 15% of PD occurs in people with a family history of the disease, and up to 2% of all PD is due to causative (usually dominant) mutations in single genes. Causative alleles can vary greatly in different populations: The G2019S mutation in Leucine-rich repeat kinase 2 (LRRK2) accounts for ~37% of familial PD in some Arab groups (43) but is rare in Taiwan (44). At the other end of the spectrum, nearly 100% of Huntington’s disease (HD) arises from triplet repeat expansions in the HTT gene encoding the huntingtin protein.

Deciding which mutation to put into a mouse to model a specific neurodegenerative disease has often been a pragmatic choice of what is the fastest mutation to make and what is likely to result in the most aggressive change found in humans to have the best chance of producing a quantifiable phenotype in mouse. However, given the range of tools now available for genome engineering such as CRISPR-Cas9, we can take a more considered approach to creating mouse models to answer specific questions about mechanism without expecting a perfect mimic of human disease.

Here, we discuss the range of genetic mouse models available. Mouse models created using recombinant adeno-associated viral (AAV) or lentiviral vectors, for example, are also giving new insight into disease mechanisms for disorders such as ALS (45, 46). This is a rapidly developing field that has implications for studying mechanisms of neurodegeneration and providing insights for downstream gene therapy approaches (47). Another area that will be extremely important for dissecting neurodegenerative disease mechanisms is the increasing use of human-mouse chimeric animals, in which specific cell types are human and can be studied in vivo in the mouse (for example, wild-type human cells are studied within the environment of the brain in a mouse model of amyloid deposition found in AD) (48). This research platform brings its own challenges including ethical issues and is a rapidly expanding area that we touch on only briefly below.

Mouse models: What is the question?

There are a few key questions that reoccur with all neurodegenerative diseases. For example, specific proteins are usually deposited in aggregates in the brain, but is protein deposition a disease-causing mechanism, a disease-response mechanism, or both? Why do neurons die? Causative genes are usually ubiquitously expressed, so why do only specific types of neurons or synapses die, usually in a well-described pattern? Does disease arise from loss or gain of function of the mutant protein or both? What determines timing of symptom onset? Does a neuron die from mechanisms solely within that neuron (cell autonomous) or do neighboring cells play a role (non–cell autonomous)? How does the disease spread? What determines time to death? Why do these disorders progress from presymptomatic to symptomatic to end stage? Why does incidence tend to increase with aging? What pathways produce variation in severity (and penetrance of genetic disease) and are these pathways routes to therapeutic modulation? What is the connection between histopathological changes and altered behavior? What is the therapeutic window? Is neurodegenerative disease reversible?

To help answer these questions, we can now create mouse models to order with exquisite accuracy (Fig. 2). These include knockin models where the human mutation is engineered into the mouse genome, genomically humanized mouse models, transgenic mice, chemically mutagenized mice, conditional mutant mice, inducible mutant mice, chromosome-engineered mutant models, and transchromosomal mutant models, each giving different types of information about neurodegenerative processes (Fig. 2). In creating a mouse model of neurodegeneration, we therefore need to consider “which question are we trying to answer?” and then “which mouse model would best provide that answer?” The two big caveats that need to be considered as we ask these questions are as follows: (i) mouse models are not necessarily predictable, and so we need to make the mouse to learn about the phenotypic outcomes, and (ii) the ultimate mouse phenotype depends on genetic background and environment.

Fig. 2 Questions that can be addressed with mouse models of neurodegenerative disease.Credit: A. Kitterman/Science Translational Medicine

For research into early neurodegenerative processes, mice expressing disease genes at physiological levels may provide slowly progressing, prodromal disease phenotypes that do not reach end stage within the life span of a mouse. However, this may not be helpful for trying to test therapeutics targeted to later stages of the disease, which is often when patients first come to the clinic. Developing biomarkers or predictors of disease (for example, blood biochemicals, neuroimaging modalities, and behavioral changes) may require access to nervous system and other tissues across disease stages using different mouse models to address diagnostic issues and to monitor therapeutic outcomes. Endophenotypes (characteristics associated with a disorder but not necessarily a direct outcome of that disorder) may help with such issues but need careful validation. Mouse models can also indicate whether disease is reversible even at late stage (49) or whether treating nonneuronal tissue is helpful. For example, we think of SMA as a neurological deficit, but therapies derived from studying a genetic mouse model of severe SMA now target the liver because preclinical studies of liver pathology showed efficacy in this model (50).

