Synergistic drug combinations tend to improve therapeutically relevant selectivity

Abstract

Drug combinations are a promising strategy to overcome the compensatory mechanisms and unwanted off-target effects that limit the utility of many potential drugs. However, enthusiasm for this approach is tempered by concerns that the therapeutic synergy of a combination will be accompanied by synergistic side effects. Using large scale simulations of bacterial metabolism and 94,110 multi-dose experiments relevant to diverse diseases, we provide evidence that synergistic drug combinations are generally more specific to particular cellular contexts than are single agent activities. We highlight six combinations whose selective synergy depends on multitarget drug activity. For one anti-inflammatory example, we show how such selectivity is achieved through differential expression of the drugs' targets in cell types associated with therapeutic, but not toxic, effects and validate its therapeutic relevance in a rat model of asthma. The context specificity of synergistic combinations creates many opportunities for therapeutically relevant selectivity and enables improved control of complex biological systems.

Main

Achieving therapeutic selectivity has been a major obstacle in drug discovery¹, leading to disappointing returns despite a surge in R&D spending². Drugs for infectious diseases often achieve selectivity by modulating pathogenic proteins with no human counterparts, but the treatment of cancer, metabolic or inflammation disorders must rely on targets that are present in both healthy and diseased tissues. This requires precise target modulation, which can be thwarted by the compensatory mechanisms available to complex biological systems^3,4,5. Overcoming this compensation often requires high drug doses that can induce unwanted effects in other tissues^6,7. Thus, although the current target-based drug design paradigm efficiently finds candidate drugs that are selective in a molecular sense, an alarming fraction have side effects in vivo that prevent their use at effective doses¹.

Synergistic combinations of two or more agents can overcome toxicity and other side effects associated with high doses of single drugs by countering biological compensation, allowing reduced dosage of each compound or accessing context-specific multitarget mechanisms^8,9,10. Therapeutically promising multitarget effects can be identified using experimental^{11,12,13,14,15,16} and theoretical^17,18,19,20 techniques. Because they actually make use of biological complexity, combinations are among the most promising avenues toward treating multifactorial diseases¹.

One empirical method for finding potential multitarget therapies is to seek synergistic responses to combined perturbing agents, like drugs, using phenotypes that arise from the concerted function of several cellular functions²¹. However, synergies between phenotypic perturbers that exceed expectations from the single agents' activities²² are relatively rare^14,18,23,24, and searching the vast space of combinations makes sense only if potential therapies are not expected to have synergistic side-effects²⁵. Previously we described a high-throughput platform to systematically test combinations of ∼3,000 approved drug ingredients, emerging therapeutics and research probes against cellular phenotypes representing diverse disease areas¹¹. Here we analyze drug combinations from 13 phenotypic screens relevant to six disease areas and perform flux balance analysis simulations of chemically inhibited metabolism in Escherichia coli bacteria, to show that synergistic combinations tend to operate in more narrow biological contexts than do single drugs. We also present examples of how therapeutically relevant selectivity can arise from the multitarget cooperativity underlying most phenotypic synergy.

Note: Supplementary information is available on the Nature Biotechnology website.

Results

We performed synergy and selectivity analyses for large data sets derived from simulations and 13 experimental screens. In each screen, many pairs of drugs were administered in two in vitro, cell-based assays: a 'test' assay served as a therapeutically relevant disease model, and a 'control' assay modeled a corresponding nondisease state. Drugs were serially diluted and all possible pairs of diluted drug combinations were tested on each assay (Fig. 1a). The measured cellular phenotypic response, T, to each drug combination was then compared to the median level, V, of 10–20 vehicle-treated samples. Based on these measurements, a dose matrix of inhibitory values, Z = (1 − T/V), can be computed for each pair of drugs in both the test and control assays (Fig. 1b). Our drug activity and combination analyses, which focus on positive inhibition values, could be generalized to account for both inhibitors and activators of the measured phenotype, but because most of our experimental drug activities are inhibitory, the metric Z defined above is sufficient.

Figure 1: Measuring selectivity bias.

