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.

---

🔬 PlatyGeno Discovery Engine Workflow

From raw unorganized DNA sequences to confident biological landmarks.

---

🧪 Technical Validation (IBD-MDB)

PlatyGeno v1.0.4 has been validated using a small part of 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 Structural DNA Landmark: Identified Feature 7393 (101bp), a reference-free discovery with no prior database matches. While AlphaFold2 yielded high structural modeling confidence (pLDDT ≈ 80), integrated validation (CPC2/RNAfold) classifies it as a novel non-coding structural landmark.

  
  AlphaFold2 structural prediction of Feature 7393. While yielding a high pLDDT score (≈ 80), integrated analysis revealed this to be a mathematical consequence of its high internal periodicity, confirming it as a structural DNA landmark rather than a protein.

• 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 ).

Important

Performance Highlight: While both modes are preserved in discovery, validation benchmarks confirm that Consensus Assembly yields statistically superior significance (E-values) and cleaner taxonomic resolution.

👉 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 - Quick Start

🚀 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

Parameter	Type	Default	Description
input_path	str	Req	Path to sequence file.
min_activation	float	5.0	Minimum signal strength.
rel_freq_max	float	1.0	Rarity cap (1.0 = All significance).
scan_end	int	None	Last read index (None for end of file).
top_n	int	-1	Max features to return (-1 for ALL).

---

⚠️ 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.

Software available under Apache 2.0 License; Documentation available under CC BY 4.0.