CPA - Compositional Perturbation Autoencoder

What is CPA?

CPA is a framework to learn the effects of perturbations at the single-cell level. CPA encodes and learns phenotypic drug responses across different cell types, doses, and combinations. CPA allows:

• Out-of-distribution predictions of unseen drug and gene combinations at various doses and among different cell types.
• Learn interpretable drug and cell-type latent spaces.
• Estimate the dose-response curve for each perturbation and their combinations.
• Transfer pertubration effects from on cell-type to an unseen cell-type.

Installation

Installing CPA

You can install CPA using pip:

pip install cpa-tools

See detailed instructions here.

How to use CPA

Several tutorials are available here to get you started with CPA. The following table contains the list of tutorials:

Dataset	Year	Description	Link
Lotollahi et al.	2023	Predicting combinatorial drug perturbations	-
Lotollahi et al.	2023	Predicting unseen perturbations uisng external embeddings enabling the model to predict unseen reponses to unseen drugs	-
Norman et al.	2019	Predicting combinatorial CRISPR perturbations	-
Kang et al.	2018	Context transfer (i.e. predict the effect of a perturbation (e.g. disease) on unseen cell types or transfer perturbation effects from one context to another) demo on IFN-β scRNA perturbation dataset	-

How to optmize CPA hyperparamters for your data

We provide a tutorial on how to optimize CPA hyperparameters for your data.

Datasets and Pre-trained models

Datasets and pre-trained models are available here.

Recepie for Pre-processing a custom scRNAseq perturbation dataset

If you have access to you raw data, you can do the following steps to pre-process your dataset. A raw dataset should be a scanpy object containing raw counts and available required metadata (i.e. perturbation, dosage, etc.).

Pre-processing steps

0. Check for required information in cell metadata: a) Perturbation information should be in adata.obs. b) Dosage information should be in adata.obs. In cases like CRISPR gene knockouts, disease states, time perturbations, etc, you can create & add a dummy dosage in your adata.obs. For example:

       adata.obs['dosage'] = adata.obs['perturbation'].astype(str).apply(lambda x: '+'.join(['1.0' for _ in x.split('+')])).values
   

   c) [If available] Cell type information should be in adata.obs. d) [Multi-batch integration] Batch information should be in adata.obs.

1. Filter out cells with low number of counts (sc.pp.filter_cells). For example:

   sc.pp.filter_cells(adata, min_counts=100)
   

   [optional]

   sc.pp.filter_genes(adata, min_counts=5)
   

2. Save the raw counts in adata.layers['counts'].

   adata.layers['counts'] = adata.X.copy()
   

3. Normalize the counts (sc.pp.normalize_total).

   sc.pp.normalize_total(adata, target_sum=1e4, exclude_highly_expressed=True)
   

4. Log transform the normalized counts (sc.pp.log1p).

   sc.pp.log1p(adata)
   

5. Highly variable genes selection: There are two options: 1. Use the sc.pp.highly_variable_genes function to select highly variable genes. python sc.pp.highly_variable_genes(adata, n_top_genes=5000, subset=True) 2. (Highly Recommended specially for Multi-batch integration scenarios) Use scIB's highly variable genes selection function to select highly variable genes. This function is more robust to batch effects and can be used to select highly variable genes across multiple datasets. python import scIB adata_hvg = scIB.pp.hvg_batch(adata, batch_key='batch', n_top_genes=5000, copy=True)

Congrats! Now you're dataset is ready to be used with CPA. Don't forget to save your pre-processed dataset using adata.write_h5ad function.

Support and contribute

If you have a question or new architecture or a model that could be integrated into our pipeline, you can post an issue

Reference

If CPA is helpful in your research, please consider citing the Lotfollahi et al. 2023

@article{lotfollahi2023predicting,
    title={Predicting cellular responses to complex perturbations in high-throughput screens},
    author={Lotfollahi, Mohammad and Klimovskaia Susmelj, Anna and De Donno, Carlo and Hetzel, Leon and Ji, Yuge and Ibarra, Ignacio L and Srivatsan, Sanjay R and Naghipourfar, Mohsen and Daza, Riza M and 
    Martin, Beth and others},
    journal={Molecular Systems Biology},
    pages={e11517},
    year={2023}
}