(a) Determining synergy using dose matrix data. A factorial dose matrix was used to sample all mixtures of two serially diluted single agents. (b) Phenotypic measurements, such as inhibition Z relative to vehicle-treated samples, were visualized over the matrix using a color scale. (c) To determine synergy, each measurement was compared to expected values derived from the single-agent data along the left and bottom edges of the matrix. Color scale as in b. (d) Synergy was also described using an isobologram, which compares the doses needed to reach 50% inhibition along an equal-effect contour (in blue) to those along a predicted contour based on a model of dose additivity (red). In this example, to reach 50% inhibition, only 4 μM of drug X and 2 μM of drug Y were required in combination, compared to >41 μM and >34 μM for the single agents. (e) Calculating selectivity. Single agent (circled horizontal rows) and fixed-dose-ratio combination curves (circled diagonals) were extracted from test and control dose matrices. The concentrations at which drugs result in 50% inhibition (Z_cut = 0.5) were used to calculate the selectivity index SI = log₁₀(C_ctrl/C_test), which measures potency shifts between assays. SI = 1 implies a tenfold potency ratio favoring the test assay. (f) Selectivity bias B was determined by comparing SI shifts across many drug combinations within the same assay pair. To avoid spurious correlations due to noise, each test matrix was split into independent copies, one to calculate SI and the other for S. The SI distributions across the screen were compared for all combinations (green distribution) and those with S > S_cut (red), and the selectivity bias B was measured as the difference between the mean SI values for the two distributions. The SI distribution of the more effective single agent is shown in black. Arrows indicate for which analyses each matrix is used.

Measuring synergy and selectivity

The synergistic effects of a drug combination are typically determined relative to a null expectation that is calculated from the single agent activities based on standard non-interaction models^22,26. Several such models are in use, the most common being Loewe dose additivity²⁷, which is the expected response if both agents inhibit the same molecular target by means of the same mechanism; Bliss independence²⁸, the multiplicative probability derived for statistically independent target eliminations; and Gaddum's noninteraction²⁶, also called “highest single agent” (HSA), which is simply the higher of the two single-agent effects at corresponding concentrations. For the test phenotype, synergy was measured using a score S = ln f_X ln f_Y ∑_doses max(0,Z_data) (Z_data − Z_model) calculated from the volume between the measured combination and HSA response surfaces with weights to account for drug dilution factors f_X,f_Y (Fig. 1c). Dose-shifting relative to Loewe additivity was measured using a combination index²⁹ CI = C_X/IC_X + C_Y/IC_Y (Fig. 1d, black arrow), which measures the fractional shift between the most potent combination doses (C_X and C_Y) and the single agents' 50% inhibitory concentrations (IC_50(X) and IC_50(Y)).

Selectivity for drug potency in the test but not in the control assays was measured using a standard selectivity index, SI. SI represents the logarithmic shift in effective concentration between the test and control phenotypes at a chosen level of inhibition (50% in the case of our analyses). SI was measured separately for the single agents and for fixed-ratio combination curves extracted from the test and control dose matrices (Fig. 1e). The synergistic selectivity of individual combinations can be described using a differential selectivity index ΔSI between the combination and single-agent curves, or a differential synergy ΔCI between the test and control assays.

The selectivity bias B for each data set was measured as the difference between the average SI of those combinations with S exceeding a cutoff S_cut, and that of the unfiltered combinations (Fig. 1f). Separate replicates were used when calculating S and SI to correct for statistical correlations due to noise, and the significance of B was estimated in two ways: (i) calculating the standard error of B, assuming normal SI statistics; and (ii) generating histograms for both sets of SI values and determining the likelihood of their being drawn from the same distribution based on Poisson counting statistics within each bin³⁰. Simulated combination screens with Gaussian random noise confirm that this approach does not lead to spurious B detections from noise alone (Supplementary Note 1 and Supplementary Data Sets), and simulations with defined synergy and selectivity signals demonstrate that this approach reliably detects introduced selectivity biases without excessive dependence on screen design parameters or the chosen analysis cutoffs (Supplementary Note 1).

Simulated bacterial metabolism

To demonstrate the selectivity bias in a comprehensive biologically relevant data set, we simulated the inhibitory effects of drug combinations on bacterial growth using a flux-balance analysis model of E. coli metabolism³¹. The flux-balance analysis model comprises ∼950 enzymes or transporters that are organized into ∼44 distinct metabolic pathways or processes (Supplementary Note 2 and Supplementary Data Sets). Reaction fluxes throughout the network were modeled for growth under two conditions, minimal aerobic acetate medium³² and minimal glucose fermentation³³, chosen to activate very different pathways in the network.

