platygeno 1.0.2

Unsupervised Gene Discovery via Evo 2 & SAE Interpretability

Unsupervised Biological Significance Mapping via Evo 2 & Sparse Autoencoders

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

---

🏗️ Technical Foundation

🔭 Scientific Philosophy: Zero-Reference Significance

PlatyGeno is a Reference-Free Microscope that detects the "Signal" of life directly from raw DNA, bypassing the need for sequence libraries or databases:

• Signals over Samples: Detects functional peaks in DNA grammar directly. If a sequence is biologically significant, the AI finds it—even if it has never been cataloged.
• Significance First: Prioritizes Activation Strength (AI "excitation") as a primary beacon for functional mapping.
• Optional Novelty Mining: Isolate "Genomic Dark Matter" (novel viruses or enzymes) by optionally filtering for rare landmarks.

🏗️ Simplified Architecture

PlatyGeno layers a "De-coding" layer on top of the Evo 2 foundation model:

1. Evo 2 (The Brain): A 7B parameter foundation model by Together AI that understands the genomic grammar of all sequenced life.
2. Sparse Autoencoders (The Interpreter): We utilize Goodfire's Sparse Autoencoders (specifically the Layer-26-Mixed expansion) to translate dense AI math into 32,768 discrete, human-interpretable biological concepts.
3. Landmark Scouter: Scans raw data to find the precise coordinates where these concept nodes fire with the highest intensity.

---

⚙️ 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)
# Automatically saves to: results/sample_Significance.csv
platygeno --input data/sample.fastq --limit 5000 --threshold 5.0

🚀 Quick Start for GitHub Clones

If you are cloning the repository for research or development, follow these three steps to run your first discovery:

# 1. Clone & Enter
git clone https://github.com/khoatran1995/PlatyGeno.git
cd 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. Install in Editable Mode
pip install -e .

# 4. Trigger Discovery (on the validation sample)
platygeno --input data/sample.fastq --limit 5000

📚 Documentation & Reference

• Technical Documentation: Deep dive into Evo 2, Sparse Autoencoders, and technical API Reference.
• Technical Architecture: Details on Mean-Pooling, Zero-Gate Discovery, and the Padding Filter.
• Validation Methodology: Detailed audit trail for clinical gene discovery.

---

🚀 Usage & API Reference

🚀 Advanced Python Discovery

Researchers can integrate the engine into custom discovery pipelines using the Python API:

import platygeno

# Advanced Discovery: Tuning parameters for a custom GPU environment
results = platygeno.discover_genes(
    input_path="data/sample.fastq",
    scan_start=0,
    scan_end=5000,
    min_activation=8.0,      # High-confidence threshold
    batch_size=32            # High-performance 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, Default).

⚙️ Hardware Optimization

PlatyGeno is optimized for high-performance discovery through its dedicated Batched Inference engine.

Batch Size Guide (--batch-size)

Parallelizing your scan is the fastest way to get results. Match this setting to your GPU VRAM:

Hardware	VRAM	Recommended Batch Size
A100 / H100	80GB	32 – 64
RTX 3090 / 4090	24GB	8 – 16
RTX 3060 / 4070	12GB	1 – 2

Out of Memory? If you encounter an OOM error, simply lower the --batch-size.

---

⚡ Performance

The following benchmarks reflect the Standard 20k Read Survey (Clinical Gut Metagenome) using the optimized PlatyGeno v1.0.2 engine.

Mode	Engine Implementation	Runtime (20k Reads)	Discovery Speed
v1.0.2 (Current)	Batched Mean-Pooling	~4.8 Minutes	🚀 100% (High Speed)

Benchmarks conducted on an NVIDIA RTX 4090 (24GB VRAM). Performance scales linearly with GPU memory and batch size.

---

🧪 Clinical Validation Case Study

🚀 Complete Validation Suite

Researchers can choose to manually run the full discovery-to-validation pipeline for detailed clinical audits:

One-Line Discovery Pipeline: python validation/discovery_pipeline.py --input data/sample.fastq

Manual Step-by-Step Walkthrough:

1. Discovery: python validation/step1_discovery.py --input data/sample.fastq --limit 5000 --threshold 8.0 (Parameters: Change --limit or --threshold to explore different sensitivity levels)

2. Significance Audit: Result saved to results/PLG_sample_Significance.csv.

3. Viral/Gene Validation: python validation/step2_blast.py --input results/PLG_sample_Significance.csv --threads 10 (Parameters: Increase --threads for faster NCBI validation)

4. Structural Check: Push novel sequences to AlphaFold 2 for 3D modeling: 👉 AlphaFold 2 (ColabFold) Online

💡 For more command-line arguments, refer to the Technical Documentation or run platygeno --help.

🧪 Validation data (IBD-MDB)

PlatyGeno includes a clinical validation set (data/sample.fastq) from the IBD Metagenomic Database. This enables researchers to verify the engine's ability to identify autonomous biological landmarks in high-complexity clinical samples with zero-reference databases.

🧪 Initial Use Cases

• General Genomic Research: Map functional landmarks (promoters, coding sequences, motifs) to identify the biological identity of any sequenced sample.
• Novel Gene Discovery: Directly target "Genomic Dark Matter" and novel viruses with reference-free significance mapping.
• Advanced Discovery Pipelines: Build automated workflows that bridge AI detection with structural modeling (AlphaFold) for high-fidelity validation.

---

⚠️ Technical Limitations & Scope

As an AI-native discovery tool, PlatyGeno’s insights are subject to several technical boundaries:

• Pre-training Bias: PlatyGeno relies on the Evo 2 foundation model. If specific genomic structures or rare taxa were significantly under-represented or excluded in the model’s pre-training corpus, the engine may demonstrate lower sensitivity for those regions.
• SAE Bottleneck: While Sparse Autoencoders provide human-interpretable "concepts," they represent a discrete compression of the 7B parameter model. Extremely subtle motifs or novel biological nuances may occasionally fall below the SAE activation threshold.
• Validation Requirement: A high significance score is a "Biological Beacon," but it is not a final proof of function. All discovery candidates should be cross-verified using structural tools (AlphaFold/ESMFold) and/or experimental assays.

---

📜 References

1. PlatyGeno (This Package):

@software{PlatyGeno2026,
  author = {Khoa Tu Tran},
  title = {PlatyGeno: Unsupervised Significance Mapping via Evo 2},
  url = {https://github.com/khoatran1995/PlatyGeno},
  doi = {10.5281/zenodo.19568630},
  year = {2026}
}

PlatyGeno is a product of ongoing research into AI-guided, reference-free metagenomic discovery. We extend our professional gratitude to Together AI for the Evo 2 foundation model and to Goodfire AI for their groundbreaking work on SAE-based interpretability. If your research utilizing this suite yields significant findings, we request that you also cite these foundational contributions as appropriate.