PARP1 rs1805407 Increases Sensitivity to PARP1 Inhibitors in Cancer Cells Suggesting an Improved Therapeutic Strategy

Personalized cancer therapy relies on identifying patient subsets that benefit from a therapeutic intervention and suggest alternative regimens for those who don’t. A new data integrative approach, based on graphical models, was applied on our multi-modal –omics, and clinical data cohort of metastatic melanoma patients. We found that response to chemotherapy is directly linked to ten gene expression, four methylation variables and PARP1 SNP rs1805407. PARP1 is a DNA repair gene critical for chemotherapy response and for which FDA-approved inhibitors are clinically available (olaparib). We demonstrated that two PARP inhibitors (ABT-888 and olaparib) make SNP carrier cancer cells of various histologic subtypes more sensitive to alkylating agents, but they have no effect in wild-type cells. Furthermore, PARP1 inhibitors act synergistically with chemotherapy in SNP carrier cells (especially in ovarian cancer for which olaparib is FDA-approved), but they are additive at best in wild-type cancer cells. Taken together, our results suggest that the combination of chemotherapy and PARP1 inhibition may benefit the carriers of rs1805407 in the future and may be used in personalized therapy strategies to select patients that are more likely to respond to PARP inhibitors.

This pseudolikelihood is convex and efficiently computable. Learning is performed using accelerated proximal gradient methods implemented in TFOCS (7). We used Nesterov's 1983 method for optimization with a maximum of 700 iterations for our stability runs and 1000 iterations for all other runs. We used a modification to the Stability Approach to Regularization Selection (StARS) method (8) on the range .1 < λ < .3 subject to an instability threshold of .05 to select the value λ = .2 which we used to learn the model presented in the results.

Directionality assesment step.
Undirected graphs learned over datasets produced by an underlying directed model tend to generate false positive edges. Indeed, when there is a "collider" in the true graph, XàZßY (i.e., X and Y are causing Z) then the learned undirected model will be X -Z -Y -X. This is because X and Y are dependent given Z or Dep(X, Y | Z).
The false positive edge X -Y can be removed if we perform a conditional independence test over all possible subsets. For example, in the simple case of XàZßY, we will find that Ind(X, Y | Ø), and so the X -Y edge will be removed and correct orientation of the XàZ and ZßY edges will be thus established. In addition for some additional undirected edge, Z -W, in the absence of the edges X -W and Y -W, we can infer the direction ZàW. This is because a ZßW true edge would have produced false positive edges X -W and Y -W. We call this the directionality assessment step. Algorithmically we follow the procedure for PC-Stable (9) except we start from the MGM graph rather than a fully connected graph, and since we do not assume acyclicity we only use orientation rule R1.
Generalized Correlation. In order to measure association between a continuous and categorical variable or two categorical variables we use the following strategy. We would like to calculate the equivalent of Pearson's product moment coefficient for each possible pairing of these variables. The general formula for Pearson's correlation between two vectors of observations, X and Y, with means X and Y and standard deviations X and Y is XY = ] where ( p = ) is an indicator function that is 1 when p = and zero otherwise, and î = One motivation for this approach is that these sample covariances turn out to be proportional to the partial gradients of negative log pseudolikelihood in a factorized (i.e. zero edges) MGM as described above with respect to the edge parameters and variable levels (see (5) supplement). Namely: and where X is the indexed by i and Y is indexed by j in the MGM and the pairs of X and Y are continuous-continuous, discrete-continuous, and discretediscrete respectively.

Computational analysis methods -Software availability
The MGM-Learn platform was developed in MATLAB and is available upon request.
Undirected graphs are learned using the MATLAB code from http://www.stanford.edu/jdl17/learningmgm.html. For the non-paranormal normalization we used HUGE (10). To quantify PARP1 isoform abundances from paired-end reads of TCGA metastatic melanoma samples we used kallisto (11) using transcript definitions from Ensembl (12).

SNP imputation on TCGA samples and NCI-60 cell lines
NCI-60 data were obtained from Cell Miner in June 2013 (http://discover.nci.nih.gov/cellminer/). For those cell lines or TCGA samples for which the identity of SNP rs1805407 was not available we used imputation to infer its identity. Using SNAP (13) we found 51 SNPs to be in perfect linkage disequilibrium (LD) with rs1805407 (R 2 = 1). Of these, 9 variants were covered by the Affymetrix SNP Array 6.0 used by the TCGA. To determine the rs1805407 genotype in TCGA samples we used birdseed calls (14) from Affymetrix Genome-Wide Human SNP Array 6.0. Only samples with a birdseed confidence less than 0.1 or where all 9 SNPs in perfect LD agreed with the birdseed call were used. Fig. S1. Association of SNP rs1805407 to response to TMZ treatment.   to be "sensitive" had at least one copy of C in this locus. Cell line was considered "sensitive" when chemopotentiation ratio was ≥ 2. S: sensitive; R: resistant.  Supplementary Figure S4. Assessment of drug interactions by Bliss independence model. Data from the median effect studies were independently analyzed by the Bliss independence method (15). Growth inhibition curves of individual agents and their combinations were first fitted to a four parameter logistic equation. Affected fractions (Fa) for concentrations of individual drugs that corresponded to their respective concentrations in the combination were then interpolated and used to compute an expected level of activity (Fa) according to Fa = FadrugA + FadrugB -FadrugA * FadrugB (15). Expected effect levels (black bars) were then compared to actual toxicity caused by the MMS/ABT-888 combination (gray bars). Observed effect levels that are larger than expected constitute synergy; effect levels smaller than expected, antagonism. The data are largely consistent with the median effect analysis, showing synergy to additivity in the SNP cell lines, and additivity to antagonism in the WT cell lines.  (16). Increased sensitivity was observed for Irofluven, an alkylating agent that inhibits DNA replication (17). For comparison purposes, we also added the PARP1 inhibitor Olaparib (which is not statistically significant when used as single agent).  Supplementary Table S4. Results from MMS treatment of cell lines with and without PARP1 inhibitor (ABT-888 or olaparib). The data from the MTT assays were expressed as mean ± standard deviation (SD). A "potentiation factor", defined as the ratio between the IC50 means of MMS treatment alone and in combination with ABT-888 or olaparib was calculated for each cell line. A potentiation factor (ratio) ≤ 1 indicates no chemo-potentiation.