Drug effects were simulated by restricting the flux of target enzymes by varying amounts corresponding to drug concentrations, at each such dose using minimization of metabolic adjustment¹⁸ to model growth responses (Supplementary Note 2 and Supplementary Data Sets). The effective concentrations for each target were used to define dosing ratios for 111,389 pairwise combinations, each of which was simulated as a fixed-ratio dosing series.

Using fermentation as the test phenotype and aerobic growth as the control, S and SI were calculated for each drug combination (Supplementary Data Sets). By grouping drug combinations based on the functions of the proteins targeted by each drug, and averaging the SI values of drug combinations that targeted the same pair of metabolic pathways, we were able to highlight pathways that distinguish fermentation from aerobic metabolism (Fig. 2a). As might be expected, drug combinations that targeted the citric acid cycle were highly selective and synergistic in aerobic conditions, whereas drug combinations that targeted pyruvate metabolism were particularly effective under fermentation. The bias (B ∼ 0.6) for the top 1% of synergistic combinations corresponds to almost a fourfold increase in potency over the single agents (Fig. 2b,c).

Figure 2: Simulation results identify drug combinations that inhibit E. coli growth in fermentation (minimal glucose) rather than aerobic (minimal acetate) conditions.

Drug effects were simulated by restricting the maximum flux rate through target enzymes, over a series of six concentrations chosen to sample each inhibitor's transition to activity. (a) Circles display the average SI of drug combinations that target enzymes in the same two functional categories. (b) Plotting selectivity (SI) versus synergy (S) for each drug combination. Combinations defined to be synergistic (S > 1) are marked in red. r, correlation coefficient. (c) The SI distributions of the most synergistic combinations (red), all combinations (green) and the more selective inhibitor in each combination (black) showed a strong selectivity bias for the top ∼1% of synergies. N_syn, number of synergistic combinations.

Experimental combination screens

For experimental validation, we analyzed 13 sets of combination data comprising 94,110 dose matrices, drawn from screens in six different disease areas (Fig. 3a, Supplementary Note 2 and Supplementary Data Sets). The screens varied considerably in design (that is, number of drugs, an 'aspect ratio' that compares the lists of drugs that were combined, sampling density for each combination and choice of assays), as well as in choice of agents (that is, mechanistic coverage, and activity in one, or selectivity across, multiple assays). Some screens tested for therapeutically relevant selectivity by comparing disease-related phenotypes to models of healthy cell viability. Others investigated mechanistic selectivity, where differential response profiles across phenotypes can highlight particular mechanisms. Up to four phenotypes were tested in each screen, resulting in two to six possible assay pairings, only some of which could be considered aligned with a therapeutic objective. We report synergy and selectivity indices across all assay pairs for each combination tested, along with a summary of statistical results for each comparison. Detailed combination response data are also provided (Supplementary Data Sets) for the specific examples of selective synergy discussed below (Table 1).

Figure 3: Selectivity bias for 13 sets of combination data focused on six disease areas.

(a) Experimental design parameters: N_agent, number of single agents tested; N_comb, number of combinations tested; N_copy, number of combinations with independent replicates; Aspec, drug list aspect ratio; Dens, dose matrix density (^* indicates that sparse matrices, similar to Fig. 1f, were tested). Z_agent and SI_agent, average activity and selectivity index across the single agents, respectively. (b) Selectivity bias for all assay comparisons. For each screen, all pairs of assays were compared, in 'forward' (circles, filled when aligned with a therapeutic objective) and 'reverse' (diamonds) order relative to the assay designations listed. The top 5% of synergistic combinations for each test assay were used to determine selectivity bias. Error bars represent 95% confidence with a sequential multiple hypothesis adjustment⁴⁷ to account for all assay comparisons in each screen.

Table 1 Individual examples of selective synergy

Representing therapeutically relevant selectivity, our viral, bacterial, inflammation and cardiovascular screens compared disease-relevant assays to human cell line proliferation as toxicity models (Supplementary Note 3). Most showed visible shifts toward positive SI for the synergistic combinations (Fig. 3b). The strongest biases occurred with the inflammation (Inflam 12x100; Fig. 3) and respiratory syncytial virus (RSV) screens, which compared single-protein readouts to broad cell-viability assays. Of the 16 assay pairs that were aligned with therapeutic objectives, 11 had significant B (P < 0.05) and for only two of those pairs did the reverse comparison yield a comparable positive bias. The main exceptions are the viral hepatitis C screen (HepC replicon; Fig. 3), where extremely selective single agents and a high incidence of pro-viral effects in the replicon assay led to skewed selectivity distributions, and the cardiovascular screen (Vascul 90 × 90; Fig. 3), where the strong single-agent selectivity left little room for additional selectivity gains.

