FibrilNet maps conserved and tissue-specific molecular environments across systemic amyloidoses

Technical Summary: FibrilNet Maps Conserved and Tissue-Specific Molecular Environments Across Systemic Amyloidoses

Problem Statement
Systemic amyloidoses are protein-misfolding disorders characterized by the accumulation of insoluble fibrillar deposits. While distinct amyloidogenic precursor proteins initiate these diseases (e.g., Transthyretin in ATTR, Immunoglobulin Light Chains in AL, Serum Amyloid A in AA), the resulting tissue pathology often shares recurrent non-fibrillar components involved in extracellular matrix organization, complement activation, lipid transport, and proteostasis. A critical unresolved question is whether these recurrent proteins form a conserved, systems-level molecular environment across different amyloid diseases, and to what extent this environment is dependent on the specific precursor protein, the tissue context, or the disease variant. Traditional network approaches often treat protein-protein interactions (PPIs) as functionally interchangeable, ignoring the biological context that may differentiate interactions within the same physical graph.

Methodology
The authors developed FibrilNet, a network framework designed to integrate experimentally defined amyloid proteomes with a human PPI graph enriched by Gene Ontology (GO)-derived semantic information.

• Data Construction: The underlying human graph comprises 17,997 proteins and 925,977 physical interactions. Each interaction is augmented with a 9-dimensional semantic vector derived from GO annotations (biological process, molecular function, and cellular component), encoding information-content-based node similarity, path similarity, and annotation availability.
• Algorithmic Approach: FibrilNet compares two diffusion strategies:
  1. Topology-only Random Walk with Restart (RWR): Standard propagation based solely on graph connectivity.
  2. Semantic-RWR: A modified propagation where transition probabilities are biased toward interactions with higher semantic compatibility (ontology alignment).
• Evaluation Strategy: The study employs a "frozen" leave-one-out module reconstruction approach. Disease-associated protein modules (derived from experimental proteomics) are used as query sets. One protein is removed at a time, and the algorithm attempts to re-rank the remaining graph nodes to recover the held-out target. Crucially, disease labels are not used to retrain the underlying representation; the evaluation tests the ability of the network structure and semantic context to recover known disease modules.
• Disease Models Analyzed:
  • Cardiac: Expanded and compact amyloid-specific proteomes for ATTR and AL.
  • Renal: Proteomic atlases for AA and ALECT2 amyloidosis.
  • Neurologic: A murine hTTR-A97S sural-nerve model mapped to human orthologs to represent neurologic ATTRv.

Key Contributions and Results

1. Superiority of Semantic Diffusion:
   Semantic-RWR consistently outperformed topology-only RWR in reconstructing experimentally defined amyloid modules.

   • In expanded cardiac ATTR, Semantic-RWR increased the Mean Reciprocal Rank (MRR) from 0.00167 to 0.05015 and Recall@100 from 0.0199 to 0.3377, improving 132 of 151 held-out targets.
   • Significant gains were also observed in renal AA (MRR 0.00249 to 0.09061) and ALECT2 (MRR 0.00239 to 0.01614).
   • The effect was robust across varying module sizes and sensitivity tiers, suggesting the amyloid environment possesses a distributed, functionally coherent organization that topology alone fails to capture.

2. Identification of a Restricted Recurrent Core:
   Analysis of compact modules across four systemic amyloidoses (ATTR, AL, AA, ALECT2) revealed a limited but significant overlap.

   • Three proteins—APCS, VTN, and TIMP3—were present in all four modules (Feature-Enrichment Core frequency = 1), defining a strict cross-disease core.
   • APOE was present in three of the four modules.
   • These proteins represent points of cross-disease convergence rather than universal causal drivers.

3. Tissue-Specificity and Precursor Separation:

   • Cardiac vs. Neurologic: Despite sharing TTR as the precursor, cardiac and neurologic ATTR modules showed minimal overlap (Jaccard similarity = 0.0569; only 19 shared proteins).
   • Precursor Limitations: Seeding the network with the precursor protein alone (TTR or LECT2) failed to strongly recover the broader downstream molecular environment in either cardiac or neurologic contexts. This indicates that the mature amyloid environment is a distributed network state not reducible to local graph proximity around the fibril-forming protein.
   • Directional Specificity: In the neurologic ATTRv model, semantic improvements were concentrated in the downregulated proteomic program (associated with cytoskeletal abnormalities and axonal dysfunction), whereas upregulated proteins showed less semantic coherence.

Significance and Claims
The paper claims that FibrilNet supports a multilayer model of systemic amyloidosis. In this model:

• A restricted conserved amyloid environment (the core) coexists with strong precursor-, tissue-, and disease-specific organization.
• The downstream molecular environment is a distributed network state that emerges from the interaction of the precursor with the specific tissue context, rather than being a simple local neighborhood of the precursor.
• Ontology-derived semantic information provides a biologically meaningful inductive bias, allowing network propagation to distinguish between functionally compatible and incompatible interactions, thereby revealing the "distributed molecular environment" of amyloid deposition.

The authors emphasize that their findings do not challenge the causal role of precursor proteins but rather clarify that the pathological environment is a complex, tissue-dependent system state. The framework offers a reproducible method for analyzing systemic amyloidoses and suggests that future analyses should account for variant- and tissue-specific molecular datasets to fully resolve the heterogeneity of these diseases.