Mouse studies don’t predict results in humans very well

If there’s one thing about science reporting that drives most researchers, particularly medical researchers, bonkers, it’s the hyping of preclinical studies in cell culture and animal models as major breakthroughs, even though the drug or treatment tested has yet to be translated into a treatment for human beings. Don’t get me wrong; sometimes a mouse study can be an important breakthrough. However, far more often than not, treatments that work in mice fail to translate to humans and don’t work or don’t work nearly as well in humans. My favorite example of this is something I studied during my fellowship, angiogenesis inhibitors. Back in the late 1990s, angiostatin and other angiogenesis inhibitors, compounds that could prevent the ingrowth of new blood vessels into tumors, were being touted as a cure for cancer. This was because for all solid tumors at least, cancers rely on the ability to induce the body to generate new blood vessels for their continued growth beyond a couple of millimeters in diameter to feed their voracious needs for oxygen and nutrients. There were spectacular results, too, in mice. My scientific hero, the late great Judah Folkman, could cure cancers in mice. He could even induce tumor dormancy, such that the tumor would shrink to a very tiny size and then remain dormant indefinitely, all using angiostatin and endostatin, two endogenous angiogenesis inhibitors that he had isolated from the urine and blood of mice. Unfortunately, in humans, although angiogenesis inhibitors do work, their effects are nowhere near as dramatic, and they have taken their place as just one type of drug among many used to treat various cancers, with no spectacular results, just incremental improvements in treatment. As Dr. Folkman quipped at the time, “If you have cancer and you are a mouse, we can take good care of you”.

So why do so many treatments that work in mice for various diseases fail to work in humans? For instance, in neuropsychiatric research, the track record of mouse models for predicting the efficacy of new treatments in humans has been particularly abysmal. There was a study published last week that suggests one reason, at least for drugs designed to treat psychiatric and neurological disorders, and that’s previously unappreciated differences in brain function. Before I get into the study itself, let’s let Sharon Begley, writing about the study for STAT News, set the stage:

Lab mice endure a lot for science, but there’s often one (temporary) compensation: near-miraculous recovery from diseases that kill people. Unfortunately, experimental drugs that have cured millions of mice with Alzheimer’s disease or schizophrenia or glioblastoma have cured zero people — reflecting the sad fact that, for many brain disorders, mice are pretty lousy models of how humans will respond to a drug. Scientists have now discovered a key reason for that mouse-human disconnect, they reported on Wednesday: fundamental differences in the kinds of cells in each species’ cerebral cortex and, especially, in the activity of those cells’ key genes. In the most detailed taxonomy of the human brain to date, a team of researchers as large as a symphony orchestra sorted brain cells not by their shape and location, as scientists have done for decades, but by what genes they used. Among the key findings: Mouse and human neurons that have been considered to be the same based on such standard classification schemes can have large (tenfold or greater) differences in the expression of genes for such key brain components as neurotransmitter receptors.

I don’t know if the team of researchers really was “as large as a symphony orchestra” (I counted 63 authors, but I didn’t double check; so I could be off by one or two), but it does take a lot of people from several different universities, such as UC-Davis, the University of Washington, Columbia University, UC-San Diego, and multiple institutes) to do a study of this type. In any event, as you can gather from this summary, even though anatomically the neurons and circuits connecting various regions of the brain in mice look very similar to the corresponding neurons and circuits in humans, at the gene expression level they are actually very different.

So here’s what the researchers did. They harvested postmortem brains from male and female donors 18–68 years of age with no known history of neuropsychiatric or neurological conditions as the controls, as well as brain tissue from patients undergoing surgery for epilepsy. From this tissue and from the whole postmortem whole brains, the researchers focused mainly on a part of the brain called the middle temporal gyrus (MTG), which is often available from epilepsy resections and thus allowed the comparison between the tissue of the epileptic subjects and the control subjects. They then subjected these specimens to single-nucleus sequencing:

Single-cell transcriptomics enables molecular classification of cell types, provides a metric for comparative analyses, and is fuelling efforts to understand the complete cellular makeup of the mouse brain¹⁸ and even the entire human body¹⁹. Single-cell RNA sequencing (scRNA-seq) of mouse cortex demonstrates robust transcriptional signatures of cell types^20,21,22 and suggests around 100 cell types per cortical area. Dissociating live cells from human brain is difficult, which makes scRNA-seq challenging to apply to this type of tissue, whereas single-nucleus RNA-seq (snRNA-seq) enables transcriptional profiling of nuclei from frozen human brain specimens^23,24. Of note, nuclei contain sufficient gene-expression information to distinguish closely related cell types at a similar resolution to scRNA-seq^25,26, but early applications of snRNA-seq to human cortex did not have sufficient depth of coverage to achieve similar resolution to mouse studies^27,28. Here, we established robust methods for the classification of cell types in human brain using snRNA-seq and compared cortical cell types to reveal conserved and divergent features of human and mouse cerebral cortex.

