Why epilepsy is a good case disease to study in further detail
Epilepsy represents a unique case disease for us to evaluate in this project for several reasons. First, it is within the field of expertise of the three clinician collaborators (neurology). Second, a large majority of drugs undergone evaluation for repurposing in epilepsy; this means we largely know which drugs are epileptogenic (exacerbate epilepsy) and which are anti-epileptic (AEDs, drugs given in epilepsy). Thus study of this disease is a good way to validate the predictions of the rephetio project. Third, the distribution of high-scoring predictions in this disease were higher than most. This allows us to characterize the association between prediction score and likelihood of successful drug repurposing (see Daniel's post below for citations). It should be noted that it is unclear if the results seen in epilepsy would necessarily generalizable to other diseases. It would also be prudent to evaluate whether diseases with a large number of high-scoring predictions are the same diseases which had a large number of input disease-drug connections (which is certainly true in epilepsy). Perhaps it is a case of more information in, more information out, and if so it would be nice to characterize this.
I believe it is important for drug-predictions within a given disease to be evaluated by an appropriate clinical specialists. The clinicians collaborating on this project (including myself) are all neurologists by training. I personally find it extremely difficult to evaluate drug predictions for non-neurologic diseases, I would estimate it requires 10x more time for me to research and evaluate each prediction in a non-neurologic disease. Epilepsy is one of the most commonly encountered neurologic diseases in clinic, and I do feel we are fit to evaluate it efficiently and expertly.
Repurposing is more advanced in epilepsy than the majority of other human diseases, we know whether most drugs improve or exacerbate epilepsy. Small animal models of epilepsy are incredibly quick and inexpensive and rather highly applicable to humans. Anti-epileptics work in a dose dependent manner to increase the seizure threshold. The seizure threshold is the amount of electric stimulation that is required to induce seizure. Experimentally, this is done in small animal models by providing graduated increasing electric stimulation until a seizure is attained. The results are generally highly reproducible and thus it requires a very small number of animals. The results are often reproducible between species, further decreasing cost. Furthermore, within the same experiment, one can determine whether a medication would exacerbate epilepsy, these are the drugs that lower the seizure threshold. Given the relatively inexpensive (time and money) nature of pharmacologic testing in epilepsy, drug companies have already attempted drug-repurposing experiments on vast arrays of drugs. Thus, we have a large degree of information about which drugs would possibly work in epilepsy. This makes epilepsy an interesting case to study when looking to validate our repurposing rankings. (there are good reviews cited by Daniel below)
Epilepsy is an outlier for prediction scores. It is the only disease where 1% of drugs had a prediction score higher than 0.35. In the attached pdf, I plotted the distribution of disease scores for each disease, and I'll make a few observations as a digression from epilepsy: hypertension and asthma also have large numbers of drugs with high prediction scores, we see that the highest single predicted scores are not in epilepsy, some diseases such as Paget's disease of bone have a few very highly predicted drugs but overall a low distribution of drug scores. It would also be prudent to evaluate whether diseases with a large number of high-scoring predictions are the same diseases which had a large number of input disease-drug connections (which is certainly true in epilepsy). This is certainly true in hypertension and asthma.
Epilepsy case study
The top 100 drug predictions for epielpsy were reviewed (henceforth called "predicted" drugs). This represents the top 7% of drugs predicted for epilepsy (93rd percentile and higher).
The attached powerpoint figure one shows the percent of predicted drugs that are known Anti-Epileptic drugs (AEDs), are known to have antiepileptic activity (raise the seizure threshold), are known to be harmful in epilepsy (lower the seizure threshold), or have no prior evidence ("unknown"). The top 1% of predicted drugs are all known AEDs (17/17, 100%); these have prediction scores greater than 0.35. Of the top 3% of predicted drugs, the overwhelming majority (45/47, 95%) are known AEDs; these have prediction scores greater than 0.09. As we get deeper into the rankings of predicted drugs, we see an increasingly larger percentage of unknown drugs and drugs that lower the seizure threshold. By the 94th percentile, about 40% of the predicted drugs are known AEDs or have known antiepileptic properties, about 47% have unknown epileptic properties, and about 13% are known to lower the seizure threshold; the prediction scores here are around 0.03 to 0.04.
