silicobio/hoike

Hōʻike: A Joint-Embedding Predictive Architecture for Transcriptome Data Generation with Diffusion Models

Silico Biosciences

Visit the GitHub repository for the full framework code: https://github.com/silicobio/hoike

Usage

## 1. Look up the condition samples for a tissue that exists in the normal reference set.
user_target_tissue = "Skin"
condition_subset = dataset.condition_df[dataset.condition_df["tissue_type"] == user_target_tissue].reset_index(drop=True)
normal_baseline_array = dataset.normal_profiles[user_target_tissue]

## 2. Generate with sampling-time normalization consistent with diffusion training.
generated_df = generate_synthetic_condition_data_consistent(
    normal_profile=normal_baseline_array,
    jepa=jepa_model,
    diffusion=diff_model,
    scheduler=scheduler,
    gene_cols=dataset.gene_cols,
    num_samples=2500,
    value_cap=condition_value_cap,
    sampling_noise_scale=1.1,
)