Transcriptomic Challenges of Animal Models for Neurological Disorders

1. Introduction

In addition to common disease indicators related to memory, cognition, behavior, and senses, medical statistics categorizes the neurological patients in large groups according to race [1], sex [2], age group [3], geography [4], and climate [5]. However, there are numerous other non-negligible influential factors like smoking [6], alcohol [7] and drug [8] consumption habits, medical history [9,10], diet [11], exposure to stress [12] and toxins [13], whose dynamic combination is unique for each human. Even monozygotic twins might be differently affected by the same disease, each of them experiencing distinct severity and development course and treatment outcomes [14,15]. Although the physician deals with the sickness of the real patient with all his/her distinctive characteristics, in medical school s/he learned about the disease [16], i.e., a simplified, yet general enough, model of the deviation from what is considered “normal”. The “normal” is defined in medical textbooks by selected sets of average characteristics of the so-called “healthy” human, although the distributions of “healthy” and “diseased” features partially overlap.

Beyond the locally imposed moral and religious rules, medical research on human subjects must also respect the ethical principles included in the Declaration of Helsinki [17] and the additional constraints imposed by the local Institutional Review Board (IRB). These mandatory rules limit exploration of molecular phenomena to tissues removed in routine surgery (eventually spread into immortalized cell cultures), blood and body waste, imaging and non-invasive investigations. Postmortem human tissue samples purchased from accredited biobanks are free of most ethical constraints but still represent a narrow spectrum of pathologies [18,19].

Therefore, one needs to replicate human diseases into animal models that, although restricted to protocols approved by the local Institutional Animal Care and Use Committee (IACUC), still allow never acceptable studies on living humans. IACUC should adhere to internationally recognized ethical principles [20] and align with the Guide for the Care and Use of Laboratory Animals (e.g., [21]. Our expertise on neurological diseases is with animal models, although in cancer research, we have also profiled surgically removed thyroid, prostate, lung and kidney tumors from living human patients. Although there is no way to exactly reproduce a human disease on an animal model, one may assume that certain recipes for physiological phenomena and their alterations are shared with some other mammals and remained unchanged during evolution. Moreover, a non-negligible advantage of the animal model is that the investigator has full control of the experimental conditions while each human is affected by a never repeatable, mostly out of our control, combination of favoring factors.

This Report details the challenges our labs faced in developing and investigating animal models of some neurological disorders over many years of research. The challenges are illustrated through our genomic investigations on hippocampus regions, hypothalamic arcuate and periventricular nuclei, frontal cortex, spinal cord, striatum, dorsal root ganglia, retina, and primary, immortalized and genetically engineered neurons, astrocytes, and oligodendrocytes. The cells and tissues were collected from in-house maintained mouse, rat and rabbit colonies that modeled astrocytoma, epilepsies, glaucoma, cerebral malaria, multiple sclerosis, neuroblastoma, neuropsychiatric lupus erythematosus, occulo-dento-digital dysplasia, and X-linked Charcot–Marie–Tooth diseases. We were also involved in confocal imaging of calcium signaling in mouse astrocytoma, behavioral studies of epileptic rats, and electrophysiological experiments on various components of the nervous systems from mice, rats, rabbits, earthworms, and frogs. In addition to caring for animal colonies and cell cultures, improving technology, and performing molecular experiments with optimized wet protocols, we developed mathematically advanced algorithms and software, and acquired bioinformatic expertise.

2. Major Modulators of Animal Models

Almost all molecular characteristics of an animal model depend on species [22,23], strain [24,25], sex [26,27,28,29], explored region [30], age/developmental stage [31,32,33]. They are modulated also by exposure to oxygen deprivation [34,35], toxins [36,37] and radiation [38], diet [39,40,41], and external stimuli [42,43,44]. Not negligible are the hormonal status [45,46], disease history [47] and treatment [48,49], and even exposure to microgravity [50]. We proved that neuronal status is also strongly dependent on the pattern of action potential [43,51]. Moreover, the subcellular localization of certain proteins not only differs between sexes but also changes during the estrogen cycle [52,53], making the female animal models much more difficult to manage and interpret than the male counterparts.

Therefore, choosing and handling the right animal model [54] is far from an easy task. Furthermore, sanitary precautions and housing conditions (e.g.,: distribution of cages and vicinity of other caged animals [55,56], chow and water abundance and quality, environmental temperature, humidity, and day-night lighting cycle [57] are also major modulators of the experimental results. Not negligible are the neurological effects of the anesthesia used for painless sacrifice of the animal. For instance, the intraperitoneal injection of urethane we used to study the neurotranscriptomic effects of microgravity [50] desynchronize the electric waves in the hippocampal dentate gyrus [58]. Therefore, to minimize the anesthesia effects, in other experiments we used either carbon dioxide [59] or just decapitation. However, even if all animal characteristics were hypothetically identical, there is still biological variability due to the stochasticity of biological responses. The variability extends also to samples collected from different regions of the same tissue as we found by profiling pathologically equally graded cancer nodules in surgically removed prostate [60] and kidney [61] cancer tumors from the same individuals. This transcriptomic heterogeneity results from the coexistence of several stem cell clones with similar yet nonidentical phenotypes within the same region, potentially generated by exposure to different local conditions [62,63].