The most widely used models of dominant neurodegenerative diseases are still transgenic mice created by injecting DNA into the pronucleus of a fertilized mouse egg. The DNA inserts randomly, often creating additional unintended mutations such that multiple founder lines are required to exclude phenotypes arising from random insertion. The DNA also usually concatemerizes, so the animal has multiple copies of the transgene and thus overexpresses the protein of interest. A study of 40 well-cited transgenic mouse strains, many of which are used in neurodegeneration research, showed that about half had mutations at the insertion site, although little phenotypic outcome appears to be associated with these insertional mutation events (51). Injected DNA constructs are often cDNAs under the control of nonendogenous promotors that are not expressed with the same pattern as the endogenous gene and are only expressed as the splice isoform encoded by the cDNA. Alternatively, bacterial artificial chromosomes (BACs) may be injected into fertilized mouse eggs to create BAC transgenic mice. These animals have the following advantages. BACs have large insert sizes (up to ~200 kb) for genomic DNA that may include the entire promoter/exon/intron architecture of a single gene (or genes); thus, they may express the complete set of human splice isoforms. BACs also tend to integrate fewer than three copies, hence ameliorating overexpression effects.

Transgenic mouse models expressing multiple copies of a mutant gene may show accelerated mid-life– or late-onset disease (if the amount of mutant protein and disease onset positively correlate), which is necessary for studying later stages of neurodegeneration. However, disease phenotypes may arise as a result of overexpression per se, rather than being caused by the effects of the mutation within the transgene DNA, as exemplified by the SOD1G93A transgenic mouse model of ALS (52). Nevertheless, this ALS mouse strain is used to model human SOD1-ALS disease. This has allowed scientists access to an animal model of rapidly progressing disease and enabled the study of later neurodegenerative processes. However, it is particularly important that strains such as the SOD1G93A transgenic mouse are compared to a transgenic mouse control overexpressing similar amounts of the human wild-type protein, which in this case would be wild-type human SOD1 (53).

The utility of transgenic mice is exemplified by the field of prion disease research. Transgenic mice overexpressing a human prion protein were instrumental in showing that bovine spongiform encephalitis (BSE) could be transmitted between mammalian species (i.e., potentially to humans). Notably, had transgenic mice overexpressing the human prion protein not been used, this result would have been missed because the time course of the disease would have been longer than the life span of the mouse (54).

In terms of gene targeting, knockout mice have DNA sequences removed such that the targeted gene is no longer functional (Fig. 2). These mice may be conditional or inducible such that the gene either loses function in response to genetic triggers (conditional) or can be inducibly expressed under experimental conditions through hormone or drug administration, either by injection or addition to the drinking water. In addition to understanding gene function, these animals can be extremely helpful for determining whether late-stage neurodegeneration is reversible by switching off the relevant gene mutation at specific time points (55).

Knockin mice have a sequence such as a mutation or mutant gene precisely targeted into their genome at a specific locus (Fig. 2). Sequences may have exogenous promoters and can be placed into “parking spots” or “safe harbors,” such as the Rosa26 locus, which are safe places to insert foreign DNA without disrupting other genes. DNA sequences may also be targeted by homologous recombination or by CRISPR-Cas9 and other technologies into their orthologous mouse locus such that they are expressed at physiological levels by the mouse gene promoter. Because gene expression is therefore lower than in transgenic models, knockin mice tend to have longer times to disease onset and a slower less aggressive course of disease, which may give researchers access to early disease stages. For example, a knockin mouse model of dominant ALS expresses the mutant protein VAPB (vesicle-associated membrane protein B/C) at physiological levels and exhibits a much slower disease course than does a transgenic animal overexpressing this protein (56). These knockin mice have dysfunctional motor neurons but no signs of motor neuron loss. Despite the lack of motor neuron death in this model, these mice give us access to early disease processes and, potentially, to biomarkers that are urgently needed for many human neurodegenerative diseases. Equally, mouse models of PD may not give the full human clinical picture but can give access to the prodromal syndrome including behavioral abnormalities (57), although it can be difficult to tease out repair and compensation responses from true disease processes found in humans.

Knockin mouse models are particularly important if the gene of interest is dose sensitive, i.e., more than the normal two copies give rise to a phenotype that is not necessarily related to the disease under study. For example, mutations in the FUS gene (fused in sarcoma) can cause ALS. Because FUS is a dose-sensitive gene, transgenic FUS mice do not model FUS-ALS well, whereas knockin models expressing pathogenic FUS mutations at physiological levels are excellent models of ALS, giving access to early-stage disease processes (55, 58). Other ALS genes, such as C9orf72 or TARDBP encoding the TDP-43 protein, are also exquisitely dose sensitive (56, 5962).

