An official website of the United States government
NCBI Bookshelf. A service of the National Library of Medicine, National Institutes of Health.
Noebels JL, Avoli M, Rogawski MA, et al., editors. Jasper's Basic Mechanisms of the Epilepsies. 5th edition. New York: Oxford University Press; 2024. doi: 10.1093/med/9780197549469.003.0049
The advent of high-throughput molecular technologies has fundamentally transformed epilepsy research. It is now possible to simultaneously measure tens of thousands of biomolecules in a stunning array of biological contexts to discern mechanistic drivers of molecular dysregulation in seizure disorders. Correspondingly, there has been an increased appreciation that the relationship between molecular perturbations (e.g., gene mutations) and their corresponding functional outcome is highly nonlinear and mediated through vast regulatory networks. Overcoming the immense complexity of biological systems to “reverse engineer” the pathological functioning of molecular networks is one of the fundamental tasks of epilepsy research over the next 10 years. This chapter discusses how to address the yet underappreciated gene–genome interactions. It provides a comprehensive overview of both specific experimental approaches and most recent advances in analytical methods to disentangle the nonlinear relationship between a complex molecular trait and its functional consequences leading to a specific phenotype, for example, epilepsy.
The advent of high-throughput molecular technologies has fundamentally transformed epilepsy research. It is now possible to measure simultaneously tens of thousands of biomolecules in a stunning array of biological contexts to discern mechanistic drivers of molecular dysregulation in seizure disorders. Correspondingly, there has been an increased appreciation that the relationship between molecular perturbations (e.g., gene mutations) and their corresponding functional outcome is highly nonlinear and mediated through vast regulatory networks (Scott et al., 2018). Overcoming the immense complexity of biological systems to “reverse engineer” the pathological functioning of molecular networks is one of the fundamental tasks of epilepsy research over the next 10 years.
The concepts of complex systems and network theories have recently entered the views on brain malfunction and epilepsy development, thereby moving away from oversimplified relationships; for example, a single gene mutation explains the disease (Klingler et al., 2021). Indeed, while many epilepsies have strong genetic risk factors, there is substantial heterogeneity among individuals bearing the same mutation (e.g., tuberous sclerosis (Caylor et al., 2018). Conversely, different mutations in the same gene can manifest as profoundly different disorders, for example, SCN1A mutations in Dravet syndrome versus GEFS+ (Escayg and Goldin, 2010). Likewise, many epileptogenic insults (genetic, brain injury, etc.) converge on specific molecular networks that predict seizure severity (Delahaye-Duriez et al., 2016), despite the profound differences in the epileptogenic mechanisms. Thus, we have a picture suggesting that the genotype-phenotype map in epilepsy is hugely complex and that the molecular pathophysiology correlating with recurring seizures is relatively conserved (Klassen et al., 2011). This counterintuitive state of affairs arises because the molecular networks within and among cells in the brain are complex systems whose behavior is highly nonlinear. That allows seemingly small differences between perturbations (e.g., distinct mutations in the same gene) to result in highly divergent functional outcomes, while distinct perturbations (e.g., traumatic brain injury vs. gene mutations) can activate pathways that converge on the same functional outcome. The central question for epilepsy research, then, is how to identify and control these complex molecular networks to prevent the emergence of symptoms in patients. In this chapter, we will discuss some of the emerging tools and techniques of molecular systems biology that allow us to map molecular networks to resolve the heterogeneity among patients and identify points of control for future therapies.
