Skip to content

Background

Research Context

The systematic identification and curation of genetic research literature on non-model organisms presents substantial practical challenges. While model organisms—such as Escherichia coli, Drosophila melanogaster, and Mus musculus—have long served as cornerstones of genetic research, studies of non-model organisms offer complementary biological insights that are not fully captured by these established systems.

Scientific Importance of Non-Model Organism Genetics

Non-model organisms play a critical role across multiple domains of biological research.

Biodiversity and Functional Diversity

Non-model organisms often harbor unique metabolic pathways, adaptive strategies, and protein repertoires that are absent from conventional model systems. These features are essential for advancing evolutionary biology, ecology, and systems-level understanding of biological complexity.

Applied Biotechnology Potential

Many non-model organisms represent largely untapped reservoirs of biotechnologically valuable traits, including novel enzymes, natural product biosynthesis pathways, and mechanisms of tolerance to environmental stresses such as salinity, drought, and disease.

Phylogenetics and Comparative Genomics

Accurate inference of phylogenetic relationships and gene family evolution requires broad taxonomic sampling. Overreliance on model organisms constrains the resolution and generalizability of comparative genomic analyses.

Addressing Taxonomic Bias

The disproportionate emphasis on model organisms introduces systematic taxonomic bias into the scientific literature. Incorporation of non-model organisms enables validation of general biological principles and promotes a more representative understanding of biological systems.

Problem Statement

Despite their scientific importance, the identification of genetic research focused on non-model organisms remains challenging.

Current Limitations

Labor-Intensive Manual Curation

To our knowledge, no fully automated or systematic method currently exists for efficiently identifying non-model organism genetics papers from large-scale literature databases. Existing approaches rely heavily on manual review, including:

  • Individual examination of titles, abstracts, and MeSH terms across thousands of publications (e.g., 4,940 papers in our case)
  • Construction of complex database queries requiring expert knowledge of biological nomenclature
  • Iterative searches across multiple resources (e.g., NCBI Gene, NCBI Taxonomy)
  • Manual extraction and structuring of gene names, gene identifiers, species names, taxonomy IDs, higher-level taxonomic classifications, and research categories

Complexity Beyond Traditional Workflows

Literature screening for non-model organism genetics involves challenges that exceed the capacity of traditional mechanical workflows, including:

  • Multiple decision points requiring domain-specific expertise
  • Autonomous task execution with adaptive query formulation
  • Integration of heterogeneous biological databases
  • Contextual interpretation that cannot be reliably captured by rule-based or simple keyword-matching approaches

These characteristics motivate the need for more sophisticated, AI-driven solutions.

Motivation

Hypothesis

While PubTator demonstrates strong performance in species recognition and gene annotation for model organisms, we hypothesize that gene annotation for non-model organisms is inherently more challenging. This difficulty likely arises from heterogeneous and inconsistent gene nomenclature, resulting in suboptimal annotation performance.

Validation of Species Recognition Performance

To contextualize this hypothesis, we examined the composition of PubTator’s species recognition evaluation datasets to assess representation of non-model organisms:

  • Analysis of the PubTator3 evaluation dataset revealed that 85.93% (104/635) of test cases correspond to non-model organisms (AIONER test dataset: https://github.com/ncbi/AIONER/blob/main/data/pubtator/SPECIES_Test.PubTator)
  • Species named entity recognition and entity linking achieved high performance in these evaluations
  • Validation results are documented in data/species_match, using scripts/check_species_match.py

Working Assumption Based on this evidence, we assume that PubTator’s species recognition performance is reliable for non-model organisms and therefore focus our investigation on potential limitations in gene annotation.

AI Agent-Based Approach

Given the complexity of multi-step decision-making, database integration, and contextual interpretation required for literature curation, we propose an AI agent-based approach as a suitable solution. AI agents are capable of:

  • Autonomously navigating complex decision trees during literature evaluation
  • Performing iterative database queries with adaptive refinement
  • Integrating information across multiple biological databases
  • Extracting and structuring relevant metadata, including gene identifiers, species information, taxonomic classifications, and research categories
  • Scaling to large literature corpora while maintaining consistent evaluation criteria

Existing Literature Annotation Systems

PubTator PubTator, developed by the NCBI, provides automated bioconcept annotation for biomedical literature. Although effective for many annotation tasks, its performance in gene annotation for non-model organisms has not been systematically evaluated.

Manual Curation Gold-standard annotations depend on expert manual curation, which remains prohibitively time-consuming and impractical for large-scale literature databases.

Rule-Based Systems Traditional keyword-based and rule-based approaches lack the flexibility and contextual understanding required to accurately identify non-model organism genetics research.

Gap in Current Approaches

To date, no existing system simultaneously provides:

  1. High-precision identification of non-model organism genetics literature
  2. Automated and scalable processing of large literature corpora
  3. Structured extraction of genetic and taxonomic metadata
  4. Domain-aware decision-making for ambiguous or borderline cases

This study addresses these gaps through the development and evaluation of AI agent-based literature curation systems specifically tailored to non-model organism genetics research.