MuSpec: an unsupervised multi-scale spatial-niche encoder recovers conserved granuloma niches and nominates host-directed repurposing candidates in human pulmonary tuberculosis

Abstract Background. Tuberculosis (TB) granulomas are spatially organised host–pathogen structures, and the spatial arrangement of immune and stromal cells within the granuloma shapes disease outcome. Spatial transcriptomics can now resolve these structures at single-cell resolution, but most spatial-niche models have been developed and validated on tumour tissue. There is a shortage of unsupervised methods benchmarked on infectious-disease granulomas, and little evidence on how such niche models behave under input noise, missing modalities, and cross-sample transfer — the conditions that decide whether a niche atlas is translational or merely descriptive. Methods. We assembled MuSpec , an unsupervised multi-scale spatial-niche encoder, and applied it to two single-cell-resolution Stereo-seq slides from human TB lung: B1 (pulmonary granuloma; Y40012NC13; 36,382 cells) and P1 (peri-lesional lung; Y40012LA11; 51,834 cells). Each cell was represented as a 39-dimensional input (7-dimensional cell-type composition from cell2location deconvolution against a 146,345-cell peripheral-blood scRNA-seq reference, 30 expression principal components, and 2 spatial coordinates). MuSpec smooths the input at three kNN scales (k = 10/20/30), rasterises each scale onto a ≤ 224² lattice, and passes the three feature maps through four deep-learning modules built on published attention-and-fusion designs — ASFF cross-scale fusion, TripletAttention, ExternalAttention, and a Fourier graph network (FGN) — to produce a 16-dimensional niche embedding; a shared encoder is trained jointly across both slides with an MSE reconstruction loss plus a spatial-smoothing loss, and niches are obtained by KMeans (K = 9). We benchmarked MuSpec against BANKSY at matched numbers of spatial domains, ran a three-seed module ablation and a 34-configuration single-factor hyperparameter sweep, and used the recovered granuloma signature for multi-library signature-reversal drug repurposing (LINCS L1000 chemical perturbations, GEO perturbations, NIBR DRUGseq 2025, and the original CMAP, merged) and AutoDock Vina docking against the host target SP110. Results. MuSpec produced a compact, sub-million-parameter model (782,251 parameters; 93.9% in ASFF) that trains in minutes on a single consumer GPU (≈ 7 GB peak). On every matched-K comparison, MuSpec embeddings separated niches far more sharply than BANKSY (Δsilhouette + 0.31 to + 0.40 on both slides); MuSpec’s own K = 9 silhouette was 0.322 (B1) and 0.285 (P1). Three-seed ablation showed that no module was removable losslessly : removing ExternalAttention lowered niche separation (Δsilhouette − 0.093) but raised robustness and cross-slide transfer (Δ of + 0.137 to + 0.362), identifying a clean separation↔robustness trade-off. For normalisation, a three-seed paired design showed that 7 of 8 robustness metrics changed sign across seeds, so we report no seed-stable difference between global and per-slide normalisation, with only a small stable (+ 0.010 ± 0.003) silhouette advantage to per-slide normalisation on B1. The hyperparameter sweep revealed a monotone separation–robustness frontier with no cliffs, on which the default setting was the robustness optimum (mean noise-robustness ARI 0.758; drop-composition ARI 0.683 on B1 and 0.518 on P1, the highest in the sweep). Joint niche characterisation recovered myeloid/granuloma (SPP1, MMP9, CTSK), airway/club (MUC5B, SCGB1A1, BPIFB1), fibrotic (SFRP2, COL1A1, INHBA) and fibroblast (IL11, TFPI2, PI16) niches, and an SP110–IFN hub analysis. A 324-gene granuloma signature (233 up, 91 down) yielded top signature-reversal hits led by a PPARγ-agonist trio (pioglitazone 90.08, troglitazone, rosiglitazone), and in-silico docking of a focused 6-ligand panel against SP110 gave the strongest affinities for withaferin A (− 7.18 kcal/mol) and nintedanib (− 6.92 kcal/mol). Conclusions. MuSpec provides an unsupervised, reproducible, computationally light route to spatially resolved niche atlases in infectious-disease tissue, and its systematic benchmarking exposes a separation↔robustness trade-off that should be reported rather than hidden. Applied to human TB lung, it recovered biologically coherent, cross-slide-conserved granuloma niches and nominated host-directed repurposing candidates.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11339003/v1
Primary Topic
Single-cell and spatial transcriptomics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

MuSpec: an unsupervised multi-scale spatial-niche encoder recovers conserved granuloma niches and nominates host-directed repurposing candidates in human pulmonary tuberculosis

Jiangdong Wu, Yuxuan Gou, Yi Peng, Yating Zhao et al.
3 links
Research Square
Single-cell and spatial transcriptomics
preprint