The epilepsy genome has expanded continuously over the past decade. While traditional linkage analysis and positional cloning in families with Mendelian inheritance identified the first genes, the genetic revolution started with the next-generation sequencing (NGS) era. Today variants in over 140 genes have been identified in idiopathic generalized epilepsy and epileptic encephalopathies (Ellis et al., 2020). This ever-expanding list implicates a large number of interconnected molecular processes, including membrane excitability, synaptic function, growth, inflammation, apoptosis, and stress response. The biggest problems are to make functional sense out of these genetic variants—how does an allele produce a phenotype—and use this information to design new treatment options. This problem of genetic complexity is not limited to epilepsy but is now understood to be characteristic of all complex diseases to such an extent that geneticists have developed the notion of polygenic (even omnigenic) inheritance, for which almost every gene has a meaningful influence on nearly every trait (Leu et al., 2019; Boyle et al., 2017; Gramm et al., 2020). While omnigenic inheritance is currently only a theoretical construct still being developed empirically, it is at minimum clear that the complexity of gene regulatory and other molecular networks is a significant source of patient heterogeneity. Future mechanistic work in epilepsy must account for this complexity and, ideally, harness it for therapeutic gain or risk prediction.
It is challenging to quantify an individual patient’s genetic risk based on common genetic variants for epilepsy due to their small effect size. In addition, analyzing a single common genetic risk variant does not consider the effect other variants may have on overall epilepsy risk. Leu et al. recently combined for the first time known common genetic variants from several large epilepsy genome-wide association study (GWAS) cohorts to calculate the polygenic risk scores in people with epilepsy and population controls. By combining the effect sizes of several thousands of common genetic variants, polygenic risk scores had been shown earlier to have predictive and diagnostic power in a variety of medical conditions, including type 2 diabetes (Udler et al., 2019), psychiatric disorders (Bogdan et al., 2018), and breast cancer (Mavaddat et al., 2019; Khera et al., 2018). The study by Leu et al. was the first to examine them in the context of epilepsy. They showed that a common polygenic variant burden for epilepsy can be measured and is differently distributed among patients with epilepsy and controls as well as between the two main epilepsy phenotypes, that is, generalized and focal (Leu et al., 2019). Polygenic risk scores for epilepsies may provide physicians with an estimate of an individual’s overall genetic risk to develop epilepsy that could aid in early diagnosis and direct treatment in the future.
Another potential avenue forward to make sense of complex genetic data in epilepsy is to search for convergences across multiple epilepsies explicitly—a final common pathway—for insight into the critical pathological process. Recent studies in this direction applied gene coexpression network analysis to tissues from patient samples and rodent model systems (Delahaye-Duriez et al., 2016; Johnson et al., 2016). For example, Delahaye-Duriez et al. explicitly searched for convergence of etiologies at the molecular level across multiple epilepsies. The authors showed conserved gene expression modules, that is, groups of coexpressed genes differentially regulated between human and rodent epilepsies and respective controls, independent of etiology (genetic vs. acquired). These results show that there are common molecular endpoints across many epilepsies, implicating a systematic dysregulation of gene expression in the brains of patients with recurrent seizures independent of the cause of epilepsy. Interestingly, using bioinformatics drug databases, these authors showed that these dysregulated modules are targetable using valproate, giving potential insight into why valproate is an effective antiseizure therapy across many etiologies. However, this study only looked at established diseases by necessity and could not resolve commonalities or differences in pathogenesis, as patient brain tissue is only available from surgery.
Unlike human studies, model systems allow controlled interventions. Thus, in model systems, it is possible to identify regulators of responses to insults. Recently, Ferland et al. used systems genetics to identify genetic variants influencing epileptogenesis after a multiple-seizure insult (Ferland et al., 2017). Their strategy was to use the natural genetic variability among a panel of inbred mouse strains (the Hybrid Mouse Diversity Panel; Lusis et al., 2016) to map loci conferring resilience and susceptibility to epileptogenesis. Importantly, they showed that the population’s major susceptibility and resilience factors were distinct from the baseline seizure susceptibility in the same mice, suggesting novel genetic variants that act early in pathogenesis. This study highlights the importance of phenotype in genetic association studies. This study did not seek to identify or validate candidate epilepsy risk genes but to identify modifiers of epileptogenesis. The latter genes could provide essential clues to the early pathogenic events and the pathways that can be targeted to mitigate poor outcomes.