A clear example of a therapeutically aligned selective combination is the antibacterial synergy against methicillin-resistant Staphylococcus aureus proliferation between ribavirin (Copegus, Rebetol, Virazole), a metabolism inhibitor, and disulfiram (Antabuse), a metabolic drug used for alcohol avoidance therapy (Fig. 4a). The combination has almost no effect on human smooth muscle cell viability. Similarly, the antiviral synergy of cepharanthine, an anti-inflammatory drug with antiviral potential³⁴, and benzamil, a potent inhibitor of ion transport channels³⁵, shows no detectable synergy against host cell viability (Supplementary Note 4). Finally, the anti-inflammation synergy between prednisolone and nortriptyline (Pamelor) against secretion of tumor necrosis factor (TNF)-α from stimulated peripheral blood mononuclear cells (PBMC) shows no corresponding increase in toxic effects as measured by PBMC metabolic viability (Fig. 5a). In all cases the synergy increases the likely safe treatment window for potential therapies.

Figure 4: Examples of therapeutically and mechanistically selective synergistic combinations, showing the control (left) and test (center) matrices, and the test isobologram (right).

(a) Ribavirin and disulfiram were each active on metabolic targets (ribavirin on inosine monophosphate dehydrogenase and disulfiram on mitochondrial aldehyde dehydrogenase 2) at high doses in both primary human smooth muscle cell viability (control, left) and S. aureus proliferation (test, center) assays, but the strong antibacterial synergy is completely absent in the human cell toxicity model (CASMC viab.). (b) In cancer cell lines, the synergy between camptothecin and LY 294002 was stronger against lung-derived H460 than colon-derived Colo-205 cells.

Figure 5: Selective synergy between glucocorticoids and tricyclic antidepressants (TCA).

(a) Synergistic activity of prednisolone (a glucocorticoid) and nortriptyline (a TCA) against TNFα secretion in peripheral blood mononuclear cells, showing the control (left) and test (center) matrices, and the test isobologram (right). (b) Mechanistic model. Glucocorticoids (red) activate the glucocorticoid receptor (GCR), which suppresses inflammatory signaling. In response to stress, lymphocytes secrete catecholamine hormones, such as norepinephrine, which suppress inflammatory signaling via beta-adrenergic receptors (ADRB2)³⁹. TCAs such as nortriptyline (red) block norepinephrine transporters (SLC6A2), which increases extracellular norepinephrine levels, with a synergistic anti-inflammatory effect when combined with glucocorticoids⁴⁰. Directly adding norepinephrine and modulating ADRB2 in combination with glucocorticoids confirms the role of this pathway in the glucocorticoid-TCA synergy (Supplementary Note 4). The in vivo therapeutic selectivity likely arises because, whereas GCR, SLC6A2 and ADRB2 are co-expressed in lymphoid cells, ADRB2 is expressed three to tenfold lower⁴¹ in tissues such as liver and pituitary that mediate major glucocorticoid-associated adverse effects⁴², weakening the norepinephrine-mediated pathway and attenuating the glucocorticoid-TCA synergy. GC, (glucocorticoid); NE, (norepinephrine). (c) Lung infiltration by eosinophils in a rat asthma model using ovalbumin challenge. Drugs used: Nor., N, nortriptyline; Budes., B., budesonide (another glucocorticoid widely used for asthma treatment); Dex., dexamethasone. (d) Expression of a corticosteroid side-effect marker tyrosine aminotransferase (TAT) in a rat liver toxicity model. Pred., P., prednisolone.

Considering mechanistic selectivity, many of our anticancer screens were designed with that goal in mind (Supplementary Note 3), and comparisons between very similar assays in the other screens also provide mainly mechanistic information. Although most of these assay pairs yield significantly positive shifts, their B tend to be weaker than those from therapeutically aligned assay pairs. The main exceptions are in the multiple myeloma screen, where substantial selectivity shifts occur between glucocorticoid-resistant MM-1R and the sensitive cell lines.

