BLEG: LLM Functions as Powerful fMRI Graph-Enhancer for Brain Network Analysis

Rui Dong; Zitong Wang; Jiaxing Li; Weihuang Zheng; Youyong Kong

cs.LGApr 1, 2026

BLEG: LLM Functions as Powerful fMRI Graph-Enhancer for Brain Network Analysis

Authors: Rui Dong, Zitong Wang, Jiaxing Li, Weihuang Zheng, Youyong Kong

Organizations: School of Computer Science and Engineering, Southeast University

Abstract

Graph Neural Networks (GNNs) have been widely used in diverse brain network analysis tasks based on preprocessed functional magnetic resonance imaging (fMRI) data. However, their performances are constrained due to high feature sparsity and inherent limitations of domain knowledge within uni-modal neurographs. Meanwhile, large language models (LLMs) have demonstrated powerful representation capabilities. Combining LLMs with GNNs presents a promising direction for brain network analysis. While LLMs and MLLMs have emerged in neuroscience, integration of LLMs with graph-based data remains unexplored. In this work, we deal with these issues by incorporating LLM's powerful representation and generalization capabilities. Considering great cost for directly tuning LLMs, we instead function LLM as enhancer to boost GNN's performance on downstream tasks. Our method, namely BLEG, can be divided into three stages. We firstly prompt LLM to get augmented texts for fMRI graph data, then we design a LLM-LM instruction tuning method to get enhanced textual representations at a relatively lower cost. GNN is trained together for coarsened alignment. Finally we finetune an adapter after GNN for given downstream tasks. Alignment loss between LM and GNN logits is designed to further enhance GNN's representation. Extensive experiments on different datasets confirmed BLEG's superiority.Code can be available at https://github.com/KamonRiderDR/BLEG.

Figures & tables

Figure 1: A illustration of our method. GNN-based methods have limited performance. LLM methods require great training cost. Our method aims to enhance GNN’s performance with much less training cost.

Figure 2: The overall framework of BLEG. (1) We prompt LLM to generate augmented text data for input graph. (2) A smaller LM is trained through instruction tuning based on generated textual data. GNN is trained together for coarsened alignment. (3) GNN tuning for certain tasks. Logit from LM is utilized to boost GNN’s representations.

Methods	HCP		ADHD		MDD		ABIDE
Methods	ACC	SEN	ACC	SEN	ACC	SEN	ACC	SEN
GCN	64.03 $\pm$ 1.21	55.51 $\pm$ 2.88	66.74 $\pm$ 1.47	31.83 $\pm$ 2.33	62.38 $\pm$ 0.37	70.88 $\pm$ 1.94	69.14 $\pm$ 0.84	64.07 $\pm$ 2.20
GAT	65.72 $\pm$ 0.67	60.91 $\pm$ 3.26	66.56 $\pm$ 0.30	32.31 $\pm$ 1.56	63.05 $\pm$ 0.35	69.30 $\pm$ 2.16	68.79 $\pm$ 0.74	63.83 $\pm$ 3.23
GraphSAGE	66.87 $\pm$ 0.32	61.32 $\pm$ 2.71	67.80 $\pm$ 0.37	31.52 $\pm$ 2.54	63.29 $\pm$ 0.27	69.48 $\pm$ 1.53	70.74 $\pm$ 0.89	65.52 $\pm$ 2.81
GraphTrans	67.46 $\pm$ 0.55	62.71 $\pm$ 3.17	67.00 $\pm$ 0.37	31.85 $\pm$ 2.54	63.37 $\pm$ 0.27	68.31 $\pm$ 1.93	68.41 $\pm$ 1.20	62.79 $\pm$ 2.75
BrainGNN	66.46 $\pm$ 2.12	62.92 $\pm$ 2.45	67.16 $\pm$ 2.01	32.75 $\pm$ 3.31	63.25 $\pm$ 1.06	68.91 $\pm$ 2.76	70.03 $\pm$ 1.69	62.80 $\pm$ 4.19

Table 1: Comparison results on public datasets. We run 10 times for each model and record the corresponding average acc

\pm

std (%). The best results are marked bold , and second best underline .

Methods	GCN	GraphTrans	BrainGNN	BrainNPT	ContrastPool	BLEG (medium)	BLEG (large)	Qwen3-8B
ACC	71.79 $\pm$ 1.07	70.50 $\pm$ 0.73	71.06 $\pm$ 1.05	70.25 $\pm$ 1.55	71.11 $\pm$ 0.73	75.38 $\pm$ 0.63	75.21 $\pm$ 1.01	75.82 $\pm$ 0.77
SEN	85.70 $\pm$ 2.46	86.07 $\pm$ 1.84	84.00 $\pm$ 1.81	83.00 $\pm$ 1.59	86.85 $\pm$ 1.94	87.03 $\pm$ 0.89	86.77 $\pm$ 1.27	87.03 $\pm$ 1.58
SPE	50.60 $\pm$ 5.41	50.92 $\pm$ 3.03	48.77 $\pm$ 2.12	49.21 $\pm$ 3.30	52.55 $\pm$ 3.91	58.31 $\pm$ 2.72	58.50 $\pm$ 1.93	58.81 $\pm$ 2.00
F1	78.50 $\pm$ 0.67	78.43 $\pm$ 0.49	77.18 $\pm$ 1.92	77.27 $\pm$ 0.74	78.43 $\pm$ 0.49	79.35 $\pm$ 1.66	80.78 $\pm$ 1.72	79.81 $\pm$ 1.58
AUC	68.15 $\pm$ 1.71	65.70 $\pm$ 1.16	66.39 $\pm$ 0.61	67.53 $\pm$ 1.03	68.70 $\pm$ 1.16	70.77 $\pm$ 1.16	70.78 $\pm$ 1.24	69.43 $\pm$ 2.03

Table 2: Comparison results on private dataset. We run 10 times for each model and record the corresponding average acc

\pm

std (%). The best results are marked bold , and second best underline .