Another class of knockin mouse is the genomically humanized model in which either whole genes in the mouse genome are replaced with human genes (63, 64) or key mouse amino acid residues are targeted by knockin constructs and changed into human amino acids in the hope of producing more accurate phenotypes of human disease. These knockin mice include the APPNL-F and APPNL-FG models of amyloid deposition (65) in which critical amino acids have been changed from mouse to human residues in the mouse APP protein, thus making amyloid-β (Aβ) fragments generated from APP more prone to aggregation.

Mutant mouse models can be generated with random mutations in the genome, such as those arising from chemical mutagenesis (66), which may give unexpected insights into gene dysfunction when the gene is expressed at physiological levels. For example, a chemically induced point mutation that turned out to be within the microRNA miR-96 was shown to give rise to progressive hearing loss in mice (67); this mutation was also found in humans with nonsyndromic progressive hearing loss (68). Other types of mouse model include those created by chromosome engineering, which enables massive blocks of DNA to be moved around the genome or the deletion or duplication of big genomic regions. This allows us to model, for example, chromosomal full or partial aneuploidy (6, 69) or large-scale copy number variation (70).

A different type of model is provided by chimeric mice, which are made from two different cell lines and can be extremely helpful in dissecting cell autonomous and non–cell autonomous processes. For example, human pluripotent stem cell–derived cortical neuronal precursor cells derived from unaffected individuals have been transplanted into the brains of both a mouse model of AD and wild-type animals. Notably, only the human neurons in the AD mice showed signs of neurodegeneration, indicating that human neurons respond to Aβ pathology differently from mouse neurons in vivo and that the neurodegeneration is non–cell autonomous. This study also showed that human neuronal death due to Aβ deposition could be dissociated from the effects of tau tangle formation (48). In another example, chimeric mice with motor neurons expressing either wild-type or mutant human SOD1 (causative for ALS) provided strong evidence that ALS is a non–cell autonomous disorder (71). Somatic mutations in which DNA changes in nongermline cells, such as neurons, are causal for neurodegeneration can be modeled by conventional transgenic mice (72) and by other types of mice with mosaic expression of key genes (73). It is likely that the use of chimeric mice to model neurodegeneration, for example, mice with a human immune system or specific human neurons, will increase greatly in the future.

Another advantage of working with mouse models to understand neurodegeneration is the ability to cross independent models and to study the double mutant progeny to tease out disease networks. An example is a recent study of demyelinating and axonal neuropathy with specific genes that modify these phenotypes in mouse models of Charcot-Marie-Tooth disease (74). Other examples include study of the effects of ApoE4 on tau-mediated neurodegeneration (75) and crossing a mouse model of amyloid deposition with a Down’s syndrome mouse to determine why Down’s syndrome gives rise to early-onset AD (76). Similarly, such classical genetic crosses allow the uncovering of new genetic interactions, such as that between mutant SOD1 in ALS models and the cytoplasmic dynein heavy chain (7779).

Clearly, a single mouse model for any given disease is unlikely to provide all of the outcomes needed to understand both the underlying early-stage and late-stage molecular mechanisms of disease and to develop therapies and biomarkers. We need a tailored range of different animal models to address different questions about disease pathogenesis and treatment, and fortunately, we are now in a position to produce the mice we need.


Genetic mouse models enable us to study phenotype in a highly controlled system under well-controlled conditions. They can help to tease out how individual variation arises and how it affects disease but only if rigorous statistical and unbiased blinded approaches to analysis are taken (80, 81). This variation may highlight key cellular pathways, as exemplified in human GWAS, and it provides a potential basis for dissecting individual responses to disease processes and therapeutics. However, as with humans, we still understand comparatively little of what causes variation in individual mice. Age, gender, and genetic background are well-known sources of variation in mouse studies. For example, a single well-defined mutation can result in markedly different phenotypes when bred onto different inbred mouse strains, which have homogeneous (within the inbred line) but different (between inbred lines) genetic backgrounds (60, 82). These include “collaborative cross” mice and “recombinant inbred” mice designed to have known variation in their genetic background to help map the genetic components of complex gene-environment interactions (83, 84). Expectedly, there are major efforts to try and maintain a standard background and avoid genetic drift over time (85). In working with different substrains of individual inbred lines, we may have expectations that substrains used in different laboratories will be similar to each other, whereas in fact they can be markedly different. For example, the widely used C57BL/6J and C57BL/6N substrains have some notable phenotypic differences such as reduced motor performance in the C57BL/6N compared to the C57BL/6J substrain (86).

