CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies

Gabriel Ocana-Santero; Marko Tvrdic

q-bio.QMOct 7, 2026

CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies

Authors: Gabriel Ocana-Santero, Marko Tvrdic

Organizations: Exin Therapeutics, Inc., 2261 Market Street STE 22614, San Francisco, California 94114, United States

Abstract

Drug discovery for neurological disease has traditionally centered on the molecules altered by disease. But the molecules that cause pathology are not necessarily the best points from which to reverse it. Here, we ask which otherwise unaltered molecular control points can be engaged to restore pathological neural circuits toward functional states. We present CircuitATLAS, a provenance-grounded systems-neuroscience knowledge graph and agentic framework for target discovery in circuitopathies. It structures literature-derived relationships across diseases, phenotypes, electrophysiology, circuits, brain regions, cell types and molecular effectors, while deliberately excluding direct disease-gene and disease-protein edges to reduce shortcut reasoning. The graph contains 3.83 million nodes and 7.66 million edges, including 5.31 million LLM-extracted relations, and incorporates structured datasets such as the Human Cell Atlas and new multimodal in vivo measurements. We then introduce an agentic workflow that reasons from measurable disease phenotypes through their circuit and cellular substrates to molecular interventions, therapeutic feasibility and clinical constraints. Finally, we introduce a human-governed in vivo lab-in-the-loop linking hypothesis generation to experimental iteration. Within this framework an agent nominated ATP1A3, the neuronal alpha3 Na+/K+-ATPase, as a control point on cortical excitability; interneuron-restricted expression of ATP1A3 abolished the beta- and gamma-band response to a focal 4-aminopyridine challenge in vivo, and the validated target was then carried into a structure-guided small-molecule campaign terminating in a defined assay to resolve the direction of modulation. CircuitATLAS thus provides a framework for discovering therapeutics based not only on what is molecularly disrupted in disease, but on what can be controlled to restore circuit function.

Figures & tables

Figure 1: Construction and structure of the CircuitATLAS knowledge graph. (A) Schematic of the literature-screening workflow. Sources shown are PubMed abstracts, PubMed Central full text, Europe PMC full text, arXiv, bioRxiv, and medRxiv. Records are passed through two-hit lexicon filtering and co-occurrence of at least two lexicon entities, yielding 15.9 million papers after screening. Retained papers are assigned to eight topic-stratified subcorpora: disease, cell type, region, circuit motif, cell electrophysiology feature, phenotype, circuit electrophysiology feature, and behavior/cognition. (B) Schematic of the six-stage distributed pipeline for knowledge-graph construction. The stages shown are datastore, candidate generation, aggregation, work-queue preparation, LLM verification, and entity resolution/KG load. Candidate mentions are aggregated by edge type, entities, confidence, and year, producing 57.7 million co-occurrences. The verification stage shows supported, unsupported, or uncertain classifications with confidence scores and quotations. (C) Circular summary of LLM-curated typed relations and a table showing example edge fields. The graph is organized into molecular, cellular, circuit, organism, and translational levels. The center indicates 105 directed relation types. The accompanying table lists stored edge attributes, including edge ID, edge type, head and tail IDs, verdict, confidence, supporting quote, supporting IDs, document title, year, snippet start/end, bucket ID, run ID, and creation time. (D) Overview of the CircuitATLAS knowledge graph. Colored node clusters are shown for 23 node types. The graph contains 3,831,997 nodes and 7,659,618 edges, including 5,306,909 LLM-extracted edges and 2,352,709 edges from canonical datasets. A schematic at the bottom shows incorporation of new literature through incremental pipeline runs and versioned graph snapshots.

Edge type	Head $\rightarrow$ Tail entity IDs	% support	Evidence quote example	PMID
cell_ephys_feature $\rightarrow$ disease [altered_in]	Burst afterhyperpolarization/slow AHP (cef276) $\rightarrow$ Alzheimer’s disease (nd1)	3/3	“During normal aging and in a mouse model of Alzheimer’s disease (AD), the slow AHP is increased, making neurons less excitable and making learning more difficult.”	15541708
drug $\rightarrow$ phenotype [ameliorates]	Capsaicin (CHEMBL294199) $\rightarrow$ Neurodegeneration (HP:0002180)	17/20	“…capsaicin alleviated other AD-type pathologies, such as tau hyperphosphorylation, neuroinflammation and neurodegeneration.”	32661266
biomarker $\rightarrow$ disease [associated_with]	Thyroxine/free T4 (HMDB:0000248) $\rightarrow$ Narcolepsy (nd230)	10/11	“Narcolepsy patients had significantly lower free thyroxine (FT4) levels in comparison to controls ( $p<0.001$ ).”	37898692
disease_model $\rightarrow$ phenotype [recapitulates]	Leaner mouse (dm90) $\rightarrow$ Dyskinesia (HP:0100660)	2/2	“The leaner mouse exhibits severe ataxia, paroxysmal dyskinesia, and absence epilepsy due to a P/Q-type calcium channel mutation.”	15514988

Table 1: Example LLM-extracted edges

Figure 2: Integration of the Human Cell Atlas into CircuitATLAS. (A) Schematic of Human Cell Atlas data integration and summary of extracted graph entities and relationships. The table shows the addition of 108 brain-region nodes, 24 cell-type nodes, 1,452 region-to-cell-type edges, 99,701 region-to-receptor expression edges, and 28,081 cell-type-to-receptor expression edges. Of 1,541 receptor genes, 1,538 were mapped to existing HGNC nodes, yielding approximately 129,234 new edges. (B, C) Heatmaps of receptor-gene expression across (B) human brain regions and (C) cell types. Rows represent brain regions or cell types, columns represent receptor genes, and color intensity indicates the fraction of cells expressing each receptor. Receptor genes are grouped by neurotransmitter system, as indicated above each heatmap.