The above study highlights the power of natural genetic diversity that exists across standard inbred laboratory mouse strains. There have been systematic efforts to increase the genetic and phenotypic diversity of research mice, including outbreeding stocks such as the Swiss-Webster and Diversity Outbred stocks (Churchill et al., 2012), as well as developing new inbred lines with alleles from diverse parental strains, such as the Collaborative Cross (Threadgill et al., 2011). Similar efforts have been made with fruit flies (Mackay et al., 2012) and are underway with laboratory rats (Ren and Palmer, 2019). These populations complement the engineered genetic diversity made available by genome editing technologies, such as CRISPR/Cas9 (Jinek et al., 2012). With improved phenotyping, these populations could help dissect the genetic complexity of epilepsy by highlighting alleles that modify outcomes throughout the disease course, from the earliest insult to established pathology.
DNA is not just a string of code. It requires a lot of signposting to bring it to life. In recent years, the essential role of epigenetic modifications in regulating gene expression and cellular differentiation has emerged (for more details, see Chapter 35, this volume). Changes in DNA methylation, covalent post-translational modifications (PTMs) of histones, and other nuclear proteins together with noncoding-RNAs define a complex language—the epigenetic code—which regulates chromatin structure and dynamics (Wu et al., 2019). A broad variety of sequencing approaches have been developed in the past decade to assay the chromatin landscape with all possible modifications across different sample types. The utility of epigenomic data as a diagnostic or prognostic biomarker as well as its role as a molecular pathomechanism in epileptogenesis has been repeatedly demonstrated (Debski et al., 2016; Kobow et al., 2013, 2019, 2020a, 2020b; Jabari et al., 2022; Hoffmann et al., 2023; Miller-Delaney et al., 2015; Brennan et al., 2016; Morris et al., 2019; Reschke et al., 2019,). Genetic and epigenetic variation, as well as environmental and lifestyle factors, work in concert to influence human health and disease (Wu et al., 2019). Unfortunately, genomics, epigenomics, transcriptomics, and proteomics in epilepsy are still being studied mainly as compartmentalized fields. Therefore, computational methods are needed to combine these different multidimensional data types to create a comprehensive view of a given disease or a biological process. In such “multiomic” studies, data-integration methods need to overcome at least three computational challenges: (1) the small number of samples compared to the large number of measurements; (2) the differences in scale, collection bias, and noise in each data set; and (3) the complementary nature of the information provided by different types of data (Wang et al., 2014). Independent of the data integration method chosen, gene prioritization, functional annotation, and epistasis (i.e., gene interaction) are significant obstacles, as are temporal dynamics, stochasticity, and cellular heterogeneity. Unfortunately, there is no “turnkey” method to address these challenges, but substantial progress is being made.
One strategy that has been used in complex neurological disorders is similarity network fusion, which integrates multiple datasets in the space of samples (e.g., patients) rather than measurements (e.g., genes). For example, in a brain tumor study, Wang et al. constructed patient networks combining mRNA expression, DNA methylation, and miRNA expression data to identify glioblastoma subtypes with differential survival profiles. Intriguingly, the method has many other applications. In the clinical domain, patient networks allow integration of very different kinds of measurements, including metabolomics data and functional magnetic resonance imaging, together with genomic, clinical, and demographic data, as long as the data can be used to identify similarity between patients (Wang et al., 2014). Similarity network fusion and related strategies essentially (e.g., “kernel machines” in machine learning) treat all measurements as features defining a patient and make predictions about any given patient by similarity to outcomes in patients with similar feature profiles. Such methods produce compelling diagnostic and prognostic models deriving from complex interactions among features with sufficient data (Cavalli et al., 2017; Wang et al., 2021). However, they can also be challenging to interpret because they do not explicitly model the relationship between each feature and the outcome.