Individual mechanistically selective synergistic combinations provide insights into the biology relevant to the tested phenotypes. The strong synergy between LY 294002 and camptothecin in lung-derived H460 but not in colon-derived Colo-205 cells suggests an interaction between the targets of these drugs (that is, phosphoinositide 3 kinase and DNA topoisomerases) that may occur only in some cancer contexts (Fig. 4b). In the screen from which this combination was drawn, we found evidence of synergy in H460 for 21 out of 32 drug combinations that targeted topoisomerase and phosphoinositide 3 kinase, 12 of which had similar levels of selectivity over Colo-205 (data not shown), suggesting that the synergy results from coordinated activity on each drug's primary target. Another synergy between manganese sulfate and the hypertension drug methyldopa, in a toxin survival assay that shows no antibacterial activity, suggests potential post-infection treatments for anthrax that work through host targets (Table 1 and Supplementary Note 4). Finally, from our cardiovascular screens, the anticancer microtubule inhibitor paclitaxel (Taxol) and the vasodilator forskolin synergistically inhibit smooth muscle cell proliferation without being toxic to endothelial cells (Table 1 and Supplementary Note 4), pointing to possible use on drug-eluting cardiovascular stents to reduce their risk of causing thrombosis³⁶. Information of this kind can be used to guide the design of cotherapeutic treatments or to prioritize indications for candidate therapies.

General findings

Across all 68 pairs tested in our experimental datasets, 41 showed significant selectivity bias (P < 0.05 after multiple-hypothesis testing correction), of which all but two were positive for synergistic combinations (Fig. 3b). The SI values for unfiltered combinations were similar to those of their single agents (Supplementary Note 3), both having a positive skew owing to the more effective compound being used in the SI calculation. When synergy cutoffs were applied, however, the SI distributions of the most synergistic 5% shifted substantially toward more selectivity, with typically 15–40% of the synergistic combinations having SI > 0.5 (threefold dose shift), compared to 6–10% for unfiltered combinations and single agents (Supplementary Note 3).

Overall, after correcting for multiple assay comparisons in each screen, the average B of all assays was significantly positive (0.104 ± 0.010 at 95% confidence; Supplementary Data Sets), and the average B increased substantially for the 16 assay pairs that were aligned with clear therapeutic objectives (0.214 ± 0.021 at 95% confidence). Although the selectivity bias was strongly influenced by screen design parameters, such as the number of compounds used and the individual activity or selectivity of single agents in the screen, B depended only weakly on the chosen synergy cutoff or whether split matrices were used to separate the S from the SI measurements (Supplementary Note 3). Finally, some assay comparisons produced asymmetric selectivity biases; in that of 30 assay pairs with B > 0 at 95% confidence, 12 did not have comparable positive bias when the assay order was reversed (Fig. 3b). This asymmetry, however, increased substantially for the therapeutically aligned assay pairs (9 of 11 pairs were asymmetric), which compared a single protein expression phenotype to a broad measure of cell viability. All of this suggests that the selectivity bias arises not from stochastic effects but from biological context specificity, where synergistic combination effects require a narrower set of biological conditions than do single-agent activities.

Discussion

There is growing enthusiasm for combination therapy^1,5,10, specifically because greater selectivity is anticipated^1,9. Nevertheless, there are concerns that synergistic therapies would induce synergistic toxicity²⁵, as would be expected if the mechanisms that cause side effects are closely related to those involved with efficacy. To date, the arguments on both sides of this question have been heuristic; to our knowledge no study before this one has established the selectivity of combination approaches. Our experiments and simulations show a statistical bias toward greater selectivity for synergistic combinations, the strength and consistency of which cannot be attributed to stochastic noise or analysis parameter choices. This suggests that synergistic combinations tend to be more specific to particular cellular phenotypes than are single drugs, in agreement with the recent finding that synthetic lethal genetic interactions are less conserved between species than are single-mutant lethalities³⁷.

This selectivity bias for synergistic combinations may be understood in terms of the complexity of biological systems, where cooperative activity operates in only some cellular contexts but not others. The rarity in our screens of combinations where both agents target the same molecule, and the disparate drug mechanisms or indications underlying each of the example combinations (Table 1), point to the synergies being largely explained by multitarget interactions, as is the case for synergistic responses in theoretical studies^17,24. Because multitarget mechanisms require their targets to be available for coordinated action, one would expect synergies to occur in a narrower range of cellular phenotypes given differential expression of the drug targets than would the activities of single agents. Moreover, one would expect this specificity to narrow further as the combination order increases, until a limit is reached determined by the complexity of the biology relevant to a phenotype³⁸.