Edge	Head node	Relation	Tail node	Quantitative evidence	Biological context	Source
HCA-1	Human CA2U-CA3U, hippocampal CA2/3	EXPRESSES	GABRB3, GABA-A receptor $\beta$ 3	99.72% of cells expressing; mean expression = 37.47; specificity = 0.081; $n$ = 11,276 cells	Region-level aggregate; ligand-gated ion channel	HCA Brain v1; Siletti et al., 2023
HCA-2	LAMP5-LHX6 and chandelier interneuron	EXPRESSES	GRIN2A, NMDA receptor GluN2A	99.8% of cells expressing; mean expression = 42.98; specificity = 0.209; $n$ = 500 cells	Human A13; ligand-gated ion channel	HCA Brain v1; Siletti et al., 2023
HCA-3	Human cerebellum	HAS_CELL_TYPE	Cerebellar granule cell	95.86% of region cells assigned; $n$ = 23,783 cells	Region-level cell-type composition	HCA Brain v1; Siletti et al., 2023

Table 2: HCA-extract edges examples

Figure 3: Integration of newly generated multimodal in vivo data into CircuitATLAS. (A) Overview of the six experimental modalities included in the in-house data-ingestion pipeline, showing acquisition details, corresponding graph readouts, and representative data for EEG, in vivo electrophysiology, markerless behavioral tracking, fiber photometry, histology, and transcriptomics. (B) Relative gamma-band power in Fmr1 knockout mice ( $n=12$ ) and wild-type controls ( $n=7$ ; $P=0.0098$ , Hedges’ $g=1.17$ ). (C) Mean open-field locomotor speed in 6-OHDA Parkinsonian mice ( $n=36$ ) and controls ( $n=15$ ; $P=0.0011$ , Hedges’ $g=-1.05$ ). (D) Relative gamma-band power following saline and 4-aminopyridine (4-AP) administration in the same animals ( $n=25$ ; $P=1\times 10^{-5}$ , Cohen’s $d_{z}=1.25$ ).

Edge	Disease-model node	Relation	Feature or phenotype node	Experimental finding	Statistical evidence	Assay and cohort
E1	6-OHDA mouse (dm31)	RECAPITULATES	Bradykinesia-related behavior (beh307)	Open-field mean locomotor speed decreased from 324 to 216 px/s ( $-33\%$ )	Two-sided Mann-Whitney U; $P$ = 0.0011; Hedges’ $g$ = $-1.05$	DeepLabCut open-field tracking; $n$ = 15 controls and $n$ = 36 6-OHDA mice
E2	Fmr1 knockout mouse (dm249)	SHOWS	Broadband gamma activity (cief54)	Relative gamma power in S1 increased from 2.6% to 8.7% ( $3.3\times$ )	Two-sided Mann-Whitney U; $P$ = 0.0098; Hedges’ $g$ = $+1.17$	32-channel in vivo LFP; $n$ = 7 wild-type and $n$ = 12 Fmr1 knockout mice
E3	4-AP acute seizure model (dm617)	SHOWS	Broadband gamma activity (cief54)	Relative EEG gamma power increased from 10.2% at saline baseline to 18.6% after 4-AP	Paired Wilcoxon signed-rank; $P=1.0\times 10^{-5}$ ; Cohen’s $d_{z}$ = $+1.25$	Overnight two-channel EEG; $n$ = 25 animals measured under both conditions

Table 3: Representative edges extracted from in-house multimodal in vivo data

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.

	Hypothesis 1	Hypothesis 2	Hypothesis 3
Indication	Fragile X syndrome	Dravet syndrome	Epilepsy
Circuit phenotype	Thalamocortical hyperexcitability	Lowered threshold for generalised synchronous discharge, from failed inhibitory restraint	Cortical hyperexcitability and failure to restrain sustained neuronal firing
Measure / polarity	Cortical evoked response amplitude and habituation slope; resting gamma power; ASSR phase-locking. Higher amplitude / lower habituation = greater gain 33	Seizure and interictal discharge rate, hyperthermia seizure threshold. Higher rate / lower threshold = greater excitability	Seizure and interictal discharge rate; cortical activity during sustained excitation. Higher activity / discharge rate = greater excitability
Corrective direction	Decrease thalamocortical gain	Decrease network excitability	Decrease network excitability
Associated phenotypes	Sensory hypersensitivity, impaired habituation, audiogenic seizures, elevated resting gamma 34 , reduced auditory phase-locking 35 , sleep spindle disruption, attention and arousal deficits, anxiety	Convulsive and myoclonic seizures, SUDEP via seizure-induced respiratory arrest, developmental regression, ataxia	Seizures, interictal discharges, impaired inhibitory restraint
Intervention logic	Corrective: reduce the excess glutamatergic thalamocortical drive	Compensatory: recruit an intact neuromodulatory brake that is not itself deficient	Compensatory: augment an endogenous activity-dependent hyperpolarizing current recruited during sustained firing

