arXiv
Cross-species EEG beats mouse seizure counts in epilepsy drug screens
Mouse seizure counts can improve without proving that a drug has corrected the brain activity linked to human epilepsy. Exin Therapeutics authors placed mouse and human EEG in a shared disease map and tested 10 model–drug combinations, with the measured neural movement correlating 0.87 with a ranking of known human efficacy. If prospective tests confirm the link, an epilepsy trial could match patients and treatments by neural state rather than diagnosis alone.
A mouse can have fewer seizures after a drug and still tell us little about whether the drug will help a person. Human epilepsy is not one single disorder, and each mouse model captures only part of it. A treatment can improve a mouse-specific measurement without fixing the brain-circuit problem that matters in patients. Drug developers therefore still lean on seizure counts and related readings that may have no clear human counterpart.
A shared map for two kinds of brains
Electroencephalography, or EEG, records the brain’s combined electrical activity. Doctors routinely use it to diagnose epilepsy, but the signal changes with the electrodes, the animal or person being recorded, the surroundings and what they are doing.
The researchers’ key idea is to make a shared map from these messy signals. In machine learning, representation learning means compressing complicated data into a smaller set of coordinates that preserve a chosen pattern. Here, the chosen pattern is biological state: healthy activity, disease activity and the changes between them.
The map does not assume that one mouse model represents every person with epilepsy. Instead, it can show which human neural states resemble which mouse models. In the epilepsy analysis, PTZ and AY9944 mice most closely resembled human absence seizures, while 4-AP mice aligned more with tonic-clonic and tonic seizures.
Measuring a drug by where the brain moves
The authors trained a model to place mouse and human EEG in the same disease space. They first tested whether this was possible in a simpler setting: humans and mice received matched sounds and images. Five of six sensory tests produced a shared direction in the map, although the oddball test did not.
They then used EEG from people with and without epilepsy alongside three mouse models. The model was trained without recordings from treated animals. That matters because the drug test could then ask a new question: after treatment, did the mouse’s activity move toward the human healthy region?
The researchers call this movement a neural-rescue score. Across ten mouse model–drug combinations, that score tracked an independently assigned ranking of how well the drugs work in people. Its correlation with the human ranking was 0.87. Conventional measures, including seizure count, high-amplitude time and time spent in seizure, had weaker associations. In the same comparison, seizure count reached 0.63 and did not meet the paper’s significance test.
The result also captured an important warning. In the AY9944 model, tiagabine moved activity away from the healthy direction. The paper notes that tiagabine can worsen absence seizures in patients. A method that only rewarded fewer visible seizures could miss that kind of failure.
The authors also report a separate test involving mice with an Fmr1 mutation and people with 16p11.2 deletions or duplications. The human groups shifted toward the mouse disease direction compared with controls, suggesting that the same kind of map can connect different models and recording methods.
From mouse cages to patient trials
If this result survives larger drug panels, independent data and prospective tests, a development lab could stop advancing a compound merely because mice have fewer detected seizures. It could instead ask whether the treated EEG moves toward the human healthy pattern linked to the intended patient group.
In an epilepsy clinic, a routine EEG might eventually place a patient near a particular neural state. That could help researchers design trials around the circuit problem a person has, rather than only the broad label “epilepsy.” It could also help choose the mouse model that best matches that patient group.
That future is not established yet. The pharmacology result was retrospective, covered only ten combinations and compared different epilepsy syndromes and clinical endpoints. The authors are employees of Exin Therapeutics; two hold company equity, and the company has filed a provisional patent application covering the method. The next test is a prediction made before a clinical result is known, using human and animal data that actually cover the disease states being targeted.