CrossmodalNet

Interpretable modeling of time-resolved single-cell gene–protein expression.

Preprint: biorxiv.org/content/10.1101/2023.05.16.541011v2

CrossmodalNet is a PyTorch-based autoencoder that predicts protein surface abundance (CITE-seq ADT) from scRNA-seq gene expression, incorporating a time/condition embedding and an adversarial discriminator to disentangle temporal variation. Saliency maps expose the gene–protein regulatory links learned by the model.

Install

pip install crossmodalnet

Optional extras:

Requires Python ≥ 3.10.

Quick start

1 · Prepare data

CrossmodalNet expects two paired .h5ad files — one for X (genes) and one for Y (proteins) — with matching cell barcodes and a time/condition column in .obs:

cite_train_x.h5ad   # AnnData: cells × genes  (raw counts or normalized)
cite_train_y.h5ad   # AnnData: cells × proteins (CLR-normalized ADT)

Both files must share the same cell index. The day (or custom) column in .obs encodes the time-point integer for each cell.

2 · Train (CLI)

crossmodalnet-train \
    -x cite_train_x.h5ad \
    -y cite_train_y.h5ad \
    --tkey day \
    -o Adam -n 500 -v \
    --save --save-dir ./out

Key flags:

3 · Train (Python API)

import torch
from crossmodalnet import CrossmodalNet, load_data, sc_Dataset

dataset = sc_Dataset(
    data_path_X="cite_train_x.h5ad",
    data_path_Y="cite_train_y.h5ad",
    time_key="day",
    preprocessing_key="tSVD",   # optional; None keeps raw counts
)
train_loader, val_loader = load_data(dataset, batch_size=256)

model = CrossmodalNet(
    n_input=dataset.n_feature_X,
    n_output=dataset.n_feature_Y,
    time_p=dataset.unique_day,   # e.g. [2, 3, 4, 7]
)

4 · Inference

import torch
from crossmodalnet import load_model, load_hparams

model = load_model(
    "out/CrossmodalNet.th",
    n_input=13431,
    n_output=134,
    time_p=[2, 3, 4, 7],
    hparams_dict=load_hparams("out/hparams.json"),
)
model.eval()

with torch.no_grad():
    pred_proteins = model(x_tensor, T=time_onehot)   # shape: (cells, proteins)

5 · Save and load

from crossmodalnet import save_model, save_hparams, load_model, load_hparams

save_model(model, path="./out")           # writes out/CrossmodalNet.th
save_hparams(model, path="./out")         # writes out/hparams.json

model = load_model(
    "out/CrossmodalNet.th",
    n_input=..., n_output=..., time_p=...,
    hparams_dict=load_hparams("out/hparams.json"),
)

6 · Saliency (gene importance)

from crossmodalnet import saliency

sal = saliency(
    counts=x_tensor,         # (cells, genes) float tensor
    times=t_tensor,          # (cells, n_timepoints) one-hot tensor
    model=model,
    genes=list(dataset.var_names_X),
    proteins=list(dataset.var_names_Y),
)
sal.compute_saliency("CD14")
sal.get_top_genes(k=50, include_TF=True)

ax = sal.plot_top_genes(topk=20)
ax = sal.plot_top_TFs(topk=20)

7 · Hyperparameter tuning

pip install "crossmodalnet[tune]"
crossmodalnet-tune \
    -x cite_train_x.h5ad \
    -y cite_train_y.h5ad \
    --trials 50 --max-t 300

Or from Python:

from crossmodalnet.tune import run_tune

results = run_tune(
    data_path_x="cite_train_x.h5ad",
    data_path_y="cite_train_y.h5ad",
    trials=50,
)

Citation

If you use CrossmodalNet in your work, please cite:

@article{yang2023crossmodalnet,
  title   = {Interpretable modeling of time-resolved single-cell gene-protein expression},
  author  = {Yang, Yongjian and others},
  journal = {bioRxiv},
  year    = {2023},
  doi     = {10.1101/2023.05.16.541011}
}

License

© 2023 Yongjian Yang, Texas A&M University. MIT-licensed — see LICENSE.