Table 4: Example of three new targets identified by CircuitATLAS

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 $n=8$ , ATP1A3 baseline $n=6$ , reference 4-AP $n=5$ , ATP1A3 4-AP $n=7$ ); $q$ 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 $\alpha$ 3 with $\alpha$ 1 and $\alpha$ 2 in parallel on an ouabain-resistant background, at sub-maximal Na + load.

Explore similar work

Sep 20, 2026cs.LG

Circuit-Diff: Factual Edit-based Intervention Method for Localizing Knowledge in Attribution Graphs

Mechanistic interpretability defines features as the fundamental units of a neural network and circuits as the weighted subgraphs that carry out its computation. Because individual neurons are polysemantic, Cross-Layer Transcoders (CLTs) were introduced as a way to approximate a model's circuits by generating an attribution graph. The nodes of that graph, however, are unlabeled features: reading a graph means pruning it and then working out by hand what each surviving node means. To make CLTs easier to use for circuit discovery, we introduce Circuit-Diff, which intervenes on the model itself with a low-rank factual edit and takes the features whose role in the attribution graph changes under that edit as related to the edited knowledge. On the edits we examine, the flagged nodes are not only detectors of the object token: read off the CLT's released feature dashboards, they include features for the history, geography and associations surrounding the old and new objects. We formalize the method, measure how reliable a frozen CLT remains after a factual edit, test the selected nodes causally by patching them on up to 24 CounterFact edits, give a case study, and release an open-source implementation built on the circuit-tracer package, together with two further tools (multi-prompt aggregation and rule-based supernode labeling).

Edward G. Friedman, Xiangchen Song

School of Computer Science Machine Learning Department Carnegie Mellon University Pittsburgh, PA 15213

Circuits Mechanistic Interpretability Causal Graph Localization

Ask PDF

Oct 7, 2026cs.AI

DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists

Drug target discovery requires distinguishing molecules that causally drive disease from those that are merely associated with it. Training and evaluating AI agents to perform this workflow end-to-end is difficult because real world biobanks lack known causal ground truth and participant-level data is access controlled. We introduce DrugTargetWorld, a framework that procedurally generates simulated biobanks, or "worlds," with known but concealed causal structure. Each world contains genotypes, proteins, health records, outcomes, and synthetic magnetic resonance imaging (MRI) for 54,000 participants. Agents must construct a disease phenotype, identify causal driver proteins, infer the beneficial direction of modulation, and optionally conduct virtual 'wet lab' experiments. We evaluated nine agents in 540 episodes across 20 cardiovascular worlds and three experimental budgets. Opus 5 and GPT-5.6 Sol achieved the highest mean composite scores, 39.98 and 35.38 of 100, respectively, and both recovered 64% of causal drivers on average. However, no agent reliably distinguished misleading non-causal proteins, and performance remained limited by the integrative judgments required to connect phenotype construction, causal evidence, and intervention decisions. By making each world's causal structure known to the evaluator but hidden from the agent, DrugTargetWorld turns end-to-end drug target discovery into a scalable training and evaluation problem with verifiable reward.

Samuel Margolis, Paul Schmiedmayer, Alan Huang +12

Department of Biomedical Data Science, Stanford University, Stanford, CA 94305, USA · Department of Medicine, Stanford University, Stanford, CA 94305, USA · Division of Computational Medicine, Department of Medicine, Stanford University, Stanford, CA 94305, USA +1

Drug Discovery Agentic Discovery

Ask PDF

Jul 27, 2026cs.CL

Neural Circuit Function Inference with LLMs

The success of connectome mapping now shifts the challenge of understanding the nervous system to the interpretation of neural circuits. Here, we devise a new automated method, LLantia (LLM automated neural circuit inference and analysis), to systematically infer neural circuit function and the role of its component neural cell types. Our approach distills descriptions of cell type function from the literature and, in combination with the connectome, then infers the function for all other cell types, which serves as a basis for subsequent neural circuit function inference. Results are structured hierarchically, with different possible circuit functions organised under multiple possible behavioural and physiological contexts, and each circuit function composed of subcircuit descriptions alongside relevant cell types to facilitate both backtracking to known, published information and support further experimental research. We illustrate our method by inferring cell type function for all cell types of the adult fruit fly brain and for select broader circuits within, and validate our findings, including by cross-checking with literature published after the release date of our analysis.

Yijie Yin, Albert Cardona

Department of Physiology, Development and Neuroscience, University of Cambridge, Cambridge, UK · MRC Laboratory of Molecular Biology, Cambridge, UK

Neural Circuits Connectome Circuits Cells

Ask PDF