IEKB: a comprehensive knowledge base for inner ear genetics integrating curated associations, cochlear interactions, Bayesian candidate prioritisation, explainable dark-gene support relations, and a scientific entity network

The paper introduces the Inner Ear Knowledge Base (IEKB), an open-access resource that unifies curated gene-phenotype-disease associations, cochlear interactions, Bayesian candidate prioritization, explainable support relations, and a multi-entity scientific network to advance inner-ear genetics research through automated curation and interactive exploration tools.

The Gist

Imagine the human inner ear as a highly complex, ancient city. This city is responsible for our hearing and balance, but it's built with thousands of tiny, intricate parts (genes) that work together. Sometimes, a part breaks, and the city goes silent (hearing loss).

For a long time, scientists have been trying to map this city. They've found some broken parts, but the information is scattered. It's like having a million different maps, each drawn by a different explorer, stored in different libraries, and written in different languages. Some maps show the streets (genes), others show the traffic patterns (interactions), and others just list the buildings (phenotypes). No single map showed the whole picture.

Enter IEKB: The "Google Maps" for the Inner Ear City.

This paper introduces a new, free, open-source tool called the Inner Ear Knowledge Base (IEKB). Think of it as a massive, super-smart digital library and navigation system that finally puts all those scattered maps into one place.

Here is how it works, using simple analogies:

1. The "Super-Sleuth" Team (AI + Humans)

Imagine you need to read 250,000 scientific papers to find clues about hearing loss. That's impossible for one person.

• The AI Agents: IEKB uses a team of "robot librarians" (AI agents) that read thousands of papers a day. They are like super-fast scanners that highlight sentences like "Gene X causes hearing loss" or "Gene Y talks to Gene Z."
• The Human Experts: But robots can make mistakes. So, real scientists (the "Senior Detectives") check the robot's work. They verify the clues, fix errors, and make sure the information is trustworthy.
• The Result: A clean, organized database of 6,000+ confirmed connections between genes and hearing problems, plus 4,000+ interactions showing how genes talk to each other inside the ear.

2. The "Sherlock Holmes" Module (Finding the Unknown)

Here is the really cool part. We know about many genes that cause hearing loss, but there are thousands of genes we don't know about yet. These are the "Dark Matter" of the ear.

• The Prediction Engine: IEKB has a special "Crystal Ball" (a Bayesian model). It looks at the genes we do know and asks: "Who are the neighbors of these known troublemakers? Who shares their habits?"
• The Guess: Based on these patterns, it generates a list of 243,000+ suspects (candidate genes) that might be causing hearing loss but haven't been proven yet. It gives them a "score" of how likely they are to be guilty.

3. The "Evidence Chain" (Explaining the Guess)

Usually, when a computer guesses something, it just gives a number. "This gene is 85% likely." But scientists need to know why.

• The Detective's Notebook: IEKB adds a layer called "Dark Relation." If the computer suspects a new gene, it doesn't just say "maybe." It says: "We suspect this new gene because it hangs out with Gene A (who we know causes deafness) and Gene B (who works in the same factory)."
• The Benefit: It provides a clear, auditable trail of evidence, so researchers know exactly which known genes support the new guess.

4. The "City Map" (The Network)

Instead of just a list of names, IEKB builds a giant, interactive 3D web.

• Imagine a spiderweb where every strand connects a gene to a disease, a cell type, or a body part.
• You can click on "Sensorineural Hearing Loss" and see the entire web of genes, pathways, and cells involved. It helps researchers see the big picture, not just isolated dots.

5. The "Conversational Assistant" (IEKB QA)

Finally, IEKB has a chatbot friend.

• Instead of typing complex search codes, you can just ask: "Which genes are linked to age-related hearing loss, and what do we know about their mechanism?"
• The AI reads the database, finds the answers, and writes a clear, citation-backed summary for you, like a helpful research assistant.

Why Does This Matter?

Before IEKB, a scientist might spend months digging through libraries to find a connection. Now, they can open IEKB, see the whole map, get a list of the most promising new suspects, and see exactly why those suspects are interesting.

In short: IEKB takes the chaotic, scattered puzzle pieces of inner-ear genetics and assembles them into a complete, interactive picture. It helps us find the broken parts of our hearing city faster, so we can fix them sooner. And the best part? It's free for everyone to use.

Technical Summary

1. Problem Statement

Inner ear genetics has expanded rapidly due to next-generation sequencing (NGS), yet the supporting evidence remains fragmented across vast literature and disparate resources. Existing databases (e.g., Hereditary Hearing Loss Homepage, Deafness Variation Database, Gene4HL) typically focus on specific layers such as loci, variants, or expression data, but fail to provide:

• Integrated biological interpretation connecting genes, phenotypes, and diseases.
• Cochlear-specific interaction networks.
• Probabilistic prioritization of novel ("dark matter") candidate genes.
• Explainable support relations linking novel candidates to known genes.
• A unified multi-entity scientific network for hypothesis generation.

Researchers and clinicians lack a single, open-access platform that unifies curated associations, interaction evidence, predictive modeling, and network exploration for inner ear research.