Genetic background effects can in fact give great biological insight as this common source of variation can be used to help identify modifier loci (87). For example, a modifier locus for human Dravet syndrome (a severe infant-onset epileptic encephalopathy) was identified using the strain-dependent epilepsy phenotype of a mouse carrying a mutation in the Scn1a gene encoding the sodium voltage-gated channel α subunit 1 (88). Another example comes from studies of the Cacna1c gene encoding calcium voltage-gated channel subunit α 1C and the Tcf7l2 gene encoding transcription factor 7 like 2, which have been implicated in human psychiatric disorders and in type 2 diabetes, respectively. Crosses of null mutations in these genes onto 30 different inbred mouse strains showed that opposite effects from the same allele could occur depending on the genetic background and that sex could also have a marked effect on phenotype (89). Thus, if we confine ourselves to studying one inbred line, or one sex, then we may miss the panoply of variation arising from a single mutation. Rather unexpectedly, the genetic background in a mouse cross can also include “inert” control elements such as the tetracycline transactivator, which in theory only exert an effect when activated but in practice may independently affect phenotype (90).

Other sources of variation such as parent-of-origin effects (when a phenotype from an allele depends on which parent contributed the allele, which may warrant phenotyping wild-type offspring of mutant mice) (91) and contributions from transgenerational inheritance (92, 93) may be important sources of variation for some phenotypes. Similarly, even when closely defined hybrid lines are used in highly controlled environments, litter-of-origin effects (i.e., the phenotype depends on which litter a mouse comes from) may arise. Thus, mice from the same litters should be distributed across experimental cages, as demonstrated in a carefully controlled therapeutic preclinical trial in mice modeling certain aspects of HD (94). This study noted sources of variability such as maternal nutritional status, which may result in pups from the same litter having similar characteristics regardless of genotype, and in fact, the authors saw a modest effect of litter of origin on a behavioral phenotype in their HD mice (94).

It is important to identify and understand the many other sources of phenotypic variation so that we can fully characterize our animals and thus better understand variations in human health and disease (95). This may be crucial for understanding gene-gene interactions and also gene-environment interactions, both of which can exert huge influences on disease etiology and progression. Here, we briefly discuss some of the conditions leading to phenotypic variability between studies.

Life history, well-being, and stress

Life history is an important determinant of disease for a mouse and for a human. For example, the level of adversity during development biases protective stress or anxiety responses to better adapt an organism to the environment into which it was born (resilience), but then, these same responses persist into adulthood (9698). Notably, this may not always be helpful later in life if there is a mismatch between the developmental and adult environments. Mice, like humans, are affected by early-life trauma, which can alter brain structure, produce epigenetic and genomic effects, and lead to neuropsychiatric phenotypes later in life (99104). Stress can also influence the in utero environment such as when effects of maternal inflammatory responses lead to brain and behavior abnormalities in offspring (105) or when prenatal stress disrupts the maternal vaginal microbiome leading to lasting effects on the gut and hypothalamus of male offspring (106) .

Stress is a hugely important determinant of pathology and behavioral sequelae in neurodegenerative disorders. Stress and the hypothalamic-pituitary-adrenal (HPA) axis are key determinants of the concentrations of circulating corticosteroid hormones that can have powerful effects on neurodegenerative phenotypes. For example, stress, either from chronic isolation or as the result of an acute challenge, can lead to increased Aβ peptide in brain interstitial fluid in a transgenic mouse model of Aβ deposition used to study AD (101).

Social interactions play a role in mouse and human well-being and can markedly affect phenotypes. Single housing of social animals such as mice causes stress and aberrant behaviors (107). Furthermore, social hierarchies and dominant/submissive relationships may have a profound impact on mouse well-being, and so, for example, the population density of a cage may affect the phenotype (108). Likewise, human relationships are important for outcomes in neurodegeneration. In an ALS patient cohort, being married was associated with an 8-month median longer survival time compared to single individuals (109). Interestingly, it now appears that the genetic status of those surrounding an individual may also be important for the individual’s well-being. Notably, in mice, several traits—from wound healing to anxiety—are affected by the genotype of cage mates (110).

