cs.AIAug 29, 2026

Agentic AI uncovers conserved cross-tissue protein co-abundance programs inaccessible to single-dataset analysis

Authors: Runyu Guan, Dehao Wu, Qiqi Xie, Yang Li, Haohan Wang

Organizations: School of Information Sciences, University of Illinois Urbana-Champaign, Champaign, 61820, IL, US · Department of Ecology and Evolutionary Biology, University of Michigan, Ann Arbor, 48109, MI, US

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Protein co-abundance clusters preserved across tissues can reveal shared disease mechanisms and candidate therapeutic targets, particularly when proteins implicated in organ-confined diseases converge in peripheral or accessible tissues. However, previous cross-tissue studies have focused on biologically pre-selected tissue pairs, leaving most possible combinations and non-obvious relationships unexplored. We present an LLM-agent framework for large-scale, evidence-grounded comparison of tissue-specific protein co-abundance networks. The framework constructs tissue networks, derives pairwise consensus clusters, and integrates evidence from expression atlases, protein interaction and complex databases, pathway annotations, disease catalogues, and literature. Applied to all 820 pairwise combinations of 41 human tissues and fluids, it identified 1,833 conserved co-abundance clusters across 406 tissue pairs. Colon, synovial fluid, blood, cerebrospinal fluid, and bone marrow were the most broadly connected tissues, while the most cluster-rich pairs were dominated by bone marrow. The analysis also highlighted non-obvious relationships: skin-bone marrow exceeded the anatomically adjacent bone-bone marrow pair, while colon-breast contained cancer-relevant clusters involving extracellular-matrix remodeling, lipid metabolism, and immune modulation. Cluster-level analyses generated further mechanistic hypotheses, including a brain-gut extracellular-vesicle/redox/serotonin-cofactor axis and a liver-bone marrow stress-response axis involving genes linked to white matter disease. These results provide a global, comparable landscape of conserved protein co-abundance and a hypothesis-generating resource for mechanistic and therapeutic exploration. Code and data are available at https://github.com/Gry1005/AgenticAI-conserved-cross-tissue-protein-co-abundance.

