scWeave: A deep learning model that bidirectionally translates between gene expression and chromatin structure at single-cell resolution
Chromatin structure and gene expression are intimately linked, yet characterizing how the two covary has proven challenging, primarily because the two modalities are rarely measured in the same cells. Recently, single-cell co-assay protocols have enabled simultaneous profiling of both modalities within the same cells, but these experiments remain costly and technically challenging. To better characterize the relationship between 3D chromatin architecture and gene expression and to enable cross-modality inference from single-modality measurements, we developed a model called scWeave that bidirectionally translates between gene expression (scRNA-seq) and 3D chromatin architecture (scHi-C) at single-cell resolution. The scWeave model employs dual autoencoders to extract separate cell-level latent representations and learns to translate between these representations using dedicated translation modules. We evaluate scWeave on six publicly available co-assay datasets spanning mouse embryonic development, mouse cortex, mouse olfactory epithelium, and human bone marrow. On held-out mouse cells, scWeave outperforms a nearest-neighbor baseline and existing methods adapted to single-cell resolution, achieving a 57% improvement in median Spearman correlation when predicting gene expression from chromatin structure and an 18.8% improvement in median HiCRep similarity in the reverse direction relative to the next-best baseline. We further show that scWeave learns cross-modally aligned latent representations at single-cell resolution, enabling cells profiled in one modality to be matched to their counterparts in the other. Finally, scWeave generalizes to entirely held-out developmental timepoints in mouse olfactory epithelium and performs well on held-out human bone marrow cells despite limited human training data. By predicting the unmeasured chromatin architecture or transcriptional state from a single measured modality, scWeave offers a route to extend the benefits of costly co-assays to the many cell types, developmental stages, and species that are currently profiled with only one modality.
- Python >= 3.10
- A CUDA-capable GPU is recommended for both training and inference.
All dependencies are installed automatically (see pyproject.toml).
git clone https://github.com/Noble-Lab/scWeave.git
cd scWeave2. Create an isolated environment with uv
uv venv
source .venv/bin/activate # Windows: .venv\Scripts\activateuv pip install -e .This pulls in all dependencies and exposes the scweave package.
python -c "import scweave; print(scweave.__version__)"Pretrained scWeave weights are hosted on the Hugging Face Hub at
gmurtaza404/scWeave:
| File | Model | Species | Chromosomes | gene_names |
|---|---|---|---|---|
scweave_figure2_mouse.ckpt |
mouse co-assay (Figure 2) | mouse | 20 | "mouse" |
scweave_figure3_mouse.ckpt |
mouse + olfactory timepoints (Figure 3) | mouse | 20 | "mouse" |
scweave_figure4_human.ckpt |
human bone marrow (Figure 4) | human | 23 | "human" |
Download a checkpoint and load it:
hf download gmurtaza404/scWeave scweave_figure2_mouse.ckpt --local-dir weightsfrom scweave import scWeave
model = scWeave.load("weights/scweave_figure2_mouse.ckpt") # GPU if available, else CPUOr fetch it directly in Python:
from huggingface_hub import hf_hub_download
from scweave import scWeave
path = hf_hub_download("gmurtaza404/scWeave", "scweave_figure2_mouse.ckpt")
model = scWeave.load(path)See examples/inference/inference.md for
detailed, step-by-step instructions for running a trained model on the HiRES brain
subset: preparing the RNA and Hi-C inputs, loading the weights, and running
prediction and matching.
In brief:
from scweave import scWeave
model = scWeave.load("scweave_figure2_mouse.ckpt")
hic_pred = model.predict_hic_from_rna(rna) # (n_cells, n_chr, 224, 224)
rna_pred = model.predict_rna_from_hic(hic) # (n_cells, n_genes)
similarity, matches = model.match(rna, hic, direction="rna_to_hic")See examples/training/training.md for detailed,
step-by-step instructions for training scWeave from scratch: preparing datasets
from source, building the training splits with prepare_dataset, and training the
model with train_translator.
Each example folder pairs a notebook with a written walkthrough:
examples/inference/— load a trained model, prepare inputs, and run prediction and cross-modal matching (notebook, walkthrough).examples/training/— prepare a dataset and train scWeave from scratch (notebook, walkthrough).
Coming soon!
scWeave is released under the Apache License 2.0.