Ambient temperature is a source of variability we rarely consider because mice are generally housed at temperatures of between 20° and 26°C. However, this may be too cold, because mice have activated thermogenesis at these temperatures to maintain their normal body temperature. This mild chronic cold stress affects tumor formation and metastasis, potentially confounding our understanding of these processes. This may also be true for neurodegenerative mechanisms (111) and for metabolic pathways (112). Notably, some researchers, including those engaged in therapeutics studies, are moving to keep their mice, inbred strains or outbred strains, in environments that as closely as possible mimic those of wild mice (113).

Circadian rhythms and sleep

Circadian rhythms can affect experimental and pathological outcomes in neurodegeneration research (114). In fact, researchers usually collect tissue samples only at specific times of the day to avoid variations in gene expression arising from circadian rhythms; for example, normal daily cycles in body temperature affect alternative splicing of transcripts (115). Similarly, behavioral testing is usually carried out at the same time of day within an experiment to minimize variability. However, as a consequence, we may miss phenotypic variation across the 24-hour cycle. This may be particularly important because disruption of normal circadian rhythms can affect biochemical pathways underlying neuropathology and the expression of their behavioral sequelae (116). Moreover, many laboratories conduct behavioral tests during the animals’ subjective day, despite the fact that the species is nocturnal.

Related to circadian effects is the importance of sleep and sleep disturbance in the development of neurodegenerative phenotypes. For example, sleep plays a critical role in the clearance of Aβ from the brain (117). Thus, sleep deprivation and sleep disruption may affect Aβ deposition and exacerbate disease phenotypes and progression by affecting Aβ removal from the brain (118). Moreover, sleep disruption per se may also directly exacerbate cognitive impairments arising from neuropathological changes, similar to its putative effects in neuropsychiatric conditions (119).

Exercise and environmental enrichment

Physical exercise plays an important role in modulating brain function. Exercise results in the release of trophic factors including brain-derived neurotrophic factor (BDNF) (120) and promotes neurogenesis in the dentate gyrus of the hippocampus (121123), a brain area implicated in neurodegeneration and its behavioral sequelae. Exercise, such as voluntary wheel running in mice, enhances glymphatic influx and improves cognitive performance on behavioral tasks (123, 124). Environmental enrichment can also have an enormous impact on the development of behavioral and histopathologial trajectories in mouse models of neurodegenerative disease. For example, the onset of cerebral volume reduction and motor deficits is delayed in an HD mouse model if the animals are exposed to a stimulating environment from 4 weeks of age (125, 126). This may parallel the concept of cognitive reserve, such that people who frequently perform cognitively demanding tasks may be less likely to suffer from deterioration of brain function (127).

Diet, the microbiome, and inflammation

An advantage of working with mice is to be able to study the effects of diet on any given phenotype (128). However, often, scientists may not even be aware of any variation in mouse diet. Many mouse chows contain soy and so may result in a relatively large intake of phytoestrogens, potentially influencing disease outcomes such as anxiety or pain. Exposure to phytoestrogens in utero also has effects on disease outcomes in adulthood. However, the amount of soy and therefore phytoestrogens can vary between batches of chow, thus potentially adding to variability in phenotypes (129). This may be relevant to modeling disorders such as ALS in which estrogen may have a protective effect (130).

Alterations in the gut microbiome, whether derived from diet or otherwise, can also have a great influence on biochemical, neural, and behavioral phenotypes (131, 132), and laboratory mice have a gut microbiome that is notably different from their wild relatives (133). Strikingly, transplantation of gut bacteria from patients with PD into transgenic mice overexpressing α-synuclein exacerbated motor deficits, microglia activation, and pathology, an effect which, in turn, could be ameliorated by antibiotic treatment (134). In a mouse model mimicking aspects of autism spectrum disorder, a species of Lactobacillus was found to reverse behavioral deficits in progeny of dams fed a high-fat diet (135). These effects were mediated via the vagus nerve and not by restoring the composition of the offsprings’ gut microbiome (135). Efforts are now underway to define the effects of the wild mouse gut microbiota in promoting host fitness (133).

