Neural circuits put genes second in brain-drug discovery
Image drawn by AI
Brain-drug discovery can begin with the faulty neural circuit rather than the gene associated with a diagnosis, making the proposed target easier to inspect and test. Exin Therapeutics used CircuitATLAS to nominate ATP1A3 and increased it in mouse cortical interneurons, reducing chemically induced abnormal electrical activity across three frequency bands. If the circuit pattern, cell-specific intervention and clinical benefit match in people, an epilepsy clinic could choose treatment by the activity it records.
Neurological drug discovery usually begins with a gene or protein linked to a disease. That is a poor starting point when different genetic problems end in the same faulty neural circuit. Patients may receive medicines or devices that quiet symptoms, without a systematic way to find the circuit control point most likely to restore normal function.
From disease name to circuit problem
A neural circuit is a connected group of nerve cells whose activity produces a function. In many neurological conditions, the important clue is not which molecule changed first, but what the circuit is doing now: firing too much, or failing to maintain inhibition. A recording can reveal that abnormal activity, but turning it into a drug target requires connecting evidence from many places—brain regions, cell types, electrical signals and molecular biology.
CircuitATLAS is an attempt to make those connections inspectable. It is a knowledge graph: a map in which things such as proteins, cells and brain regions are linked by recorded relationships. Each link is meant to carry its source and supporting quotation, so a researcher can check the path rather than accept an AI answer on trust.
The map was built from a large literature search and other datasets. Its 3,831,997 nodes and 7,659,618 edges connect 23 kinds of entities through 105 kinds of relationships. But size does not make every link true. A second model judged 17.8% of a sample of 4,050 edges not to be clearly stated in the cited passages. The graph can expose evidence, while still inheriting errors, publication bias and missing negative results.
A target found by following the circuit
CircuitATLAS changes the order of the search. It starts with a disease-related activity pattern, then asks which circuit and cell population produce it, and only afterward looks for a molecule that could change that activity. The authors excluded direct disease-to-gene and disease-to-protein links during target generation, to stop the system from taking the familiar shortcut from a disease name to an associated molecule.
The workflow produced different proposals for three conditions: SLC17A6 for Fragile X syndrome, HTR1A for Dravet syndrome and ATP1A3 for epilepsy. ATP1A3 is part of a sodium-potassium pump in nerve cells. When a cell fires repeatedly, sodium builds up inside it. The pump uses that signal to move positive charge out, helping make the cell less likely to fire again. In an inhibitory interneuron, increasing ATP1A3 might therefore help hold back runaway activity.
That was a prediction, not a treatment. In mice given the chemical 4-aminopyridine to create excessive cortical activity, the team increased ATP1A3 in interneurons. The median change in beta-band electrical power fell from 11.16-fold to 1.33-fold. Low- and high-gamma activity also fell, while delta, theta and alpha activity did not change. The reported comparisons had q=0.0087 and a Cliff’s delta of -1.0 for each of the three affected bands.
The experiment makes an AI-generated target challengeable: the proposed mechanism can be tested in an animal instead of remaining a paragraph in a report. It does not show that CircuitATLAS discovers treatments better than an experienced neuroscience team. The authors say they have not yet tested the system prospectively or compared it with expert-led or other computational approaches.
Figure 4: A circuit-first evidence-gated agentic workflow for cross-scale target discovery.A user initiates a campaign with a central nervous system disease and therapeutic modality, after which subsequent inputs may be provided in free form. A coordinating PI agent manages the workflow, applies gates, identifies failures, and re-triggers earlier stages when evidence is insufficient. The circuit agent identifies disease-relevant activity phenotypes, circuits, regions, and cell populations using CircuitATLAS. After a gate requiring sourced and directionally coherent evidence, the molecular agent searches for effectors capable of modulating the selected circuit phenotype while direct disease-to-molecule information is masked to prevent shortcut reasoning. The modality agent evaluates the feasibility of engaging the proposed target, including intervention, vector, promoter, and campaign-level constraints, and the clinical agent assesses existing trials, biomarkers, safety, and translational evidence. An adversarial sceptic independently challenges the reasoning produced at each specialist stage. Three memory layers maintain the read-only semantic knowledge graph, an episodic campaign graph written by the PI agent, and a compacted working transcript. The workflow terminates in ranked target dossiers containing the supporting evidence, assumptions, uncertainties, and proposed route to validation; dossiers that fail predefined evidence or feasibility gates are rejected or returned for further search.To illustrate the hypotheses generated by this workflow, we applied the agent to Fragile X syndrome, Dravet syndrome and epilepsy (Table 4). In Fragile X syndrome, the workflow identified excessive thalamocortical