halima014/bioinformatics-scout

Bioinformatics Literature Scout

A multi-agent pipeline that automates literature discovery for bioinformatics and ML research. Given a research query, the system autonomously plans a search strategy, runs parallel literature searches across PubMed and ArXiv, and synthesizes findings into a structured research brief — replicating in minutes what would take a researcher hours of manual searching.

Built with the OpenAI Agents SDK and deployed on Hugging Face Spaces via Gradio.

How It Works

[image]

Why "Scout"?

A scout goes ahead to survey the terrain so you don't have to. This tool scouts the literature — it searches PubMed and ArXiv across multiple angles of your research question, then brings back a structured summary. You skip the hours of manual searching and go straight to reading the papers that matter.

Agent Roles

1. Planner Agent

• —Input: Your raw research query
• —Output: SearchPlan — 5 targeted sub-topic searches with reasoning
• —Model: GPT-4o-mini
• —Purpose: A broad query like "single-cell ATAC-seq transfer learning" is too vague for a single search. The planner decomposes it into precise queries covering methods, applications, datasets, benchmarks, and recent advances.

2. Search Agents (x5, parallel)

• —Input: One search term from the plan
• —Tools: pubmed_search, arxiv_search
• —Output: SearchResult — top papers with title, abstract, DOI, relevance score
• —Model: GPT-4o-mini
• —Purpose: Each agent searches both PubMed and ArXiv, then selects and scores the most relevant papers. Running 5 in parallel reduces total search time from ~50s to ~10s.

3. Synthesis Agent

• —Input: Original query + all collected papers
• —Output: Structured markdown research brief
• —Model: GPT-4o (stronger model for synthesis and reasoning)
• —Purpose: Reads all ~30 papers, identifies patterns, and writes an actionable brief with themes, gaps, and next steps.

Tools

The agents use two tools — Python functions that call real scientific APIs:

Tech Stack

Key Patterns

• —Parallel async agents — asyncio.create_task() + as_completed() for 5 concurrent searches
• —Structured output — output_type=SearchPlan forces typed, parseable responses via Pydantic
• —Tool use — @function_tool wrapping real scientific APIs (PubMed, ArXiv)
• —Sequential pipeline with parallel middle stage — plan → [search x5 parallel] → synthesize
• —Live status updates — async generator yields progress to Gradio UI at each stage

Setup

# Clone
git clone https://github.com/Mituvinci/bioinformatics-scout.git
cd bioinformatics-scout

# Install dependencies
uv sync

# Configure API keys
cp .env.example .env
# Edit .env and add:
#   OPENAI_API_KEY=sk-...
#   NCBI_EMAIL=your@email.com   (any email — no NCBI account needed)

# Run
uv run python src/app.py

Example

Query: "transformer models single cell genomics 2023 2024"

The pipeline will:

1. 1.Plan 5 sub-topic searches (e.g., "scBERT single cell language model", "attention mechanism gene expression")
2. 2.Search PubMed and ArXiv in parallel for each sub-topic
3. 3.Synthesize a brief including:
4. 4.Executive summary of the field
5. 5.Top papers by sub-topic with DOI links
6. 6.Emerging themes (e.g., foundation models, multi-omics integration)
7. 7.Research gaps
8. 8.Suggested follow-up queries

Deployment — Hugging Face Spaces

Deployed on Hugging Face Spaces (free CPU tier). No GPU needed — all compute happens on OpenAI's API.

API keys are stored as Hugging Face Secrets (Settings → Variables and Secrets), never in the code.

Author

Halima Akhter — PhD Student, Computer Science Specialization: ML, Deep Learning, Bioinformatics GitHub