Who can see your viewing activity?
For the mini-batch methods, does this assume approximate balance among the underlying groups? The subsampling would seem to be sensitive to losing low-occupancy clusters. And how are the individual minibatch clusterings combined? Is this like a 'bootstrap' version of doing k-means?
To follow on Miranda's question, how sensitive is the method to the order of sampling?
Thanks for the nice talk, cool implementation. The single-cell field makes extensive use of community-detection algorithms like Louvein. Is the general strategy the one of using k-means to ermine the larger dataset structure until data is mall enough to run those preferred methods from the community? Or do you think that clustering pipelines could be fully based on k-means?
Great talk, @Stephanie!
Wonderful talk Stephanie thank you
@Miranda @Adriana — great questions! Let me know if I sufficiently answered your questions, or if you have any follow up ones.
Fantastic talk thank you for kicking off this session! I also like to start by reminding people that millions of cells are now commonplace per experiment - great reminder that we’re v quickly heading towards billions
@Gioele — thank you for the question! It’s really up to you as the user what you choose to use. I agree that the field has gravitated towards Louvein, but we have found it to be slow and *very memory intensive* compared to mbkmeans. You would be encouraged use mbkmeans in any way you would means. I haven’t tried to explore your idea of using mbkmeans to do rough clustering and then trying the others for more granular clustering. But yes, clustering pipelines could be absolutely be fully based on k-means
@Ruben — Thank you!
@Russell @Adam — thank you for the opportunity to present!
@Stephanie - Thank you for the answer, and thanks again for the cool talk.
Maria Abou Chakra
Great talk @Geoffrey. I find your approach interesting and exciting to predict trajectories. My question is regarding about the assumption that cells do not arise from multiple lineages. For instance in the hematopeoetic system, dendritic cells are now thought to arise from two different “stem” lineages. knowing that is there a way to expand the approach to relax the assumption and include multiple possible trajectories?
Thank you Geoffrey for the very interesting and extremely clear talk! You seem built a cool model of the cell state dynamics that does not necessarily involves gene modelling (assuming ergodicity I guess). But then also show an (somewhat conceptually distinct) example where you determine a gene regulatory network from the state transition model. Can you elaborate on how you achieve this second task starting from the solution of the first? Thank you
Thank you Geoffrey - very nice blend of rigorous mathematical theory and application. Thanks for joining the session
@Russell — thank you for the opportunity to speak!
@Gioele and @Maria — thanks for the great questions!
Thank you Gioele - great and very information-rich talk! What you use pseudo-age is it lineage-specific? does that de-couple the gene dynamics to some extent when comparing multiple lineages?
Thank you for the question Adam! The pseudo age is just the time variable “imputed” using the JSD kNN graph. The pseudo age is used to impose a constraint and to avoid shortcuts that can exist despite building pretty good kNN graphs.
Great talk! Do the weights you obtain from the regression on the scRNAseq data correlate with the strengths of the motif hits in your ScATACseq analysis?
@Adam Regarding the second part of the question, we do not think we reached a sampling deep enough to resolve automatically the independent lineages of all the neutrons and glial types. We can do that only in ensemble of related lineages. There is much more work to do to fully disentangle the lineages, however we are pretty confident on the subtypes we described. It will take more research to figure out all the detailed links
Thanks for the great talk! How sensitive are the gene perturbation results to the inferred GRN? Do you think that false positive gene-gene relationships impact your results?
@Gioele got it! Thanks again, great work. Look forward to discovering what going deeper gives. (Also great synergy between talks - going deeper => data getting even BIGGER => need better dim reduction & clustering ;) )
Also, might a sigmoidal formulation of the regression be more appropriate, to account for saturation of TF binding sites?