Orthoformer Datasets

📌 Overview

Orthoformer is a large-scale genomics dataset designed to support function-centric foundation modeling of microbial and viral genomes.

Unlike conventional sequence-based models that infer biological roles from nucleotide or protein context, Orthoformer represents each genome by its orthologous group composition and abundance, treating functional units rather than sequences as the basic biological vocabulary.

The dataset was constructed from approximately three million prokaryotic and viral genomes, each encoded as a high-dimensional functional profile capturing:

• biochemical identity
• gene family dynamics
• evolutionary conservation
• pathway-scale metabolic capacity

These representations enable learning a functional embedding space that provides an alignment-free measure of genomic similarity, supporting robust taxonomy, phylogenetic analysis, and detection of functional convergence across microbial lineages.

The same functional embeddings generalize beyond evolutionary structure to predict:

• biosynthetic gene cluster abundance
• ecological niche differentiation
• organism-level phenotypes
• marker-gene associations

Together, the Orthoformer dataset establishes a function-first framework for microbial genomics, offering a scalable alternative to sequence-centric datasets for studying microbial evolution, function, and ecology.

---

📂 Dataset Structure

The repository is organized into three functional splits:

Split	Purpose	Description
foundation_model_dataset	Pretraining	Large-scale collection of microbial and viral genomes represented by orthologous group composition and abundance, used for self-supervised function-centric foundation model training
Downstream_Tasks_dataset	Fine-tuning	Labeled genome-level datasets for functional, metabolic, and ecological prediction tasks, including niche, biosynthetic capacity, and phenotype inference
Orthoformer_eval_dataset	Evaluation	Benchmark datasets for evaluating functional embeddings on taxonomy, phylogeny, functional convergence, and biological consistency

---

📊 Dataset Statistics

Split	Size	Max Sequence Length
foundation_model_dataset	~3M sequences	2048
Downstream_Tasks_dataset	Task dependent	Task dependent
Orthoformer_eval_dataset	Benchmarks	Task dependent

---

🚀 Accessing the Dataset

You can download the dataset via Hugging Face using Git + Xet (recommended for large files):

# Install git-xet (Linux)
curl -sSf https://raw.githubusercontent.com/huggingface/xet-core/main/git_xet/install.sh | sh
git xet install

# Clone the dataset
git clone https://huggingface.co/datasets/jackkuo/Orthoformer

If you only want the metadata without large files:

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/jackkuo/Orthoformer

---

🧬 Intended Use

The Orthoformer dataset is intended for training and evaluating foundation models for microbial functional genomics, including:

• Functional embedding of microbial and viral genomes
• Alignment-free phylogeny and taxonomy
• Functional convergence and evolutionary analysis
• Metabolic and pathway-level phenotype prediction
• Biosynthetic gene cluster and ecological niche inference

It is specifically designed for function-centric modeling, where orthologous gene groups are treated as the fundamental representation units instead of raw nucleotide or amino acid sequences.

---

📜 License

This dataset is released under the MIT License.

---

📖 Citation

If you use this dataset, please cite:

@dataset{xxx,
  title = {Orthoformer: xxx},
  author = {xxx},
  year = {2025},
}

---

🔗 Related Resources

• Model: https://huggingface.co/jackkuo/Orthoformer
• Code: https://github.com/JackKuo666/Orthoformer

---