To boil it all down, doing single-cell sequencing of the genome would be the best approach, but in the brain apparently separating single cells from the brain without killing them is a technical challenge; so the authors used the next best method, single-nucleus sequencing. The disadvantage is that nucleus doesn’t contain all the messenger RNA (mRNA) being made, as the mRNA is transported out of the nucleus to be translated into protein. However, based on previous research, in this case single-nucleus sequencing is as good as single-cell sequencing for determining the differences in gene expression between cell types in the brain. As for the rest, RNA-Seq is a next generation sequencing technique that allows the sequencing of every mRNA sequence detectable in a given RNA isolate from cells or tissue; it’s become so sensitive that it is now possible to perform RNA-Seq on single-cell and single-nucleus RNA isolates.

Now consider the sheer magnitude of this study. Nuclei were isolated from eight donor brains, with most coming from postmortem donors (n = 15,206) and a minority (n = 722) from layer (L)5 of MTG removed during neurosurgical procedures:

In total, 15,928 nuclei passed quality control, including those from 10,708 excitatory neurons, 4,297 inhibitory neurons and 923 non-neuronal cells. Nuclei from each broad class were iteratively clustered as described²⁶ (see Methods). Clusters were generally robust to different iterative clustering methods and were distinguished from nearest neighbours by at least 30 differentially expressed genes and at least one, and often more, binary markers.

Translation: After sequencing, the researchers noted that the various cell types could be robustly identified by various mathematical clustering methods used to analyze genomics data.

The purpose of this study was to develop and validate methods and a system for using snRNA-Seq to catalog the cell types in the MTG and later elsewhere in the brain. Given that, it’s not important to go through all the details of the cell types and their locations that were documented, especially for a non-neurologist or non-neuroscience geek. A few observations were interesting, though. For example, excitatory neurons were more widely distributed than expected, with most excitatory neuron types not being restricted to single layers. Three types were more common in L2-L3, while ten types were enriched in L3-L6, leading the authors to note that this “heterogeneity implies that anatomical laminar location alone is insufficient to predict neuron type, although it remains to be seen whether this is a feature of MTG or human cortex generally”.

When the authors compared these single-nucleus gene expression profiles to those of cells in the same region of mouse brain, they noted several differences. Indeed, they found that neurons in mice and humans that have been generally considered to be the same based on anatomy, structure, histology, and standard classification schemes can actually have ten-fold or even greater in the expression of genes for very important proteins, such as neurotransmitter receptors. (Neurotransmitters are the peptides that neurons use to communicate with each other.) For example:

Comparison of homologous types showed a mix of conserved and divergent expression. The Sst Chodl type (Inh L3–L6 SST NPY in human) had conserved expression overall, but 18% of genes had highly divergent expression (defined conservatively here as a more-than-tenfold difference), including many marker genes. OPCs also had conserved expression and 14% highly divergent genes. Two thirds of all genes analysed (9,748) had divergent expression in at least one of 37 homologous types, and many had expression changes restricted to one type or class. Non-neuronal types had the most divergent expression (3,643 genes with more than tenfold difference), supporting increased evolutionary divergence of non-neuronal expression patterns between human and mouse¹⁷

Translation: Nearly 20% of genes had very different expression between mouse and human, with up to a greater than ten-fold difference, and at least two-thirds of genes analyzed had different expression between mouse and human. This difference was particularly notable for the gene encoding serotonin receptors:

Serotonin receptors exhibit highly divergent expression between species: four of seven G-protein-coupled receptors and both ionotropic receptor subunits (HTR3A and HTR3B) were in the top 10% most-divergent genes (Fig. 6e). The most-divergent gene families include neurotransmitter receptors, ion channels, extracellular matrix elements and cell-adhesion molecules. Among the top 3% most-divergent genes (Supplementary Table 5), the collagens COL24A1 and COL12A1 and glutamate receptor subunits GRIK1 and GRIN3A were expressed in different cell types between species and were validated to have different laminar distributions in human and mouse (Fig. 6f, g). The cumulative effect of so many differences in the cellular patterning of genes with well-characterized roles in neuronal signalling and connectivity is certain to cause many differences in human cortical circuit function.