Increasing evidence shows that the molecular characteristics of individual cells depend on cellular environment in a heterogeneous tissue [64,65] and change dramatically when studied outside their original tissue. We demonstrated the impact of cellular environment on several functional pathways by profiling the transcriptomes of astrocytes and oligodendrocytes when co-cultured in insert systems vs. cultured separately [66,67,68].

Over one hundred single-cell sequencing methods have been developed so far to determine separately the omics of the several cell types composing a tissue (e.g., [69,70,71]) but it is difficult to estimate how much the separation method of cell phenotypes alters the real transcriptome. Moreover, transcripts’ counting appears sparse in some platforms that might bias the expression ratios between two cell phenotypes (Iacobas, unpublished observations). 2D and 3D coculture systems [72,73,74] where two or more cell types are grown with or without direct physical contact, are now used by many groups trying to understand the intricacy of the nervous system physiopathology [75,76,77]. Nonetheless, integrating multicellular features of a hetero-cellular tissue into a mathematical model faces enormous theoretical difficulties and requires powerful computer techniques like those used in machine learning approaches [78].

The characteristics of the specimen studied are also strongly influenced by the culture preparation technique [79,80] and applied stimuli [43,81,82,83], hormones [84,85] and chemical treatment [86,87]. In some investigations [66,67,68], we have used immortalized mouse oligodendroglial precursor cell lines (Oli-neu) obtained through the standard protocol [88,89], while in others we used primary cultures of mouse and rat neurons, astrocytes, oligodendrocytes and microglia [50,67,79]. Therefore, we can affirm that, although preferred for their stability, immortalized cell lines [90,91,92] have substantially modified physiology with respect to the primary cells. The purity of cell culture is extremely important, not only because the result would be averaged expression levels over all present phenotypes, but the “infectant” might interact and thus modify the transcriptome of the investigated cells. However, even in pure cultures, cells are not identical. For instance, in cultured rat β-pancreatic cells, we found that only part of them expressed connexin 36 and in cultured HL-1 cardiomyocytes some cells were contracting spasmodically, while others were not (Iacobas, unpublished observations). When possible, it is also preferable to synchronize the cultured cells to detect the dynamics of the cell cycle. As such, one should be careful when extending the findings from either primary or immortalized cell cultures to the original in situ hetero-cellular tissue.

3. Genetically Engineered Animal Models

3.1. General Considerations

It is a widespread belief that many diseases are frequently associated with altered sequence (e.g.,: [93,94,95]) or/and expression level (e.g.,: [96,97,98,99] of certain critical genes, termed (gene) biomarker(s) (e.g.,: [98,99,100,101]. Therefore, it is assumed that, by engineering the same alteration in the genome of an animal, one might reproduce the key features of that human disease (e.g.,: [102,103]). It is also assumed and already included in clinical trials that a cure may be provided by restoring the normal sequence/expression level of the biomarker gene(s) (e.g., [104,105]).

The problem is that many disorders are multifactorial, involving alteration of two or more genes. For instance, mutations of six genes (SNCA, LRRK2, Parkin, PINK1, DJ-1, ATP13A2) are blamed for Parkinson disease [106], while NF-κB, NLRP3 inflammasome, and mTOR signaling pathways are significantly altered following spinal cord injury [107]. A few selected multi-gene diseases were simulated in double knockout mouse models like the Gfap-/-Vim-/- mouse with attenuated glial scar formation in post-stroke [108]. Moreover, effective mutations can be located within several exons, such are the 65 known mutations of GJA1 identified as responsible for the autosomal-dominant ODDD [109].

Despite being adopted by most genomicists, the biomarker paradigm has major flaws [110]. When even pathologically graded cancer nodules from the same tumor exhibit substantial differences in transcriptomic topologies [60], it is hard to accept that distinct individuals exhibiting similar symptoms should have the same set of gene mutations and/or transcriptomic signature. Hence, we have proposed replacing the gene biomarker paradigm with the genomic fabric paradigm [34] that provides the most theoretically possible comprehensive characterization of the transcriptome and allows identification of personalized gene master regulators [111].

The main challenge with the use of genetically engineered animals and cells is that the targeted mutation goes over the ~3 million other mutations present in every cell at any given time. About one in every 1000 nucleotides is randomly mutated just because of the stochastic nature of the chemical reactions involved in DNA replication [112], although most of these mutations have practically no effect because they occur in non-coding regions. Moreover, in addition to the gene whose expression level is experimentally altered, expressions of hundreds of other genes are also regulated through their interlinkages in functional pathways. To make things much more difficult, the combination of the affected genes is not only unique to each individual and even to each histopathologically distinct region to the tissue, but also changes over time Furthermore, each region exhibits distinct strengths in the homeostatic control of transcript abundances, gene networks, and functional pathways. We encountered these difficulties in all (below described) genetically engineered animal models used in our neuro-genomics studies.

