SPACE-Pro

The four-species model described in SPACE: A foundation model for transferable genomic learning across species.

Model

The encoder supports human, mouse, C. elegans (ce11) and D. melanogaster (dm6). Input length is 196,608 bp. Functional-profile supervision is provided for human and mouse; the additional species use masked nucleotide-composition learning. The checkpoint configuration also contains output-head entries for ce11 and dm6; their presence does not imply functional-profile supervision.

Loading

Use the accompanying code repository's examples/extract_embeddings.py, with this repository downloaded into a local checkpoint directory. The original checkpoint uses custom SPACE classes and a legacy model_type tag; it is not a standard AutoModel checkpoint. The code is available at ZhuJiwei111/SPACE-Pro.

The upload consists of pytorch_model.bin (2,376,882,394 bytes), config.json this model card and LICENSE. The checkpoint is provided as an inference-only release. Optimizer state, RNG state, training arguments and logs are not part of the inference release.

Validation and use

All 420 state tensor names and shapes match the packaged SPACE-Pro implementation. Full-length synthetic inference produced finite (1,896,3072) embeddings on an NVIDIA A40: 0.80 seconds for forward computation and CPU transfer, with 4.75 GiB peak CUDA-allocated memory, excluding checkpoint loading and file saving. A fresh Python 3.10 environment installation and full-length demo also passed on 9 September 2026; the full demo command took 30.59 seconds. This model supports genomic representation research; benchmark performance does not establish clinical validity or causal regulatory mechanisms. Downstream fine-tuning and frozen probing use different protocols and should be compared accordingly.

Selected targets: code ZhuJiwei111/SPACE-Pro, model IsaacZHU/SPACE-Pro, and data IsaacZHU/SPACE-Pro-data.

License

The SPACE-Pro weights, configuration and author-provided model documentation are licensed under the Apache License 2.0. Third-party datasets and baseline weights are not covered by this grant and retain their own terms.