Serotonin (5-hydroxytryptamine, or 5-HT), as some readers might know, is a molecule with multiple functions in different parts of the body, where it can function as a hormone, growth factor, or neurotransmitter, among other functions. For instance, it’s stored in platelets, which release it when binding to a clot, where it causes constriction of the blood vessels at high concentrations. (As an interesting aside, at lower concentrations it is a vasodilator.) In addition, serotonin also plays a role in wound healing. In the brain, however, it is primarily a neurotransmitter, where its activity is believed to impact a number of functions, including mood, sexual function, appetite, sleep, memory and learning, temperature regulation, and some social behavior. The most widely known role of serotonin (among the general public, at least) is its role in depression, because of the class of drugs known as selective serotonin reuptake inhibitors (SSRIs), which are used for the treatment of social phobia, anxiety disorders, panic disorders, obsessive-compulsive disorders (OCD), major depression, irritable bowel syndrome (IBS) and eating disorders.

Basically, when two neurons communicate, one will release certain neurotransmitters, like serotonin, into the space between them (the synapse). A neurotransmitter will diffuse from one neuron (the presynaptic neuron) through the synapse to the postsynaptic neuron, where it will bind to its receptor, thus activating a signal in the postsynaptic neuron. After the signal is sent, neurons get rid of the extra neurotransmitter in the synapse through protein transporters that take them up into the cell again after they’ve bound to the receptor and been released. SSRIs inhibit the activity of specific serotonin transporters, which allows the serotonin to linger in the synapse for longer, thus letting it stimulate its receptor longer.

Because of this, in the discussion, the authors note:

Our results demonstrate species divergence of gene expression between homologous cell types, as shown at the single-gene¹⁵ and gross-structural level¹⁶. These differences are likely to be functionally relevant, as divergent genes are associated with connectivity and signalling, and many cell-type markers show divergent expression. Notably, serotonin receptors are the second-most-divergent gene family, challenging the use of mouse models for many neuropsychiatric disorders that involve serotonin signalling⁵².

And conclude:

These observations quantitatively frame the debate of whether human cortex is different from that of other mammals^10,11, revealing basic transcriptomic similarity of cell types punctuated by differences in proportions and gene expression between species that are likely to influence microcircuit function. Furthermore, these results help to resolve the paradox of failures in the use of mouse for preclinical studies despite conserved structure across mammals^52,53, and highlight the need to analyse human brain in addition to model organisms. The magnitude of differences between human and mouse suggests similar profiling of closely related non-human primates is necessary to study many aspects of human brain structure and function. The enhanced resolution afforded by these molecular technologies also shows great promise for accelerating our mechanistic understanding of brain evolution and disease.

Or, as the lead investigator Ed Lein of the Allen Institute for Brain Science in Seattle said in STAT:

“All of the drugs people are trying to develop act on receptors or other molecules,” said neurobiologist Ed Lein of the Allen Institute for Brain Science in Seattle, who led the study, published in the journal Nature. “If the neurotransmitter receptor you’re hoping to target isn’t used in the same cells in humans that it is in mice, then your drug will hit the wrong circuit” and not have the same effect in patients as in lab rodents.

The bottom line is that there are substantial similarities between mouse and human brains, but this study has unveiled substantial difference that were unexpected, in which the neurons in the mouse version of human brain structures use different neurotransmitter receptors. The results of this study will help scientists interpret results in mouse studies better and, hopefully, apply them to human neurophysiology more accurately. In the meantime, whenever you see a study, look for whether it was just done in mice. There’s even a Twitter feed, @justsaysinmice, whose main purpose is to point out hyped studies that were only done in mice:

IN MICEhttps://t.co/QUp4R2oIKh

— justsaysinmice (@justsaysinmice) August 23, 2019

IN MICE https://t.co/3jwoX1m80S

— justsaysinmice (@justsaysinmice) August 13, 2019

https://t.co/gh4ZlkfSno

— justsaysinmice (@justsaysinmice) April 16, 2019

You should follow @justsaysinmice while we’re waiting to see what this new study contributes to making animal models more predictive of human response.