🧬 Gene Set Function Discovery with LLM-based Agents and Knowledge Retrieval

Discovera is an interactive, agent-based system that integrates traditional bioinformatics tools with large language models (LLMs) and retrieval-augmented generation (RAG) to support hypothesis generation and mechanistic discovery in functional genomics.

Presented at ICLR 2025 MLGenX Workshop

🔗 Read the Paper · 🖼️ View the Poster

What It Does?

Bridges the gap between computational analysis and interpretability in biomedical research. It is designed to assist researchers—regardless of their coding expertise—in:

• Interactively exploring gene sets associated with complex phenotypes
• Conducting functional enrichment analyses
• Summarizing mechanistic hypotheses using evidence from literature
• Formulating data-grounded biological insights

Key Features

• ⚙️ Modular System: Combines established tools (e.g., GSEApy, INDRA) with custom modules for extensibility.
• 📖 LLM Integration: Uses LLMs for natural language reasoning and explanation generation.
• 🔎 Retrieval-Augmented Generation: Grounds summaries in real literature to improve accuracy and transparency.
• 💬 Chat Interface: Enables intuitive, dialogue-based exploration of hypotheses and gene set functions.

Use Case Example

In our initial deployment, this agent was used in the context of endometrial carcinoma (EC) to:

• Analyze gene sets linked to phenotypic features
• Perform enrichment analysis on the resulting sets
• Summarize literature-supported mechanisms of action

System Architecture

The system consists of:

• Enrichment tools (e.g., GSEApy)
• INDRA for biological statement synthesis
• LLM-enabled prompt orchestration
• Chat-based user interface

Getting Started

Prerequisites

• Python 3.8+
• GSEApy
• INDRA
• OpenAI or similar LLM API access

Getting Started

Prerequisites

• Python 3.8+
• Docker & Docker Compose
• OpenAI API key (for LLM integration)

Installation and Setup

Clone the repository:

git clone https://github.com/VIDA-NYU/discovera.git
cd discovera

Configure Beaker

Copy the configuration template:

cp .beaker.conf.template .beaker.conf

Open .beaker.conf in a text editor and add your OpenAI API key, if using OpenAI:

api_key = "your-api-key-here"

This configuration will be used when launching Docker.

Running with Docker

To use Discovera with the Beaker context setup:

docker compose build
docker compose up -d

Once running, navigate to http://localhost:8888 and select discovera.

Example Workflow

Try the following steps to see how BKD-Agent assists in biomedical discovery:

1. Load Gene Expression Data

   • Load gene_expression.csv into a Pandas DataFrame.

2. Gene Set Enrichment Analysis (GSEA)

   • Use the following parameters:
     • gene_sets: GO_Biological_Process_2023
     • hit_column: hit
     • corr_column: corr
     • min_set: 5
     • max_set: 2000

3. Identify Lead Genes from Top Pathway

   • Extract lead genes from the most statistically significant pathway.

4. Refine Gene List Based on Correlation

   • Identify top 20 most correlated genes, excluding those already in the lead gene set.

5. Retrieve Documented Gene Relationships

   • Example gene pair: ["CTNNB1", "GLCE"]

6. Summarize Gene Pair Relationship Types and Frequencies

   • Example pairs:
     • ["CTNNB1", "GLCE"]
     • ["CTNNB1", "NOTUM"]

7. Construct Gene Network Graph

   • Nodes = genes, edges = literature-backed relationships weighted by frequency/strength

8. Extract Gene Relationship Excerpts

   • Retrieve textual evidence supporting the relationships.

9. Contextualize Gene Relationships in Disease (Endometrial Carcinoma)

   • Analyze how gene interactions relate to the disease context.
   • Use GSEA results, pathway data, and literature to generate hypotheses.
   • Ask:

     Can you summarize these excerpts in the context of prospective endometrial carcinoma?
     What hypotheses can be drawn from this analysis?

10. Suggest Future Research Directions

    • Propose:
      • Novel hypotheses
      • Drug targets or pathway interventions
      • Further bioinformatics studies (e.g., single-cell RNA-seq)

Extending BKD-Agent: Adding Tools

Currently, the agent has one tool: query_gene_pair defined in:

src/discovera/agent.py

To add more tools:

• Copy the format of query_gene_pair.
• Tools are registered with @tool from Archytas.
• Archytas expects arguments as string variable names, not DataFrames directly.
• For example, this will correctly pass a variable and a string:

query_gene_pair({{ dataset }}, target="{{ target }}", method="{{ method }}")

• The underlying code executed lives in:

procedures/python3/query_gene_pair.py

Prompt Modification

There are two main areas to adjust the agent's behavior:

1. Context Management

   • File: src/discovera/context.py
   • Modify the auto_context function to alter background knowledge or enumerate tools.

2. Prompt Customization

   • File: agent.py
   • The main system prompt is written in the BKDAgent docstring.

Use these to tweak how the agent interprets user queries, formats responses, and integrates with tools.