Affinage: Genome-Scale Mechanistic Gene Annotation from the Published Literature

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Understanding gene function is crucial for biological research, but this information is often fragmented or outdated in current databases.
• Existing methods for synthesizing literature on gene function are time-consuming, expensive, and difficult to scale.
• Large language models (LLMs) offer potential for literature synthesis but require efficient, reproducible methods.

Purpose of the Study:

• To develop and validate Affinage, a novel LLM pipeline for automated gene function annotation from primary literature.
• To create a reusable, structured knowledge base of human gene mechanistic functions.
• To identify the proportion of the human proteome that remains mechanistically uncharacterized.

Main Methods:

• Affinage employs a two-pass LLM approach: a biologist-designed reading pass for experimental evidence extraction and a synthesis pass for mechanistic reasoning.
• The pipeline processes scientific literature to generate annotations for protein-coding genes.
• Annotations are validated against existing curated databases and assessed by an LLM judge.

Main Results:

• Affinage successfully annotated 19,293 human protein-coding genes, providing mechanistic insights for thousands previously lacking functional descriptions.
• The pipeline outperformed curated databases in accuracy for 99.1% of compared genes.
• Affinage identified that 10% of the human proteome remains mechanistically uncharacterized.

Conclusions:

• Affinage offers a scalable, reproducible, and accurate method for gene function annotation, significantly enhancing biological knowledge discovery.
• The pipeline serves as a continuously updated, literature-grounded census of gene function, guiding future research.
• This work demonstrates the power of domain-expert-guided LLM pipelines for improving public biological data resources.