3.2. Occulo-Dento-Digital Dysplasia (ODDD)

ODDD is believed to be caused by mutations in the GJA1 (gap junction alpha 1) gene that encodes connexin 43 (Cx43), the most important gap junction channel forming protein in astrocytes [113,114]. Neurological manifestations of ODDD include isolated ataxia combined with spasticity [115,116], microphthalmia, microcornea, glaucoma, cataracts, and optic neuropathy [117]. ODDD was simulated in genetically engineered C57Bl/6j, C3H/HeJ, FVB mice (e.g., [118]). I130T/+ and G60S/+ mutations of Gja1 reproduce most of the ODDD features [119,120,121] in mouse models. We profiled the gene expression in C57Bl/6j and 129/SvEv mice whose Gja1 gene was: i) knocked out (Cx43KO, Cx43^-/-) through homologous recombination [122,123], ii) conditionally knocked out with hGFAP-Cre:Cx43f/f (method in [31]) only in the brain [24], or iii) knocked down through siRNA [79]. In addition to Cx43-/- mice, we also used brains and astrocytes from Cx43+/- mice [124,125] and performed several other transcriptomic experiments with truncated carboxy terminal of Cx4388n various loci [126]. However, although the brain looks anatomically normal, the Cx43KO mouse is not a good model for ODDD since the mutant dies at birth because of a developmental heart abnormality [127]. Therefore, we have profiled the brains of Cx43KO neonates taken from C-section on G17 day. Even though our mice with a genetically altered Gja1 did not age to exhibit ODDD features, the experiments provided valuable information about the major roles played by Gja1 in modulating several functional pathways and its interaction with membrane purinergic receptors [128]. Nonetheless, in all experiments with these mice, we found that, together with the targeted Gja1, expressions of hundreds of other genes that are involved in a wide diversity of physiological functions were significantly altered. Moreover, we found that the transcriptomic alterations strongly depended on the mouse strain (e.g., C57Bl/6 vs 129/SvEv) [24].

3.3. X-Linked Charcot–Marie–Tooth Syndrome (CMT1X)

CMT1X is a demyelinating disorder characterized by muscle weakness and sensory neuropathy [129,130]. It is believed that CMT1X is caused by loss-of-function mutations affecting the GJB1 (gap junction beta 1) gene, located on chromosome X, that encodes connexin 32 (Cx32) the gap junction channel forming protein in oligodendrocytes (brain) and Schwann cells (peripheral nervous system) [131]. Several CMT1X-causing missense mutations of GJB1 have been identified so far, including: p.Arg15Trp, p.Val63Ile, p.Leu89Val, p.Ala96Gly, p.Arg107Trp, p.Arg142Gln, p.Arg164Trp, p.Arg164Gln, p.Pro172Ala, p.Asn205Ser, p.Val13Glu, p.Glu186Gly, and p.Met194Ile [132]. The disease was reproduced in several mouse models (e.g.,: [133], by knocking out Gjb1 or by CRISPR/Cas9 genome editing to induce mutations in p.T55I, p.R75W [134], or R15Q [135] in Gjb1. Although over 100 other genes were blamed for CMT1X [136], GJB1 remains the main suspect and animal modeling are focusing on manipulating this gene [136].

Because males have only one chromosome X while females have two, GJB1 pathological alterations occur more frequently, at younger ages and with higher severity in men than in women with Cx32 deficiency [137]. We compared the brain transcriptomes of Cx32KO and wild-type C57Bl/6 male mice and found that together with Gjb1, hundreds of other genes were significantly regulated in Cx32KO mice. Moreover, the significant overlap between the brain regulomes of Cx43KO and Cx32KO mice indicates a pan-glial transcriptomic continuity in the brain [138] although the two genes are expressed in different cell types. However, there is no known overlap of the ODDD and CMT1X syndromes.

3.4. Temporal Lobe Epilepsy (TLE)

TLE is a group of chronic brain disorders characterized by recurring seizures (abnormal, paroxysmal changes in the electrical activity) of the brain [139], excitotoxicity, neuronal loss and cognitive decline [140,141]. A frequently used rat model is obtained by lithium–pilocarpine-induced status epilepticus [142]. Also in use are the kainic acid model [44,143], obtained like the pilocarpine model and flurothyl model [144] by administration of a chemoconvulsant. Other models are obtained by traumatic brain injury [144], electrical kindling [145] and genetic manipulation of the sodium channel gene SCNA1 [146]. The efficiency of cannabidiol treatment of TLE was tested on a mouse kindling model [147].

It was reported that GJD2 (gap junction delta2) which encodes connexin 36 (Cx36), the main neuronal gap junction channel forming protein, plays a major role in epilepsy [148,149]. This finding stimulated the use of quinine, a blocker of the interneuronal Cx36 gap junction channel, as an effective suppressor of seizures [150,151]. Our transcriptomic study of Cx36KO mouse brain revealed hundreds of other genes being significantly regulated in addition to Gjd2. We also found that the Cx36KO mouse brain regulome was largely different from those of the Cx43KO and Cx32KO mice [122,138], indicating a lack of transcriptomic interaction between neurons and glial cells.

