Automated Parameter Extraction for Biologically Realistic Neural Networks: An Initial Exploration with Large Language Models
Keywords: Large Language Models, Knowledge Graphs, Computational neuroscience, Neural model construction
Abstract: In computational neuroscience, extracting parameters for constructing biologically realistic neural models is a resource-intensive task that requires continuous updates as new research emerges. This paper explores utilizing large language models (LLMs) in automating parameter extraction from scientific literature for biologically realistic neural models. We utilized open-source LLMs via Ollama to construct KGs, capturing parameters such as neuron morphology, synapse dynamics, and receptor properties. SNNBuilder \cite{Gutierrez2022}, a framework for building spiking neural network (SNN) models, serves as a key validation example for our framework. However, the methodology we outline here can extend beyond SNNs and could applied to systematic modelling of the brain.By experimenting with different prompting strategies—general extraction, in-context hints, and masked prompting—we evaluated the ability of LLMs to autonomously extract relevant data and organize it within an expert-base or data-driven ontology, as well as to infer missing information for neural model construction. Additionally, we implemented retrieval-augmented generation (RAG) via LangChain to further improve the accuracy of parameter extraction through leveraging external knowledge sources. Analysis of the the generated KGs, demonstrated that LLMs, when guided by targeted prompts, can enhance the data-to-model process, paving the way for more efficient parameter extraction and model construction in computational neuroscience.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 14082
Loading