Omics Data Discovery Agents

This paper presents an agentic framework leveraging large language models and containerized tools to automatically retrieve, extract, and re-analyze omics data from biomedical literature, thereby transforming static publications into a scalable, executable resource for automated data reuse and cross-study discovery.

The Gist

Imagine the world of biomedical research as a massive, chaotic library. Inside this library are millions of books (scientific articles) about how our bodies work at a microscopic level (omics data). The problem? Most of these books are written in a secret code, and the actual "ingredients" (the raw data) needed to cook the recipes inside are scattered in different rooms, hidden in footnotes, or locked in basements that no one knows how to open.

Because of this, even if a scientist publishes a groundbreaking discovery, other scientists can't easily check their work or use that data to make new discoveries. It's like finding a recipe for a perfect cake, but the list of ingredients is missing, or the instructions are written in a language you don't speak.

Enter the "Omics Data Discovery Agents."

Think of these agents as a team of super-intelligent, tireless librarians and chefs working together. They are powered by advanced AI (Large Language Models) and are designed to solve the library's chaos. Here is how they work, using some simple metaphors:

1. The Detective Librarian (Finding the Clues)

First, the agents act like detectives. They scan thousands of scientific articles. Instead of just reading the title, they dive deep into the text, the footnotes, and the "supplementary files" (which are like the hidden back pages of a book).

• What they do: They look for clues like "We used this specific machine," "The data is stored in this online vault," or "Here is the code we used."
• The Result: They turn a messy, unstructured book into a neat, organized card in a database. They can tell you exactly where the ingredients are, even if the author didn't put them in a standard spot.

2. The Master Chef (Recooking the Dish)

Once the agents find the "ingredients" (the raw data), they don't just leave them there. They act as master chefs who can re-cook the dish from scratch.

• The Challenge: Sometimes the original recipe says, "Add a pinch of salt," but doesn't say which salt or how much.
• The Solution: The agents read the article carefully to figure out the exact settings (like the type of machine used or the software version). They then use a special "kitchen" (a secure, containerized computer environment) to run the analysis again.
• The Proof: When they re-cooked a dish from a famous paper, the taste (the results) was 63% identical to the original. This proves they can follow the instructions well enough to get the same outcome.

3. The Matchmaker (Connecting the Dots)

This is the most magical part. Imagine you have three different books about liver disease written by different authors in different countries. A human might miss the connection because the books look different.

• The Agent's Superpower: The agents can read all three books, understand the meaning behind the words (not just the keywords), and realize, "Hey, these three stories are actually talking about the same problem!"
• The Discovery: They combined data from these three different studies and found a consistent pattern: certain proteins were behaving the same way in liver fibrosis across different species (mice and humans). This is a discovery that would have been incredibly hard for a human to spot manually because the data was buried in different formats.

Why Does This Matter?

Currently, reusing scientific data is like trying to build a house using bricks that are scattered across a field, with no instructions on how to stack them. It takes months of manual work.

This new system turns the library into a smart, searchable, and executable resource.

• Before: "I hope someone else has the data I need, and I hope they remember how they analyzed it."
• After: "Ask the AI agent to find all liver fibrosis studies, download the data, re-analyze it using the original methods, and tell me what the common patterns are."

The Bottom Line

This paper introduces a system that turns static, dusty scientific papers into living, breathing, and reusable tools. It automates the boring, difficult work of finding and cleaning data, allowing scientists to focus on the big questions: What does this mean for human health?

It's like upgrading from a library where you have to manually search every shelf for a single sentence, to a library where a robot can instantly find every relevant sentence, cook the data into a new format, and serve you a fresh cup of coffee (or in this case, a new scientific discovery) on demand.

Technical Summary

Here is a detailed technical summary of the paper "Omics Data Discovery Agents" by Hutton and Meyer.

1. Problem Statement

The biomedical literature contains a vast and growing collection of omics studies (proteomics, transcriptomics, etc.), yet the underlying data remains largely functionally inaccessible for computational reuse.

• Data Fragmentation: While raw data is increasingly deposited in public repositories (e.g., PRIDE, MassIVE, GEO), essential information required to reproduce results (processing parameters, code, sample annotations) is scattered across unstructured main text, supplementary files, and separate code repositories.
• Reproducibility Barriers: Reprocessing raw data requires significant domain knowledge to configure quantification tools correctly. Without standardized distribution of quantitative data, manual reanalysis is often impractical.
• Data Decay: Data availability declines over time; retrieval success rates drop by ~17% annually as articles age due to personnel turnover or storage limits.
• Limitations of Current Tools: Existing automated curation relies on static metadata (keywords) or AI assistants that summarize text but cannot autonomously locate, download, reprocess, and integrate raw datasets.

