Introduction

What is CellCap?

CellCap is a variational autoencoder for analyzing single-cell perturbation data. The primary goal of CellCap is modeling the correspondence between cellular identity and perturbation response. To understand this correspondence, CellCap was built in several explainable components:

1. Basal state z_basal: basal state z_basal can be understood as cell state in latent space where only intrinsic cellular identity is preserved.

2. Perturbation program w: each perturbation program has its latent representation w that has the same dimension as basal state z_basal and explains the transcriptional activation or deactivation after perturbation.

3. Perturbation key K: matching a perturbation key K with a basal state z_basal through attention mechanism can establish the correspondence between cellular identity and perturbation response. The output attention score A further represents the relevance of each basal state z_basal to this perturbation program w.

4. Program relevance H: similar to attention score A telling the relevance of each basal state z_basal to a perturbation program w, H presents the relevance of a perturbation condition to a perturbation program.

Is CellCap right for your data?

Before you proceed to run CellCap, it’s always important to ask if CellCap is right for your perturbation data. For CellCap to accurately model the correspondence between cellular identity and perturbation response, there are a few underlying assumptions to be considered:

1. The variation within unperturbed cells (control group) is primarily driven by cellular identity, and all possible cell states should have their representatives in the control group.

2. After perturbation, cells still remain their cellular identities that are explained by cell states. Loss of cellular identities due to perturbation would obscure the correspondence between cellular identity and perturbation response.

3. Perturbation will induce transcriptional activation or deactivation, but this change will not establish new cell states that are not present in control group.

4. If perturbation induces cells to establish new cell states, these perturbed cells should still preserve their old cellular identities.

What biological investigations do CellCap aim to address?

Understanding the correspondence between cellular identity and perturbation response would facilitate multiple biological investigations. We list a few suggested questions here, and users can use explainable components in CellCap to address them.

1. Do perturbations behave similarly or differently: Use of H alone could describe a general relationship between any two perturbations. The same question can be asked again under a specific cellular context. Combining A and H can be used to understand if different perturbations cause similar cellular responses in given a cell population.

2. What transcriptional responses can a perturbation induce: H also tells which perturbation programs w are relevant to a perturbation. Each perturbation program w then aims to reveal a set of genes that are concurrently activated or deactivated in perturbed cells.

3. Is a certain transcriptional response posed by a perturbation is cell-state specific: given perturbation program w of interest, basal state z_basal can be sorted by matching perturbation key K. This is the key question CellCap tries to address: the correspondence between cellular identity and perturbation response.

4. How large is this transcriptional response in this cell state: given a perturbed cell that was treated with a perturbation, CellCap first infers its basal state z_basal. Then, multiplying A and H for this perturbed cell can reveal how large perturbation program is in a cell state.

More information

For detailed scope and discussion about CellCap, please see our manuscript.