To dissect the relationships among features requires techniques that explicitly capture correlations among measurements. One such method is multiomics factor analysis (MOFA), a computational method for discovering the principal sources of variation in multiomics data sets. MOFA infers a set of (hidden) factors that capture biological and technical sources of variability. It disentangles axes of heterogeneity that are shared across multiple modalities and those specific to individual data modalities. The learned factors enable various downstream analyses, including identifying sample subgroups, data imputation, and the detection of outlier samples. Argelaguet et al. applied MOFA to integrate somatic mutations, RNA expression, DNA methylation, and ex vivo drug responses in patients with leukemia (Argelaguet et al., 2018).
Single-cell multiomics technologies can reveal cellular heterogeneity at multiple molecular layers within a population of cells and demonstrate how this variation is coupled or uncoupled between the captured molecular layers. The data sets generated by these techniques can enable a deeper understanding of the critical biological processes and mechanisms driving cellular heterogeneity and how they are linked with normal development and disease etiology (Chappell et al., 2018).
Mosaicism is the natural condition of all somatic tissues, including the brain. Several groups have identified somatic mutations in sporadic forms of epileptic disorders associated with structural brain lesions, mainly in focal cortical dysplasia (FCD) type 2 (Leventer et al., 2015; Lim et al., 2015; Nakashima et al., 2015; Møller et al., 2016; D’Gama et al., 2017; Ribierre et al., 2018; Sim et al., 2018; Baldassari et al., 2019; D’Gama et al., 2015; Lee et al., 2021) but also in other cortical malformations (Bonduelle et al., 2021; Kobow et al., 2020a) and low-grade epilepsy-associated tumors (Koh et al., 2018; Slegers and Blumcke, 2020; Hoffmann et al., 2023). However, from sequencing studies in bulk tissue or cerebrospinal fluid (Ye et al., 2021; Kim et al., 2021), it remained unclear which cell type was affected by, for example, the reported mTOR pathway-related mutations and whether distinct topographic distributions of small circumscribed FCD and large hemispheric malformations (i.e., hemimegalencephaly [HME]) corresponded to somatic mutations in different neuronal or glial subtypes. Whole-genome sequencing of amplified DNA from single cells may help identify mutational heterogeneity in normal and diseased tissues. D’Gama et al. used sequencing of enriched neuronal (NeuN+) and non-neuronal (NeuN-) cell populations from patients with somatic brain mutations (D’Gama et al., 2017). The authors were able to show that pathogenic somatic mutations in FCD and HME were always present in neurons but only variably present in glia. With the use of conditional mouse models, it was further shown that mTOR pathway activation in excitatory neurons and glia, but not in the interneuron lineage, was sufficient to model cortical dysplasia. Moreover, the time and place of origin of the underlying mutations were suggested to indicate lesion differences between FCD type 2 and HME.
Similar studies in ganglioglioma or mild malformation with oligodendroglial hyperplasia (MOGHE) have identified cell-type-specific somatic BRAF and SLC35A2 variants. Isolating neurons and glial cells from ganglioglioma with laser capture microdissection and subsequent targeted sequencing of BRAF in enriched cells, Koh et al. showed that the BRAF mutation existed in both cell types. Their data imply that the mutation arose from neural progenitor cells during early brain development, which differentiated into neuronal and glial lineage cells. A mouse model with somatic Braf mutation induced in neural progenitor cells during embryonic corticogenesis presented with focal lesions with dysmorphic, but non-neoplastic, neurons with cortical dyslamination, as well as neoplastic glial cells, and marked expression of CD34 similar to the histopathological features of ganglioglioma. Ninety percent of animals were epileptic. The authors proved that epileptic seizures noted in Braf mutant mice were attributable to the neuronal, but not glial, Braf mutation arising during early brain development (Koh et al., 2018). Using digital droplet polymerase chain reaction in microdissected cells, Bonduelle et al. showed SLC35A2 variants to be enriched in oligodendroglial cells and heterotopic neurons in the white matter in surgical MOGHE samples. Again these data argue that the mutation arose early in development targeting glio-neuronal progenitors (Bonduelle et al., 2021). However, it remains unclear yet how a brain somatic loss-of-function mutation in the UDP-Galactose transporter SLC35A2 and subsequent impairment of protein/sphingolipid glycosylation in the cell mediate seizures.
