search

IGS / gEAR

The gEAR Portal was created as a data archive and viewer for gene expression data including microarrays, bulk RNA-Seq, single-cell RNA-Seq and more.
https://umgear.org
GNU Affero General Public License v3.0
10 stars 5 forks source link

Issue with clustering #669

Open gear-portal-team opened 2 months ago

gear-portal-team commented 2 months ago

From: Maxime Moog

Email: maxime.moog@tuni.fi

Server IP: 10.142.0.14

Msg: Hello, I have been trying to use your platform to analyze some data and no matter my characteristics or the data I use, I am not able to complete the Louvain clustering step, ending in error. Do you have any tip on how to get this step done or is there any required set up to have? Thank you

Tags: ['']

Screenshot: None

adkinsrs commented 2 months ago

@JPReceveur this probably relates to what I told you recently about Louvain running out of memory. I'm thinking maybe we should just change the clustering algorithm flavor from "vtraag" to "igraph"

https://scanpy.readthedocs.io/en/stable/generated/scanpy.tl.louvain.html

RLC-DCPPC commented 2 months ago

Plan to discuss on the 20th

adkinsrs commented 1 month ago

Another thing is that when we redesign the scWorkbench page, I would like to switch over to the "leiden" clustering method since it seems to be the recommended version now

https://scanpy-tutorials.readthedocs.io/en/latest/pbmc3k.html#clustering-the-neighborhood-graph

adkinsrs commented 1 month ago

Quickly ran a "leiden" test with the same parameters as "louvain", and the tSNE and UMAP results were identical in both clustering algorithms (identical per algorithm, not projection). This was also using "igraph" flavor in louvain (leiden has no flavor options). Tested using the Kelley cochlear epithelium dataset with 5 neighbors and 5 PCs.