### Cell culture and siRNA transfection

Umbilical cords were collected with written informed maternal consent and the approval of the Cambridge (UK) Research Ethics Committee. Human Umbilical Vein ECs (HUVECs) were isolated by collagenase digestion, as previously described [31]. Cells were cultured in fully supplemented media without antibiotics (basal EBM-2 with a propriety mix of heparin, hydrocortisone, epidermal growth factor, fibroblast growth factor, vascular endothelial growth factor, 2% foetal calf serum (FCS) Lonza, Cambridge, UK), at 37°C/5% CO_{2}. To carry out siRNA transfection, HUVEC pools consisting of 10 biological isolates (of equal contribution) were prepared using passage 3 cultured cells. The HUVEC pools were plated in 6-well plates at 2.5 × 10^{5} per well and left for 24hrs until approximately 70% confluent. siRNA transfection was carried out using pools of four siRNA duplexes from Dharmacon Inc (Lafyette, Colorado, USA) and the SiFectamine transfection reagent (ICVEC, London, UK) according to the manufacturer’s instructions.

### RNA processing and microarray preparation

RNA was extracted using TRIzol® reagent (Invitrogen, London, UK). RNA quality was assessed using the Agilent 2100 bioanalyser. Biotin labelled cRNA was generated and hybridised on the CodeLink Human Uniset 20K microarrays following the manufacturer’s instructions (Applied Microarrays, Tempe, Arizona, USA, formally supplied by GE Healthcare).

### Quantitative PCR

cDNA was synthesised from 1μg of total RNA using the QuantiTect reverse transcription kit (Qiagen, Crawley, UK), following the manufacturers protocol. Quantitative PCR was carried out using an ABI 7700 sequence analyser (Applied Biosystems, Warrington, UK). Reactions were performed using the Applied Biosystems universal master mix according to the manufacturers instructions. The Taqman probe / primers used were: *VASH1* (Hs00208609_m1), *SOX18* (Hs00746079_s1), *PTX3* (Hs00173615_m1), *FAM78A* (Hs00604618_m1), *PPARA* (Hs00231882_m1), *SLC7A2* (Hs00952727_m1), *BDNF* (Hs00542425_s1), *MTSS1* (Hs00207341_m1), *BTG2* (Hs00198887_m1), *TNFSF-12* (Hs00356411_m1), *FLT4* (Hs01047677_m1) and *NTRK2* (Hs00178811_m1), all from Applied Biosystems.

### Previosuly generated microarray datasets used in this study

The siRNA targeting of 351 different mRNA transcripts, chosen for their importance in EC biology, including transcription factors, signalling molecules, receptors and ligands is described by Hurley et. al. [11]. The microarray data from these 351 siRNA experiments is available from Gene Expression Omnibus, reference GSE27869.

The generation of triplicated microarray data from an eight time point HUVEC SFD time course has been described previously [18]. Briefly, HUVEC RNA was extracted at time points 0, 0.5, 1.5, 3, 6, 9, 12 and 24 hours after survival factor withdrawal (i.e. transfer from complete media to basal EBM-2 media with no supplements apart from 2% charcoal stripped serum), and hybridised onto CodeLink Human Uniset 20K microarrays. The raw and normalised triplicate time course microarray data has been deposited in Gene Expression Omnibus, accession number GSE23067.

### Data processing

CodeLink microarray quality was assessed using the CodeLink Expression analysis software v4.0. The array data were filtered to remove probes that did not contain “Good” flags in 90% of the arrays, as measured by the CodeLink Expression analysis software. The log base2 (log2)–transformed apoptosis time course data and 351 siRNA disruptant data were then both normalised using the Loess method [32, 33]. For the disruptant dataset log2 ratios against a virtual median array were calculated and these ratios were then z-transformed within each microarray prior to network inference.

For the SFD time course data, we selected transcripts concordantly regulated in abundance across the timecourse to used for GRN generation as previously described [18]. Briefly, log ratios between each time point and the first time point were calculated for all transcripts. For each transcript at each time point z-scores were then calculated by subtracting the log2 ratios from the mean of log2 ratios for that time point, and dividing by the standard deviation of log2 ratios for that time point. Transcripts were then selected that had −2 ≤ z ≤+2 at ≥ *two* adjacent time points in the triplicate data set. This analysis was repeated using the last time point instead of the first time point, and the union of the RNA lists prodced by the analyses that used the first and last timepoints was taken as the final list of concordantly expresed RNAs. For comparison to this z-score method, ANOVA was used to identify RNAs significantly differntially expressed at *two* adjacent time points relative to either the first or last time point, and the empirical Bayes method of Tai and Speed [34] was also applied. In addition, a statistically more complex method was used to identify RNAs significantly differntially expressed across the timecourse; generalised estimating equations with a Markov correlation model were fitted to the timecourse data. Contrasts were used to identify linear relationships and quadratic trends within the data using Matlab's GEEQBOX toolbox (http://www.mathworks.com/products/matlab/). Thresholds for concordant regulation were set using an absolute linear coefficient of >21 (and linear q value <0.01) OR an absolute quadratic coefficient of > 7 (and q value <0.01).

