🚨🚨🚨 If you're a bioinformatician who loves: 🧬 #SingleCell #LongReadSequencing & #AlternativeSplicing 🚀 Collaborating and travelling Apply to work with us at the spectacular @BIMSB_MDC in Berlin + @UPFbiomed-@CRGenomica in Barcelona. Qs: https://t.co/ZikhH1zEGM Please RT!🙏

(20) And of course big thanks to my supervisors (@adameykolab and @KharchenkoLab), our experimentallist-magician Alek (@AlekErickson), and all of the collaborators (@EmmaRAndersson, @ratz_michael, @FrisenJonas). It was cool :)

(19) We tried to make the data as accessible as possible, so now you can find all you need to reproduce our results or test your hypothesis! Feel free to ask any questions and contribute to the packages :) Also, you can check my slides about this project:

(18) At the end of the day, it was a very fun journey and a good illustration of why it’s important to look at the data and be extremely careful with it (hi, gorilla! https://t.co/BQ5raPgtcM). A useful example from my PhD for my future academic career, ha-ha.

(17) Unlike CoSpar [ https://t.co/5Hp6ELnPyk], clone2vec isn’t designed to identify biasing gene expression programs in progenitors itself, but the output might be used for that purpose — and our whole study is about the identification of such programs.

(16) In the preprint, we carefully explored different properties of our algorithm (see Supplementary Note: https://t.co/QZdVAgnCpZ) and showed that it’s quite robust to subsampling of the data and can help to identify patterns that are barely visible to the human eye.

(15) We also decided to develop a small exploratory tool that will map clones from clonal embeddings to the gene expression space — clones2cells — so we can be sure that the output of our tool makes (at least some) sense. Feel free to play with it! https://t.co/SuI2r91dCb

(14) At the end, we have a vector representation of clones that we can explore with familiar to every scRNA-Seq researcher techniques — cluster and visualize them, compare compositions of different clones between conditions, and so on. Here is the package:

(13) At the end of the training process, clones with similar contexts will have similar weights in this neural network — and these weights will be used as embedding for clones themselves. (For those who’re familiar with word2vec: clones are words, and contexts are defined by kNN)

(12) In a core of clone2vec is a skip-gram neural network, in which based on one-hot encoded clonal label we’re trying to predict probability to observe other clones close to it (for each clonally labeled cell we take k nearest clonally labeled cells).

(11) Can we do it cluster-free? Can we introduce some metrics between clones based on the similarity of their location on PCA embedding? After a trial and error procedure of different ideas, we ended up with a word2vec-inspired approach that we called clone2vec. How does it work?

(10) But the amount of fates is quite big, and clones are small, and the resulting representations weren’t robust to the clone size and subsampling of the data. Also, even in the example above we see that clones occupy only a small part of, for example, cartilage.

(9) “Hmm, it might indicate something important about the data we analyze.” We then asked if it was possible to somehow identify such clones with similar behavior. Like, cluster them?.. In the beginning, we tried to represent each clone in cells' fates space.

(8) So, at E7/E8 we’re infecting different cells with different behavior (and possibly even the tree structure). How can we study it? At this moment we found some interesting examples of clones distributed in very similar domains of the gene expression space.

(7) Methods like CLiNC [ https://t.co/FqbLJ36PAH] can be used to find possible deviations from the reconstructed tree, but in our data amount of clones is not so big for the amount of fates we’re trying to study, and the method gave us some results that weren’t really robust.

(6) We know this fact for a while, but imagine if we don’t — and just reconstruct the tree between cell types based on the all clones from the data (both from the trunk and the head). It will show us a reconstructed tree, but it will be a superposition of different trees.

(5) A good illustration in this case is the origin of mesenchyme in the face and trunk. In the trunk, neural crest-derived cells are mostly related to CNS fates, but in the face, they’re closer to mesenchyme, because almost all of the facial mesenchyme is NC-derived itself.

(4) How can it be possible? Of course, because of NMPs, these cells give us some information about relationships between almost all of the fates we see. But how can we be sure that the same tree structure is true for every part of the body? A short answer — we can’t.

(3) What can we do with this data? Firstly, we can try to do what almost all developmental biologists dream of — reconstruct cell type trees. We did it — and got a pretty reasonable tree. Also, we observed a significant amount of clones sharing neuronal and mesodermal fates.

(2) If you haven't read a thread from the previous tweet, here is the study design briefly: we have a scRNA-Seq dataset of ecto- and mesodermal derivatives at E13 day of mouse development, and on top of that, we have information about which cells are clonally related (from E7/E8)