Of the top 100 drugs, 69 are primary AEDs (AEDs that are primarily used in epilepsy), another 7 have known antiepileptic properties. Thus over three quarters (76/100, 76%) are anti-epileptics or known to have antiepileptic properties. An important minority of drugs lower the seizure threshold (15/100, 15%). The remaining drugs are unknown in epilepsy (9/100, 8%). All drugs which are not AEDs are dicussed below:
Drugs known anti-epileptic properties
The seven drugs with known anti epileptic properties include three carbonic anhydrase inhibitors (methazolamide, acetazolamide, and diclofenamide), one calcium channel blocker (verapamil), and one sodium channel blocker (Ranolazine). The other two medications are indapamide and modafinil, the mechanism of action of these two drugs in increasing the seizure threshold is unknown.
Drugs that induce seizures
The 15 predicted drugs that are known to induce seizures include five are tricyclic antidepressants (TCAs) (amitriptyline, imipramine, nortriptyline, clomipramine, desipramine) three are monoamine oxidate inhibitors (MAOIs) (isocarboxazid, phenelzine, amoxapine), and two are antipsychotics (clozapine and Loxapine) and one is an antihistamine (cyproheptadine). The reason amitriptyline was chosen was explored in the discussion above, and highlights one of the limitations of our method. The TCAs are known to act on GABA-R, which is known to be important in epilepsy (also the TCAs treat migraine and many migraine drugs are also AEDs). The limitation of our method is that it does not differentiate the direction of effect on GABA-R (agonistic versus antagonistic), and thus our method suggests TCAs may treat epilepsy. I suspect the reason MAOIs and antipsychotics were chosen is similar. The other predicted drugs that are known to lower the seizure threshold are antipyrine (an analgesic), dalfampridine (K-channel antagonist), baclofen (inhibits the transmission of both monosynaptic and polysynaptic reflexes at the spinal cord level, possibly by hyperpolarization of primary afferent fiber terminals), and memantine (anti-NMDA).
Drugs with unknown effect on epilepsy
Halogenated ethers (Sevoflurane, Desflurane, and Enflurane): These are inhaled anesthetics. Some of the halogenated ethers are used in refractory status epilepticus (RSE), such as isoflurane (classified as an AED). Desflurane is also used in RSE, this should likely be reclassified as an AED before we make our final figures/tables. Enflurane and sevoflurane withdrawal can also cause seizures, suggesting that these drugs may have anti-epileptic properties.
There are multiple case reports of sevoflurane-provoking seizure-like activity, particularly in children  and where high concentrations are used in conjunction with hypocapnea . In high concentration, enflurane exhibits periods of suppression with paroxysmal epileptiform discharges in cats and rats . There have been multiple reports of seizure activity in humans after enflurane anaesthesia [5, 6]. Isoflurane has well-characterized anticonvulsant properties. Both isoflurane and desflurane can be used in refractory status epilepticus  [the worst end of the epilepsy spectrum].
Glutethimidie: This is potentially a good choice, it is a GABA agonist anesthetic which makes it a good candidate for repurposing for epilepsy. The prediction score was 0.08 and the percentile was 96th. It is an anesthetic (similar to thalidomide).
Acamprosate: This drug appears to work in promoting alcohol abstinence by acting as a GABA agonist, which makes it a good candidate for repurposing for epilepsy. The side effect profile is also reasonable. Further, if it is demonstrated to have antiepileptic, it may serve a dual benefit for recovering alcoholics (promoting alcohol abstinence while also perhaps preventing alcohol-associted seizures). The prediction score was 0.03 and the percentile ws 93rd.
Chlorthalidone: This is a blood pressure medication that acts on Na-Cl channels, potentially related to epilepsy. Other blood pressure medications (albeit of different classes) have been repurposed for epilepsy, as described in the section above). The prediction score was 0.04 and the percentile was 95th.
Phenazopyridine: a urinary analgesic which is excreted rapidly in the urine and is unlikely to have any clinical effect.
Antineoplastic agents (dabrafenib and bortezomib): These are poor choices due to the adverse side effect profiles.
Of note, I did not consider the drug Quinidine barbiurate, as this is a combination drug which is poorly handled by our analysis pipeline.
Here are relevant papers I found while learning about the pharmacotherapy of epilepsy.
