Figure 1.
Schematic of Bayesian network reverse engineering and Monte Carlo simulation.
A. Genetic, gene expression and phenotypic data are prepared for modeling by formatting followed by investigation to select the appropriate data transformation for the particular data type. Confounding factors and other explanatory variables are considered and modeled if appropriate. B. Likely fragments for network reconstruction are identified by scoring all 2-, 3- and 4-variable combinations with the constraint that SNPs are causally upstream. There are too many scored combinations to consider all during network reconstruction. The fragments that had the most likely Bayesian scores for each individual variable were identified and retained for network reconstruction. C. Parallel global network sampling constructs an ensemble of 1024 network structures that explain the data. The probabilistic directionality computed by the Bayesian framework allows inferences to be made about what lies upstream and downstream of particular phenotypic variables D. Diversity in network structures identified during network reconstruction captures uncertainty in the model. Hypotheses are extracted from the network ensemble by completing Monte Carlo simulations of “what-if” scenarios. Down-regulating the blue transcript would be expected to impact both TJ and SJ, while leaving CRP unchanged. The change in these phenotypic parameters would further predict reactive transcription changes in the liver. The separation of upstream and downstream components can identify potential intervention points and markers.
Figure 2.
Consensus topology of network ensembles for pre-treated and TNF-α blocker treated data.
A. Snapshot of the network ensemble at 2.5% consensus topology generated from pre-treated subjects. Pain does not appear to be controlled by measures extracted from whole blood mRNA profiling in pre-treated subjects. B. Snapshot of the network consensus topology at 2.5% consensus generated from TNF-α blocker treated data.
Table 1.
Category 1 Transcripts from the post-treatment model.
Table 2.
Category 1 Transcripts from the pre-treatment model.
Table 3.
Category 2 Transcripts from the post-treatment model.
Table 4.
Category 2 Transcripts from the pre-treatment model.
Table 5.
Category 3 Transcripts from the post-treatment model.
Table 6.
Category 3 Transcripts from the pre-treatment model.
Table 7.
Significantly enriched immune response related GO terms for transcripts from untreated model.
Table 8.
Significantly enriched immune response related GO terms for transcripts from anti-TNF-α treated model.
Figure 3.
10-fold knockdown of IL32 modulates number of swollen joints.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of cytokine IL32 in A. pretreated subjects and B. TNF-α blocker treated subjects. The effects are only predicted in pre-treated patients, suggesting a dependence on TNF-α signaling. The largest predicted effect is to modulate the number of swollen joints.
Figure 4.
Swollen joints predicted to be modulated by LASS5 in pre-treated patients only.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of ceramide synthase, LASS5, in A. pretreated subjects and B. TNF-α blocker treated subjects. The modulation of swollen joints is only predicted in pre-treated patients, suggesting a dependence on TNF-α signaling.
Figure 5.
Tender joints predicted to be modulated by WARS in pre-treated patients only.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of tryptophanyl-tRNA synthetase, WARS, in A. pretreated subjects and B. TNF-α blocker treated subjects. The modulation of tender joints is only predicted in pre-treated patients, suggesting a dependence on TNF-α signaling.
Figure 6.
Modulation CD86 predicted to affect both tender and swollen joint counts in TNF-α treated patients.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of CD86, the target of abatacept (CTLA4-Ig), in A. pretreated subjects and B. TNF-α blocker treated subjects. The modulation of tender joints is only predicted in TNF-α treated patients, suggesting both a mechanism that is independent of TNF-α signaling that could be exploited for subjects that do not respond well to TNF-α blocker therapies.
Figure 7.
RAP2C predicted to modulate both tender and swollen joint counts in TNF-α treated patients.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of RAP2C, a recently described ras G-protein, in A. pretreated subjects and B. TNF-α blocker treated subjects. The simulations suggest that this novel gene can modulate both tender and swollen joint count in TNF-α treated subjects. RAP2C may provide insight into novel TNF-α independent signaling pathways in RA.
Figure 8.
GON4L predicted modulate both tender and swollen joint count in TNF-α treated patients.
Plots of simulated number of tender joints, swollen joints, pain and C-reactive protein concentrations in response to a 10-fold knockdown in gene expression of GON4L, recently described as a novel factor in B-cell differentiation, in A. pretreated subjects and B. TNF-α blocker treated subjects. The simulations suggest that this novel gene can modulate both tender and swollen joint count in TNF-α treated subjects. GON4L may provide insight into novel TNF-α independent signaling pathways in RA.