3.5. Juvenile Myoclonic Epilepsy (JME)

JME, a subsyndrome of idiopathic generalized epilepsy [152], was simulated in several transgenic mouse models including Efhc1 (EF-hand domain containing 1)-deficient [153], Kcnc1-p.Arg320His/+ [154] and Brd2+/- [155]. Interestingly, none of the mutations of CILK1 (ciliogenesis associated kinase 1)/ICK (intestinal cell kinase), blamed for JME in humans [156], reproduced the JME phenotype in mice [157], questioning the validity of the gene biomarker paradigm. The diversity of the animal models is justified by the multi-gene JME etiology. Moreover, in our hands, the haploinsufficiency in bromodomain containing 2 (Brd2+/-) JME mouse model reveled sex specific behavioral traits [158] that we explain by the observed sex differences in the networking of neurotransmission genes [159].

3.6. Neuropsychiatric Lupus Erythematosus (NPSLE)

Patients with NPSLE present cognitive impairment, psychosis, anxiety, mood, and movement disorders, confusional state, memory loss, seizures, stroke, and headache [160]. The interaction of the tumor necrosis factor (TNF)-like weak inducer of apoptosis (TWEAK or TNFSF12) with its cognate receptor, FN14 (TNFRSF12A, expressed in astrocytes, microglia, brain microvascular endothelial cells, and neurons), is believed to activate pro-inflammatory cytokine production [161]. We profiled the transcriptomes of brain cortices and hippocampi of MRL/+, MRL/lpr (that manifest lupus-like phenotype [162] and MRL/lpr-Fn14 knockout (Fn14ko backcross generation #8) adult female mice, all sacrificed at diestrus of the estrogen cycle to minimize the biological variability caused by hormonal activity. The experiments revealed significant alterations in the chemokine and PI3K/AKT signaling pathways [163] and an interesting link between neuroinflammation and neurodegeneration [164] Currently in use is also the tetramethylpentadecane (known also as pristane)-induced NPSLE mouse model [165] with various reporters to visualize the cellular dynamics [166].

4. Chemical, Hormonal, Viral and Mechanical Induction of Neurological Diseases in Animal Models

4.1. Multiple Sclerosis (MS)

MS is an inflammatory demyelinating disease that alters the neuronal circuits in the brain and spinal cord, resulting in paralysis and death ([167,168]. MS symptoms were satisfactorily replicated in rodents’ adoptive transfer experimental autoimmune encephalomyelitis (AT-EAE) [169,170]. Current MS therapy includes administration of immunomodulators and immunosuppressants, but it just alleviates the symptoms and reduces relapse frequency without stopping progression of disability [171]. AT-EAE, the golden standard MS murine model, is induced by injecting myelin basic protein dissolved in sterile phosphate-buffered saline and emulsified with incomplete Freund’s adjuvant supplemented with Mycobacterium tuberculosis emulsion in the rodent spinal cord [172]. The disease develops after 8 - 10 days, the animals manifesting ascending paralysis, starting with the hind limbs.

Our study on spinal cords of SJL/J adult, clinical index 4 (hind- and front-limb paralysis) female AT-EAE mice revealed that Gja1 is down-regulated, and its downregulation correlates with a substantial increase in the populations of dystrophic neurons and monocytes [173]. We also found that, together with the immune response, the expression control and the coordination of several genes included in the functional pathways Ca2+-signaling, cell cycle, cytoskeleton, energy-metabolism, RNA-processing, and transport of ions and small molecules were also altered [174].

4.2. Glaucoma (GL)

GL is a group of eye diseases leading to blindness that is caused by degeneration of the retinal ganglion cells (RGC), sometimes triggered by damage to the optic nerve [175]. The disease was induced in domestic chicks by exposure to continuous intense light [176], in rabbits by lensectomy, vitrectomy, and transvitreal photocoagulation of the ciliary processes [177] and in rats by injection of magnetic microspheres [178] or hydrogel [179] into the anterior chamber. It was also obtained in rats by external ocular compression through circumlimbal suture [180] or by crushing the optic nerve [181], and in mice by subretinal or intravitreal injections of fluorescently labeled mitochondria [182].

We explored the retinal transcriptomes of adult Lister Hooded rats two weeks after optic nerve crush and compared them with those of sham-operated controls to experience similar surgical stress. Experiments have shown about 60% reduction in the number of viable RGCs and significant alterations in the expression level, control and networking of genes involved in the complement cascade and Notch signaling functional pathways [183]. However, excepting exposure to intense light, the other methods to induce glaucoma in animal models are rarely the cause of the disease in humans and therefore RGC might follow a different trajectory [184]. Moreover, in addition to the species, the relevance of the animal model depends also on the strain [185].