The overall health status of the mouse colony can greatly affect the cellular components of the innate and adaptive immune system, which may affect phenotypes in mouse models (136). Central nervous system (CNS) inflammation has long been established as important in the progression of neurodegenerative disorders (137). Recently, it has been found that systemic inflammation also plays a key role (138). The adaptive pathways by which signals of systemic inflammation are communicated to the brain have now been well described, and it is becoming increasingly clear that excessive or prolonged activation of these pathways is detrimental to the brain (105) and can exacerbate neurodegenerative conditions. Thus, health status may be an important determinant of the rate of progression of pathology in a given mouse model of neurodegenerative conditions. This may relate to a much bigger story about how the health status of human subjects affects disease etiology and progression. In particular, infection and inflammation may be key drivers of disease processes, including those underlying neurodegenerative disorders (73, 139).

Making sense of behavior

A further major source of variation across laboratories and across studies comes from the behavioral phenotyping of mice. Often, the use of different behavioral tasks and different test protocols in different laboratories means that researchers are not actually assessing the same behavioral processes in their mouse studies, and hence, different results can be obtained. Even small differences in experimental protocol (such as the amount or nature of any pretraining, duration or nature of stimuli, intertrial interval, and motivational state) can have a major influence on the way rodents might solve a particular task and hence the sensitivity of the animals to a given experimental manipulation. This reflects the multiple memory mechanisms and cognitive processes underlying complex behaviors in both rodents and humans. Again, understanding the sources of this variability across studies, in terms of behavioral outcomes, may shed light on disease processes.

A key related question when using genetically modified mouse models of neurodegenerative disorders is whether the behavior we are studying in rodents is accurately modeling the appropriate behavior in humans. Therefore, it is essential to identify the psychological process that is disrupted in a given mouse model to determine whether this is the same psychological process that underlies impairment in human patients. A clear understanding of these psychological processes will also greatly aid in identifying the underlying neural circuits and mechanisms that are affected in the animal model. Importantly, this requires the characterization of the mouse model across several behavioral tests, which allows the precise nature of the impairment to be inferred by comparing what the mouse can and cannot do. It is not possible to determine the key psychological process that is disrupted in a disease model by studying mice performing a single task.

In a number of situations, classical rodent assays of cognitive behavior may not always model the human cognitive process as intended (Fig. 3). For example, do deficits in Morris water maze performance in mouse models of neurodegeneration always indicate memory impairment (140)? Is working memory studied in rodents undertaking a win-shift maze task (for example, the radial maze or T-maze task) really the equivalent of working memory in humans measured by the N-back task or the digit span task (141)? Instead of using working memory, rodents often may solve these tasks using familiarity judgments based on short-term habituation processes (141). Does contextual fear conditioning in mice measure episodic memory or conditioned anxiety (a change in emotional state)? There are potential problems with the contextual fear conditioning test including generic issues with interpreting freezing data in animals that exhibit marked locomotor hyperactivity, which is found in some mouse models of neurodegenerative disease. Examples from basic science experiments have highlighted these issues, but the questions raised from these studies also likely apply to tests using neurodegenerative mouse models. This demonstrates the importance of extending behavioral tests to provide a more comprehensive and precise description of any cognitive and psychological deficits in the animal model.

Fig. 3 Problems associated with behavioral tests in mouse models of neurodegeneration.

(A) In the Morris water maze task, mice are trained to find a hidden escape platform that is submerged 1 to 2 cm below the surface of the water but remains in the same spatial location relative to external cues. Difficulties associated with testing mice in this water maze task include thigmotaxis (swimming next to the wall), floating behavior, and fatigue after prolonged swim times, all of which may be more pronounced in genetically modified animals. The question is whether impairments in water maze performance in some mouse models of neurodegenerative disease reflect the deficits in learning and memory that the water maze test is supposed to measure. (B) Spatial working memory in rodents is tested using a T maze, but this may not equate to tests (N back, digit span) used in humans to measure working memory. During the sample run (left) in the T-maze task, mice must enter one of the two arms of the maze to obtain a food reward (red circle). After a short delay, the animal is then given a choice run (right) between both arms of the maze and must go to the previously unvisited arm to gain a second reward (red circle). (C) Contextual fear conditioning in mice is often used to test hippocampus-dependent, episodic-like memory. In this task, mice are placed in a chamber and receive a mild shock to the foot. The mouse is later returned to the chamber, and the amount of freezing behavior it exhibits is measured. Potential problems with the contextual fear conditioning test include interpreting freezing behavior in mice that exhibit locomotor hyperactivity, a feature of some mouse models of neurodegeneration (156).

