💬 NLP

Neural Circuit Function Inference with LLMs

The paper introduces LLantia, an automated method that leverages large language models to distill literature-based cell type functions and integrate them with connectome data to systematically infer and hierarchically structure the functions of neural circuits and their component cell types, as demonstrated and validated in the adult fruit fly brain.

Original authors: Yijie Yin (Department of Physiology, Development and Neuroscience, University of Cambridge, Cambridge, UK, MRC Laboratory of Molecular Biology, Cambridge, UK), Albert Cardona (MRC Laboratory of Molecu

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yijie Yin (Department of Physiology, Development and Neuroscience, University of Cambridge, Cambridge, UK, MRC Laboratory of Molecular Biology, Cambridge, UK), Albert Cardona (MRC Laboratory of Molecular Biology, Cambridge, UK, Department of Physiology, Development and Neuroscience, University of Cambridge, Cambridge, UK)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand how a city works just by looking at a map of every single street, alley, and sidewalk. You know where the roads connect, but you have no idea what happens at the intersections. Is that traffic light for a school zone? Is that alley a shortcut for delivery trucks or a dead end? This is the current state of neuroscience. Scientists have finally finished mapping the "wiring diagrams" of tiny brains, like that of a fruit fly, showing exactly which neurons touch which others. But knowing the connections is only half the battle; the real mystery is figuring out what those connections actually do. How does a specific tangle of wires turn a smell of vinegar into a decision to fly toward it?

To solve this, researchers are using a new kind of digital detective: Large Language Models (LLMs). Think of these as super-smart AI readers that have devoured almost every science paper ever written. They are excellent at finding patterns in text, but they can't usually look at a wiring diagram. The big question is: Can we teach an AI to look at a map of connections and read the history books about what those parts do, then use that knowledge to guess the function of the parts nobody has studied yet? If we can do this, we could instantly understand how complex circuits work without needing to run thousands of slow, expensive experiments for every single neuron.

The AI Detective: LLantia

Meet LLantia (LLM automated neural circuit inference and analysis), a new digital tool created by researchers Yijie Yin and Albert Cardona. Their goal was to build a system that could take the massive, completed wiring map of an adult fruit fly's brain and automatically figure out what every single type of neuron is doing, even if no scientist had ever written a paper about that specific neuron before.

The Problem: A Library of Lost Pages
The fruit fly brain has about 140,000 neurons, but when grouped by their shape and connections, they form about 9,000 distinct "cell types." Scientists have studied some of these types, but the information is scattered like puzzle pieces in different boxes. Some papers say a neuron helps with smell; others say it helps with movement. Meanwhile, the wiring map shows that these neurons are densely connected—like a crowded subway station where 80% of the stations are linked within just five stops. Trying to manually figure out what happens when you combine all these signals is impossible for a human to do quickly.

The Solution: A Two-Step Magic Trick
LLantia works in two main stages, acting like a very organized librarian and a brilliant detective rolled into one.

Step 1: The Librarian (Reading the Clues)
First, the system goes on a scavenger hunt through scientific literature. It starts with a few "seed" papers and uses a citation graph (a map of which papers reference which others) to find every related study. It downloads the text and uses an AI to extract the specific functions of neurons mentioned.

The Challenge: Scientists often use different names for the same neuron (like calling a "dog" a "canine" or a "puppy"). The AI has to be smart enough to realize that "MBON-α'1" and "MBON15" are the same character in the story.
The Result: The system successfully distilled the functions of 411 unique cell types from thousands of papers, turning long, complex sentences into short, clear summaries.

Step 2: The Detective (Solving the Mystery)
Next, the system takes the wiring map and the summaries it just read. It creates a custom "prompt" (a detailed instruction) for a powerful AI model. This prompt says: "Here is a specific neuron. Here is who it talks to. Here is what those neighbors are known to do. Based on this, what is this neuron's job?"

The Logic: If a neuron receives signals from "smell detectors" and sends signals to "flight controllers," the AI infers that this neuron is likely a "smell-to-flight converter."
The Output: Instead of a simple guess, the AI produces a structured report. It suggests the neuron's function, the specific behaviors it might control (like "hunting for food when hungry"), and even lists the specific connections that support this idea. It organizes these guesses into a hierarchy, showing how small circuits fit into bigger ones.

