greenelab/czi-rfa

Application to "Collaborative Computational Tools for the Human Cell Atlas" https://chanzuckerberg.com/initiatives/rfa

czi-rfa

Application to "Collaborative Computational Tools for the Human Cell Atlas" https://chanzuckerberg.com/initiatives/rfa

I just decided to write one of these. I am interested in methods to define the gene expression latent space for single cells, and to connect that back to large numbers of bulk transcriptomes. Drop a line via issues if you're interested in networking our proposals. Please also note if you'd be interested in reviewing some text via PR. I'd love to operate under a pull request + review model if I can find sufficient reviewers.

Thanks! Casey

Other related applications

• Elana Fertig
• Loyal Goff

czi-rfa's Issues

Aim 3 - Arjun's Data and Experimental Design

As @cgreene noted in #3

We will need more detail from Arjun to provide an example. I think you should make a hypothetical example of the latent space arithmetic experimental design. What about doing it as single cell data + bulk perturbation & comparing to single cell + perturbation? That seems like the cleanest design to test your hypothesis.

We will use this issue to discuss the nature of the data and experimental design to test the hypothesis:

that the latent space will preserve vector arithmetic operations between state transitions within benchmark perturbation expression datasets.

The issue can be closed once the experimental design is updated in aim 3.

RFA Bullet Points

Copied from the RFA

The goals of this RFA include, but are not limited to:

• Developing standard formats and analysis pipelines for genomic, proteomic, and imaging data, in forms that enable consistent use of these pipelines by numerous experimental labs
• Identifying and solving common challenges for web-based interactive visualization of cellular and imaging data
• Developing user tools that allow scientists and physicians to extract and analyze data organized by genes, cells, or tissues of interest
• Supporting analytical methods and machine learning approaches to solving problems such as multimodal integration, inference of state transitions and developmental trajectories, and representation of spatial relationships at the cellular or molecular level
• Generating curated benchmark datasets from new or existing data for evaluating computational methods and designing future analysis competitions
• Developing new computational approaches to comparing and normalizing genomic and imaging data across assays, subjects, and species
• Generating experimental datasets that directly address computationally-guided questions in quality control, reproducibility, or multimodal integration

Use zero-inflated reconstruction loss

Single-cell RNA-seq data is sparse/zero-inflated. There are several works [1-3] showing advantages of modeling zero-inflated distances between cells over L-1/L-2 distances.

For the VAE in aim 1, I expect that using zero inflated loss rather than euclidean reconstruction loss would work significantly better.

[1] https://genomebiology.biomedcentral.com/articles/10.1186/s13059-015-0805-z
[2] https://arxiv.org/abs/1610.05857
[3] https://genomebiology.biomedcentral.com/articles/10.1186/s13059-017-1188-0

Garmire Application

@lanagarmire mentioned to me that she will be putting in an application for a visualization & workflow system. I will mention this in our collaborative network too.