Figures & tables

Figure 1: Pairwise tissue comparison across 41 human tissues and biofluids. Lower triangle: number of shared protein co-abundance clusters between each tissue pair. Upper triangle: expression similarity (Pearson rr of log-transformed mean iBAQ values). Tissues are ordered by hierarchical clustering of the shared cluster count matrix.
Tissue or fluidCountTissue or fluidCount
Colon40Ileum14
Synovial fluid39Adrenal13
Blood39Heart13
Cerebrospinal fluid38Adipose13
Bone marrow38Cell line13
Liver37Stem cells13
Table 1: Number of tissues or fluids sharing at least one conserved protein co-abundance cluster with the reference tissue
SourceCategoryTerm#Genespp /FDR
GOCCMitochondrion23.39×10−43.39\times 10^{-4}
GOBPMitochondrial calcium ion homeostasis—2.61×10−22.61\times 10^{-2}
KEGGPathwayPeroxisome—4.55×10−24.55\times 10^{-2}
GOMFL-xylulose reductase (NADPH) activity15.00×10−25.00\times 10^{-2}
GOMFPeptide serotonyltransferase activity15.00×10−25.00\times 10^{-2}
GOMFPeptide dopaminyltransferase activity15.00×10−25.00\times 10^{-2}
Table 2: Liver–Adipose co-abundance Cluster 1 (7 proteins: ANXA6, DCXR, IDH2, LACTB, PECR, PPM1A, TGM2) — Integrated Enrichment Profile. HPA: 7/7. hu.MAP 3.0: none. STRING PPI: 0/21 pairs (0%).
Figure 2: Cross-tissue co-abundance cluster sharing connectivity of 41 human tissues and biofluids. Each bar represents the number of other tissues sharing at least one conserved protein co-abundance cluster with the reference tissue. Tissues are colored by significance (binomial test, BH-FDR corrected): hub tissues (blue, FDR << 0.05) show significantly elevated connectivity, isolated tissues (orange, FDR << 0.05) show significantly reduced connectivity, and neutral tissues (light blue) show connectivity consistent with random expectation (dashed line, expected =19.8=19.8 ).
Figure 3: Pairwise cross-tissue conservation of protein co-abundance clusters. The top 10 (left) and bottom 10 (right) tissue pairs ranked by shared co-abundance cluster count are shown. Significance was assessed using a degree-preserving network rewiring permutation test ( n=1,000n=1{,}000 permutations) with Benjamini-Hochberg FDR correction. Colored bars indicate significant enrichment (blue, FDR << 0.05) or depletion (dark orange, FDR << 0.05) of shared co-abundance clusters relative to degree-matched random networks.
Figure 4: Protein co-abundance network of Cluster 0 from the brain–gut consensus analysis. Nodes are colored by functional group. Gold rings denote the EV/cargo-sorting and complement-associated ensemble (CD9, VPS4B, ATP6V1A, TRIM25, UBE2N, C3), which shares only 2 of 190 possible pairwise STRING interactions and has no corresponding complex in hu.MAP 3.0. Nodes with a blue outer ring indicate proteins with STRING-documented interactions that fall below the co-abundance threshold and are therefore absent from the consensus network (VPS4B, PDCD6, TRIM25, UBE2N); nodes carrying both rings belong to both categories. The amber dashed edge between SPR and TXNRD1 represents a novel co-abundance association (weight =0.926=0.926 ) with no prior STRING or hu.MAP 3.0 evidence. Edge opacity is proportional to co-abundance weight.
Figure 5: Protein co-abundance network of Cluster 0 from the liver–bone marrow consensus analysis. Nodes are colored by functional group. Gold rings denote proteins independently implicated in heritable white matter diseases (AIMP1: Pelizaeus-Merzbacher-like disease; CAPN1: hereditary spastic paraplegia type 76; EIF2B5: vanishing white matter disease) whose co-occurrence in a single co-abundance cluster was identified without prior disease-focused selection. Blue edges indicate STRING high-confidence interactions (3 of 253 possible pairs, FDR<0.05\text{FDR}<0.05 ); no complexes containing any co-abundance cluster members are recorded in hu.MAP 3.0. Edge opacity is proportional to co-abundance weight.
Figure 41: Overview of the LLM-agent pipeline. Stage 1: tissue-specific co-abundance network construction with agent-driven threshold refinement. Stage 2: pairwise consensus network, Leiden community detection, and module filtering. Stage 3: RAG-based module annotation against HPA, STRING, hu.MAP 3.0, GO, and KEGG. Stage 4: agent-driven module-function inference, gene-level profiling, and novelty assessment via Open Targets and Europe PMC.
SourceCategoryTerm#Genespp /FDR
GOCCExtracellular exosome141.85×10−81.85\times 10^{-8}
GOCCExtracellular vesicle—2.24×10−82.24\times 10^{-8}
GOCCExtracellular organelle—2.25×10−82.25\times 10^{-8}
GOCCExtracellular membrane-bounded organelle—2.25×10−82.25\times 10^{-8}
GOCCExtracellular space155.11×10−75.11\times 10^{-7}
GOCCExtracellular region—1.74×10−51.74\times 10^{-5}
Table 51: Brain–Gut Module 0 (20 proteins) — Integrated Enrichment Profile. HPA: 20/20. hu.MAP 3.0: none. STRING PPI: 2/190 pairs (1.05%).
SourceCategoryTerm#Genespp -value
GOCCCytosol132.13×10−42.13\times 10^{-4}
GOCCExtracellular exosome114.88×10−44.88\times 10^{-4}
GOMFProtein-disulfide reductase activity32.55×10−32.55\times 10^{-3}
GOMFDisulfide oxidoreductase activity34.19×10−34.19\times 10^{-3}
GOBPNucleobase-containing small molecule metabolic process68.17×10−38.17\times 10^{-3}
GOMFOxidoreductase activity, sulfur group donors31.10×10−21.10\times 10^{-2}
Table 52: Liver–Bone Marrow Module 0 (23 proteins) — GO/KEGG Functional Enrichment. HPA: 23/23. hu.MAP 3.0: none.
CategoryTerm#GenesFDR
KeywordAcetylation162.58×10−52.58\times 10^{-5}
TISSUESDigestive gland141.10×10−31.10\times 10^{-3}
WikiPathwaysMethionine de novo and salvage pathway32.20×10−32.20\times 10^{-3}
TISSUESViscus175.40×10−35.40\times 10^{-3}
KeywordRedox-active center31.10×10−21.10\times 10^{-2}
ComponentExtracellular exosome111.60×10−21.60\times 10^{-2}
Table 53: Liver–Bone Marrow Module 0 (23 proteins) — STRING Network Enrichment & Protein–Protein Interactions. STRING PPI: 3/253 pairs (1.19%).

