Variational autoencoder for multimodal mosaic integration and transfer learning

This repository contains implementations of scVAEIT for integration and imputation of multi-modal datasets. scVAEIT (Variational autoencoder for multimodal single-cell mosaic integration and transfer learning) was originally proposed by [Du22] for single-cell genomics data. scVAEIT is a deep generative model based on a variational autoencoder (VAE) with masking strategies, which can integrate and impute multi-modal single-cell data, such as single-cell DOGMA-seq, CITE-seq, and ASAP-seq data. scVAEIT has also been extended to impute single-cell proteomic data in [Moon24], though it is also applicable to other types of data. scVAEIT is implemented in Python, and an R wrapper is also available.

Check out the example folder for illustrations of how to use scVAEIT:

Example	Language	Notebooks
Imputation of ADT		imputation_1modality.ipynb
Imputation of RNA and ADT		imputation_2modalities.ipynb
Integration of RNA, ADT, and peaks		integration_3modalities.ipynb
Imputation of RNA		imputation_scRNAseq.ipynb
Imputation of peptides		imputation_peptide.ipynb

Reproducibility Materials

The code for reproducing results in the paper [Du22] can be found in the folder Reproducibility materials. The large preprocessed dataset that contains DOGMA-seq, CITE-seq, and ASAP-seq data from GSE156478 can be accessed through Google Drive.

Dependencies

The package can be installed via PyPI:

pip install scVAEIT

Alternatively, the dependencies can be installed via the following commands:

mamba create --name tf python=3.9 -y
conda activate tf
mamba install -c conda-forge "tensorflow>=2.12, <2.16" "tensorflow-probability>=0.12, <0.24" pandas jupyter -y
mamba install -c conda-forge "scanpy>=1.9.2" matplotlib scikit-learn -y

If you are using conda, simply replace mamba above with conda.

Prameters

Network parameters

In the example, basically, the network is operated in two levels of blocks:

• The feature levels: the number of genes $n_g$, the number of adts $n_a$, the number of peaks $n_p$. The related parameters are dim_input_arr and uni_block_names (meaning that they have the same length).
• The subconnection level: the number of genes, the number of ADTs, the number of peaks in chrom 1 $n_p^1$, the number of peaks in chrom 2 $n_p^2$, etc. The related parameters include dist_block, dim_block_embed, dim_block_enc, dim_block_dec, and block_names.

We explain the parameters as below:

• dim_input_arr represents the size of input features. In the example, it is simply $[n_g, n_a, n_p]$.

• dim_block represents the number of subconnected features in all modalities (assuming that the features have been rearranged accordingly). In the example, it is $[n_g, n_a, n_p^1, n_p^2, \ldots]$.

• dist_block: There are four distributions implemented: 'NB', 'ZINB', 'Bernoulli', 'Gaussian' for negative binomial, zero-inflated negative binomial, Bernoulli, and Gaussian, respectively. However, only 'NB' and 'Bernoulli' were tested and used to generate the results for the paper. 'Bernoulli' is used for ATAC-seq data, and 'NB' is used for genes and proteins.

• dim_block_embed represents the embedding dimension of the binary mask. For example, dim_block_embed = [1, 2, 3, ...] means the mask will be embedded into a continuous vector of dimension 1 for block 1,  and so on.

• dim_block_enc represents the structure of the first latent layer of the encoder. Using skip-connection helps reduce memory and computation complexity.  In the example, dim_block_enc = np.array([256, 128] + [16 for _ in chunk_atac]) means that the genes will be embedded into a vector of dimension 256, the adts will be embedded into a vector of dimension 128, and so on.  For block i, we have a sub-network that takes both the features of size dim_input_arr[i] and the mask embedding of size dim_block_embed[i] and outputs a vector of size dim_block_enc[i]. After that, the embedding vectors in all blocks will be concatenated into a vector as the input to the encoder. 

• Similarly, dim_block_dec represents the structure of the last latent layer of the decoder. For block i, we have a sub-network that takes latent features of size dim_block_dec[i] and outputs a vector (the predicted features) of size dim_input_arr[i].

• dimensions and dim_latent specify the network structure in the middle. For example, dimensions = [256, 128] and dim_latent = 32 mean that we have a network $n{in}-256-128-32-128-256-n{out}$ where $n_{in}$ is the sum of dim_block_enc, and $n_{out}$ is the sum of dim_block_dec.

Hyperparameters

Some of the important hyperparameters are:

• beta_unobs represents that weight for unobserved features.
• p_feat represents the probability of masking for the individual features. The larger value of p_feat encourages imputation ability but also requires more training epochs to have a good performance. But the influence of it is not large when training for enough epochs, so we recommend fixing fix p_feat as any reasonable value, e.g. 0.2. 
• p_modal represents the probability of masking out one modality. You can just leave it as a uniform.

In our experiments, the results were not sensitive to the above parameters. So you can just use reasonable values as in the example, except the following parameter requires some care depending on your data:

• beta_modal represents the importance of each modality. You run the model on your dataset for a few epochs and pick beta_modal such that the likelihoods (which will be printed during training) of all modalities are roughly in the same order. Notably, the number of peaks is generally very large, so its likelihood will have a higher value. And that is why you can see it has a small weight 0.01, in the example where beta_modal = [0.14,0.85,0.01].

References

• [Du22] Du, J. H., Cai, Z., & Roeder, K. (2022). Robust probabilistic modeling for single-cell multimodal mosaic integration and imputation via scVAEIT. Proceedings of the National Academy of Sciences, 119(49), e2214414119.
• [Moon24] Moon, H., Du, J. H., Lei, J., & Roeder, K. (2024). Augmented Doubly Robust Post-Imputation Inference for Proteomic data. bioRxiv, 2024-03.