The anti-inflammatory synergy between prednisolone and nortriptyline against TNF-α secretion in PBMCs (Fig. 5a) illustrates how multitarget activity can lead to therapeutic selectivity. Prednisolone inhibits glucocorticoid receptors (GCR), and nortriptyline inhibits a separate autocrine pathway via norepinephrine transporters (SLC6A2) and beta-adrenergic (ADRB2) receptors^39,40 (Fig. 5b). The synergy probably operates via the primary drug targets because it remained even when related drugs were substituted (Supplementary Note 4 and Supplementary Data Sets; also S > 5 standard errors for 59 of 63 other drug combinations targeting GCR and SLC6A2 at sufficient concentrations). Because the ADRB2 receptors are more highly expressed in lymphocytes⁴¹ than in the liver and pituitary cells⁴¹ that mediate major glucocorticoid toxicities⁴², we would expect the synergy with tricyclic antidepressants to increase the therapeutic window of a glucocorticoid over those toxicities (Supplementary Note 4).

To test this in vivo, we used a rat asthma model to evaluate a similar drug combination (nortriptyline and budesonide, another glucocorticoid that is widely used for asthma treatment) at individually subtherapeutic doses. This combination performed better than single drugs and comparably to dexamethasone, a high-dose steroid control (ANOVA, P < 0.05 over single agents; Fig. 5c, Supplementary Note 4 and Supplementary Data Sets). Anti-inflammatory synergy of nortriptyline with prednisolone was confirmed in a rat pain model (Supplementary Note 4 and Supplementary Data Sets). This amplified anti-inflammatory effect does not show a corresponding rise in glucocorticoid-associated toxicity in rats at similar doses (Fig. 5d), as has also been seen with a related anti-inflammatory synergy, prednisolone with the cardiovascular agent dipyridamole (Persantine)^43,44. Toxicity of high-dose prednisolone was confirmed (ANOVA P < 0.05). These combinations represent a multitarget approach toward the long-sought “dissociated steroid,” where the anti-inflammatory activity of glucocorticoids can be separated from chronic side effects⁴⁵.

The increased specificity of combinations over single agents has implications for drug discovery and bioengineering. In medical contexts, the selectivity bias reinforces the potential of chemical combinations for network polypharmacology^1,8 by reducing concerns that synergistic side effects would make selective combinations too rare²⁵. For example, given performance typical for our screens (Fig. 3), a disease with ∼100 useful chemical agents would be expected to have ∼2–5 that are selective, with a threefold potency shift between two assays, but the ∼250 pairwise combinations representing the top 5% of synergistic combinations would be expected to yield 40–80 more treatments with similar or increased selectivity (Supplementary Note 3). For bioengineering, selective synergies provide opportunities for optimizing conditions in reactors, producing fuels, synthetic materials or pharmaceutical products⁴⁶, for example, by introducing combinations of chemical ingredients and comparing phenotypic measurements that track metabolic production of a desired chemical and toxic byproducts that limit a reactor's performance. There is much to be gained by expanding the notion of a target from a single biomolecule to the right set of nodes in a complex biological network.

Methods

Chemical handling and quality control.

The chemical library was archived in robotically accessible vials to which diluent (dimethyl sulfoxide or water) was added in preparation for addition to 384-well master plates by a Tecan Freedom liquid-dispensing robot. Liquid transfers to dilution and assay plates were handled using a PerkinElmer MiniTrak station adapted for the combination high-throughput procedure. Each 384-well assay plate contained multiple-dose matrix blocks, with serial dilutions at fixed ratios from the top concentration for each agent. Additional wells were reserved for transfer and untreated control wells. Compound mixtures were then added to the biological assay.

The contents of each plate were tracked in an automated laboratory information management system, using integrated barcode scanners in the liquid handling equipment, and stored in an Oracle database. Plates lacking transfers or with insufficient dynamic range (usually signal-to-noise ratio < 5 between untreated controls and a cell-free background) were rejected and repeated. The remaining plates were inspected using custom quality-control software. Individual wells that fell outside the expected range for normal assay readouts, or which were discontinuous with their neighbors, were marked for exclusion. Finally, the single-agent wells in each combination block were visually inspected for consistency across the experiment, and combination blocks containing the most discrepant single agents were marked for exclusion.

Measuring synergy and selectivity.

