VenusGenome-85M

VenusGenome-85M

VenusGenome-85M is a 12-layer, width-768 genome encoder with 85,455,416 parameters. It contextualizes ordered gene-level protein representations and returns 768-dimensional gene- and genome-level embeddings.

Usage

Install the inference-only release:

git clone --branch release/inference-v1 \
  https://github.com/zhangliang-sys/VenusGenome.git
cd VenusGenome
pip install -e '.[proteins]'

from venusgenome import VenusGenome

model = VenusGenome.from_pretrained("85m", device="cuda")
outputs, genome, preprocessing = model.encode_genome_files(
    "genome.gff",
    "genome.faa",
    esm_device="cuda",
)

This path matches GFF CDS features to FAA proteins, generates per-protein ESM-2 esm2_t12_35M_UR50D representations, and runs VenusGenome. Precomputed [n_genes, 480] protein embeddings can still be passed to encode_genome. The model returns gene_level, mean_pool, and global_token arrays. Genomes longer than 4096 genes are processed in contiguous chunks.

Checkpoint

This repository contains encoder-only BF16 safetensors from the mask-75% pretraining run at step 40,000. Decoder and optimizer state are not included. The architecture is recorded in config.json, and SHA256SUMS records the weight checksum.

Limitations

The model does not consume raw nucleotide sequence directly. Its outputs depend on upstream gene calling, genomic ordering, and ESM-2 protein embeddings. Chunking removes attention across chunk boundaries. Evaluate performance on the target organism distribution and downstream task.

The current repository has no declared open-source model-weight license.