GPU-ENABLED DIFFERENTIAL DEPENDENCY NETWORK ANALYSIS OF LARGE - - PowerPoint PPT Presentation

▶
gpu enabled differential dependency network analysis of
SMART_READER_LITE
LIVE PREVIEW

GPU-ENABLED DIFFERENTIAL DEPENDENCY NETWORK ANALYSIS OF LARGE - - PowerPoint PPT Presentation

Jul 14, 2023 99 likes •418 views

GPU-ENABLED DIFFERENTIAL DEPENDENCY NETWORK ANALYSIS OF LARGE DATASETS Gil Speyer May 9, 2017 Nvidia GTC S7254 Transcriptomic Data Personome.com Biological Pathways Nature.org Differential expression analysis Class labels C1 C2

slide-1
SLIDE 1

GPU-ENABLED DIFFERENTIAL DEPENDENCY NETWORK ANALYSIS OF LARGE DATASETS

Gil Speyer May 9, 2017 Nvidia GTC – S7254

slide-2
SLIDE 2

Transcriptomic Data

Personome.com

slide-3
SLIDE 3

Biological Pathways

Nature.org

slide-4
SLIDE 4

Differential expression analysis

C1 C2

Class labels Differential Expression GSEA

Pathway p-value P1 0.00002 P2 0.0001 … … Pk 0.01

Pathway analysis

slide-5
SLIDE 5

Differential Coexpression and DDNs

Biolayout.org

slide-6
SLIDE 6

EDDY: discovery of differential dependency

Gene Set DB Gene set G1, G2, … Gi = {gi,1, gi,2, …} ¡

g1 ¡ g2 ¡ g3 ¡ g4 ¡ g5 ¡

known ¡ interaction ¡ inferred ¡ dependency ¡

GDNi ¡ Known gene-gene Interactions Gj: DDN

C2-specific dependency C1-specific dependency Common dependency

Prior ¡knowledge ¡

Gene set p-value G1 0.00001 G2 0.00005 … … Gk 0.007

C1 ¡

¡

C2 ¡

¡ g1 ¡ g2 ¡ g3 ¡ g4 ¡ g5 ¡

GINi ¡

P G DC1

( )

Network Likelihoods C1 C2

…

Statistical ¡ ¡ test ¡

…

Jung ¡et ¡al., ¡NAR ¡2014 ¡ Speyer ¡et ¡al., ¡PSB ¡2016 ¡

slide-7
SLIDE 7

Transcriptomic Data

slide-8
SLIDE 8

EDDY + CTRP-CCLE

– Identifies pathways enriched with differential dependency between sensitive and non-sensitive cancer cell lines, as in DDNs – Discover mediators of drug sensitivity, i.e. potential targets?

Cell lines chemical sensitivity for 481 small compounds RNAseq of 935 cancer cell lines

slide-9
SLIDE 9

Drug Sensitivity

10 15 0.0 0.1 0.2 0.3

auc

  • ●
  • ●
  • ●
  • ●
  • ●
  • ●
  • 200

400 600 800 BRD−K12502280

  • Sensitive (125)

Uncertain (596) NOT sensitive (126)

  • BRD−K12502280

[TG−101348 −− inhibitor of Janus kinase 2]

Viability (% DMSO) Concentration (log)

Dose-Response Curve

Area Under Curve

IC50 ¡ EC50 ¡ Cell line chemical sensitivity for 481 small compounds across >900 CCLs

slide-10
SLIDE 10

EDDY-CTRP Workflow

  • ●
  • ●
  • ●
  • ●
  • ●
  • ●
  • ●
  • ●
  • ●
  • 200
400 600 800 BRD−A04287157
  • Sensitive (209)
Uncertain (419) NOT sensitive (209) [ELOCALCITOL −− agonist of vitamin D receptor]

sensitive non-sensitive

P G DC1

( )