gain and nominated partial suppression of SLC17A6 (VGLUT2) in excitatory thalamic relay neurons as a corrective intervention. Reducing vesicular glutamate loading was predicted to lower quantal transmission at thalamocortical synapses and thereby reduce cortical gain, although the essential role and broad subcortical expression of VGLUT2 make partial and anatomically restricted modulation an important constraint. In Dravet syndrome, the agent proposed partial suppression of the somatodendritic HTR1A autoreceptor in dorsal raphe serotonergic neurons, with the aim of releasing these neurons from autoinhibition and increasing forebrain serotonin tone. This strategy has clinical precedent in the indication, since fenfluramine is approved for seizures associated with Dravet syndrome and acts in part through serotonergic mechanisms, although the workflow also surfaced potentially divergent effects of 5-HT1A signalling across seizure phenotypes. In epilepsy, CircuitATLAS nominated ATP1A3, the neuronal α\\alpha3 Na+/K+-ATPase, as a control point on cortical excitability, with potentiation predicted to increase activity-dependent hyperpolarizing current and preserve inhibitory restraint during sustained firing. We selected ATP1A3 for experimental validation and carried this hypothesis into the in vivo and structure-guided campaign described in the next section. Together, these examples illustrate how CircuitATLAS identifies targets through their capacity to control pathological circuit states rather than through their molecular association with disease. Source: CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies（arXiv）Figure 5: An in vivo lab-in-the-loop, from a graph-derived control point to a chemistry campaign.(A) Campaigns enter at one of four points and are planned by an experiment-design agent that addresses the laboratory through a Model Context Protocol server, exposing 25 tools and 20 resources over a structured record of 81 versioned procedures, operator competence, equipment state, colony and consumable stock, reserved capacity and cost per subject. Designs pass deterministic gates before human approval; an assay the laboratory does not hold enters as a capability request. (B) Pump turnover exports net positive charge and depends on intracellular Na+, so the hyperpolarizing current it generates is recruited in proportion to the Na+ that firing accumulates (inset). (C) Mean cortical LFP power spectra for AAV9.mDlx.ATP1A3 and the reference vector at baseline and under focal 4-AP, canonical bands shaded. Baseline power differs between arms, so the primary contrast is within-animal. (D) Absolute band power, one point per animal, with medians (reference baseline n=8n=8, ATP1A3 baseline n=6n=6, reference 4-AP n=5n=5, ATP1A3 4-AP n=7n=7); qq values are Benjamini–Hochberg corrected across bands for the within-animal fold-change contrast. (E) Structure-guided campaign on the validated target, with one representative compound and predicted binding score per site, sites ordered by the prior probability that occupancy raises turnover. Docking predicts binding, not polarity. (F) The polarity screen: 48 wells across the four site classes, thallium flux on human α\\alpha3 with α\\alpha1 and α\\alpha2 in parallel on an ouabain-resistant background, at sub-maximal Na+ load.We used ATP1A3 to illustrate how a target hypothesis can be advanced through this process. CircuitATLAS nominated the neuronal α\\alpha3 Na+/K+-ATPase as a control point on cortical excitability. Pump turnover exports net positive charge and is recruited by the intracellular Na+ load that firing generates, producing an activity-dependent hyperpolarizing current (Figure 5B)36. ATP1A3 is restricted to neurons and strongly expressed in GABAergic populations37, while loss of function is epileptogenic38 and pharmacological pump inhibition is convulsant39, together suggesting potentiation as the therapeutic direction. The unresolved question was whether increasing ATP1A3 in inhibitory interneurons would strengthen inhibition during sustained firing or instead suppress their output. Source: CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies（arXiv）
What a real drug would still have to prove
The next step is not simply to find a molecule that sticks to ATP1A3. A docking calculation can predict binding, but it cannot tell whether the molecule boosts or blocks the pump. The campaign produced 250 predicted binders, including 50 compounds that could be ordered directly. The needed tests must show the desired direction of action and must avoid closely related sodium-potassium pumps in other tissues.
If those tests succeed, a drug-discovery laboratory could reject a candidate that merely scores well on a computer and advance one that changes the neuronal pump in the intended way. Further ahead, an epilepsy clinic could match a patient’s repeatable pattern of runaway cortical activity to a therapy aimed at that circuit abnormality rather than choosing mainly from the diagnosis. That scene depends on prospective human trials showing that the recording, the intervention and clinical benefit travel together. It also depends on delivering a treatment to the right human cells: AAV vectors can target defined cell populations in experimental animals, but cell-specific delivery and durable control in people remain difficult.
CircuitATLAS is a preprint from Exin Therapeutics dated 7 October 2026. The company has filed a provisional patent application related to the work, and both authors are company co-founders and shareholders. The single mouse validation is a useful test of the idea, not yet proof that the idea can reliably produce better medicines.
