GENERanno: A Genomic Foundation Model for Metagenomic Annotation

<p align="center">
  <picture>
    <img alt="Gener" src="figures/logo.png" width=50%>
  </picture>
</p>

<h1 align="center">GENERanno: A Genomic Foundation Model for Metagenomic Annotation</h1>

## 📰 News

* 🤗 **[2026-02-10]** Our expert model for eukaryotic genome annotation `GENERanno-eukaryote-1.2b-cds-annotator-preview` is now available on [HuggingFace](https://huggingface.co/GenerTeam/GENERanno-eukaryote-1.2b-cds-annotator-preview)!
* 📑 **[2025-06-05]** Our paper is now available on [bioRxiv](https://www.biorxiv.org/content/10.1101/2025.06.04.656517v1)!
* 🤗 **[2025-05-10]** Our expert model for metagenomic annotation `GENERanno-prokaryote-0.5b-cds-annotator` is now available on [HuggingFace](https://huggingface.co/GenerTeam/GENERanno-prokaryote-0.5b-cds-annotator)!
* 🤗 **[2025-02-11]** Our models `GENERanno-prokaryote-0.5b-base`,
  `GENERanno-eukaryote-0.5b-base` are now available on [HuggingFace](https://huggingface.co/GenerTeam/)!

## 🔭 Overview

<div align="center">
  <img src="figures/model_overview.png" />
  <img src="figures/eukaryotic_annotation.png" width="80%" />
  <img src="figures/bacteria_annotation_summary.png" />
</div>

In this repository, we present GENERanno, a genomic foundation model featuring a context length of 8k base pairs and
500M parameters, trained on an expansive dataset comprising 386 billion base pairs of eukaryotic DNA. Our evaluations
demonstrate that the GENERanno achieves comparable performance
with [GENERator](https://huggingface.co/GenerTeam/GENERator-eukaryote-1.2b-base) in benchmark evaluations,
including [Genomic Benchmarks](https://huggingface.co/datasets/katielink/genomic-benchmarks/tree/main), [NT tasks](https://huggingface.co/datasets/InstaDeepAI/nucleotide_transformer_downstream_tasks_revised),
and our newly proposed [Gener tasks](https://huggingface.co/GenerTeam), making them the top genomic foundation models in
the field (2025-02).

Beyond benchmark performance, the GENERanno model is meticulously designed with its specialization in gene annotation.
The model efficiently and accurately identifies gene locations, predicts gene function, and annotates gene structure,
highlighting its potential to revolutionize genomic research by significantly enhancing the precision and efficiency of
gene annotation processes.

Please note that the GENERanno is currently in the developmental phase. We are actively refining the model and will
release more technical details soon. Stay tuned for updates!

In this repository, you will find the following model checkpoints:

| Model Name                       | Parameters | Data |     Category     |                    Status                     |
|----------------------------------|:----------:|:----:|:----------------:|:---------------------------------------------:|
| `GENERanno-prokaryote-0.5b-base` |    0.5B    | 715B |    Prokaryote    | [Available](https://huggingface.co/GenerTeam/GENERanno-prokaryote-0.5b-prokaryote-base) |
| `GENERanno-prokaryote-0.5b-cds-annotator` |     0.5B     | 890B |    Prokaryote    |[Available](https://huggingface.co/GenerTeam/GENERanno-prokaryote-0.5b-cds-annotator)|
| `GENERanno-eukaryote-0.5b-base`  |    0.5B    | 386B |    Eukaryote     | [Available](https://huggingface.co/GenerTeam/GENERanno-prokaryote-0.5b-eukaryote-base) |
| `GENERanno-eukaryote-1.2b-cds-annotator-preview` |     1.2B     | 1T |    Eukaryote    |[Available](https://huggingface.co/GenerTeam/GENERanno-eukaryote-1.2b-cds-annotator-preview)|

## 🎯 Quick Start

### Dependencies

* Clone this repo, cd into it

```shell
git clone https://github.com/GenerTeam/GENERanno.git
cd GENERanno
```

* Install requirements with Python 3.10

```shell
pip install -r requirements.txt
```

> If your network cannot access huggingface.co normally, we recommend using the following mirror:
> ```shell
> export HF_ENDPOINT=https://hf-mirror.com
> ```

### Downstream

#### Coding DNA Sequence (CDS) Annotation

You can run CDS annotation on the [cds-annotation dataset](https://huggingface.co/datasets/GenerTeam/cds-annotation/) using the unified CLI interface below.