Network Likelihoods C1 C2

…

EDDY

Pathway DB # genes p-val Rewiring FIBRINOLYSIS B 12 0.0062 0.56 IL4 B 11 0.0074 0.67 DCC MEDIATED ATTRACTIVE SIGNALING R 13 0.0086 0.66 GRANULOCYTES B 14 0.0093 0.53 P130CAS LINKAGE TO MAPK SIGNALING FOR INTEGRINS R 15 0.0101 0.61 DSCAM INTERACTIONS R 11 0.0105 0.6 TRAF3 DEPENDENT IRF ACTIVATION R 14 0.0169 0.73 CARM1 B 13 0.0173 0.68 SPRY REGULATION OF FGF SIGNALING R 14 0.0183 0.56 THE NLRP3 INFLAMMASOME R 12 0.0185 0.72

Gene Set DB Known gene-gene Interactions

CTRP CCLE

slide-11
SLIDE 11

Pathway Summary

  • ~9,000 drug-pathway pairs identified
  • http://biocomputing.tgen.org/software/EDDY/CTRP

Top 10 most frequent pathways associated with drug response

slide-12
SLIDE 12

TG 101348, a JAK2 inhibitor

Sensitive Resistant BOTH PRIOR NEW

TG101-348

  • A JAK2 inhibitor
  • EPONFKB Pathway
  • Essentiality mediators

JAK2, CDKN1A

slide-13
SLIDE 13

TG 101348, a JAK2 inhibitor

Sensitive Resistant BOTH PRIOR NEW

slide-14
SLIDE 14

Nutlin-3 Response

Non-Sensitive Sensitive BOTH PRIOR NEW RB Pathway Essentiality mediators

  • CDK4, CDK1

High specificity mediators

  • TP53
slide-15
SLIDE 15

Nutlin-3 Oncoprint

188 resistant samples 176 sensitive samples

slide-16
SLIDE 16

GPU Implementation

EDDY usage

slide-17
SLIDE 17

EDDY internals

… … … …

K1 K2

Unique set of networks Independence test on each edge

slide-18
SLIDE 18

… … … …

K3 assessed from Condition 1 samples assessed from Condition 2 samples

EDDY internals

Likelihood Distribution

slide-19
SLIDE 19

GPU Implementation

Independence test on each edge

slide-20
SLIDE 20

GPU Implementation

Unique set of networks

slide-21
SLIDE 21

GPU Implementation

Distribution of network scores

slide-22
SLIDE 22

GPU Verification

slide-23
SLIDE 23

GPU Performance

CPU: Intel Xeon 2.3 GHz with 33 GB of RAM GPU: NVIDIA Quadro K6000 902 MHz with 12 GB DDR5

slide-24
SLIDE 24

GPU Implementation

Level 1: Avg. fraction of possible edges

slide-25
SLIDE 25

GPU Implementation

Level 2: Avg. number of networks

slide-26
SLIDE 26

GPU Implementation

Level 3: Number of significant pathways

slide-27
SLIDE 27

TCGA Pan-Cancer

PIK3CA WT Mutation PRIOR NEW

  • PIK3CA 465

Mutation and 4289 Wild Type (WT)

  • PI3K Activation

pathway

  • PIK3CA not a node
  • f significant

alteration

slide-28
SLIDE 28

Single Cell Transcriptomic Data

slide-29
SLIDE 29

Single Cell Transcriptomic Data Analysis

REGULATION OF THE FANCONI ANEMIA

FANCD2 is an important protein in DNA double strand break repair. It stays connected to ATR. In the mesenchymal cells, FANCD2 disconnects from this network, but now ATR rewires with ATM, which has a similar activity to ATR. This is suggestive of a switch in the type of DNA repair mechanisms and possible lack of DNA double strand repair in mesenchymal subtype

Non-Mesenchymal Mesenchymal BOTH PRIOR NEW

slide-30
SLIDE 30

Future Directions and Acknowledgements

  • Compute the Cure award
  • Single-cell RNAseq dataset
  • drop-out events
  • more outliers
  • eddy-gpu will be available
  • n github once manuscript

is submitted (this month)

  • TGen Biocomputing
  • Seungchan Kim
  • Juan Jose Rodriguez
  • Tomas Bencomo
  • TGen CTD2
  • Michael Berens
  • Harshil Dhruv
  • Jeff Kiefer