Thread by @lpachter

Interested in "integrating" multimodal #scRNAseq data? W/ @MariaCarilli, @GorinGennady, @funion10 & Tara Chari we introduce biVI, which combines the scVI variational autoencoder with biophysically motivated bivariate models for RNA distributions. 🧵 1/
biorxiv.org/content/10.110…

One of the clearest cases for "integration" is in combining measurements of nascent and mature mRNAs, which can be obtained with every #scRNAseq experiment. Should "intronic counts" be added to "exonic counts"? Or is it better to pick one or the other? 2/

This important question has been swept under the rug. Perhaps that is because it is inconvenient to have to rethink #scRNAseq with two count matrices as input, instead of one. How does one cluster with two matrices? How does one find marker genes with them? 3/

One approach could be to adapt a method such as totalVI (@adamgayoso et al.), which is an integration method built on scVI for CITE-seq data. nature.com/articles/s4159…
Such a method could take as input two matrices, but would not utilize the biological relationship between them. 4/

There is a way. Figure 1 of our preprint summarizes the idea. But first let's talk about adapting scVI to work like totalVI, except for nascent and mature mRNA counts. This is what is shown in panel (a) of our Figure 1. 5/.

"Generative" means that scVI learns (via a neural network) parameters for negative binomial distributions modeling gene counts. This is useful in practice (see nature.com/articles/s4158…) however the use of negative binomial distributions reflects a supposition about the data. 6/

The supposition is absent a mechanistic rationale, i.e. there is no interpretation to the negative binomial distributions, the merely label a black box. One may not care to make the black box transparent, though giving it meaning is necessary for "integration". But how? 7/

@GorinGennady has been thinking about the bursty model of transcription and how it relates #scRNAseq data for several years (see, e.g. sciencedirect.com/science/articl…). His idea was to use a variational autoencoder to parameterize distributions arising from a CME. 8/

The model we used in depicted in panel (b). A telegraph model driving bursty transcription generates nascent mRNAs that are processed and subsequently degraded. 9/

This brings us to biVI, shown in panel (c). We replace estimates of negative binomial parameters for each gene, with mechanistically motivated parameters based on the model described in panel (b). The scVI black box is now an interpretable, open, transparent box. 9/

The math underlying this method is not easy, and requires solving the chemical master equation for a non-trivial model. We did this using another neural network (RHS of panel b). This neural network takes as input parameters and outputs steady state distributions. 10/

A detailed description of how this works is in another preprint with @MariaCarilli and @GorinGennady.

In fact, the biVI application motivated the neural network CME solver. We believe it will find many other applications. 11/

To validate our implementation we first tested it on simulated data, where the sim is not just of counts matching a distribution, but a mechanistic sim of nascent and mature mRNAs. The "sanity check" worked well. biVI is much better at recovering the ground truth than scVI. 12/

Notably, "scVI" here is the adaptation of totalVI framework to work with two count matrices. I.e., to be fair, we are feeding scVI both matrices. Current common practice is to use scVI with one count matrix. Even so, while scVI performed reasonably, it's better to use biVI. 13/

What about biological data? We ran biVI on the BICCN primary motor cortex data that we analyzed in a publication last year, i.e. @sinabooeshaghi et al.: nature.com/articles/s4158….
14/

The information biVI outputs is interesting on several levels. Our Fig. 3 shows that distinct cell types are separated by the inferred parameters, and that novel markers can be detected (that are blind to an analysis even with two matrices, but that is not mechanistic). 15/

BTW, running biVI is straightforward. We demonstrate how to do it in a @GoogleColab notebook contributed by @funion10: github.com/pachterlab/CGC…

The repository contains the code for biVI, and also notebooks to generate the figures and results in the preprint. 16/

Note that in addition to implementing the bursty model, we also implemented the constitutive model, and an "extrinsic model" which production rates are random (from a Gamma distribution). 17/

In summary, we propose an answer to how one should use "intronic" and "exonic" counts together in a #scRNAseq analysis. There is much more to do: biVI can be extended to include a linear decoder for interpretability of the latent space, as in LDVAE academic.oup.com/bioinformatics… 18/

The underlying mechanistic model can be extended to account for different technical artifacts, as well as to include other modalities, for example protein quantifications as in totalVI. @MariaCarilli will be pursuing these ideas in the future. 19/

There is more work to be done on how to accurately count nascent* and mature mRNAs. A recent preprint w/ @kreldjarn, @DelaneyKSull, @GuillaumOleSan & @pmelsted addresses this with more to come soon. 20/ biorxiv.org/content/10.110… *this terminology simplifies the underlying biology.

Finally, a shout out to @MariaCarilli and @GorinGennady who led the project, and with whom it has been a pleasure to work with, and learn from. 21/21

For a chemical engineering perspective on this work see