scLAMBDA

Modeling and predicting single-cell multi-gene perturbation responses with scLAMBDA

We present scLAMBDA, a deep generative learning framework designed to model and predict single-cell transcriptional responses to genetic perturbations, including single-gene and combinatorial multi-gene perturbations.

By leveraging the embedding capabilities of large language models, scLAMBDA effectively predicts genetic perturbation outcomes for unobserved target genes or gene combinations. Its disentangled representation learning framework enables the modeling of single-cell level perturbations by separating basal cell representations from salient representations associated with perturbation states, allowing for single-cell level generation and in silico prediction of perturbation responses.

Installation

scLAMBDA can be installed from from GitHub:

git clone https://github.com/gefeiwang/scLAMBDA.git
cd scLAMBDA
conda env update --f environment.yml
conda activate sclambda

Normally the installation time is less than 5 minutes.

Quick Start

Vignettes

Users can follow the example for training and evaluating scLAMBDA on the Perturb-seq dataset from Norman et al. (https://www.science.org/doi/10.1126/science.aax4438). The multi-cell line training code can be found in the example, which uses essentail gene perturbations from four cell lines (https://www.sciencedirect.com/science/article/pii/S0092867422005979; https://www.nature.com/articles/s41588-025-02169-3).

Preprocessing

scLAMBDA accepts logarithmized data stored in an AnnData object as input. The target genes should be specified in adata.obs['condition']. For single-gene perturbation datasets, this field should contain entries such as "gene_a" for perturbed cells and "ctrl" for control cells. For two-gene perturbation datasets, entries should follow the format "gene_a+gene_b" for cells with two target genes, "gene_a+ctrl" for cells with one target gene, or "ctrl" for control cells.

To run scLAMBDA, a dictionary of gene embeddings in the form gene_embeddings = {"gene_name": np.array} must also be provided as input.

The function can be used for simulating a dataset splitting:

n_split = 0
# for single-gene perturbation
adata, split = sclambda.utils.data_split(adata, split_type='single', seed=n_split)
# for two-gene perturbation
adata, split = sclambda.utils.data_split(adata, seed=n_split)
# use all available perturbations in training:
adata, split = sclambda.utils.data_split(adata, split_type='all_train', seed=n_split)

Training

To train the scLAMBDA model, use:

model = sclambda.model.Model(adata, 
                             gene_embeddings,
                             model_path = 'path_to_save_dir/models')

model.train()

Note that for single-gene perturbations, multi_gene = False should be included in the sclambda.model.Model() parameters.

• (Optional) By setting use_tg_coord = True, the model leverages the indices of perturbation target genes to improve the target-gene effect prediction.

For multi-cell line training, context embeddings encoding cell line or cell type information should be included using the following script:

model = sclambda.model.Model_context(adata, 
                                     gene_embeddings,
                                     ct_embeddings,
                                     model_path = 'path_to_save_dir/models')

model.train()

For this utility, three columns should be included in adata.obs: the target genes specified in adata.obs['condition'], the cell types specified in adata.obs['cell_type'], and a combined column adata.obs['cell_type+condition'], which can be created as follows:

adata.obs['cell_type+condition'] = [adata.obs['cell_type'].values[i] + '_---_' + adata.obs['condition'].values[i] for i in range(adata.shape[0])]

should be included.

(Optional) Reliability plot

We have developed two reliability scores to help assess the reliability of prediction results without requiring access to ground-truth data. To compute these scores for a set of test perturbations pert_test, run

rs_df = sclambda.utils.compute_reliability_score(adata, 
                                                 pert_test,
                                                 model.pert_emb,
                                                 model.pert_delta)

A diagnostic scatter plot generated by plt.scatter(rs_df['rs_rep'].values, rs_df['rs_cons'].values) visualizes the two reliability metrics jointly, where unseen perturbations located toward the upper-right region are more likely to yield reliable prediction results.

Prediction

Once trained, scLAMBDA can be used for in silico predictions. With a list of test perturbations pert_test, you can perform in silico perturbation using:

res = model.predict(pert_test, return_type = 'cells')

which will return the prediction results on control cells used in training. With a NumPy array containing the gene expression matrix of control cells ctrl_exp, the in silico perturbation can be performed on the input control cells:

res = model.predict(pert_test, ctrl_exp = ctrl_exp, return_type = 'cells')

Alternatively, with basal states generated from a standard normal prior, you can generate new perturbed cells with a given sample size n_cells using:

res = model.generate(pert_test, n_cells = 10000, return_type = 'cells')

Setting return_type='mean' in both modes returns the prediction results as the mean gene expression.

Citation

Gefei Wang, Tianyu Liu, Jia Zhao, Youshu Cheng, Hongyu Zhao. Modeling and predicting single-cell multi-gene perturbation responses with scLAMBDA. bioRxiv 2024; doi: https://doi.org/10.1101/2024.12.04.626878.