Explore similar work

CardsList
Oct 15, 2026

spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics

Spatial proteomics (SP) captures the molecular composition and spatial organization of tissues at single-cell resolution, opening a direct view of the cellular ecosystems that drive disease. Protein panels, acquisition platforms, and tissue contexts differ from one cohort to the next, and a coherent picture of tissue biology can only emerge from models trained to bridge this heterogeneity. Foundation models offer the natural path forward, but their training depends on large, harmonized corpora that span platforms, panels, and cohorts. No public resource currently approaches that scale; available data sit in small, narrowly scoped releases with incompatible formats and conventions. We present spora\texttt{spora}, a large-scale multimodal SP dataset of 13,546 standardized multiplex images from 5,477 patients across 35 cohorts and five major SP technologies. Curated with biomedical experts, spora\texttt{spora} provides model training-ready formats together with cell and tissue segmentations, structured clinical metadata, and, where available, paired H&E enabling the training of virtual staining models. It is complemented by spora [io]\texttt{spora [io]}, a unified interface for tiling, sampling, and standardization across all cohorts and modalities, and spora [bench]\texttt{spora [bench]}, a benchmark suite spanning cell phenotyping, diagnostic and therapeutic prediction as well as virtual staining tasks across twelve cohorts that places current SP foundation models on common ground against task specific baselines.

Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain +9

EPFL · University Hospital Zurich · University of Zurich +2

Multimodal DatasetMedical Image Benchmarks
AskPDF
Mar 21, 2026

Beyond Expression Similarity: Contrastive Learning Recovers Functional Gene Associations from Protein Interaction Structure

The Predictive Associative Memory (PAM) framework posits that useful relationships often connect items that co-occur in shared contexts rather than items that appear similar in embedding space. A contrastive MLP trained on co-occurrence annotations--Contrastive Association Learning (CAL)--has improved multi-hop passage retrieval and discovered narrative function at corpus scale in text. We test whether this principle transfers to molecular biology, where protein-protein interactions provide functional associations distinct from gene expression similarity. Four experiments across two biological domains map the operating envelope. On gene perturbation data (Replogle K562 CRISPRi, 2,285 genes), CAL trained on STRING protein interactions achieves cross-boundary AUC of 0.908 where expression similarity scores 0.518. A second gene dataset (DepMap, 17,725 genes) confirms the result after negative sampling correction, reaching cross-boundary AUC of 0.947. Two drug sensitivity experiments produce informative negatives that sharpen boundary conditions. Three cross-domain findings emerge: (1) inductive transfer succeeds in biology--a node-disjoint split with unseen genes yields AUC 0.826 (Delta +0.127)--where it fails in text (+/-0.10), suggesting physically grounded associations are more transferable than contingent co-occurrences; (2) CAL scores anti-correlate with interaction degree (Spearman r = -0.590), with gains concentrating on understudied genes with focused interaction profiles; (3) tighter association quality outperforms larger but noisier training sets, reversing the text pattern. Results are stable across training seeds (SD < 0.001) and cross-boundary threshold choices.

Jason Dury

Eridos AI

Contrastive LearningBioinformatics
AskPDF
Jun 23, 2026

SP-Mind: An Autonomous Reasoning Agent for Spatial Proteomics Analysis

Spatial proteomics enables single-cell-resolution characterization of protein expression within tissue architecture, playing a critical role in understanding tumor microenvironments and guiding precision medicine. However, current analysis workflows remain fragmented, requiring expert manual orchestration of heterogeneous tools and limiting research scalability and reproducibility. We present SP-Mind, the first autonomous AI agent designed to unify the spatial proteomics analysis pipeline, from raw multiplexed tissue imaging to downstream phenotype discovery. Equipped with expert-curated biological analysis skills and specialized computational tools, SP-Mind converts natural-language queries into end-to-end analytical workflows without task-specific fine-tuning. To rigorously evaluate its capabilities, we introduce SP-Bench, a comprehensive benchmark spanning diverse tissue types, comprising 102 tasks across 18 distinct categories. Through extensive evaluation on SP-Bench and established downstream tasks, SP-Mind achieves state-of-the-art performance compared to existing open-source biomedical agent baselines.

Yucheng Yuan, Yuanfeng Ji, Zhongxiao Li +1

Department of Computer Science, Stanford University, Stanford, USA · Department of Radiation Oncology, Stanford University, Stanford, USA

AI Agent BenchmarksTool-Using AgentsScientific Workflow AutomationAutonomous Research Agents
AskPDF