4.3. Catamenial Epilepsy (CE)

CE, also known as menstrual seizures, is a cyclical change of seizures frequency at various stages of the menstrual cycle [186] owing to the interdependence between β-estradiol concentration and synaptic transmission. CE was partially induced in rodent female models by extended exposure to high levels of progesterone and estrogens followed by rapid decline [187] or by administration of exogenous hormones in ovariectomized rats [188]. Recent models use optogenetics, consisting in stimulation of membrane ion channels of parvalbumin-positive interneurons neurons by blue light pulses [189].

To quantify the influence of the sex hormone on the epilepsy development, we profiled the hippocampal dentate gyrus transcriptomes of intact and ovariectomized (OVX) 8–9-week-old female Sprague-Dawley rats with kainic acid-induced status epilepticus. Kainic acid was injected through mini-osmotic pump implants [190] and the success of castration was confirmed by measurement of vaginal impedance. The estrogen protection of neurotransmission was quantified by comparing the group that received 17β-estradiol benzoate with the group that received only sterile peanut oil, both with intact females [45]. From the perspective of estrogen production, OVX female rats recapitulate the menopausal state despite being forced into a young female state, whereas in women, it occurs naturally at mature ages. However, the epilepsy in menopausal women is not triggered by kainic acid, so the dynamic of the disease progression in humans is most likely different from what was observed in rats [191].

4.4. Infantile Spasms (IS)

IS (epileptic spasms during infancy or West syndrome [192]) start in the first year of life, the infant presenting stereotypical spasms, chaotic brain hypsarrhythmia and developmental delay [193]. Most likely caused by hypothalamic dysfunction, IS infants have decreased concentrations of adrenocorticotrophic hormone (ACTH) and cortisol in cerebrospinal fluid. Therefore, ACTH is the FDA-approved first-line treatment of IS [194]. There are several genetically engineered mouse models of IS in use including: Arx (Aristaless-related homeobox) Knock-In (Arx(GCG)10+7), Arx Conditional Knock-Out (cKO), APC (Adenomatous polyposis coli) cKO, and Tsc1+/ (Tuberous sclerosis 1) [195,196]. IS was also induced in rats by infusing tetrodotoxin (TTX) into the developing hippocampus [197].

We preferred the two-hit model where IS is triggered by postnatally administrating N-methyl-D-aspartic acid (NMDA) to prenatally primed with betamethasone rat pups. The model is based on the observation that prenatal exposure to corticosteroids like betamethasone in difficult pregnancy affects children’s mental development [198,199]. We explored the transcriptomic effects of prenatal (G15) intraperitoneal injection with betamethasone on synaptic transmission in the prefrontal cortex and in the arcuate and paraventricular hypothalamic nuclei of Sprague Dawley rats of both sexes [200]. From postnatal day 12 (P12) to P15, male and female pups who received either betamethasone or saline (G15) prenatally were injected with either NMDA to trigger the spasms or saline. The IS positive rats were then randomly divided into 3 treatment groups: saline, ACTH or PMX53 (a potent, selective, cyclic hexapeptide, C5ar1 antagonist [201]). We found a significant sex dichotomy with males exhibiting more transcriptomic alterations of the (glutamatergic, GABAergic, cholinergic, dopaminergic, and serotonergic) synaptic transmission but also more efficient recovery following the anti-inflammatory treatment [47,200,202]. Importantly, the transcriptomic effects were significantly different in paraventricular nuclei collected from the same animals, indicating regional brain specialization in neurodegenerative diseases [29].

4.5. Germinal Matrix Hemorrhage - Intraventricular Hemorrhage (IVH)

IVH, a major neurologic complication of 20% of premature (< 1500g at birth) infants [203], is responsible for brain injury (including post-hemorrhagic hydrocephalus), cerebral palsy (10%), and mental retardation [204]. IVH was induced so far in mouse [205,206], rat [207,208] and rabbit [209,210] animal models. We used the standard protocol for rabbits delivered prematurely by C-section at G29 (32 days full-term) that were injected intraperitoneally 3-hour postnatal with either 50% glycerol (6.5g/kg) or saline [211]. Presence and severity of IVH were assessed 24h post injection by head ultrasound. Oligodendrocyte precursor cells were isolated from coronal slices cut at the level of the head of the caudate nucleus from the frontoparietal region. The IVH rabbit model was used to determine the efficacy of the peroxisome proliferator activated receptor-γ (PPAR-γ) in enhancing myelination, and the efficacy of the 17β-estradiol treatment to restore the hippocampal dentate gyrus development [212]. We found that estrogen treatment increases the number of calbindin interneurons and prox1 neurons in the hippocampus dentate gyrus [213].

4.6. Cerebral Palsy (CP)

Children with cerebral palsy [214], astrogliosis and motor impairment present ataxia (lack of muscle coordination in voluntary movement), spasticity (stiff or tight muscles and exaggerated reflexes), weakness in one or more arms or legs, walking on the toes, a crouched gait, or a “scissored” gait [215]. Postnatal administration of glucocorticoids (GC) in premature infants for the treatment of lung disease was associated with CP and neurodevelopmental delay [216]. CP was induced in mouse [217], rat [218] and rabbit [219].

