Estimating　brain　effective　connectivity　network　from　functional　magnetic　resonance　imaging　(fMRI)　data　is　a　key　research　topic　in　the　fields　of　neurology　and　brain　science,　as　this　data　can　play　an　important　role　in　early　diagnosis　of　brain　diseases.　However,　estimating　a　large-scale　brain　effective　connectivity　network　from　fMRI　data　remains　challenging　due　to　the　high　computational　complexity　and　complex　data　distributions.　In　the　present　study,　we　proposed　a　novel　method　for　estimating　large-scale　brain　effective　connectivity　network　based　on　multi-subgraph　attention　diffusion　(MsAD-LEC).　This　method　comprises　two　modules:　a　multi-resolution　cluster-based　subgraph　partition　module　and　an　attention　diffusion-based　brain　effective　connectivity　estimation　module.　In　the　partition　module,　a　partial　correlation　analysis　is　first　conducted　to　determine　the　conditional　independent　relationships　among　brain　regions,　to　accurately　construct　an　undirected　causal　skeleton.　The　method　then　applied　the　multi-resolution　Louvain　algorithm　to　partition　the　undirected　causal　skeleton　into　multiple　subgraphs　and　cross-subgraph　edges.　This　partition　can　transform　a　large-scale　estimation　problem　into　multiple　smaller-scale　estimation　problems,　thus　reducing　the　overall　computational　complexity.　In　the　estimation　module,　the　method　first　applies　forward　diffusion　to　normalize　the　data　distributions　to　address　the　data　distribution　shifts　potentially　caused　by　the　preceding　subgraph　partition.　The　method　then　involves　inverse　diffusion　with　multi-head　self-attention　to　capture　long-term　brain　region　dependencies,　to　accurately　estimate　subgraph　and　cross-subgraph　effective　connectivity,　to　obtain　a　large-scale　brain　effective　connectivity　network.　Experimental　results　demonstrated　that　the　MsAD-LEC　can　accurately　estimate　large-scale　brain　effective　connectivity　networks　and　scales　in　hundreds　of　brain　regions.　©　2024　Elsevier　B.V.