Panel of Table 3 — Table 3: $k$ -shot few-shot experiments. We run ten times and record average accuracy.

Panel of Figure 3 — Figure 3: (a)–(b) $k$ -shot experiments on different datasets. (c) Ablation studies on different datasets. (d) Biomarker visualizations on ABIDE, (e) ADHD and (f) zhongdaxinxiang datasets.

Figure 4: Few shot results on different ratios.

Figure 5: Text generation for ZDXX dataset from LM.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Prompt design for given FC graph.

Figure 7: An example for prompt response from LLM of ABIDE dataset.

Figure 8: Preprocess of fMRI data and construction for FC dataset.

Datasets	Tasks	Samples	Nodes ( $\mathcal{\|V\|}$ )	Classes	Categories
ABIDE	ASD diagnosis	618	90	2	{HC, ASD}
ADHD	ADHD diagnosis	938	90	2	{HC, ADHD}
HCP	Gender classification	1039	90	2	{Male, Female}
Rest-meta-MDD	MDD diagnosis	2165	90	2	{HC, MDD}
zhongdaxinxiang	MDD diagnosis	520	90	2	{HC, MDD}

Appendix

Table 4: More details of datasets.

Types	Parameters	Values
Training Arguments	Epochs	5
	Dataset Length	4760
	Batch Size	32
	Learning Rate	$5\times 10^{-5}$
LM Settings	Model	BioGPT
	Parameters	347M (1.57GB)

Appendix

Table 5: Hyper-parameter settings for instruct-tuning.

Types	Parameters	Values
Training Arguments	Epochs	150
	Early Stop	50
	Batch Size	$[32,64,128]$
	Learning Rate	$[0.0005,0.0001]$
	Weight Decay	$[0,0.0001]$
	LoRA rank	64

Appendix

Table 6: Hyper-parameter settings for supervised fine-tuning.

Explore similar work

FCN-LLM: Empower LLM for Brain Functional Connectivity Network Understanding via Graph-level Multi-task Instruction Tuning
Mar 1, 2026Xingcan Hu, Wei Wang, Li XiaoBrain Network Analysis Functional Connectivity
University of Science and Technology of China
When Language Models Meet NeuroGraphs: Exploring Enhanced Agentic LLM Framework Towards Brain Network Analysis
Jul 24, 2026Jiaxing Li, Rui Dong, Muyao Tang +1LLM Interpretability LLM Agents
School of Computer Science and Engineering, Southeast University
SABER: A Semantic-Aligned Brain Network Analysis Framework via Multi-scale Hypergraphs
Jul 2, 2026Yidan Xu, Xiangmin Han, Rundong Xue +1Medical Diagnosis Hypergraph Neural Networks
Hangzhou Dianzi University, China · Tsinghua University, China · Xi’an Jiaotong University, China

Mar 1, 2026J/K move · Enter open · S save

FCN-LLM: Empower LLM for Brain Functional Connectivity Network Understanding via Graph-level Multi-task Instruction Tuning

Xingcan Hu, Wei Wang, Li Xiao

University of Science and Technology of China

Brain Network Analysis Functional Connectivity Multimodal Large Language Models Instruction Tuning Multimodal Pretraining

Open paper AskPDF arXiv

Large Language Models have achieved remarkable success in language understanding and reasoning, and their multimodal extensions enable comprehension of images, video, and audio. Inspired by this, foundation models for brain functional connectivity networks derived from resting-state fMRI have shown promise in clinical tasks. However, existing methods do not align FCNs with the text modality, limiting the ability of LLMs to directly understand FCNs. To address this, we propose FCN-LLM, a framework that enables LLMs to understand FCNs through graph-level, multi-task instruction tuning. Our approach employs a multi-scale FCN encoder capturing brain-region, functional subnetwork, and whole-brain features, projecting them into the semantic space of LLM. We design multi-paradigm instruction tasks covering 19 subject-specific attributes across demographics, phenotypes, and psychiatric conditions. A multi-stage learning strategy first aligns FCN embeddings with the LLM and then jointly fine-tunes the entire model to capture high-level semantic information. Experiments on a large-scale, multi-site FCN database show that FCN-LLM achieves strong zero-shot generalization on unseen datasets, outperforming conventional supervised and foundation models. This work introduces a new paradigm for integrating brain functional networks with LLMs, offering a flexible and interpretable framework for neuroscience.

Abstract

Explore similar work

FCN-LLM: Empower LLM for Brain Functional Connectivity Network Understanding via Graph-level Multi-task Instruction Tuning

When Language Models Meet NeuroGraphs: Exploring Enhanced Agentic LLM Framework Towards Brain Network Analysis

SABER: A Semantic-Aligned Brain Network Analysis Framework via Multi-scale Hypergraphs