Thread by @simocristea

Amazing week for #DeepLearning in #spatial #singlecell biology, with 2🔥new Graph Neural Networks methods!

1.STELLAR🇺🇸 @jure: a cell type annotation & discovery atlas-type framework
2.NCEM🇪🇺 @fabian_theis: an approach to infer cellular communication patterns

Deep dive below🧵

But first, some background.

Spatial molecular biology has actually been around since the 70s. @lpachter's wonderful book-like article "Museum of Spatial Transcriptomics" comprehensively discusses history, tech & methodology advances in the past 50 years.
nature.com/articles/s4159…

Nevertheless, recent advances in single cell molecular technologies (brought by e.g. @10xGenomics & @AkoyaBio) have facilitated the high-throughout profiling of (groups of) single cells in their tissue context across embryogenesis, normal tissue development & disease progression.

The most popular high-throughput profiled data types are #SpatialTranscriptomics & #SpatialProteomics.

But the repertoire (pun!) is quickly diversifying, e.g. a very recent paper measures the transcriptome & TCR sequence of cells in situ🤯(Slide-TCR-seq)
cell.com/immunity/fullt…

To note is that most technologies do not in fact achieve single-cell resolution. Rather, they record the molecular signal from a spatial region spanning few tens of cells.

This signal can be either
1.approximated as corresponding to a single cell
2.computationally deconvoluted

Now, back to this week's first highlight, STELLAR, coming from the @Stanford labs of @jure, @GarryPNolan & @SnyderShot.

STELLAR proposes a reference-based automated annotation & discovery atlas-style framework of cell types in unseen settings.

nature.com/articles/s4159…

STELLAR combines 2 sources of information to automatically infer & discover cell type labels, by the logic that:
1.Molecular similarity means functional similarity (assumption underlying entire #scRNAseq field)
2.Spatial physical proximity means functional similarity (histology)

Graph Neural Networks (GNNs) intrinsically capture the combination of these two streams of information: molecular similarity & histology, such that cells that are spatially close together and that have similar levels of gene or protein expression are embedded close to each other.

Two main features differentiate STELLAR from existing space-aware clustering methods:

1. It also requires a spatially annotated reference, with cells labeled according to their cell types (needs to be pre-existing)
2. It does automatic annotation, without relying on human input

In other words, human curation is replaced by the reference. With more & more spatial data to be generated, one can easily see how such an approach is optimal, as the human cost is fixed. In contrast, for manual annotation, the human cost increases with number of new datasets.

Using the given reference, STELLAR learns spatial & molecular signatures that define cell types. Then, it transfers this learning to an unseen & unannotated dataset,which can even belong to a different donor or tissue!

STELLAR also identifies new cell types, unseen in reference!

In the paper, STELLAR is used to annotate CODEX multiplexed fluorescent microscopy & MERFISH spatial transcriptomics data.

It annotates 2.6 million cells labeled with 54 protein markers from the human intestine of 8 different donors, saving hundreds of hours of manual annotation

Let's move on to this week's most recent highlight: NCEM, coming from the lab of @fabian_theis in Munich 🇩🇪 & led by @AnnaCSchaar & @davidsebfischer.

NCEM is a graph-based statistical model that infers spatial cell-cell communication patterns.

nature.com/articles/s4158…

NCEM goes beyond the state of the art for most scRNAseq cell-cell communication inference methods nowadays, which is assessing ligand and receptor molecular signatures across cell types and inferring interactions whenever signals from both types of signatures are identified.

NCEMs are graph neural networks that predict a cell’s observed gene expression vector from its cell type label and its spatial neighborhood (niche). The model explicitly considers both statistical dependencies of gene expression, as well as cell-cell communication events.

3 types of such graphs with increasing complexity in modeling cell communication are proposed: linear, nonlinear and conditional variational autoencoder (CVAE), with various data-dependent predictive performances.

In some instances (but not all!), linear models do well enough.

To address the constraints of limited experimental capture of ligand–receptor pairs, a really nice graph kernel extends the model of receptor activity in cells conditioned on observed ligand expression.

NCEM works on data generated w. e.g. CODEX, MERFISH,MIBI-TOF.

It can also be applied to Visium,following deconvolution (w. e.g. cell2location @OliverStegle). The variation of expression within cell types across spots depends on their inferred composition
nature.com/articles/s4158…

For ex, applied on a deconvoluted VISIUM lymph node data, NCEM identified a bidirectional dependency of B cells and follicular dendritic cells (FDC) that indicates positive feedback between both cell types in germinal center organization, as well as B cell-mast cells interactions

You can start exploring these 2 brand new tools on your own spatial dataset, they both are publicly available & well documented

STELLAR: snap.stanford.edu/stellar/
NCEM: ncem.readthedocs.io/en/latest/

Lastly:

Biology is naturally graph-like & graphs modeling has been around in biology for as long as anybody remembers.

Building on this & on the promise of #DeepLearning in biology, biological applications of Graph Neural Networks (GNNs) will become more & more popular.

If these two papers made you curious to understand GNNs better, here's two fun, easy, yet engaging & complex, online resources I recommend getting started with (@distillpub):

distill.pub/2021/gnn-intro/
distill.pub/2021/understan…