In the manuscript and in the further discussion of this topic, for simplicity and better precision, we chose to call any drug with anti-seizure properties as an anti-ictogenic drug (AIGD), this includes known AEDs as well as drugs which are used as ancillary drugs in the treatment of status epileptics. We chose to call drugs which lower the seizure threshold ictogenic drugs (IGD).
Of the top epilepsy predictions, 77 were AIGDs, 8 were UNKDs, and 15 were IGDs (notebook). Of the 25 disease-modifying epilepsy drugs in PharmacotherapyDB, 23 were in the top predictions. Propofol and vigabatrin were the two that did not make the top 100 predictions and were respectively in the 90.6th and 83.7th percentile for epilepsy predictions.
I created the following visualizations of the top predictions. First, the plot of predicted probability versus compound's rank (ignoring quinidine barbiurate):
Here is a more complex version of the above plot which uses a sliding window that looks 7 predictions ahead/behind to compute category frequency within the window. The plot shows that top predictions are primarily AIGDs with some UNKDs and IGDs creeping in for lower predictions.
Casey Greene: The axes in this figure are unnatural to me. If you rotated this it would be much easier for me to naturally interpret.
Daniel Himmelstein: @caseygreene, I suspect you're talking about the second (two panel) figure. I agree. Additionally, the second figure adds great complexity with no more information compared to the first. I posted it here, but plan to proceed with only a version of the first image with compound labels removed..
Casey Greene: Yes - the second portion. I think that it could be interpretable with predicted prob on the y-axis. I do agree that the first is much clearer.
Enflurane is the inhalation agent that anesthesiologists most often avoid when caring for patients with epilepsy, because it lowers seizure threshold. In children and adults with no history of epilepsy, enflurane can cause epileptiform activity with concomitant facial or appendicular myoclonus or generalized tonic-clonic movements [1, 2, 3]. In epilepsy patients, the extent but not the frequency of spike activity on the electrocorticogram is increased . Epileptogenic foci may be activated during epilepsy surgery [4, 5]. As the depth of anesthesia is increased with enflurane, the EEG demonstrates high-voltage spikes and spike and slow-wave complexes, the spikes with burst suppression.
The mechanism of enflurane-induced hyperexcitability in humans is unclear. In animals, enflurane inhibits synapses and stimulates excitatory neuronal transmission in cortical and subcortical areas  (page 555).
Although low enflurane concentrations (1.0–1.5%) administered to a normocarbic patient (arterial partial pressure of carbon dioxide [PaCO2] equals 40 mm Hg) are not frequently associated with seizure activity , increasing enflurane concentrations (2–3%) or hyperventilating an anesthetized patient enhances seizure activity. Hyperventilation to a PaCO2 of 20 mm Hg from 40 mm Hg is associated with seizure activity at a 1% lower enflurane concentration. Because hyperventilation is frequently used by neuroanesthesiologists to decrease cerebral blood flow and intracranial pressure, enflurane is avoided when hyperventilation is indicated. An increase in PaCO2 from 40 mm Hg to 60 mm Hg increases the minimum enflurane concentration at which seizures occur by 1% .
Generalized tonic-clonic and myoclonic seizures can occur within the immediate postoperative period and, potentially, for a few days after enflurane anesthesia. The role of other CNS-active drugs remains uncertain in these cases . The convulsant effects may result from enflurane's organic and inorganic nonvolatile fluorinated metabolites .
Although anesthesiologists consider diazepam and thiopental to be anticonvulsants and use them extensively to treat seizure activity, there is some evidence that these drugs may potentiate enflurane-related epileptiform activity in humans . Nitrous oxide (N2O) does not alter epileptiform activity induced by enflurane .
@pouyakhankhanian do you think this is sufficient evidence to classify enflurane as a IGD? It's likely this information was overlooked due to the poor discoverability of these old studies.
Pouya Khankhanian: @dhimmel I would still classify as unknown because there is conflicting data. For example, it is exhibits periods of suppression with paroxysmal epileptiform discharges in cats and rats, so suppression (anti-epileptic activity) is the norm and the epileptiform discharges (ictogenic activity) is sporadic.