Credit: A. Kitterman/Science Translational Medicine

In the Morris water maze swim test, for example, it is imperative to include the appropriate control tasks that are well matched in terms of sensorimotor and motivational task demands (visual acuity, emotionality, stress, swim time, difficulty, etc.). In addition, a battery of spatial memory tests also needs to be included to allow the generality of any findings from the water maze swim test to be ascertained under different behavioral conditions that are unaffected by performance factors such as thigmotaxis (continually swimming close to the side wall of the pool) and floating behavior. Both of these behaviors present confounds in water maze studies conducted in mouse models of neurodegeneration (140). The distance that the platform is located from the side wall of the pool may thus be an important experimental variable in such studies (Fig. 3) (142). If a water maze deficit in a mouse model of neurodegeneration is due to thigmotaxis (possibly as the result of altered anxiety), then any drug treatment developed from such an assay may be unlikely to remedy cognitive symptoms in patients. Conversely, however, increased thigmotaxis or enhanced floating behavior likely reflects important behavioral responses by the mice to a given aversive situation. Understanding these behaviors and the contribution that a particular genetic change or its neuropathological sequelae make to these phenotypes may provide important information about emotionality phenotypes (such as depression and anxiety) that can occur in neurodegenerative conditions in human patients.

Phenotyping is a group activity

Importantly, the best way of working with mice, which also likely involves using the fewest animals, is to work collaboratively with scientists with expertise in different areas—from genetics to physiology/endocrinology, from transcriptomics to developmental/behavioral biology, and from protein chemistry and neuroimaging to immunology—in concert with clinical experts for each neurodegenerative disease. Multidisciplinary teams enable us to maximize what we learn from each model including pleiotropic effects (for example, separating these from comorbidities) and help us to avoid maintaining too narrow a focus on, for example, just one specific cell type. Each mutant mouse strain should undergo broad phenotyping of the sort provided by the International Mouse Phenotyping Consortium (81) and in-depth investigations into disease mechanisms by disease specialists.

Metabolomic studies provide a good example of why studying neurodegeneration should not only be the work of neuroscientists. One of the most distressing aspects of ALS and HD is weight loss, and this may correlate with poor clinical outcome. Furthermore, low body mass index is thought to be a risk factor for ALS (143). In ALS, we know that there are lipid disturbances in both humans and mouse models, but why these arise remains unknown. This could turn out to be an important avenue of research for those developing disease-modifying treatments for ALS.


Despite the hundreds of mouse models of human neurodegenerative disease, we still have no cure for any major form of neurodegeneration and only extremely limited treatments for just a handful of these diseases. Many insightful articles have been written about moving forward from mouse studies of disease mechanisms and the difficulties of translating therapies that modulate disease in mice to successful human clinical trials (27, 125128, 144, 145). Here, we suggest that variability in both mouse and human not only may be an important factor underlying current translation failures but also may be a source of insights into disease mechanisms.

What is translation?

The translation slogan “from bench to bedside” encompasses two separate processes: (i) the discovery of mechanisms of disease pathogenesis and (ii) the development of new therapies and their testing in clinical trials. In our view, mouse models have great utility for dissecting biological processes but so far have not proven particularly useful or practical for producing treatments that work in humans, with a few notable exceptions including the success of antisense oligonucleotide therapies for treating SMA (146). Currently, therefore, the major and proven utility of mouse models lies in identifying disease mechanisms. Their potential utility (or lack thereof) as drug screening tools may only become apparent once we have a better understanding of both pathomechanisms and the reasons why current attempts at translation have so often failed. We argue that the answer to the latter question may partly lie within our appreciation of the variability in mouse studies and also the great variability within the human clinical population.

From mouse and mechanism to medicine

In biomedical studies, we generally only look at a snapshot of disease. For technical or financial reasons, we study genetic mutations in one or two inbred mouse strains only or, worse yet, in ill-defined outbred mouse strains that are usually a lot more inbred than expected (147) [although outbred strains certainly have their uses (148) and abuses (149)]. We tend to study single gene mutations, although most human diseases are a mix of genes and environment. Even rare monogenic diseases may show synergistic effects between unrelated genetic loci (150).

As scientists, our cultural viewpoint is to minimize variation within our studies to maximize our chances of identifying notable differences between experimental and control conditions. To this end, we generate genetically modified mice on homogeneous genetic backgrounds and maintain these animals under constant environmental conditions. It follows that any important discoveries arising from these animal studies, particularly with respect to therapeutics that modulate biochemical pathways and networks, may only be relevant to a limited subgroup of the clinical population.

