Liatir

Appearance

Geneformer V1 10M ​

Geneformer is a single-cell transcriptome foundation model designed to learn gene-network context from many human single-cell transcriptomes.

What it does ​

Geneformer can produce cell representations and supports downstream workflows such as cell or gene classification, embedding extraction, and in-silico perturbation analysis.

Current status in Liatir ​

Geneformer V1 10M is installable as a signed Runtime Box. Liatir pins the official V1 checkpoint and matching Genecorpus-30M dictionaries to one upstream revision instead of following the repository's changing default model.

Expected inputs ​

  • Human AnnData .h5ad data with non-negative raw counts in .X.
  • Ensembl gene IDs in var["ensembl_id"], or Ensembl IDs as var_names.
  • Optional obs["n_counts"]; Liatir computes totals from .X when absent.
  • Optional obs["filter_pass"] to include only passing cells.

Expected outputs ​

  • Embedded AnnData with obsm["X_geneformer"].
  • Lightweight CSV preview and JSON summary.
  • Input warnings, metrics, Jobs, Results, and reproducible provenance.

The current Liatir slice produces cell embeddings only. Gene-network inference, classification, fine-tuning, and in-silico perturbation are not exposed by this runtime.

Hardware and installation ​

Measured CPU throughput was not viable for the current product target. The published targets are macOS arm64 Metal and Linux x86_64 CUDA 12.9; Windows x86_64 runs the Linux CUDA target through WSL2, which needs an NVIDIA GPU visible inside WSL2. Liatir does not substitute a CPU or different CUDA target when one of these exact environments is unavailable.

The input matrix is normalized per cell to 10,000 counts, scaled by the official Genecorpus-30M gene medians, converted to the V1 rank-value encoding, and capped at 2,048 gene tokens. Cell embeddings are mean-pooled from the second-to-last hidden layer. The original input file is never modified.

Official source ​