Acamprosate for epilepsy
Epilepsy was the top prediction for acamprosate — an approved treatment for alcohol dependence. From the results of our initial project report [1, 2]:
Given this high precision (77%), the 8 compounds of unknown effect are promising repurposing candidates. For example, acamprosate — whose top prediction was epilepsy — is a taurine analog that promotes alcohol abstinence. Support for this repurposing arose from acamprosate's positive modulation of the GABAᴬ receptor and inhibition of the glutamate receptor. If effective against epilepsy, acamprosate could serve a dual benefit for recovering alcoholics who experience seizures from alcohol withdrawal.
And from the discussion:
Accordingly, we hope certain predictions will spur further research, such as trials to investigate the off-label use of acamprosate for epilepsy.
More recently, @pouyakhankhanian identified a 2008 study (Farook et al.) that provides evidence in a single animal model that acamprosate treats epilepsy . The study found that both acampasate and diazepam (positive control) reduced "handling induced convulsions":
We pick up that acamprosate binds to the GABAA receptor as a positive modular. Furthermore, acamprosate binds to the glutamate receptor as an antagonist. Felbamate and rufinamide — which treat epilepsy — are also glutamate receptor antagonists/inhibitors. The network support provides similar reasoning to the authors of Farook et al.:
It was anticipated that diazepam would reduce alcohol withdrawal-induced seizures and this was the case. It is of particular interest that acamprosate had similar anticonvulsant actions with that of diazepam. The underlying mechanism or mechanisms for acamprosate's action are still not well understood. Early on, it was believed that the actions of acamprosate were via its effects on the inhibitory neurotransmitter GABA but this was based primarily on structural analysis with little empirical evidence to support this. Some of the strongest recent evidence supports the actions of acamprosate on glutamate function and/or receptors as the putative mechanism .
Accordingly, we continue to be excited the use of acamprosate to treat seizures from alcohol withdrawal syndrome.
Source/target edge contributions in epilepsy predictions
We've recently developed a new way to assess which nodes or relationships are contributing to a specific prediction. @pouyakhankhanianmentioned:
I'd love to see the weight given to various nodes in the top predictions for epilepsy, especially the ones in the top 100 which were not classified as AEDs.
For each prediction, we calculated source / target edge contributions and deployed them as tables in our Neo4j Browser guides. Here we will evaluate which source edges contributed to the 100 top epilepsy predictions. To do this, we summed source edge contributions across the 100 predictions, replacing the specific source compound with the generic "Compound" (notebook, dataset).
In total, 1,667 source edges contributed to the predictions. Here are the top 10:
Compound—includes—Decreased Central Nervous System Disorganized Electrical Activity
The top edge (Compound—includes—Decreased Central Nervous System Disorganized Electrical Activity) represents a pharmacologic class which overlaps with the indication of epilepsy. We attempted to avoid including such pharmacologic classes, but our solution was imperfect. Nonetheless, this edge is only responsible for 6% of the top 100 epilepsy predictions. Overall, the predictions drew on a multitude of edges: 22 source edges each contributed over 1%, while 166 contributed over 0.1%.
We also computed the contribution of each target edge to the top epilepsy predictions (dataset). In Hetionet v1.0, epilepsy has 531 relationships, of which 375 contributed to at least one prediction. Wanted to get the data up here — a more detailed analysis will follow.
Chemical similarities between of the top epilepsy predictions
Hetionet v1.0 includes a Compound–resembles–Compound relationship based on chemical similarity. Resembles relationships were included for compounds with a Dice coefficient ≥ 0.5 .
We can use the Hetionet Browser at https://neo4j.het.io to visualize the chemical similarity between the top epilepsy predictions. First, set a Cypher parameter for epilepsy_predictions by executing:
To see the network: run the following (with auto-complete turned on):
WHERE c.name in $epilepsy_predictions
And to get a table with the number of similar epilepsy predictions for each compound, run:
WHERE c0.name in $epilepsy_predictions AND c1.name in $epilepsy_predictions
RETURN c0.name AS compound_name, count(rel) AS n_similar_compounds
ORDER BY n_similar_compounds DESC, compound_name
This analysis shows that our top epilepsy predictions include many structurally similar compounds as well as many structurally dissimilar compounds. 29 of the 100 compounds resemble diazepam. Indeed, the benzodiazepines form the largest strongly connected component. The second largest strongly connected component appears to consist of barbiturates, such as phenobarbital which has 14 similar compounds. Halogenated ethers (e.g. enflurane) form a disjoint connected component.