We studied the effects of postnatal administration of glucocorticoids dexamethasone and betamethasone on the forebrains of preterm (GD29) rabbits. The experiments have shown that postnatal GC induces hypomyelination, gliosis and neurologic deficits [220].

4.7. Cerebral Malaria (CM)

CM, a severe neurological manifestation of Plasmodium falciparum infection is associated in 50% of cases with cognition deficit, memory impairment, visual ataxia, seizures, hemiplegia, psychiatric disorders, and deficient motor coordination [221,222,223]. The most popular mouse model is obtained through intraperitoneal injection of blood infected with Plasmodium berghei ANKA, a single-celled parasite from the subgenus Vinckeia [224,225].

Our gene expression study with CM simulated in 5-week-old C57BL/6j female mice [226] infected with Plasmodium berghei ANKA revealed substantial transcriptomic alteration of genes involved in chromatin remodeling, cell development, negative regulation of apoptosis, lipid metabolism, hydrolase activity, walking, and regulation of muscle contraction. However, despite their accessibility and affordability, murine models of CM are still far from human CM. Therefore, efforts are made to refine organoids, spheroids and organs-on-chip with human cells collected from various organs affected by CM [227].

5. Experimental Issues

5.1. Technical Noise

From sample collection to results delivery any molecular investigation is affected by the technology limited resolution to detect small concentrations, distorted outcomes by saturation at high concentrations, and inherent technical noise that produces nonuniform background signals. Both resolution and saturation limits are continually improved by technological advancements in gene expression platforms. They can also be optimized through sample preparation protocols (adjusting sample dilution and the amounts of fluorescent markers), more efficient processing kits, and adjustments of scanning parameters.

RNA-sequencing is based on converting the mRNAs obtained by rRNA depletion/mRNA enrichment of total RNA extracted from the tissue of interest into fragments of complementary DNAs (cDNAs) which are then aligned to standard or de novo genome. There are several potential noise-generating steps in library preparation, clustering and sequencing. Part of it is caused by the impossibility of fully eliminating hybridization errors and washing out the no longer needed fragments.

Microarray technology relies on selective hybridization of fluorescently labeled cDNAs obtained by reverse transcribing of total RNA with tens of thousands of ~1000bp DNA clones (the cDNA method) or with 50 – 70 mer oligonucleotide sequences spotted on the treated surface of a glass slide. Because of this, the microarray can detect only the transcripts that stably hybridize to the spotted sequences and are not removed by the washing. Moreover, inherent slight differences on the deposited cDNAs or oligonucleotides amounts occur among the microarray domains printed on the glass slide with different pins that would bias the results if not corrected by an appropriate normalization procedure.

The protocol for quantitative reverse transcription polymerase chain reaction (qRT-PCR) involves reverse transcription of RNA into cDNA, followed by quantitative PCR to measure RNA levels by comparing to a standard curve. As such, qRT-PCR is affected by technical noise during RNA extraction, purification, cDNA conversion, and amplification.

A gene may be transcribed into numerous mRNAs differing from each other by what exons were used, how much their normal succession was preserved, whether their nucleotide sequences were integral or truncated, and whether portions of exons from other genes were translocated. By evidence, local conditions are neither uniform nor constant, depending on various physical and chemical interactions at the site of transcription. Therefore, the distribution of gene transcripts depends on the animal strain, sex, tissue, disease development, and response to treatment and/or other external stimuli. As a result, a microarray study might miss some very important transcripts in the investigated tissue and condition. On the other side, increasing the number of distinct transcripts picked by the RNA-sequencing is limited by the increasing time and price of the sequencing.

Almost all our experimental studies with qRT-PCR, microarrays or NextGen RNA-sequencing started with profiling technical replicas to determine the technical noise of the method. We found that all microarray and sequencing platforms were affected by 25 – 35% technical noise when respecting the manufacturer protocol, and even qRT-PCR, the so-called “golden standard” for gene expression profiling, is affected by noise [228] (~12% Iacobas, unpublished results). Figure 1 shows how the observed expression ratio of a real 2x up regulation and of a real expression equality are affected by the technical noise, potentially transforming into false negative or false positive results.

By evidence, (in)/validating regulation of one or a few genes with qRT-PCR out of tens of thousands quantified by a high throughput (microarray or RNA-sequencing) platform has no statistical significance. Therefore, if needed, results of high throughput platform should be checked only with another high throughput platform using technical replicas of the investigated specimens. Owing to lower price and possibility to optimize the protocol, we preferred Agilent microarrays [229] (where the noise was reduced to 15 – 20% with our wet strategy and normalization algorithm), even though microarrays profiled redundantly only the 44k transcript variants printed on the slide.