Single-cell mRNA-sequencing can be used to assess the molecular states of specific cell types at defined time points (Pfisterer et al., 2020; Ayhan et al., 2021). This can help to understand functional differences of cells within local networks and defined brain structures. If also high temporal resolution is provided, we can appreciate molecular programs underlying cell-type diversity in the developing and adult neocortex and relate these molecular programs to developmental disease. Telley et al. recently surveyed the transcriptional identity of cells early in mouse brain development. They showed that as a neuroprogenitor transitioned to new states, it produced daughter neurons that reflect those new states. The neuron’s postmitotic differentiation program is overlaid onto these parentally supplied programs, driving the emergence of specialized neuronal cell types in the neocortex (Telley et al., 2019). The authors provide a database summarizing these cell-type-specific developmentally regulated molecular programs (http://genebrowser.unige.ch/telagirdon/#query_the_atlas). Integrating these programs with, for example, genetic data from developmental epilepsies associated with cortical malformations may enhance our understanding of the biological relevance of certain mutations in a specific developmental time window. Despite the great promises for knowledge gain from existing technology and advances made concerning data analysis (Adossa et al., 2021), real single-cell matched multiomics studies in human and experimental epilepsy are still missing.
The genomic DNA consists of more than 3 billion nucleotides within each cell, spanning over 2 m in length. Packaging this genomic material within the micrometer-sized nuclear space requires extensive folding (Tang et al., 2015), which is presumed to be both specific and functional. However, details regarding general folding principles, distinct topologies, and relationships to gene activity are still largely unknown. Starting from 2011, the development of Hi-C-sequencing and related technologies enabled the study of the three-dimensional (3D) genome at ever-improving resolution. During chromosome conformation capture, chromatin is cross-linked with formaldehyde and then digested and re-ligated so that only DNA fragments that are covalently linked together form ligation products. The ligation products contain the information of where they originated from in the genomic sequence and where they reside, physically, in the 3D organization of the genome. This method has the power to explore the biophysical properties of chromatin as well as the implications of chromatin structure for the biological functions of the nucleus. Indirect evidence for an essential role of 3D genome organization in epilepsy comes from identifying mutations in chromodomain helicase DNA binding protein 2 (CHD2), one of the more common causes of epileptic encephalopathy (Carvill et al., 2013; Suls et al., 2013). CHD2 is an ATP-dependent chromatin-remodeling enzyme that facilitates disassembly, eviction, sliding, and spacing of nucleosomes (Narlikar et al., 2013). While we keep gaining insights into the biological function of CHD2 (Shen et al., 2015; Meganathan et al., 2017) as well as into the phenotypic spectrum of CHD2 mutations, the mechanisms leading to recurrent seizure development in CHD2 epileptic encephalopathy remain vaguely understood.
Chromosomes are organized in a variable but nonrandom manner inside the nucleus occupying so-called chromosome territories. Chromosome arrangements seem specific to cell and tissue type and can change during processes such as differentiation and development. Although this organization is rather probabilistic than absolute, it appears to be related to functional compartmentalization of the nuclear space with active and inactive genome regions separated from each other, which possibly enhances the efficiency of gene expression or repression (Meaburn and Misteli, 2007). Topologically associating domains (TADs) are the fundamental structural unit thought to guide distal regulatory elements like enhancers to their cognate promoters (Szabo et al., 2019). CTCF and Cohesin help to maintain TAD boundaries. Mutations affecting TADs and the boundary regions separating them lead to genomic rearrangements that can alter gene expression and disease in multiple ways (Lupiáñez et al., 2016). SNVs in TAD boundaries can have haplotype-specific differential effects on chromosome configuration, influencing gene expression and providing mechanistic insights into functions associated with disease susceptibility (Tang et al., 2015; Zhang et al., 2018). Moreover, SNVs in enhancer elements can lead to gain or loss of function of the target gene (Spielmann and Mundlos, 2016).