All other bioinformatic manipulations used the R software package, (http://www.R-project.org), and unless otherwise stated, multiple testing corrections were applied using the Benjamini and Hochberg method. Gene ontology/pathway enrichment analyses were carried out using Fatigo software [35], GeneSetDB [36], GATHER (http://gather.genome.duke.edu/) and IPA (Ingenuity systems, http://www.ingenuity.com).

### Apoptosis bayesian GRN generation and analysis

Apoptosis Bayesian networks were generated using the methods of [37], with some modifications. Given the relative sizes of the time-course and siRNA data sets, a dynamic GRN generated from the time-course data was used as a prior for GRN generated from the siRNA data as described [37].

When estimating the time-course GRN from the apoptosis time course microaray data, a method of bootstrapping was applied to the data. With 8 time points (obtained 0, 0.5, 1.5, 3, 6, 9, 12 and 24hrs after serum withdrawal) and 3 replicate microarray time course experiments, there are 3^{8} = 6561 possible combinations to create combinatorial apoptosis time course datasets. With such a large number of combinations, it is not computationally viable to fit regression curves through all combinations. Therefore the time course data used for network estimation was generated from the random resampling of 25 of the possible 6561 combinations as follows: Let *D* be the combinatorial time course data of all genes. If *D*(*c*) is the 8 time points, with each time point consisting of one of 3 replicates, then *D*(*c*) can be randomly resampled with replacement 25 times from the 6561 combinations so that *D*(*c*) (1 ≤ *c* ≤ 6561). The bootstrap sample can therefore be defined as *D*
^{*} = {*D*
^{*} (1),…., *D*
^{*} (25)}. Using this sample of 8 x 25 = 200 randomly resampled microarrays, the apoptosis GRN was estimated. This bootstrapping procedure was repeated 100 times to generate 100 different GRNs; *Ĝ*
^{*}
_{
T
}
^{ 1},… . *Ĝ*
^{*}
_{
T
}
^{ 100}, where *Ĝ*
^{*}
_{
T
}
^{
B
} is the estimated graph based on the *B*-th bootstrap sample. To estimate the reliability of the edges to be used as prior information, the bootstrap probability of each edge was calculated as follows: the reliability of the edge between the *i*-th gene to the *j*-th gene (termed the bootstrap probability) is
. A bootstrap probability threshold value was set at P = 0.8 and only those edges that passed this threshold value were included in the prior, *Z*
_{1}.

As described [37], a second prior (named the "array prior", Z_{2}), was also generated. This prior was based on the up- or down-regulation of the abundance of all mRNAs, represented as z-scores, analysed by the microarrays following siRNA-medaited targeting of the 351 genes.

Priors *Z*
_{1} and Z_{2}, were used when inferring a static Bayesian network based on the disruptant dataset [37]. Again bootstrap resampling of the microarrays (100 times with replacement) was applied to improve the reliability of edges included in the final network. The GRNs were viewed and analysed using Cell Illustrator 5.0, freely available software that can be downloaded from http://www.cellillustrator.com.

### Quantification of apoptosis

Passage 3 HUVEC pools comprising equal numbers of cells from 10 independant isolates were plated at 5 × 10^{3} cells per well in a 96-well plate and cultured for 24 hrs before siRNA transfection. Three different pools of 10 isolates were used for each assay. Cells were then left for a further 24hrs before treatment with either survival factor deprived conditions of basal media without supplements (EBM-2) or fully supplemented media without antibiotics (EGM-2) for 24 hrs. Active caspase-3 and −7 were quantified using the Caspase-Glo 3/7 assay system, following the manufacturer’s instructions (Promega, Southampton, UK). The ADP:ATP ratio was calculated using the Apo Glow assay (Lonza, Cambridge, UK), according to the manufacture’s protocol. Assays were carried out using a Fluostar Optima luminometer (BMG Labtech, Aylesbury, UK). Statistical analysis was carried out using a paired two-tailed t-test.