It is not surprising, therefore, that when attempts are made to translate findings in animal models to the much more heterogeneous human clinical population (which varies in terms of age, gender, genetic background, environment, and life history), these attempts have failed consistently. To put it another way, the homogeneity of approach that benefits reproducibility and increases statistical power in preclinical studies (and yet that is still remarkably difficult to carry out in mice) may come at the cost of reducing generalizabilty when it comes to translation of therapeutics from mouse to human (151). This issue is not exclusive to animal models of neurodegeneration but instead reflects a general problem in making the crucial step from preclinical studies in animals to clinical trials for many human disorders, as has been highlighted for the stroke field (151, 152) and neurodegeneration (153). Researchers have made suggestions for improving this process, such as performing mouse studies across several independent laboratories to mimic multicenter clinical trials in humans. Nevertheless, at the heart of this issue remain both the inherent variability of preclinical mouse studies and the pronounced variability of human clinical patient populations. Crucially, therefore, embracing rather than rejecting variability in our mouse studies, and then understanding both its sources and its underlying mechanisms, could be of great benefit for successful translation to clinical patient subgroups. Moreover, it seems likely that questions regarding the utility (or otherwise) of mouse models for determining pathomechanisms relevant to human neurodegenerative conditions, and ultimately as drug screening tools, can only be answered once we have addressed this central issue of variability.

Although currently we have few treatments for human neurodegeneration, and a relatively limited understanding of the mechanisms underlying neuronal death, we are also racing into a new era of personalized medicine. By working with the appropriate mouse models to address specific questions and by examining causes of variation in these models, we can help to identify diverse mechanisms of neuronal dysfunction and neuronal death. Variation may be particularly important for neurodegenerative disorders given the heterogeneous nature of many of these conditions and given the current drive to identify biomarkers allowing stratification of patient subgroups, precision medicine, early diagnosis, and response to therapeutic intervention.

It is impossible to standardize all experimental conditions across laboratories, but instead, we need to recognize variation and use it as a source of insight. Variation has hitherto been seen as a problem and something that should be diminished and reduced at all costs. Here, we suggest that variation could be an opportunity that may allow us to understand and identify disease mechanisms and risk factors and, at the same time, to elucidate treatment strategies on an individual by individual basis. Embracing and understanding variation may be of great benefit for translation. Of course, identifying the sources of variation will not be straightforward, but nevertheless, we suggest the following agenda for working with mouse models of neurodegeneration.

We believe mouse models should be studied in response to a specific need, for example, animals with slow disease progression and relatively mild phenotypes would be best used for understanding early disease pathogenesis; these animals may not be useful for clinical trials of therapeutics developed for treating late-stage human disease (Fig. 2). Both scientists and funding agencies should take a more sophisticated approach to each mouse model and not expect complete recapitulation of the human disease. Mouse models often may be better suited as tools for understanding the role of a specific gene or protein, or a specific aspect of the disease process, rather than as a primary drug-screening tool.

To capture how variation arises, we need considerably more detail than is often provided in research papers (95). In an attempt to ensure that mouse phenotypes are fully defined in the literature, the U.K. National Centre for the Replacement, Refinement, and Reduction of Animals in Research (NC3Rs) introduced a checklist of information required for any paper describing animal research called the ARRIVE (Animal Research: Reporting of In Vivo Experiments) guidelines (154). Undertaking animal studies in accordance with the ARRIVE guidelines should mean some sources of variability are clearly reported (154), and so, we can begin to understand their effects on mouse phenotypes. Furthermore, peer review should be rigorous to ensure accurate reporting of such crucial experimental details. However, more than anything, as a scientific community, we need to change our attitude to the variability in our data. We need to create long-term metadata repositories in which data are deposited and made available to the wider community for critical analysis and further phenotypic screening. New approaches to working with mice are on their way. For example, home cage analysis enables us to capture aspects of behavior not seen when humans are around (155). New informatics approaches including machine learning could help us to find new phenotypes from the immense amount of data generated by home cage analysis. Developing ontologies for human-mouse phenotypes, in combination with new imaging and computational methods to help cross-reference between mouse and human data sets, will greatly enrich our understanding of human (and mouse) biology and will help to elucidate neurodegenerative processes but only if we also take variation into account.


Acknowledgments: We thank the reviewers and G. Schiavo for insightful comments and our many colleagues worldwide who work in different capacities with mouse models to understand and treat human neurodegenerative disease. Competing interests: D.M.B. is an unpaid member of Eli Lily’s Centre for Cognitive Neuroscience.

Stay Connected to Science Translational Medicine

Navigate This Article