MicroGenomer: A Foundation Model for Transferable Microbial Genome Representations Enabling Multi-scale Genomic Understanding and Ecophysiological Trait Prediction

Introduction

MicroGenomer is a foundation model for transferable microbial genome representations, enabling multi-scale genomic comprehension and ecophysiological trait prediction. It adopts a hierarchical training strategy that integrates large-scale genomic sequence pre-training (234.5 billion base pairs), domain-specific mid-training with the GTDB-curated marker gene set, and task-specific post-training. Surpassing existing gene-scale encoders, MicroGenomer generates robust embeddings at the whole-genome scale.

Schematic Diagram

Figure 1. Overview of MicroGenomer.

GPU Hardware Requirements

Since the model leverages FlashAttention for computational acceleration to enhance the efficiency of genomic representation extraction, we recommend the following hardware configurations to ensure optimal model performance:

• NVIDIA GPUs with Ampere architecture or newer are required (e.g., RTX 30/40 series consumer-grade GPUs, A10/A100/A800 professional-grade GPUs).
• A minimum of 16GB VRAM is recommended (32GB or higher for batch processing of genomic data).

Quick Start

Download the GitHub Repository

Download the GitHub repository and extract the files to a designated folder.

Data Description

The input data required for MicroGenomer model inference and genomic representation extraction is formatted as a comma-separated values (CSV) file, which exclusively contains gene sequences with annotated Coding Sequences (CDS) — the core genomic regions that encode functional proteins in microbial genomes. The input data format is as follows:

• genome_id: Unique identifier for the current genome
• aa_seq: DNA sequence corresponding to CDS in the current genome
• unique_id: Unique ID for identifying the DNA sequence

Example:

unique_id,aa_seq,genome_id
RS_GCF_013393365.1@TIGR00168,TTGCCGTCCGTAG...GTAGTAAATAA,RS_GCF_013393365.1
RS_GCF_013393365.1@TIGR00382,TTGGCAAAAGATA...TGAAATCGTAA,RS_GCF_013393365.1
...

Installation

# Create a Conda Python environment
conda create -n MicroGenomer python=3.10
conda activate MicroGenomer
git clone https://github.com/BGIResearch/MicroGenomer.git
cd MicroGenomer
pip install -r requirements.txt

Usage

Foundation Model Weights

• Download the weights for MicroGenomer.
• Place the MicroGenomer-470M and downstream_tasks folders in the weights/ directory.

Run Inference for Different Downstream Tasks

Six downstream tasks available: maximum growth rate, oxygen tolerance, salinity tolerance, optimal pH, optimal temperature and probiotic prediction. It also supports extracting embeddings.

bash run.sh \
    --input_path '/path/to/input/' \
    --output_dir '/path/to/output/' \
    [--task 'downstream_task'] \ # Options: test_growth, test_oxygen, test_salinity, test_pH, test_temperature, test_probiotic, extract_embed or none
    [--level 'level_of_probiotic']  \ # For test_probiotic task only. Options: family, genus.

• --input_path: Path to the input file/folder.
• --output_dir: Path for saving output files.
• --task: Downstream task option.
• --level: Level of probiotic prediction. Default is family.

The input path can be a single CSV file or a folder. If it is a folder, all CSV files in that folder will be processed in batches. If the task parameter is extract_embed or is empty, the model will only extract embedding and will not perform any specific downstream tasks.

Docker Deployment

• First, pull the Docker image:

docker pull sunhaotong0605/microgenomer:0.1

• Run Inference for Different Downstream Tasks:

docker run --rm --gpus all \
    -v /path/to/input/:/data/input \
    -v /path/to/output/:/data/output \
    -e INPUT_PATH="/data/input" \
    -e OUTPUT_PATH="/data/output" \
    sunhaotong0605/microgenomer:0.1 \
    sh -c 'bash run.sh --input_path $INPUT_PATH --output_dir $OUTPUT_PATH --task test_growth'

Please replace /path/to/input/, /path/to/output/, and the downstream task option(e.g., test_growth) with your actual paths and desired task.

License

This project is licensed under the MIT License. See LICENSE for more details.

Citation

Kang Q, Guo Y, Hu B, et al. MicroGenomer: A Foundation Model for Transferable Microbial Genome Representations Enabling Multi-scale Genomic Understanding and Ecophysiological Trait Prediction. bioRxiv. doi: https://doi.org/10.64898/2025.12.28.696777.