Unfortunately, as most frequently performed in epilepsy, whole-exome sequencing studies miss genetic variation that affects structural regulatory features such as enhancers, promoters, TAD boundaries, or regulatory noncoding RNAs. At the time of writing this chapter, only a single study had explored the 3D genome structure in a rodent model of epilepsy. It addressed how neuronal (hyper-) activity changed chromatin structure, accessibility, and downstream function. Fernandez-Albert et al. identified time-dependent immediate and more lasting changes in chromatin occupancy in response to status epilepticus (SE) (Fernandez-Albert et al., 2019). Despite the short time window explored, covering only the immediate response to SE, their results already support the notion of a genomic memory in epileptogenesis in the form of architectural modifications durably influencing specific transcriptional activity.
Epilepsy is by no means alone in the genomics data deluge, and we can use the experience of other domains to predict the potential for progress in epilepsy research. Artificial intelligence (AI) is an umbrella term for algorithmic methods to perform complex tasks that have historically required human judgment. Paradigmatic among these tasks is computer vision, where an algorithm is trained to recognize objects within a digital image or video. Over the past decade, a significant innovation in AI called deep learning has resulted in computer vision systems with an expert-level performance at various image recognition tasks, including analyzing biomedical images (reviewed, e.g., in Carin and Pencina, 2018). In parallel with these advances in computer vision, deep learning has emerged as a general-purpose framework for a wide array of complex data analysis tasks, which are poised to significantly influence epilepsy research for three distinct, but interrelated reasons: (1) automated analyses allow unprecedented scale for systematic and rigorous phenotyping; (2) deep learning is well adapted to generating predictive models for challenging data types; and (3) deep learning allows unbiased phenotype discovery beyond classifying and scoring features that are currently appreciated.
The most straightforward applications of deep learning to epilepsy research are about scale. In the clinical domain, deep learning has been applied to sleep scoring (Malafeev et al., 2018; Fiorillo et al., 2019), seizure detection (Craik et al., 2019), computed tomography (CT) (Chilamkurthy et al., 2018), magnetic resonance imaging (MRI) (Jin et al., 2018; Bernhardt et al., 2015; Spitzer et al., 2022), histopathology, and lesion classification (Kubach et al., 2020; Jabari et al., 2022). In the preclinical setting, deep learning systems have been developed to identify and score rodent ultrasonic vocalizations (Coffey et al., 2019), track and score behavior (Arac et al., 2019; Mathis et al., 2018), and estimate animal pose (Pereira et al., 2019; van Dam et al., 2020). The aim is to approximate a time-intensive human analysis using a high-throughput, automated pipeline in each of these applications. With these systems approaching human-level performance, one can analyze complex traits for tens of thousands of subjects. Thus, it is now possible to perform imaging genomics analysis, that is, genome-wide association studies using medical images rather than a categorical diagnosis, to identify genetic disease risk factors (Wang et al., 2012). Deep learning could dramatically expand imaging genetics, by providing novel data-driven phenotypes for genetic association studies. Going forward, in the clinical setting, we can now envision capitalizing on patient databases with tens of thousands of electroencephalogram records, MRI scans, tissue micrographs, genomic data, lab results, psychiatric evaluations, and patient history. We are now empowered by deep learning frameworks to extract rigorous phenotypes at a scale that gives us the statistical power to connect these variables. In the preclinical setting, we can envision markedly improved phenotyping for disease models. For example, we can anticipate automated scoring of rodent seizures through the standard Racine stages with pose estimation systems and automated behavioral scoring. Likewise, we will be able to extract valuable behavioral features from home-cage monitoring systems that can provide clues to comorbidities without extensive human scoring. These advantages of scale synergize with new tools for model system development (e.g., genetic diversity panels and genome editing).
