• Bayesian causal modeling identified clinical and molecular causal drivers (prognostic biomarkers) of OS for mCRC. The molecular drivers were validated in independent cohorts.

• 1o side, ECOG score, AST, LDH, HGB, and metastases (intra-abdominal, and liver) were the top clinical drivers of OS.

• BRAF & RAS mutations, CMS4, and angiogenesis/ ECM remodeling signature were top molecular drivers of OS.

• Consistent with previous studies, ALOX5 and CDX2 were identified as causal driver genes of OS.

• A molecular pathway between 1o side and OS was identified. Investigation into the molecular underpinnings of sidedness in driving OS is currently in progress.

• The availability of the measures for the drivers at baseline will allow better risk stratification at initiation of treatment.

• Additional research, including prospective studies, is necessary to confirm these findings.

Support: U10CA180821, U10CA180882, U10CA180888; Eli Lilly and Company, Genentech, Pfizer. ClinicalTrials.gov Identifier: NCT00265850

• CALGB 80405 is a recently-completed phase III clinical trial of FOLFOX and FOLFIRI with randomly assigned cetuximab (cet) or bevacizumab (bev) in metastatic CRC (mCRC) patients.

• Hypothesis-free machine learning approaches to this study dataset can provide valuable insights into mCRC prognosis and management of mCRC progression.

• Causal modeling identifies the set of conditional dependencies between variables leading to outcomes.

• We built multivariate causal models of mCRC and examined the network drivers of mCRC survival.

• Using our Bayesian causal machine learning platform REFSTM, an ensemble of 128 network models were built for overall survival (OS) of mCRC.

• The ensemble enables estimation of model uncertainty and identification of key drivers by model consensus.

• Simulations were performed on the ensemble to identify causal drivers of OS after accounting for confounders. Causal effect was quantified by median hazard ratio (HR). For continuous variables, 3rd & 1st quartile values were used to compute HR.

• Analysis of NanoString data:• Consensus molecular subtypes (CMS) were computed using published code

(Guinney et al., Nat. Med. 2015) on GitHub. • Molecular clusters were computed using consensus clustering.

• Patients with both KRAS wild-type and mutant tumors were included and those who received both cet and bev treatments were excluded. Molecular data from primary tumors were included.

• Two independent cohorts (N=117 for mutations, N=206 for nanostring data) were withheld and used for causal drivers validation.

CAUSAL MODELS• Model1: Clinical variables only (N=1463, 68 variables)• Model2: Clinical+molecular variables without raw nanostring data (N=430, 84 vars)• Model3: Clinical+all molecular variables (N=430, 900 vars)

BACKGROUND RESULTS

METHODS

CONCLUSIONS

Causal Modeling of CALGB 80405 (Alliance) Identifies Network drivers of Metastatic Colorectal Cancer Rahul K. Das,1 Leon Furchtgott,1 Daniel Cunha,1 Fang-Shu Ou,2 Federico Innocenti,3 Heinz-Josef Lenz,4 Jeffrey Meyerhardt,5 Kelly Rich,1 Jeanne Latourelle,1 Donna

Niedzwiecki,6 Andrew Nixon,7 Eileen M. O’Reilly,8 Diane Wuest,1 Boris Hayete,1 Iya Khalil,1 Alan Venook,91GNS Healthcare, Cambridge, MA; 2Alliance Statistics and Data Center, Mayo Clinic, Rochester, MN; 3University of North Carolina, Chapel Hill, NC; 4University of Southern California, Los Angeles, CA; 5Dana-Farber Cancer

Institute/Partners Cancer Care, Boston, MA; 6Alliance Statistics and Data Center, Duke University, Durham NC; Duke Cancer Institute, 7Duke University Medical Center, Durham, NC; 8Memorial Sloan Kettering Cancer Center, New York, NY; 9University of California San Francisco, San Francisco, CA

Model 1: Clinical Causal Drivers of OS • 1o side, ECOG performance score, concentrations of aspartate aminotransferase (AST), hemoglobin (HGB),

absolute neutrophil counts (ANC), lactate dehydrogenase (LDH) and metastases at intra-abdominal, lung, and liver were the strongest causal drivers of OS.