2. Methodology

The authors developed the Inner Ear Knowledge Base (IEKB) using a rigorous, hybrid human-AI curation pipeline. The construction process involved the following stages:

A. Data Curation Pipeline

• Gene Universe Construction: Defined a consensus universe of 16,563 human genes (retaining those with mouse orthologs) as the search space.
• Literature Retrieval: Queried PubMed and Europe PMC using alias-aware queries for each gene combined with 11 inner-ear context terms (e.g., "cochlea," "hair cell," "vestibular"). This yielded 250,696 deduplicated records.
• Agent-Assisted Extraction: Utilized an automated agent (LLM-based) to process abstracts and full texts. The agent extracted structured data constrained by Pydantic schemas into two categories:
  1. Gene–Phenotype–Disease Associations: Including inheritance, mechanism, cell type, and evidence level.
  2. Cochlear Gene–Gene Interactions: Including interaction type, directionality, pathway, and experimental context.
• Human-in-the-Loop & Expert Review:
  • Automated outputs underwent hybrid cleaning (normalization, vocabulary mapping).
  • Researchers randomly sampled ~10% of batches for error correction.
  • Final Expert Review: All 6,051 associations and 4,102 interactions were reviewed by inner-ear genetics experts, cross-referenced with ClinGen and HHL, and assigned confidence tiers (Tier 1: literature-supported; Tier 2: inferred/requiring validation).

B. Predictive Modeling (The "Dark Matter" Framework)

To identify novel gene-phenotype associations not yet in the literature, IEKB implemented a Bayesian inference framework:

• Inputs: Curated seed genes, STRING protein interaction graphs, and HGNC gene-group annotations.
• Feature Engineering: Three features were computed for candidate genes:
  1. Direct Neighbor Score: Weighted sum of connections to seed genes in the STRING network.
  2. Random Walk with Restart (RWR): Network propagation score to measure global connectivity.
  3. Gene-Group Overlap: Jaccard-like ratio of shared gene-group memberships.
• Model: A class-balanced logistic regression model transformed these features into a posterior probability score.
• Output: 243,071 candidate associations across 13,229 genes and 43 phenotypes.

C. Explainable Support (Dark Relation Layer)

To make predictions actionable, a supervised Dark Relation layer was built:

• Goal: Rank known genes that provide the strongest evidence for a specific "dark matter" candidate.
• Method: A ranking model (Logistic Regression or LightGBM) trained on query-centric pairs (phenotype + target gene).
• Features: Included network scores (RWR, Adamic-Adar), gene-group overlaps, and evidence weights.
• Output: An auditable evidence chain providing a calibrated score and a natural-language summary of why a known gene supports a candidate.

D. Knowledge Graph & Web Implementation

• Network Construction: Built a multi-entity graph with ~4,000 entities (genes, diseases, pathways, cell types) connected by 28,616 deterministic edges and 83,712 literature-derived links.
• Architecture: A client-side web application (Next.js) using sql.js to run SQL queries entirely in the browser against a local SQLite database, eliminating backend server dependencies.
• IEKB QA: A companion conversational interface using a multi-phase agent pipeline to retrieve structured evidence and synthesize citation-backed natural language answers.

3. Key Contributions

1. First Unified Resource: IEKB is the first openly accessible platform integrating curated associations, cochlear interactions, probabilistic prioritization, explainable support relations, and a scientific network for inner ear genetics.
2. Scale and Coverage:
   • Curated 6,051 gene–phenotype–disease associations from 2,494 genes.
   • Cataloged 4,102 cochlear gene–gene interactions.
   • Prioritized 243,071 novel candidate associations across 43 phenotype categories.
3. Explainable AI: Unlike "black box" predictors, IEKB provides a Dark Relation layer that explicitly ranks known genes supporting a novel candidate, offering feature-level decomposition and natural language explanations.
4. Accessibility: All data is released under CC BY 4.0 without registration, available in CSV, JSON, SQLite, and XLSX formats, with a browser-based query engine.

4. Results

• Performance Metrics:
  • The Bayesian prioritization model achieved a global AUC-ROC of 0.8603 and AUC-PR of 0.1674.
  • Per-phenotype AUC-ROC ranged from 0.6981 (sensorineural hearing loss) to 0.9758 (infection-induced hearing loss).
  • High precision was confirmed against expert-curated gold standards (Tier 1 records).
• Network Analysis:
  • Identified hub entities such as Organ of Corti (degree 1,188) and Hair Cells (degree 871).
  • Key pathway hubs included Oxidative Stress Response and FGF Signaling.
• Coverage Comparison: IEKB covers 97.5% of ClinGen hearing-loss genes and 94.9% of HHL entries, while adding 2,301 genes absent from these reference sets.
• Case Studies: Successfully highlighted novel candidates (e.g., RBM5 for sensorineural hearing loss) and visualized complex interplays in Usher syndrome.

5. Significance

• Hypothesis Generation: By combining curated evidence with predictive "dark matter" modeling, IEKB transforms from a passive catalog into an active tool for generating testable hypotheses about novel hearing loss genes.
• Transparency: The "Dark Relation" layer addresses the "black box" problem in AI-driven genomics by providing auditable evidence chains, allowing researchers to validate predictions against domain expertise.
• Workflow Integration: The integration of bibliometrics, network visualization, enrichment analysis, and conversational AI (IEKB QA) within a single, client-side accessible platform significantly lowers the barrier to entry for researchers and clinicians.
• Future-Proofing: The system is designed for continuous updates via an automated agent framework, ensuring the database remains current with the rapidly evolving literature.

In summary, IEKB represents a significant advancement in biomedical informatics for otology, bridging the gap between fragmented literature, structured knowledge, and predictive AI to accelerate the discovery of genetic causes of hearing loss.