In total, 75 of the 100 compounds resemble at least one other compound. The median number of similar compounds was 6. Finally, 40 compounds had 2 or fewer similar compounds. This analysis shows our top epilepsy predictions pick up clusters of similar compounds, but not to the exclusion of dissimilar compounds.
Categorizing the top epilepsy predictions
This post will investigate the categories of compounds in the top 100 epilepsy predictions. For this analysis (notebook), We'll use three categorizations of compounds: DrugBank categories, third-level ATC Codes, and DrugCentral pharmacologic classes.
DrugBank annotates compounds with categories "based on pharmacological action" . Compounds can have zero or more categories. 84 of the 100 compounds had at least one annotated category. The top eight categories are shown below (full table here):
In total, the 100 epilepsy predictions covered 64 DrugBank compounds. Less common categories included stimulants (modafinil), vasodilators (verapamil), potassium channel blockers (dalfampridine), and antipruritics (cyproheptadine).
Third-level ATC Codes
The Anatomical Therapeutic Chemical (ATC) Classification System is a classification of drugs produced by the WHOCC . The third-level indicates a therapeutic/pharmacological subgroup. DrugBank compounds are annotated with zero or more ATC codes. 91 of the 100 compounds had at least 1 ATC code. Together, the predictions included 26 third-level ATC codes.
See the full table here. The most common codes in the 100 predictions were antiepileptics (25 compounds), hypnotics and sedatives (21 compounds), anxiolytics (12 compounds), "anesthetics, general" (8 compounds), antidepressants (8 compounds), and antiglaucoma preparations and miotics (3 compounds). Less common codes included antiarrhythmics (quinidine), urologicals (phenazopyridine), and antiinfectives (primidone).
DrugCentral Pharmacologic Classes
DrugCentral contains pharmacologic classes of six types (Chemical/Ingredient, Mechanism of Action, Physiologic Effect, Established Pharmacologic Class, Pharmacological Action, Application) from 3 publicly-available source (FDA, MeSH, and ChEBI) [3, 4]. We included three of the six types in Hetionet v1.0. This analysis however will look at all 6 types.
See the full table here. 92 compounds had at least 1 pharmacologic class. Overall, the epilepsy predictions included 206 distinct pharmacologic classes. Due to the large size and diverse nature of this table, it's difficult to quantify. However, it's a good place to work backwards from class to compounds for the epilepsy predictions. For example, you can search for "Monoamine Oxidase Inhibitors" to find the compounds isocarboxazid and phenelzine.
The top epilepsy predictions pick up on several drug categories with known anti-ictogenic properties, such as anesthetics, GABA modulators, and of course known anti-epileptics. In addition, the predictions cover categories of drugs that wouldn't traditionally be considered, such as the antimalarial quinidine. In addition, we pick up on classes of AEDs whose primary indication is not epilepsy, such as the carbonic anhydrase inhibitors acetazolamide, diclofenamide, methazolamide, zonisamide that have traditionally been used for glaucoma, but are also effective anti-ictogenic drugs [5, 6]. @pouyakhankhanian those are my thoughts, you likely have more!
Quantifying the biological evidence behind the top epilepsy predicitons
In a previous comment, I introduced aggregating individual path contributions for all 100 top epilepsy predictions. This post will go into further depth on the topic — specifically, we're trying to understand what biomedical evidence in Hetionet supports our top epilepsy predictions. These analyses were performed by this notebook.
Contributions by metapath
First, we'll aggregate path contributions by their metapath. Here we're asking what types of paths contributed to the top epilepsy predictions (source).
Note that the total contribution of all paths across all predictions is 100. So for example, 6,358 Compound–binds–Gene–binds–Compound–treats–Disease (CbGbCtD) paths provide 20.6% of the support for the top epilepsy predictions combined. Illustrating the importance of integrative drug repurposing in epilepsy, 10 different metapaths each provided at least 1% of the total support.
Next, we aggregated metapath contributions by their source metaedge. This analysis asks what types of edges inform the algorithm about compounds for treating epilepsy?
