Unsupervised Biological Significance Mapping via Evo 2 & Sparse Autoencoders

PlatyGeno identifies genomic landmarks directly from raw sequence data. By leveraging the Evo 2 foundation model, it isolates biologically significant DNA structures (promoters, coding sequences, precise motifs) based purely on AI confidence—without requiring labels, databases, or BLAST.

🧪 Technical Validation (IBD-MDB)

PlatyGeno v1.0.4 has been validated using the clinical IBD Metagenomic Database dataset. For complete statistical data and methodology, see the PlatyGeno Technical Audit and the PlatyGeno Technical Supplemental.

Summary of Results:

• Novel Genomic Landmarks: Identified Feature 7393, a 101bp sequence with no prior database matches and a high-confidence structural model (AlphaFold2 best prediction; pLDDT ≈ 80).

  
  Representative best structural prediction (AlphaFold2) of the novel Feature 7393 discovered autonomously by PlatyGeno.

• Statistical Correlation: Verified a Pearson correlation of r = 0.84 (p < 10^-50) between sequence length and match significance.

• Resolution Gain: Consensus assembly provided a 10³⁸ increase in E-value confidence over isolated 60bp fragments.

• Taxonomic Profile: 72% of high-activation discoveries successfully cross-validated with target gut microbiota.

🏗️ Technical Foundation

PlatyGeno operates as a Reference-Free Microscope, detecting the "Signal" of life directly from genomic grammar.

🔭 The Discovery Core

• AI-Native Interpretation: We use a Sparse Autoencoder (SAE) to translate the complex DNA "grammar" understood by Evo 2 into 32,768 human-interpretable biological concepts (e.g., promoters, viral motifs).
• Peak Pinpointing (Layer 26): The engine intercepts signals at Layer 26 to identify the exact coordinate where a biological feature fires with the highest intensity.
• Dual-Mode Discovery: Preserves both narrow Precision Snippets (separatedly high-interest DNA clips) and Consensus Assemblies (overlapping sequences from multiple reads of the same feature pieced together ).

👉 For a full hierarchical deep-dive into the methodology and validation trail, see Technical Architecture.

⚙️ Setup & Installation

⚙️ Installation & Quick Start

PlatyGeno requires a CUDA-enabled GPU (RTX 3090, 4090, A100, or H100).

# 1. Install the core package
pip install platygeno

# 2. Install high-performance GPU kernels (Mandatory for speed)
pip install ninja # for faster installation of flash-attn
pip install flash-attn --no-build-isolation

# 3. Verify & Run Discovery (on the validation sample)
platygeno --input data/sample.fastq --limit 5000 --threshold 5.0

🚀 Quick Start for GitHub Clones

# 1. Clone & Enter
git clone https://github.com/khoatran1995/PlatyGeno.git
cd PlatyGeno

# 2. Install High-Performance Kernels & editable package
pip install flash-attn --no-build-isolation
pip install -e .

# 3. Trigger Discovery
platygeno --input data/sample.fastq --limit 5000

📚 Documentation

API Reference: Details on Evo 2 integrations and technical Python parameters.

🚀 Usage & API Reference

🚀 Advanced Python Discovery

Researchers can integrate the engine into custom discovery pipelines:

import platygeno

# Advanced Discovery: Tuning parameters for clinical audits
results = platygeno.discover_genes(
    input_path="data/sample.fastq",
    scan_end=5000,
    min_activation=8.0,      # High-confidence threshold
    batch_size=32            # GPU-optimized batching
)

# View discovered biological features
print(results[['feature_id', 'feature_name', 'activation', 'sequence']])

platygeno.discover_genes() Reference

⚡ Performance

📜 References

⚠️ Technical Limitations

• Pre-training Bias: Sensitivity depends on the Evo 2 pre-training corpus.
• SAE Bottleneck: Discrete compression may miss extremely subtle biological nuances.
• Validation Requirement: High significance is a "Beacon," not final functional proof.

📜 References

@software{PlatyGeno2026,
  author = {Khoa Tu Tran},
  title = {PlatyGeno: Unsupervised Significance Mapping via Evo 2},
  url = {https://github.com/khoatran1995/PlatyGeno},
  doi = {10.5281/zenodo.19581708},
  year = {2026}
}

Thanks to Together AI (Evo 2) and Goodfire AI (SAE interpretability) and the IBD Metagenomic Database (IBD-MDB). Please cite the relevant references when using PlatyGeno.