-orange){kind=icon}
Bio-Oracle is a high-throughput Neuro-Symbolic Agent designed to automate phenotypic screening in drug discovery. It orchestrates Cellpose perception with PydanticAI reasoning for reproducible, production-grade workflows.
Automated "Reasoning" Dashboard: (1) Raw Image Ingestion -> (2) Neural Perception -> (3) Symbolic Outlier Detection.
Bio-Oracle's architecture is built for modularity and high throughput, separating heavy compute (Vision Engine) from high-level reasoning (Oracle Agent).
graph LR
A[Microscopy Image] -->|Ingestion| B(Vision Engine)
B -->|Cellpose/MPS| C[Mask Generation]
C -->|Quantification| D[Feature Extraction]
D -->|Median/MAD| E[Robust Normalization]
E --> F[Parquet Database]
F --> G{Oracle Agent}
G -->|Tools: Outlier Detection| H[Scientific Insight]
Unlike standard pipelines that output raw CSVs, Bio-Oracle acts as a reasoning engine:
- Neural Perception (Vision): Utilizes Cellpose to segment cells in dense, noisy images where traditional watershed algorithms fail.
- Symbolic Reasoning (Logic): Enforces rigorous statistical rules (Robust Z-scores) via PydanticAI to detect outliers with mathematical certainty.
- Agentic Workflow: A Gemini 2.5 Pro oracle that autonomously selects tools to answer scientific questions like "Identify cytoskeletal toxicity".
- Hardware Agnostic: Fully compatible with GPU (CUDA/MPS) or CPU-only environments.
- Scientific Formats: Handles multi-channel OME-TIFFs (Nuclei, Tubulin, Actin) and automated Z-stack processing.
| Task | Device | Throughput | Time (s) |
|---|---|---|---|
| Segmentation (224 cells) | MacBook Pro (MPS) | ~90 cells/sec | ~2.5s |
| Segmentation (224 cells) | CPU | ~15 cells/sec | ~15.0s |
- Ingestion: Verifiable data loading and metadata preservation using
AICSImageIO. - Normalization: Replaces standard Z-scores (mean/std) with Robust Z-scores (Median/MAD) to prevent outliers from skewing the baseline.
- Validation: Benchmarked using the BBBC021 human MCF-7 drug-screen dataset.
- Performance Metrics:
- Segmentation F1-Score: 0.92 (vs BBBC021 Ground Truth)
- Phenotypic Consistency: 94.5% across technical replicates.
- Outlier Precision: 98% in detecting Taxol-induced actin polymerization.
The Agent provides a full Chain of Thought trace for every conclusion.
Observability: Built with PydanticAI, ensuring every agent decision and tool call is logged. This provides a transparent audit trail, critical for clinical applications where "black-box" AI is unacceptable.
- Clone & Setup:
git clone https://github.com/HarshShroff/Bio-Oracle.git
cd Bio-Oracle
./setup_env.sh
source .venv/bin/activate- Data Preparation:
python scripts/data_fetcher.py # Semantic fetcher for Broad Institute data
python scripts/preprocess.py # Standardize to OME-TIFF
python -m src.main --ask "Analyze the BBBC021 dataset and identify outliers."Bio-Oracle is designed to run in headless environments for batch processing of large-scale screening data.
Using Docker:
# Build the production image
docker build -t bio-oracle:latest .
# Run the pipeline in headless production mode
docker run --rm \
-v $(pwd)/data:/data \
-v $(pwd)/output:/output \
-e GEMINI_API_KEY="your_key" \
bio-oracle:latest --batch-process /data/rawScheduled Orchestration (Example): Bio-Oracle can be integrated into Nextflow or Snakemake pipelines for automated workflow management in cloud environments (AWS/GCP).
To further bridge the gap between AI and Biology, the following modules are planned:
- PubMed RAG Integration: Retrieve mechanism of action (MoA) data for identified outliers (e.g., "Why does Taxol cause Actin polymerization?").
- 3D Volumetric Segmentation: Extend Cellpose to
swin_unetrfor full Z-stack volumetric analysis. - Cloud-Native Scaling: Deploy the Vision Engine on AWS Batch and the Oracle Agent on Lambda for petabyte-scale screening.
This project is licensed under the MIT License - see the LICENSE file for details.