Did it Work? The Proof
The researchers didn't just hope it worked; they tested it rigorously.

The "Future" Test: They checked their results against a brand-new scientific paper published after the AI was trained. The AI had correctly inferred that a specific neuron (LoVPN) acts as a "visual salience detector" that only works when other senses (like smell) are quiet. This matched the new paper's findings perfectly, proving the AI wasn't just memorizing old facts but actually reasoning about the circuit.
The "Memory" Test: To ensure the AI wasn't just recalling the answer from its training data, the researchers hid the names of the neurons and removed the known function descriptions. When the names were gone, the AI's guesses became vague and generic. This proved that the AI was actually using the wiring map and the context of the other neurons to figure out the answer, not just recalling a fact.
The "Human" Test: When they compared the AI's summaries to those written by human experts, the computer-generated descriptions were surprisingly similar in meaning, even if the wording was different.

What It Found
Using this method, the team generated function hypotheses for all 9,000 cell types in the fruit fly brain. They also mapped out six major "circuits" (like the path from smell sensors to the navigation center). The AI discovered common themes, such as:

Cross-inhibition: Neurons that stop one behavior (like singing a love song) when another, more urgent behavior (like running from a predator) is needed.
AND Gates: Neurons that only activate when multiple conditions are met (e.g., "I am hungry" AND "I smell food" AND "I am not mating").

The Bottom Line
LLantia doesn't replace scientists; it gives them a massive head start. It turns a mountain of disconnected data into a structured, searchable map of "what does what." By automating the process of connecting the dots between a neuron's wiring and its job, it allows researchers to skip the guesswork and go straight to designing experiments to test these new ideas. The authors suggest that this approach could be scaled up to even larger brains as our maps get better, turning the overwhelming complexity of the nervous system into a solvable puzzle.

Technical Summary: Neural Circuit Function Inference with LLMs (LLantia)

Problem Statement

The rapid acceleration of connectome mapping, particularly in Drosophila melanogaster, has shifted the primary challenge from data acquisition to the interpretation of neural circuit function. While the fruit fly connectome contains approximately 140,000 neurons, these can be reduced to a connectivity matrix of roughly 9,000 distinct cell types based on morphology and synaptic connectivity. However, the sheer density of connectivity—where ~80% of cell types are synaptically connected within five hops—creates a combinatorial explosion of potential input-output pathways. This complexity, combined with sparse and scattered literature covering only a fraction of cell types, hinders the formulation of testable models and experiments. Existing approaches to literature aggregation rely heavily on manual curation or commercial tools lacking open APIs, making it difficult to systematically integrate known functional data with structural connectome data at scale.

Methodology: The LLantia Framework

The authors introduce LLantia (LLM automated neural circuit inference and analysis), an automated pipeline designed to infer the function of neural cell types and the circuits they comprise by combining connectome structure with literature-derived functional descriptions. The framework operates through three primary stages:

1. Automated Literature Mining and Function Distillation

Citation Graph Traversal: Starting from 114 hand-picked "seed" papers, the system uses the Semantic Scholar API to traverse the citation graph (upstream and downstream, 2 hops).
Relevance Filtering: A small LLM (Qwen3.5-4B) filters papers based on titles and abstracts for relevance to Drosophila neural function, following an initial keyword filter.
Extraction and Normalization: Full-text PDFs are retrieved (via Semantic Scholar, Unpaywall, OpenAlex) and processed by a large LLM (Qwen3.6-35B) to extract cell type names and their functions.
Name Matching: Extracted names are matched to connectome identifiers using synonym lists and partial string matching.
Summarization: For cell types appearing in multiple papers, the LLM consolidates findings into concise descriptions (max 30 words) to fit subsequent prompt constraints.

2. Single Cell Type Function Inference

To infer the function of a query cell type, the system constructs a structured prompt containing:

Connectome Data: Pre-computed signed effective connectivity (excitatory/inhibitory) from sensory neurons and known functional cells to the query, and from the query to known cells.
Local Context: Direct pre- and postsynaptic partners and their connectivity weights.
Functional Knowledge: A compendium of distilled function descriptions for all "known" cell types participating in the query's polysynaptic pathways.
Constraint: The known function of the query cell type is explicitly excluded to prevent data leakage and test inference capability.
Output: The LLM (Qwen3.6-35B) generates a structured JSON response detailing the behavioral context, circuit logic, coherence assessment, inferred computation, key supporting connections, and intermediate computations.