Nevertheless, one very important challenge for any optics-based transcriptomic experiment is the choice of fluorescent tags to be efficiently attached to nucleotides used in the reverse transcription. In our experiments, we used both the pair of organic cyanine dyes Cy3 (green emission)/Cy5 (red emission) and the pair Alexa Fluor^® 555 (green)/Alexa Fluor^® 647 (red). The cyanide pair is affected by different stability and fluorescence intensity of the two fluorophores. While the Alexa pair has the stability advantage across multiple scanning, the cyanides provide a slightly better foreground-to-background ratio [230]. Because of the differences between the green and red dyes, the traditional protocol recommends either profiling each sample twice by swapping the dyes or using a red-labeled universal reference to refer to both tissues to be compared [231]. Each of these recommendations doubles the experimental costs either by twice profiling the same sample or by using a totally not interesting specimen (the reference). We avoided such waste by adopting the “multiple yellow strategy” where biological replicas were hybridized against each-other on two-color microarray slides. Our strategy eliminates the bias between the two dyes because, instead of red-to-green ratio we considered red-to-red and green-to-green by normalizing the net signal of any transcript to the median expression of all transcripts in that microarray like in a single-color experiment. Figure 2 presents the flow-chart of our computational procedure in microarray experiments).

5.2. Analysis Issues

Traditional analysis is limited to characterizing the transcriptome only by the gene expression profile, like reducing the blueprint of a supercomputer to a list of the electronic components, ignoring how these components are wired and what voltages are applied to each of them. Therefore, we characterize each transcript gene by the independent measures AVE (average expression level), REV (relative expression variation) and COR (expression correlation with each other gene) across four biological replicates (defined in Appendix). Compared to traditional analysis limited to only the average expression levels AVE of N genes, our strategy considers also N REVs and N(N-1)/2 CORs the entire information available, increasing the amount of data by several orders of magnitude.

The additional independent measures, REV and COR provide valuable biological insights on cellular transcriptomics. REV (midinterval of the chi-square estimate of the coefficient of variation) quantifies through the derived measure REC (Relative Expression Control) the strength of the homeostatic mechanism to limit the random fluctuation of the gene expression resulting from the stochasticity of the chemical reactions involved in gene transcription. High positive RECs indicate genes that are critical for cell physiology, while large negative RECs point to genes used as vectors of adaptation to continuously changing environment [29]. By comparing the REC of the disease state with the corresponding control, one determines cell priority changes in controlling expression of important genes.

COR is the pair-wise Pearson correlation coefficient of the (log₂) expression levels of two genes across biological replicates. COR analysis determines the statistically significant transcriptomic networks of functional pathways in the actual condition and their remodeling during the disease development or in response to a treatment. Traditional correlation network analyses [232,233,234] cluster the genes according to their expressions in time series, various experimental conditions and other unmatched samples allegedly to reveal system-level properties of prognostic genes. For instance, Oldham et al. 2012 [235], clustered the genes expressed in caudate nucleus, cerebellum, primary motor cortex, and prefrontal cortex in 36 healthy and 44 patients across five grades of Huntington Disease, matched as age and sex.

COR-determined networks are closer to the biological reality than the constructed functional pathways by text-mining software like Pathway Analysis [236], DAVID [237], KEGG [238] or GenMapp/MAPPFinder [239] which are the same regardless of race/strain, sex, hormonal activity, age, diet, climate or other influential factors. Moreover, the gene “wiring” of such inferred pathways is unique (no alternative interlinks) and rigid (does not change during progression of the disease or in response to treatment). Importantly, calcium signaling and several other signaling pathways are considered the same in all tissues.

Expression of a gene is usually considered as significantly regulated when comparing two conditions D (diseased) vs. C (control) if the p-value of the heteroscedastic t-test of means’ equality is below a certain value, usually 0.05. We added the condition of the absolute fold-change |x| to exceed a cut-off value CUT computed for each transcript to incorporate both the biological variability across biological replicates and the technical noises of the probing spots in the used microarrays (see Appendix).

6. Discussion

This Review presents issues and sources of errors we faced in decades of profiling the gene expression on cells and tissues collected from animal models of several human neurologic diseases. Online links to publicly available experimental protocols and results of our studies are provided in the References section for readers interested in reanalyzing our raw data or comparing them with expression profiles of other models (like those mentioned for every disease discussed). Even though limited to 12 major pathologies induced in a few strains of mice, rats and rabbits, the above discussed aspects might be common to transcriptomic studies on almost all animal models of disorders affecting the human nervous system. We have learned firsthand that, beyond animal species, strain, sex, age and explored tissue, factors like hormonal status, diet, exposure to stress, toxins, radiation, microgravity, treatment and other external stimuli and housing conditions determine the results.

Nevertheless, although mimicking the main features of the human disease, an animal model accounts only for a part of the complex, yet not completely known, etiology of that disorder. Despite practically all neuro-diseases are multi-factorial, most of the transgenic animal models had only one gene (termed biomarker) experimentally manipulated (e.g., [240]). Double transgenic (e.g., [241,242]) or even triple transgenic (e.g., [243,244] animal models are more rarely used. Moreover, in addition to the induced mutation(s), transgenic animals present at any time millions of other mutations simply caused by the stochastic nature of DNA replication chemistry and downstream ripple effects of genetic manipulation. Also, together with the targeted gene, expression levels of hundreds of other genes are spontaneously regulated owing to their networking in functional pathways.