##### Basic usage

```bash
# Eukaryotic genome annotation
python src/tasks/downstream/cds_annotation.py --organism eukaryote

# Prokaryotic genome annotation
python src/tasks/downstream/cds_annotation.py --organism prokaryote

# Enable BF16 for faster inference (recommended if supported)
python src/tasks/downstream/cds_annotation.py --organism eukaryote --bf16
```

##### Custom input

By default, each `--organism` preset uses a built-in example input.
You can override it with your own FASTA or Parquet file:

```bash
# Parquet input
python src/tasks/downstream/cds_annotation.py \
  --organism eukaryote \
  --input hf://datasets/GenerTeam/cds-annotation/examples/fly_GCF_000001215.4.parquet

# FASTA input
python src/tasks/downstream/cds_annotation.py \
  --organism prokaryote \
  --input hf://datasets/GenerTeam/cds-annotation/examples/Escherichia_coli_genome.fasta
```

##### Performance options

```bash
# Use all available GPUs (default)
python src/tasks/downstream/cds_annotation.py --organism eukaryote

# Use a specific number of GPUs
python src/tasks/downstream/cds_annotation.py --organism eukaryote --gpu_count ${NUM_GPUS}

# Enable BF16 for faster inference (recommended if supported)
python src/tasks/downstream/cds_annotation.py --organism eukaryote --bf16
```
Note: BF16 improves inference speed on supported hardware (e.g. A100) with minimal impact on accuracy.

#### Sequence Understanding (Classification/Regression)

To run the sequence understanding task
on [Gener Tasks](https://huggingface.co/datasets/GenerTeam/gener-tasks), [Prokaryotic Gener Tasks](https://huggingface.co/datasets/GenerTeam/prokaryotic-gener-tasks), [NT Tasks](https://huggingface.co/datasets/InstaDeepAI/nucleotide_transformer_downstream_tasks_revised), [Genomic Benchmarks](https://huggingface.co/katarinagresova), [DeepSTARR Enhancer](https://huggingface.co/datasets/GenerTeam/DeepSTARR-enhancer-activity),
you can use the following arguments:

* Gener Tasks / Prokaryotic Gener Tasks
    * `--dataset_name GenerTeam/gener-tasks` or `--dataset_name GenerTeam/prokaryotic-gener-tasks`
    * `--subset_name gene_classification` or `--subset_name taxonomic_classification` or ...
* NT Tasks
    * `--dataset_name InstaDeepAI/nucleotide_transformer_downstream_tasks_revised`
    * `--subset_name H2AFZ` or `--subset_name H3K27ac` or ...
* Genomic Benchmarks
    * `--dataset_name katarinagresova/Genomic_Benchmarks_demo_human_or_worm` or
      `--dataset_name katarinagresova/Genomic_Benchmarks_human_ocr_ensembl` or ...
* DeepSTARR Enhancer Activity
    * `--dataset_name GenerTeam/DeepSTARR-enhancer-activity`
    * `--problem_type regression`

on following command:

```shell
# Using single GPU
python src/tasks/downstream/sequence_understanding.py \
    --model_name GenerTeam/GENERator-eukaryote-1.2b-base \
    --dataset_name ${DATASET_NAME} \
    --subset_name ${SUBSET_NAME} \
    --batch_size ${BATCH_SIZE} \
    --problem_type ${PROBLEM_TYPE} \
    --main_metrics ${MAIN_METRICS}

# Using multiple GPUs on single node (DDP)
torchrun --nnodes=1 \
    --nproc_per_node=${NUM_GPUS} \
    --rdzv_backend=c10d \
    src/tasks/downstream/sequence_understanding.py

# Using multiple GPUs on multiple nodes (DDP)
torchrun --nnodes=${NUM_NODES} \
    --nproc_per_node=${NUM_GPUS_PER_NODE} \
    --rdzv_backend=c10d \
    --rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
    src/tasks/downstream/sequence_understanding.py

# Using DeepSpeed or Full Sharded Data Parallel (FSDP)
torchrun --nnodes=${NUM_NODES} \
    --nproc_per_node=${NUM_GPUS_PER_NODE} \
    --rdzv_backend=c10d \
    --rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
    src/tasks/downstream/sequence_understanding.py \
    --distributed_type deepspeed # or fsdp
```

## 📚 Datasets

* [Eukaryotic Gener Tasks](https://huggingface.co/datasets/GenerTeam/gener-tasks)
* [Prokaryotic Gener Tasks](https://huggingface.co/datasets/GenerTeam/prokaryotic-gener-tasks)
* [CDS Annotation](https://huggingface.co/datasets/GenerTeam/cds-annotation)

## 📜 Citation

```
@article{li2025generanno,
	author = {Li, Qiuyi and Wu, Wei and Zhu, Yiheng and Feng, Fuli and Ye, Jieping and Wang, Zheng},
	title = {GENERanno: A Genomic Foundation Model for Metagenomic Annotation},
	elocation-id = {2025.06.04.656517},
	year = {2025},
	doi = {10.1101/2025.06.04.656517},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/06/05/2025.06.04.656517},
	journal = {bioRxiv}
}
```

## 📈 Benchmark Performance

### Sequence Understanding (Classification/Regression) — `GENERanno-prokaryote-0.5b-base`
![benchmarks](figures/prokaryotic_gener_tasks.png)

### Sequence Understanding (Classification/Regression) — `GENERanno-eukaryote-0.5b-base`
![benchmarks](figures/eukaryotic_benchmarks.png)