3. Hierarchical Circuit Function Inference

For broader circuits, the pipeline employs a multi-step hierarchical prompting strategy:

Pathfinding: The Connectome Interpreter identifies all polysynaptic pathways between defined source and target cell types.
Layered Representation: The circuit is transformed into a layered graph (based on information flow) to reduce token redundancy, listing cell types by their synaptic distance from sources.
Iterative Synthesis:
1. Pairwise: Infer functions for specific input-output pairs.
2. Aggregation: Synthesize functions for all paths converging on a target or diverging from a source.
3. Global Synthesis: Combine these summaries to define the function of the entire circuit across multiple behavioral and physiological contexts (e.g., hunger, mating state).

Key Results and Validation

Literature Extraction Accuracy

The automated extraction of cell type functions was validated against manually curated datasets. Embedding similarity analysis (using Qwen3-Embedding-8B) showed that descriptions for the same cell type generated by the automated pipeline were significantly more similar to human-curated descriptions than to descriptions of different cell types (Wilcoxon $p < 1.51 \times 10^{-35}$ ).

Inference Validation

Quantitative: For 488 cell types with known functions, the LLM-inferred descriptions (excluding the known function from the prompt) were significantly closer to the literature-derived ground truth than to descriptions of other cell types ( $p < 1.30 \times 10^{-33}$ ).
Memory vs. Analysis Control:
- When known function descriptions were removed from the prompt, performance dropped, indicating the model relies on the provided context rather than just memorized cell type names.
- When cell type names were replaced with dummy names (preserving connectivity and functional descriptions of partners), the model maintained high performance, proving the inference is driven by structural and functional analysis rather than name-based retrieval.
Prospective Validation: The method was applied to infer the function of the LoVPN cell type (LTe72) using the LLM (Qwen3.6), which was released in April 2026. The analysis was conducted before the publication of Coban, Harris et al. (June 2026). The LLM predicted a "context-dependent visual salience detector" gated by other sensory modalities (olfaction, thermosensation), which aligned precisely with the subsequent experimental findings regarding LHPV6m1-mediated gating reported in the June 2026 paper.

Circuit Analysis Applications

The framework was applied to six broad circuits in the fruit fly brain (e.g., Lateral Horn to Central Complex, MBONs to Descending Neurons). The system generated hierarchical descriptions identifying:

Context-Dependent Logic: Circuits operating differently under specific states (e.g., hunger vs. satiety).
Cross-Inhibition: Mechanisms where specific cell types (e.g., LPLC2) inhibit others (e.g., pMP2) to enforce behavioral exclusivity.
AND Gates: Requirements for multiple simultaneous conditions (e.g., food odor presence + lack of courtship priority) to trigger specific motor outputs.

Significance and Claims

The paper claims that LLantia provides a scalable, automated method to bridge the gap between connectome structure and functional understanding. Its primary contributions are:

Systematic Inference: It moves beyond manual curation to systematically infer functions for all ~9,000 cell types in the fly connectome.
Backtracking and Transparency: The hierarchical, structured output allows researchers to trace circuit-level hypotheses back to specific subcircuits, synaptic connections, and the original literature, facilitating experimental validation.
Contextual Nuance: By explicitly modeling behavioral and physiological contexts, the method infers how circuits operate under specific conditions, addressing the limitation of static connectivity maps.
Open Science: The authors release the software suite (LLantia), the Connectome Interpreter, and all generated data (including hypotheses for the FAFB/FlyWire connectome) as open-source resources, enabling the community to apply these methods to other connectomes.

The authors maintain a modest stance, acknowledging that the framework depends on the existence of prior literature for at least some component cell types and that the free-text nature of the outputs requires further systematic validation by human experts or automated functional agreement checks. However, they demonstrate that modern open-weight LLMs can effectively perform inference on neural circuits when guided by structured prompts containing connectivity and distilled functional knowledge.