• Clustering of NanoString data revealed three molecular clusters with upregulation of different signatures: (1) WNT-signaling, (2) Angiogenesis & ECM remodeling, (3) Immune infiltration.

• BRAF mutation, RAS mutation, CMS4, and angiogenesis signature were the top molecular drivers of OS.

• Causal effects of 10 side on OS was found to be driven by a molecular pathway.

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●axon guidance

leukocyte cell−cell adhesion

positive regulation of peptidyl−tyrosine phosphorylation

cell−cell signaling

positive regulation of endothelial cell migration

motor neuron axon guidance

positive regulation of apoptotic process

angiogenesis

patterning of blood vessels

response to cytokine

response to progesterone

response to mechanical stimulus

cellular response to extracellular stimulus

cell surface receptor signaling pathway

chemokine−mediated signaling pathway

response to hypoxia

cell adhesion

extracellular matrix organization

platelet degranulation

positive regulation of monocyte chemotaxis

6 9 12 15nOverlap

GeneRatio●

●

●

0.10.20.3

10

15

20

25

30NegLogq

Fig. 2: Reverse Engineering: Consensus Subnetwork to OS Fig. 3: Top Causal Drivers from Simulations

Fig. 6: Consensus Subnetwork to OS Fig. 7: Top Causal Drivers from Simulations

Fig. 5: Over-represented GO Biological Processes in Angiogenesis Cluster

Model 3: Causal Driver Genes of OS • ALOX5 and CDX2 were among the top causal driver genes of OS. • The causal genes in the molecular pathways leading to OS are involved in ECM remodeling and angiogenesis, thereby

corroborating the findings from Model 2.

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●extracellular matrix disassembly

angiogenesis

cell maturation

in utero embryonic development

positive regulation of transcription, DNA−templated

heart development

axon guidance

B cell receptor signaling pathway

epidermal growth factor receptor signaling pathway

osteoblast differentiation

patterning of blood vessels

embryonic digit morphogenesis

fibroblast growth factor receptor signaling pathway

negative regulation of apoptotic process

neurotrophin TRK receptor signaling pathway

positive regulation of collagen biosynthetic process

positive regulation of stem cell proliferation

proximal/distal pattern formation

Fc−epsilon receptor signaling pathway

positive regulation of interferon−gamma production

2.5 5.0 7.5 10.0 12.5nOverlap

10

15

20

25

30NegLogq

GeneRatio●

●0.10.2

Fig. 9: Consensus Subnetwork to OSFig. 8: GO Biological Processes where Causal Driver Genes are Over-represented

Fig. 10: Top Causal Drivers from Simulations

Validation of Causal Drivers of OS• Identified causal drivers were validated in independent cohorts

using univariate Cox proportional hazard model. HR, 95% CI, and p-value are shown in the plots below.

++

+ +

++++++++++++++++++++++

++++++++

+ ++ +++++ +

0.00

0.25

0.50

0.75

1.00

0 25 50 75 100Time (Months)

Surv

ival

pro

babi

lity Strata

++

ALOX5=HighALOX5=Low

29 9 2 2 0

177 106 45 6 0ALOX5=LowALOX5=High

0 25 50 75 100Time (Months)

Stra

ta

Number at risk

++

+++++++++++++++++++++

+++++++++ ++ +++++ ++ +

+0.00

0.25

0.50

0.75

1.00

0 25 50 75 100Time (Months)

Surv

ival

pro

babi

lity Strata

++

Angio=NoAngio=Yes

172 103 44 7 0

34 12 3 1 0Angio=YesAngio=No

0 25 50 75 100Time (Months)

Stra

ta

Number at risk

+

+

+

+++ + +++++++++ +++ + ++

+

0.00

0.25

0.50

0.75

1.00

0 25 50 75 100Time (Months)