Since gene expression profile is strongly dependent on the cellular environment, it is also important to select the most homogeneous part of the tissue to investigate and caution the interpretation when using either primary or immortalized cell cultures. Despite being cheaper, free of IACUC or IRB constraints and much easier to profile, primary or immortalized cell monocultures are the last choice for them not reproducing with enough fidelity the remodeling of functional pathways as happens in the regular hetero-cellular tissue [245].

Although the actual technology is far better than what we used at the time, a good amount of actual knowledge was generated by studies like ours and we believe it is useful to discuss their trustworthiness. Nevertheless, methods such are Frozen Immunolabeled Nuclei Sequencing [246], Single-Cell Combinatorial Fluidic Indexing [247], neuronal cultures in 3D thin gel [248], spatial transcriptomics [249] and Nerve-on-a-Chip [250] to mention just a few, are more suitable to characterize the high complexity of the nervous structures under controlled environment but these spectacular technologies still need refinement and validation with technical replicates.

7. Conclusions

Notwithstanding the limited accuracy partially caused by the oversimplified etiology and the inherent technical noise, the animal models of human neurological disorders are still very important tools to decipher the fundamental molecular mechanisms of simulated pathology. In addition to mouse, rat and rabbit, neurological diseases were also studied in pig, dog, cat, horse, goat, and other mammalian models [251,252,253,254,255,256,257,258,259,260,261,262,263]. A notable advantage of the animal model is the possibility to control most of the favorably factors of the simulated disease in biological replicates, while every human is who s/he is, different from everybody else, even from a monozygotic twin [264,265,266].

Appendix A.1. The Three Independent Measures of the Gene Expression

$$\forall i,j=1÷N\left(genes\right)\&\forall c=condition/region$$

$${AVE}_i^{\left(c\right)}\equiv {\displaystyle \frac{1}{4}}{\displaystyle {\sum }_{k=1}^4}\left({{\displaystyle \frac{{\propto }_{i;k}^{\left(c\right)}}{〈{\alpha }_{i;k}^{\left(c\right)}〉}}}_{i;k}^{\left(c\right)}\right)with{a}_i^{\left(c\right)}={\displaystyle \frac{1}{4}}{\displaystyle {\sum }_{k=1}^4}{\propto }_{i;k}^{\left(c\right)}$$

(1)

$${REV}_i^{\left(c\right)}\equiv {\displaystyle \frac{{\sigma }_i^{\left(c\right)}}{2{AVE}_i^{\left(c\right)}}}\left(\sqrt{{\displaystyle \frac{r_i}{{\chi }^2\left(\beta ;r_i\right)}}}-\sqrt{{\displaystyle \frac{r_i}{{\chi }^2\left(1-\beta ;r_i\right)}}}\right)$$

(2)

$where:=stadrddeviation,r=freedomdegrees,2=chi-squarescore$

$${COR}_{i,j}^{\left(c\right)}\equiv correl\left({\mathit{log}}_2\left({{\displaystyle \frac{{\propto }_{i;k}^{\left(c\right)}}{〈{\alpha }_{i;k}^{\left(c\right)}〉}}}_{i;k}^{\left(c\right)}\right),{\mathit{log}}_2\left({{\displaystyle \frac{{\propto }_{j;k}^{\left(c\right)}}{〈{\alpha }_{j;k}^{\left(c\right)}〉}}}_{i;k}^{\left(c\right)}\right)\right);k=1÷4$$

(3)

Appendix A.2. Significant Expression Regulation

Absolute fold-change and p-value composite criterion

$$\forall i,j=1÷N\left(genes\right)\&C=control,D=disease$$

$$\left\{\begin{array}{c}\left|x_i^{\left(D/C\right)}\right|>{CUT}_i^{\left(D/C\right)}\\ p_i^{\left(D/C\right)}<0.05\end{array}\right.,where:$$

(4)

$$x_i^{\left(D/C\right)}=\left\{\begin{array}{c}{\displaystyle \frac{{AVE}_i^{\left(D\right)}}{{AVE}_i^{\left(C\right)}}},if{AVE}_i^{\left(D\right)}>{AVE}_i^{\left(C\right)}\\ -{\displaystyle \frac{{AVE}_i^{\left(C\right)}}{{AVE}_i^{\left(D\right)}}},if{AVE}_i^{\left(D\right)}\le {AVE}_i^{\left(C\right)}\end{array}\right.$$

(5)

$${CUT}_i^{\left(D/C\right)}=1+\sqrt{{\displaystyle \frac{2}{100}}\left({\left({REV}_i^{\left(C\right)}\right)}^2+{\left({REV}_i^{\left(D\right)}\right)}^2\right)}$$

(6)