Despite their origins in computer vision, deep neural networks are being applied widely to data outside imaging data. One recent application in genomics is the DeepSEA platform developed by Zhou and Troyanskaya (2015). DeepSEA is a variant effect predictor for genomic sequences. Leveraging hundreds of genomic data sets measuring chromatin accessibility, transcription factor binding sites, and epigenetic histone marks, Zhou et al. trained DeepSEA to predict hundreds of functional features based on DNA sequence alone. Their model improved significantly over previous models, in part because the deep neural network architecture allows multiscale modeling of chromatin structure that integrates from the nucleotide level up to the kilobase-pair level. As an example of the power of this framework, they show through “in silico mutagenesis” that DeepSEA could accurately predict known DNase hypersensitive alleles from ENCODE data, despite DeepSEA only being trained on the human reference genome. These data suggest that we can expect dramatically expanded functional annotation of noncoding portions of the genome in the near term. This latter point will likely prove critical for epilepsy genetics, as many epilepsy-associated single-nucleotide polymorphism variants are of unknown but likely regulatory significance. Furthermore, these results highlight the reach of deep learning systems beyond computer vision to other complex data tasks.
A third significant advantage of many deep learning systems is that they allow unbiased phenotype discovery. In the same way that high-throughput molecular technologies (e.g., RNAseq) provide high-dimensional readouts for comparing experimental groups, neural networks generate up to millions of quantitative features in the process of making an ultimate prediction. These quantitative features can be analyzed to identify the novel properties of the samples that the network is using for predictions. For example, one of the authors (JMM) recently used features from a deep neural network to identify novel histopathological features between mice with a different genetic risk for age-related kidney disease (Sheehan et al., 2019). While not related to epilepsy, this application is a proof-of-concept that DNN-based phenotyping can highlight novel traits that experts miss. In epilepsy, we can expect applications of DNNs to provide novel insights into histopathology, structural and functional imaging, and electrographic traits and relate these findings to clinical outcomes.
In a striking recent application of neural networks to systems biology, Ma et al. developed a neural network called DCell that was trained to map a yeast cell’s genotype to its growth rate and whose internal architecture was modeled on the hierarchy of biological processes encoded by the Gene Ontology (Ma et al., 2018; Gene Ontology, 2015). The authors showed that, although the system as a whole was only trained to predict growth rates from genotypes, the system’s internal states, which nominally represented biological processes, actually correlated with the activity levels of those processes across genotypes. Thus, while the training data to the network only comprised genotypes and growth rates, the self-organization of the network during training learned the underlying molecular physiology relating genotypes to growth rates. Therefore, an intriguing possibility in epilepsy is that we could use targeted mutagenesis (in neural cell culture, zebrafish, or mice) to identify the underlying molecular mechanisms of seizures or seizure-like events by modeling the genotype-phenotype map as a deep neural network. Such an approach could dramatically expand the functional characterization of human epilepsy alleles and the pro-/ anti-convulsive annotation of the genome.
The above represents potential new roads to epilepsy research in the coming decade as AI tools become more deeply embedded in systems biology. However, prognostication aside, what is clear is that capitalizing on this powerful suite of biological and analytical tools to cure epilepsy will require highly collaborative, multidisciplinary teams with expertise ranging from biochemistry to computer science to psychology to medicine. To this end, perhaps the most critical shift over the next 10 years in epilepsy research will not be the new types of data but the types of scientists that join the forefront of epilepsy research.
This is an open access publication, available online and distributed under the terms of a Creative Commons Attribution-Non Commercial-No Derivatives 4.0 International licence (CC BY-NC-ND 4.0), a copy of which is available at https://creativecommons.org/licenses/by-nc-nd/4.0/. Subject to this license, all rights are reserved.
Your browsing activity is empty.
Activity recording is turned off.
See more...