Surv

ival

pro

babi

lity Strata

++

BRAF=NoBRAF=Yes

106 45 14 2 0

11 2 0 0 0BRAF=YesBRAF=No

0 30 60 90 120Time (Months)

Stra

ta

Number at risk

+

+++++

++++++++++++++ ++

+++ ++ ++++

+

+ ++ +++ ++

++ +0.00

0.25

0.50

0.75

1.00

0 25 50 75 100Time (Months)

Surv

ival

pro

babi

lity Strata

++

CDX2=HighCDX2=Low

112 77 31 5 0

94 38 16 3 0CDX2=LowCDX2=High

0 25 50 75 100Time (Months)

Stra

ta

Number at risk

HR = 3.28 (1.67, 6.43)p = 0.0002

HR = 2.01 (1.35, 2.99)p = 0.0004

HR = 0.56 (0.41, 0.77)p = 0.0002

HR = 1.87 (1.22, 2.89)p = 0.0038

Fig. 11: Kaplan Meier Survival Curves

Model 2: Molecular Causal Drivers of OS Fig.4: Molecular Clusters

●●

●

●●●●

●

●●●

●●●

●●●

●

●●●●●●●●

●

●●

●

●

●

●

●●

●

●●●●

●

●

●

●

●

●●

●

●●●●

●●

●

●

●

●

●

●

●●●●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●●

●

●

●●

●

●●●

●●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●●

●

●

●

●●●

●

●

●

●●●

●●

●

●

●

●

●

●

●

●●●

●

●

●

●●

●●●

●●

●

●

●

●

●●●●

●●●

●●

●●●●●●

●

●●●

●

●●●●

●

●●●●

●

●

●

●●

●●

●

●

●

●

●

●●

●

●●

●●

●

●●●

●

●●●●●●

●

●●●●

●●

●●●●●●●●

●

●●●●●●●●●●●●●

●

●

●●●●●●●

●●●●●●●●●●●●●●

●

●

●●●●●●●●●●●●

●●●●●●

●

1

2

3

BRAF

_MUT

ATIO

NRA

S_MUT

ATIO

N

CMS4

ANGIO

GENES

IS

LDH

ANC

AST

SIDE

METS:

LIVER HGB

Causal Drivers of Overall Survival

Haza

rd R

atio

●

●●●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●●●●

●

●●●●

●

●

●

●●

●

●●●

●●●

●

●

●

●●

●

●

●

●●

●●●●●

●●

●●

●

●

●

●●

●

●

●

●

●●●

●

●

●●●●●

●

●

●●

●●

●

●●

●

●

●

●●●

●●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

1.0

1.5

2.0

2.5

ALOX5LDH

BRAF_MUTA

TION

DTX3VIM

PTPRC

HSPD1

COL6A3

SIDECDX2

Causal Drivers of Overall Survival

Haza

rd R

atio

Fig. 1: Schematic of REFSTM Reverse Engineering & Forward Simulation Workflow

●

●●

●

●

●●

●

●●

●●●●●●●

●

●●

●●●

●

●●

●

●

●●●●●●●●●

●

●

●●●●

●●

●

●●

●

●

●

●●●

●●●

●●●●

●

●

●●

●●●●●●●●●●

●●●●

●●●

●

●

●●●●●●

●

●

●

●

●●●●●

●

●●

●

●●●

●●●●●●●●●

●

●●●●●●●●●

●●●

●●●●

●

●●

●●●●●●●●●

●

●

●●●

●●●●●●●●●●●●●●●●●●●

●

●●●●●●●●●●●

●●

●

●

●

●

●

●●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●●●●

●

●

●●

●

●●

●

●

●

●●

●

●

●

●

●

●●

●●●

●

●

●

●●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●0.9

1.2

1.5

1.8

SIDE

ECOG_SCORE

ABDOMINAL_METS

AST

LUNG_METS

HGBANC

LDH

LIVER_METS

NEUTROPHIL (%)

Causal Drivers of Overall Survival

Haza

rd R

atio