We see that 43.8% of the support for predicted epilepsy treatments came from their targeted genes. We see that our method is leveraging many different types of information on compounds to make its epilepsy predictions. Note that which diseases a compound treats (excluding epilepsy of course) did not play a major role, as we suppressed many of these features .
If we aggregate instead by target metaedge, we can understand what aspects of epilepsy our algorithm leveraged:
In the context of top predictions, our algorithm understood epilepsy primarily by its known treatments (76%) and genetic associations (22%).
Contribution by target gene groups
43.8% of the support for epilepsy predictions came from paths that start with a Compound—binds—Gene metaedge. Let's dig deeper here. Are all the compounds binding to the same family or functional group of genes? To answer this question, I aggregated all paths by the gene node in their source relationship. Paths without a gene as the first intermediate node dropped out.
However, an issue arose where many genes belonged to the same family or functional group, such as encoding proteins for the same receptor. For example, see the following symbol to name mappings:
To resolve this issue, I truncated gene names at their first comma not in parentheses. Note that this this approach does not appear to work with current HGNC gene names, but does appear to work for our version of Entrez Gene [2, 3]. I also truncated gene names after the word "receptor" or "anhydrase". I also investigated grouping genes by their HGNC family, but this became complicated since genes can belong to many families with varying levels of specificity.
The top 8 gene groups — all groups with contribution > 1% — are shown below (full table here):
We see that 15.3% of the support for the top epilepsy predictions is based on the predicted compound binding to GABAᴀ receptors . In addition, 5.6% can be attributed to compounds binding cytochrome P450 proteins, which metabolize many drugs and are important for the pharmacokinetics of AEDs . Compounds that bound sodium channels — an established therapeutic target for epilepsy  — contributed 4.6%. The list continues, attesting to the ability of our method to detect and leverage bonafide targets for treating epilepsy [7, 8, 9, 10]. For example, the next genes groups were glutamate receptors , the calcium channel [12, 13], carbonic anhydrases [14, 15], cholinergic receptors [16, 17], and the potassium channel [18, 19].
In total, 90 gene groups provided support, 11 of which contributed over 0.5%.
Contribution by Side Effect node
117,720 CcSEcCtD paths provided 4.4% of the support for our epilepsy predictions. These paths traverse 1,137 specific side effects. If we restrict ourselves to Compound—causes—Side Effect source edge contributions, we can rank side effects by their aggregate contribution to our epilepsy predictions. The top five side effects reflect known adverse adverse of AEDs. In order, this approach highlights Ataxia (0.069% of total support) , Nystagmus (0.049%) , Diplopia (0.045%) , Somnolence (0.044%) , and Vomiting (0.043%) .
Contributions by Anatomy node
Finally, Hetionet contained Disease–localizes–Anatomy relationships automatically generated from MEDLINE co-occurrence. Epilepsy ended having relationships with 24 anatomies (anatomical structures, e.g. tissues) that contributed to the predictions through Compound–binds–Gene–expresses–Anatomy–localizes–Disease (CbGeAlD) paths. By aggregating CbGeAlD paths by their Anatomy node, we computed the contribution of each anatomy to our epilepsy predictions. What caught my eye is that anatomies that are more relevant to parthenogenesis and drug efficacy were near the top of the list. Examples include the telencephalon, forebrain, and nervous system. The anatomies that are related to epilepsy but primarily through its symptoms were near the bottom of the list. These examples include the tongue and meninx. This particular observation isn't well controlled, but I wanted to mention it as a future avenue of research.
Our top epilepsy predictions derive from a diverse set of supporting evidence that is supported by the current understanding of epilepsy. Aggregating path contributions allows us to evaluate the overall contribution of each network component, for a given context.
Daniel S Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, Sergio E Baranzini (2016) Cold Spring Harbor Laboratory Press. doi:10.1101/087619
V. Law, C. Knox, Y. Djoumbou, T. Jewison, A. C. Guo, Y. Liu, A. Maciejewski, D. Arndt, M. Wilson, V. Neveu, A. Tang, G. Gabriel, C. Ly, S. Adamjee, Z. T. Dame, B. Han, Y. Zhou, D. S. Wishart (2013) Nucleic Acids Research. doi:10.1093/nar/gkt1068