2. Methodology: The Agentic Framework

The authors propose an agentic framework that transforms unstructured publications into executable, queryable research objects. The system consists of three core components:

A. System Architecture

1. Article Ingestion & Metadata Extraction:

   • Retrieves full-text articles from PubMed Central (PMC) Open Access.
   • Uses a separate Large Language Model (LLM) (e.g., GPT-5) to extract unstructured metadata (keywords, data locations, processing parameters) from text and supplements.
   • Security Design: To prevent prompt injection, the text processing LLM is isolated from the main agent. It returns JSON-formatted responses which are processed by safe functions before the main agent sees the data.
   • Generates text embeddings for semantic similarity search.

2. Data Retrieval & Quantification (via Model Context Protocol - MCP):

   • MCP Servers: Analytical tools are exposed to agents via MCP servers, preventing agents from writing custom code for every task and ensuring reproducibility through containerization.
   • Tool Categories:
     • Data Retrieval: Downloads raw data from PRIDE, MassIVE, GEO.
     • Pipeline Selection: Identifies appropriate quantification tools (e.g., DIA-NN for DIA, MaxQuant for DDA) based on article descriptions.
     • Execution: Launches containerized pipelines (Apptainer) with parameters extracted from the article.
   • Workflow: The agent parses the article to extract parameters (enzyme specificity, mass tolerances, software versions), generates configuration files, executes the pipeline, and validates output dimensions against reported values.

3. Cross-Study Reasoning:

   • Agents identify semantically similar studies using cosine similarity of abstract embeddings.
   • They assess data compatibility (e.g., distinguishing between liver cancer vs. fatty liver studies).
   • Agents synthesize findings by re-quantifying raw data where necessary and performing cross-study differential expression analysis.

3. Key Contributions

• Automated Extraction from Unstructured Text: Moving beyond static metadata harvesting to extract processing parameters and data locations directly from full-text and supplementary materials.
• Executable Literature: Transforming static papers into "research objects" where raw data can be automatically re-quantified using parameters derived from the publication.
• MCP-Driven Reproducibility: Utilizing Model Context Protocol servers to expose containerized analysis tools, ensuring that agents execute standardized, auditable pipelines rather than ad-hoc scripts.
• Semantic Cross-Study Integration: Demonstrating the ability to identify compatible datasets across different publications and species, performing meta-analyses that were previously impossible at scale.

4. Results

The system was evaluated on 4,210 articles (focusing on proteomics and liver fibrosis) and a benchmark set of 39 proteomics publications.

• Metadata Extraction:
  • Achieved 80% precision in identifying standard repository links (PRIDE, MassIVE, GEO).
  • In the benchmark set, the system achieved 0.91 precision and 0.89 recall for dataset identification (excluding ambiguous cases).
• Automated Reanalysis (Quantification):
  • Successfully extracted parameters for a DIA proteomics study (Taneera et al.) and re-quantified the data.
  • Version Control: The agent correctly identified software version discrepancies (e.g., DIA-NN v1.8.1 vs. v2.3.1) and adjusted parameters accordingly.
  • Validation: When matching the article's preprocessing steps, the reanalysis achieved a 63% overlap in differentially expressed proteins (DEPs) compared to the original study.
• Cross-Study Synthesis:
  • The agent successfully identified three semantically similar studies on liver fibrosis (Cheng et al., Jirouskova et al., Devos et al.).
  • It re-quantified raw data, compared results across species (mouse vs. human), and identified 6 consistently upregulated proteins (including CLU, TGFBI, AMBP) across all three studies, despite these proteins not being the primary focus of the original articles.
  • The system correctly refused to integrate incompatible datasets (e.g., liver cancer vs. fatty liver).

5. Significance and Future Impact

• Scalable Data Reuse: This framework establishes a foundation for converting the static biomedical literature into a dynamic, executable resource, enabling automated data reuse at a scale impossible for human curators.
• Enhanced Reproducibility: By automating the retrieval of raw data and the execution of specific quantification pipelines, the system reduces the "black box" nature of omics analysis and allows for independent verification of results.
• New Discovery Pathways: The ability to perform cross-study meta-analyses on re-quantified data reveals consistent biological patterns (e.g., protein regulation in fibrosis) that are often missed when relying solely on the conclusions reported in individual papers.
• Security Considerations: The authors acknowledge risks regarding prompt injection (e.g., malicious instructions hidden in supplementary files) and propose a "separation of concerns" architecture where the agent does not directly execute code generated from potentially hostile text, mitigating security risks while maintaining functionality.

In conclusion, this work demonstrates that LLM agents, when coupled with containerized tools and structured protocols (MCP), can autonomously navigate the complexities of omics data discovery, extraction, and reanalysis, effectively bridging the gap between published literature and computable data.