OmicsFM tissue-specific attention networks (30 human tissues, proteomics and transcriptomics)

Tissue-specific protein association networks produced by OmicsFM, a foundation model for molecular expression profiles pretrained on proteomics, bulk transcriptomics and single-cell transcriptomics (github.com/CompOmics/OmicsFM). The networks are the model's attention, extracted zero-shot from the deep proteome and transcriptome profiles of 30 healthy human tissues (Wang et al. 2019, ProteomeXchange PXD010154). They capture which proteins the model relates to each other in a given tissue context and can be used as tissue-conditioned functional association networks: ranking candidate interactions, detecting pathway-level modules, comparing tissues, or as a prior for downstream analyses. Files tissue_networks_proteomics.zip (16.9 GB) — 60 matrices: 30 tissues from the proteomics (iBAQ) profiles, each with the learned-identity model (tissue.npz) and the ESM-C sequence-embedding model (tissue_esmc.npz) tissue_networks_transcriptomics.zip (14.5 GB) — 30 matrices from the matched transcriptome profiles Each .npz holds sim, the symmetric attention matrix over the proteins detected in that tissue, and proteins, the UniProt accessions indexing its rows and columns: import numpy as np d = np.load('tissue_networks/proteomics_ibaq/liver.npz') sim, proteins = d['sim'], d['proteins'] These networks underlie the tissue-specificity, pathway-community and network-enrichment analyses of the OmicsFM manuscript; the experiment bundles that read them are deposited separately.