Slide 1

Slide 1 text

BayesSpace Leonardo Collado Torres lcolladotor.github.io 2020-10-22

Slide 2

Slide 2 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT https://twitter.com/raphg/status/1318244963927289857?s=20

Slide 3

Slide 3 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT https://www.biorxiv.org/content/10.1101/2020.09.04.283812v1.article-metrics

Slide 4

Slide 4 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Resources ● Pre-print: https://www.biorxiv.org/content/10.1101/2020.09.04.283812 v1 ○ Has 2 supplementary files ● Bioconductor package ○ http://bioconductor.org/packages/BayesSpace ○ https://github.com/edward130603/BayesSpace ○ https://edward130603.github.io/BayesSpace/ ● Processing code: https://github.com/msto/spatial-datasets

Slide 5

Slide 5 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT 2 main features: enhanced clustering + sub-spot level analysis

Slide 6

Slide 6 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Main logical/processing steps: expression → PCs → sub-spot level modeling → high res PCs→ predicted expression

Slide 7

Slide 7 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Main parameter: number of clusters. Can be derived from the data too

Slide 8

Slide 8 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure 7 from our pre-print https://www.biorxiv.org/content/10.1101/2020.02.28.969931v1.full.pdf

Slide 9

Slide 9 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Remembering ARI: from our pre-print https://www.biorxiv.org/content/10.1101/2020.02.28.969931v1.full.pdf

Slide 10

Slide 10 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure 2: they do a better job! ^^

Slide 11

Slide 11 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure S1: their w i point towards outliers

Slide 12

Slide 12 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Some notes: positive and not so positive ● Bioconductor-friendly ● Will be a part of Bioconductor 3.12 ● We can **use** it ● BayesSpace (right now) is only meant for 1 image at a time ● How long does it take to run? ● Their analysis code is actually not public (or I haven’t found it) >.<

Slide 13

Slide 13 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure 3: can use marker genes & get some profiles at sub-spot level

Slide 14

Slide 14 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure 4: works for other tissue organizations too (non-laminar) & when you don’t know the number of clusters (q)

Slide 15

Slide 15 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure 5: sub-spot level model works with simulated data (from their own model)

Slide 16

Slide 16 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Figure S5: choosing the number of PCs is important too. Here they use PCs 1-7 for melanoma, 1-15 for DLPFC. Check outliers, think about computing time vs biology.

Slide 17

Slide 17 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT stLearn Uses deep learning on images https://stlearn.readthedocs.io/ en/latest/stSME_clustering.ht ml#Human-Brain-dorsolateral- prefrontal-cortex-(DLPFC)

Slide 18

Slide 18 text

LIEBER INSTITUTE for BRAIN DEVELOPMENT Some questions ● Will stLearn improve with improved images? ● How sure are we that w i captures outliers? ● How much will it be affected by artifacts? ○ Can we use it to QC spots? ○ If we drop spots, can we still run sub-spot? My guess: yes ● How would you apply it to all 12 DLPFC images?