MuSpec: an unsupervised multi-scale spatial-niche encoder recovers conserved granuloma niches and nominates host-directed repurposing candidates in human pulmonary tuberculosis

Jiangdong Wu, Yuxuan Gou, Yi Peng, Yating Zhao, Chunhong Xiao, Tingting Luo, Qifeng Li, Jing Chen, Yang Yang, Jie Chen
preprint en

Abstract

Abstract Background. Tuberculosis (TB) granulomas are spatially organised host–pathogen structures, and the spatial arrangement of immune and stromal cells within the granuloma shapes disease outcome. Spatial transcriptomics can now resolve these structures at single-cell resolution, but most spatial-niche models have been developed and validated on tumour tissue. There is a shortage of unsupervised methods benchmarked on infectious-disease granulomas, and little evidence on how such niche models behave under input noise, missing modalities, and cross-sample transfer — the conditions that decide whether a niche atlas is translational or merely descriptive. Methods. We assembled MuSpec , an unsupervised multi-scale spatial-niche encoder, and applied it to two single-cell-resolution Stereo-seq slides from human TB lung: B1 (pulmonary granuloma; Y40012NC13; 36,382 cells) and P1 (peri-lesional lung; Y40012LA11; 51,834 cells). Each cell was represented as a 39-dimensional input (7-dimensional cell-type composition from cell2location deconvolution against a 146,345-cell peripheral-blood scRNA-seq reference, 30 expression principal components, and 2 spatial coordinates). MuSpec smooths the input at three kNN scales (k = 10/20/30), rasterises each scale onto a ≤ 224² lattice, and passes the three feature maps through four deep-learning modules built on published attention-and-fusion designs — ASFF cross-scale fusion, TripletAttention, ExternalAttention, and a Fourier graph network (FGN) — to produce a 16-dimensional niche embedding; a shared encoder is trained jointly across both slides with an MSE reconstruction loss plus a spatial-smoothing loss, and niches are obtained by KMeans (K = 9). We benchmarked MuSpec against BANKSY at matched numbers of spatial domains, ran a three-seed module ablation and a 34-configuration single-factor hyperparameter sweep, and used the recovered granuloma signature for multi-library signature-reversal drug repurposing (LINCS L1000 chemical perturbations, GEO perturbations, NIBR DRUGseq 2025, and the original CMAP, merged) and AutoDock Vina docking against the host target SP110. Results. MuSpec produced a compact, sub-million-parameter model (782,251 parameters; 93.9% in ASFF) that trains in minutes on a single consumer GPU (≈ 7 GB peak). On every matched-K comparison, MuSpec embeddings separated niches far more sharply than BANKSY (Δsilhouette + 0.31 to + 0.40 on both slides); MuSpec’s own K = 9 silhouette was 0.322 (B1) and 0.285 (P1). Three-seed ablation showed that no module was removable losslessly : removing ExternalAttention lowered niche separation (Δsilhouette − 0.093) but raised robustness and cross-slide transfer (Δ of + 0.137 to + 0.362), identifying a clean separation↔robustness trade-off. For normalisation, a three-seed paired design showed that 7 of 8 robustness metrics changed sign across seeds, so we report no seed-stable difference between global and per-slide normalisation, with only a small stable (+ 0.010 ± 0.003) silhouette advantage to per-slide normalisation on B1. The hyperparameter sweep revealed a monotone separation–robustness frontier with no cliffs, on which the default setting was the robustness optimum (mean noise-robustness ARI 0.758; drop-composition ARI 0.683 on B1 and 0.518 on P1, the highest in the sweep). Joint niche characterisation recovered myeloid/granuloma (SPP1, MMP9, CTSK), airway/club (MUC5B, SCGB1A1, BPIFB1), fibrotic (SFRP2, COL1A1, INHBA) and fibroblast (IL11, TFPI2, PI16) niches, and an SP110–IFN hub analysis. A 324-gene granuloma signature (233 up, 91 down) yielded top signature-reversal hits led by a PPARγ-agonist trio (pioglitazone 90.08, troglitazone, rosiglitazone), and in-silico docking of a focused 6-ligand panel against SP110 gave the strongest affinities for withaferin A (− 7.18 kcal/mol) and nintedanib (− 6.92 kcal/mol). Conclusions. MuSpec provides an unsupervised, reproducible, computationally light route to spatially resolved niche atlases in infectious-disease tissue, and its systematic benchmarking exposes a separation↔robustness trade-off that should be reported rather than hidden. Applied to human TB lung, it recovered biologically coherent, cross-slide-conserved granuloma niches and nominated host-directed repurposing candidates.

Research Square
People's Hospital of Xinjiang Uygur Autonomous Region (CN), Beijing Children’s Hospital (CN), Xinjiang Second Medical College (CN)
Good health and well-being
Single-cell and spatial transcriptomics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon