Source-linked AI summary
Predicting multicellular function through multi-layer tissue networks
Marinka Zitnik, Jure Leskovec
TL;DR
Predicting protein functions that depend on specific tissues remains challenging. OhmNet learns hierarchy-aware protein representations across tissue-specific interaction networks, outperforming alternative approaches and transferring functions to unannotated tissues.
Problem
Existing approaches do not adequately predict gene–function relationships specific to particular tissues, although protein functions often depend on tissue context.
Method
OhmNet learns unsupervised protein features across tissue-specific network layers while encoding their dependencies through a multiscale tissue hierarchy.
Results
OhmNet outperforms alternative approaches by up to 14.9% on multi-label classification and transfers cellular functions to unannotated tissues.
Takeaways & Limitations
Hierarchy-aware, task-independent features support cellular function prediction across tissues and provide insights into tissue-specific protein actions.
Takeaways & Limitations
OhmNet assumes that dependencies between network layers are given as a hierarchy rather than a graph.
Abstract
from arXiv · showhide
Motivation: Understanding functions of proteins in specific human tissues is essential for insights into disease diagnostics and therapeutics, yet prediction of tissue-specific cellular function remains a critical challenge for biomedicine. Results: Here we present OhmNet, a hierarchy-aware unsupervised node feature learning approach for multi-layer networks. We build a multi-layer network, where each layer represents molecular interactions in a different human tissue. OhmNet then automatically learns a mapping of proteins, represented as nodes, to a neural embedding based low-dimensional space of features. OhmNet encourages sharing of similar features among proteins with similar network neighborhoods and among proteins activated in similar tissues. The algorithm generalizes prior work, which generally ignores relationships between tissues, by modeling tissue organization with a rich multiscale tissue hierarchy. We use OhmNet to study multicellular function in a multi-layer protein interaction network of 107 human tissues. In 48 tissues with known tissue-specific cellular functions, OhmNet provides more accurate predictions of cellular function than alternative approaches, and also generates more accurate hypotheses about tissue-specific protein actions. We show that taking into account the tissue hierarchy leads to improved predictive power. Remarkably, we also demonstrate that it is possible to leverage the tissue hierarchy in order to effectively transfer cellular functions to a functionally uncharacterized tissue. Overall, OhmNet moves from flat networks to multiscale models able to predict a range of phenotypes spanning cellular subsystems
1 Introduction
Protein functions vary across tissues, but existing interaction-network methods generally assume cellular functions are constant across organs and lack tissue specificity. OhmNet addresses this gap by learning unsupervised protein features across tissue-specific networks linked through a multiscale tissue hierarchy.
- Protein functions can differ across tissues, making unified analyses of tissue-specific cellular functions important for biology, disease interpretation, and therapeutics.
- Existing computational methods largely assume cellular functions are constant across tissues, limiting accurate tissue-specific functional maps.
- OhmNet learns unsupervised protein features from tissue-specific interaction networks organized as layers linked by a hierarchy modeling dependencies among tissues.
- Its task-independent representations are designed to generalize across downstream prediction tasks and tissues, rather than being optimized for one specific task.
- The study evaluates OhmNet on 107 genome-wide tissue-specific protein interaction layers across cellular-function classification, cross-tissue transfer learning, and network-embedding visualization.
2 Related work
Network feature-learning methods primarily use matrix factorization or neural network embeddings, while existing approaches struggle to represent multi-layer biological networks without distorting their structure. Related biological methods include graphlet-based neighborhood measures and task-dependent deep architectures.
- Network feature learning: Automatic network feature learning methods are broadly categorized into matrix factorization and neural network embedding approaches.Matrix factorization represents networks as data matrices and projects them into low-dimensional spaces using linear or nonlinear techniques.
- Network feature learning: Neural network embeddings use random-walk neighborhoods and objective functions designed to preserve local and global network structure.These methods address limitations of matrix factorization by embedding nodes into low-dimensional vector spaces.
- Multi-layer networks: Existing multi-layer embedding methods either analyze layers independently or aggregate them into one weighted network, altering topology and node importance.The related-work discussion identifies multi-layer representation as a challenging problem for neural network embedding methods.
- Biological network methods: Graphlet-based approaches predict protein functions by counting small subgraph patterns in each node’s extended network neighborhood.Graphlet degree vectors provide a systematic description of local network structure.
- Task-dependent architectures: Task-dependent graph-specific deep architectures directly optimize downstream prediction objectives, such as cellular function prediction in a particular tissue.The paper distinguishes its approach from these architectures by noting their dependence on a specific prediction task.
3 Feature learning in multi-layer networks
OhmNet formulates multi-layer network feature learning as maximum-likelihood optimization that maps nodes in each layer to low-dimensional representations. It combines layer-specific neighborhood objectives with hierarchy-based regularization, learning shared features across biological contexts and at multiple scales.
- Model formulation: OhmNet represents a multi-layer network as K edge-specific layers over nodes and learns functions f_i: V_i → R^d for node embeddings.Each f_i is equivalently a |V_i| × d parameter matrix.
- Hierarchical objectives: OhmNet models inter-layer dependencies with a directed-tree hierarchy, encouraging proteins in nearby layers to share similar features through parent-child regularization.The recursive regularization makes a node’s features in each hierarchy element similar to its features in that element’s parent.
- Single-layer objectives: Its single-layer objectives embed nodes with similar network neighborhoods or structural roles close together by predicting sampled neighborhood members from learned features.Neighborhoods are sampled using appropriately biased random walks on each layer.
- Multiscale representations: The hierarchy enables feature representations at multiple scales by learning mappings for both leaf layers and internal hierarchy elements.Internal elements provide representations for shared higher-level biological contexts.
- Computational efficiency: OhmNet models inter-layer dependencies in O(|M|N) time rather than O(K^2N) for a fully pairwise model.Here, K is the number of layers, N the number of nodes, and |M| the number of hierarchy elements.
4 Tissue-specific interactome data
The study integrates experimentally supported human protein interactions, tissue-specific gene mappings, and ontology-based tissue relationships into a 107-layer tissue interactome. It also curates tissue-specific cellular functions across the tissue hierarchy for downstream evaluation.
- Tissue hierarchy: The tissue hierarchy is a directed tree of 219 BRENDA Tissue Ontology terms, with 107 tissues as blue leaf nodes.Edges point from children to parents through “develops_from”, “is_a”, “part_of”, and “related_to” relationships.
- Tissue hierarchy: Tissue mappings from HPRD to the BRENDA Tissue Ontology were retrieved from Greene et al., and hierarchical relationships were determined from the ontology’s directed acyclic graph.Examples include muscle, adrenal cortex, bone marrow, and spleen.
- Tissue-specific network construction: The global unweighted human PPI network contains 21,557 proteins and 342,353 experimentally supported physical interactions.Interactions based on gene expression and evolutionary data were excluded.
- Tissue-specific network construction: Tissue-specific PPI layers were constructed by labeling global-network edges as specifically co-expressed when both proteins were tissue-specific or one was tissue-specific and the other ubiquitous.This criterion was applied separately to each of the 107 tissues.
- Tissue-specific functions: The dataset contains 584 tissue-specific cellular functions covering 48 tissues, with functions assigned to one or more tissue-hierarchy leaves.Functions were manually curated from Gene Ontology biological processes, and annotations were propagated along the ontology hierarchy; only functions with at least 15 annotated proteins were considered.
5 Results
OhmNet’s unsupervised, task-independent embeddings support cellular-function prediction across tissue-specific networks, outperforming benchmark approaches by modeling multiscale tissue hierarchy. The hierarchy also enables transfer to unannotated tissues and meaningful visualization of tissue-specific protein features.
- Cellular-function prediction: OhmNet learns separate unsupervised feature vectors for every node in every layer, followed by one-vs-all linear classifiers for each function.The classifiers use modified Huber loss with elastic net regularization and cross validation.
- Cellular-function prediction: 10.7% is OhmNet’s improvement over GeneMania, attributed to modeling multiscale tissue organization that GeneMania cannot use to weight layers.The comparison concerns cellular-function prediction across tissue networks.
- Transfer learning: Transferability decreases as tree-based distance between source and target tissues increases, while predictions use hierarchy-distance-weighted averages of source-layer classifiers.The weighting is mathematically equivalent to OhmNet’s regularization, and anatomically close tissues are most informative.
- Transfer learning: 7% is the average decrease in AUC for transfer tasks to tissues treated as having no cellular-function annotations.For the forebrain, the performance difference is only 5.2%, supported by nine closely related tissue layers including the cerebellum and midbrain.
- Network visualization: OhmNet produced meaningful brainstem and brain network layouts, clustering same-colored points and separating regions despite an objective independent of visualization.PCA and NMF visualizations were not meaningful because proteins from the same tissue did not cluster together.
6 Conclusion
OhmNet learns task-independent protein features from a 107-tissue multi-layer network while modeling tissue interdependence across a hierarchy of biological scales. The approach supports accurate cellular function prediction, function transfer to unannotated tissues, and tissue insights, while future work will address graph-based dependencies and additional disease-related tasks.
- Conclusion: OhmNet learns state-of-the-art task-independent protein features from a multi-layer network spanning 107 tissues and dozens of biological scales.The method models tissue interdependence up and down a tissue hierarchy.
- Conclusion: The learned features achieve excellent cellular function prediction accuracy, transfer functions to unannotated tissues, and provide insights into tissues.
- Future work: Future work will extend OhmNet from hierarchy-based layer dependencies to graph-based dependencies in biological scenarios.The current approach assumes dependencies between layers are provided as a hierarchy.
- Future work: Because the learned protein features are independent of downstream tasks, future evaluation will test gene-disease association prediction and disease pathway detection.
B Brainstem tissues C Brain tissues
The case study visualizes brain tissue-specific protein interaction networks within a two-level BRENDA Tissue Ontology hierarchy. Nine brain tissues are represented as hierarchy leaves, each associated with its own tissue-specific network.
- Brain tissues: The brain tissue hierarchy has two levels and is specified using the BRENDA Tissue Ontology.
- Brain tissues: Nine brain tissues form the leaves of the hierarchy, with each leaf linked to a tissue-specific protein interaction network.
- Brainstem tissues: The figure separately visualizes brainstem-specific protein interaction networks.