Practical overview
When it fits
Use it when a research workflow needs a shared, configurable package for training or applying models to RNA expression and DNA sequence data, rather than a single-purpose model endpoint.
Who it is for
- Researchers developing or comparing foundation models for single-cell RNA or DNA data.
- Developers building configurable omics model-training workflows.
- Teams applying listed RNA or DNA checkpoints to the package's documented prediction and embedding tasks.
Documented uses
- Pretrain single-cell RNA models from h5ad or TileDB data.
- Pretrain DNA models from reference or variant-imputed genomes.
- Fine-tune or evaluate models for listed tasks such as cell-type annotation, perturbation prediction, promoter prediction and regulatory-region analysis.
- Generate embeddings and cell-type predictions from single-cell RNA data, or embeddings from DNA-sequence input.
Requirements & limits
The repository is public and lists an Apache-2.0 license. Installation instructions recommend Python 3.12 and uv; certain optional capabilities require additional dependencies.
- RNA and DNA workflows are both included, but their inputs and prediction commands differ.
- The RNA prediction example expects an h5ad file with raw counts and gene symbols; the DNA inference example expects train, test and development CSV files containing DNA sequences.
- RNA and DNA checkpoints are listed through IBM Research’s Hugging Face collection.
- vLLM support is an optional plugin documented for biomed-rna LLaMA models.
- The package uses anndata, Hugging Face Transformers, PyTorch Lightning and Hydra.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install the package from its repository
Create a Python 3.12 environment and install the cloned package using the documented uv steps.
Before you start
- A Git client and access to the repository.
- uv and Python 3.12.
- curl or libcurl may be required by the hic-straw dependency.
Steps
Clone the BiomedSciAI/biomed-multi-omic repository and change into its directory.
Create a virtual environment with the command below.
uv venv .venv -p3.12Activate it with
source ./.venv/bin/activate.Install the package with the command below.
uv pip install -e .
Expected outcome
The package is installed in the created environment.
Guide limitations
- The documented activation command is for a Unix-style shell.
- Optional capabilities require additional dependencies; the package notes optional extras for bulk RNA, benchmarking, tests, notebooks and vLLM.
Generate embeddings and cell-type predictions from an h5ad file
Run the documented scRNA prediction command with a listed RNA checkpoint.
Before you start
- An installed package and an h5ad file containing raw counts and gene symbols.
- A checkpoint from the IBM Research Hugging Face collection.
Steps
Set
MY_DATA_FILEto the path of the h5ad file.Run
bmfm-targets-run -cn predict input_file=$MY_DATA_FILE working_dir=/tmp checkpoint=ibm-research/biomed.rna.bert.110m.wced.multitask.v1.Use the RNA inference-inspection tutorial to review resulting embeddings and cell-type predictions.
Expected outcome
The documented workflow produces scRNA embeddings and cell-type predictions for inspection.
Guide limitations
- The command relies on the specified input data format and checkpoint; the README directs users to the RNA tutorials for further details.
- The notebook route requires the optional notebook dependencies.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- BiomedSciAI/biomed-multi-omic · Official homepage title · retained in its source language