Dose matrices were assembled from replicate combination blocks on experimental plates. Some of the data showed systematic variations across the plate, most likely due to temperature or humidity gradients during incubation. Raw phenotype measurements T from each treated well were converted to inhibitions Z = (1 − T/V) relative to the median V of 10–20 vehicle-treated wells arranged around the plate. Positive modulators of a measured endpoint would best be represented by alternative expressions for Z (e.g., log(T/V) for growth or fitness), but as the vast majority of our chemical agents are inhibitors of our measured endpoints we have limited the analyses in the present work to inhibitions. Standard error estimates σ_Z for each median inhibition were also calculated, based on the quadrature sum of a minimum acceptable 3% error, the median absolute deviation (MAD) of the corrected vehicle data on each plate normalized to their median, and the MAD between replicate inhibition data between plates, using an empirical conversion from MAD to s.d.⁴⁸.

Synergy is determined by comparing the combination's response to those of the single agents^22,26. A key consideration for clinical combinations is whether a mixture is more potent than the drug-with-itself Loewe additivity level²⁷, the point at which there is some benefit over simply increasing component drug doses. Comparisons are made using an isobologram and a combination index²⁹, for example, CI₅₀ = (C_X/IC_50(X)) + (C_Y/IC_50(Y)), where (C_X/IC_50(X)) for a mixture is the ratio of the X compound's concentration C_X in a 50% effective mixture to its 50% inhibitory concentration IC_50(X) when applied alone. If a single agent does not reach the chosen effect level, we set the effective concentration to the top tested dose, and all CI values become upper limits. The combination index is best suited for analyzing fixed dose-ratio combination curves. For our experiments using dose matrices, we often measure the synergy using a synergy score S = ln f_X ln f_Y ∑_doses max(0,Z_data) (Z_data − Z_Loewe), between measured effects Z_data and a Loewe additive surface Z_Loewe derived from the single-agent curves⁴⁹. This synergy score is a positive-gated, effect-weighted volume over Loewe additivity, adjusted for variable dilution factors f_X,f_Y. Other combination reference models can be used to calculate S, and in this work we are referring all synergy scores to the 'highest single agent' Z_HSA = max(Z_X,Z_Y), the maximum single-agent response at corresponding concentrations²⁴. Combinations with strong synergies above the reference model that also occur at high effect levels score well, and regions of negative synergy on a surface do not cancel out those with positive synergy. It is worth noting that S does not have a natural scale, so it is only useful for relative comparisons within a screen, rather than as an absolute measure of strong synergy.

Selectivity is calculated by comparing across dose matrices for different measured endpoints. The selectivity index SI compares the potency of an agent or combination at the same effect level for two assays (Fig. 1). Given a pair of response matrices for the test and control assays, we measure inhibitions Z_test and Z_ctrl, and define a fixed cutoff level Z_cut. We then find the effective concentrations C_cut at level Z_SI = 0.5 × max(Z_test) for each of the single agents and for fixed-ratio curves of the combination, where the combined concentration in molar units is taken as the sum of the component concentrations, and define a selectivity index SI = log₁₀(C_ctrl/C_test). If a single agent does not reach the chosen effect level, we set the effective concentration to the top tested dose, and SI values become lower limits. Each fixed ratio curve is extracted from diagonals in the test matrix, using bilinear interpolation to determine the corresponding control responses. For each test-control pairing, SI is reported for the most selective diagonal with more than two dose samples covered by both matrices. We also report the differential selectivity ΔSI = SI_comb − SI_best between the combination and the more effective single agent, and a combination index difference ΔCI = CI_ctrl − CI_test.

Measuring selectivity bias and its significance.

To measure the selectivity bias for synergistic combinations, we compared distributions of selectivity index SI between populations of combinations for each assay pair in a screen. The synergistic combinations were chosen to be those with a synergy score S in excess of a chosen cutoff S_cut (usually capturing the top 5% of the synergistic combinations). Their SI distribution was compared to that of the overall set of combinations via a selectivity bias B (Fig. 1).

Because the B can be subject to regression-to-the-mean effects, where stochastic noise generates meaningless selectivity for synergistic combinations, we calculated the synergy and selectivity using distinct replicate matrices (Fig. 1 and Supplementary Note 1). By doing this, stochastic noise in one copy will not affect the selectivity calculation in the other. Although most of our screens had replicates, some screens collected only one. To ensure consistency, we analyzed all screens by separating each combination into distinct matrices using alternate dosing points from the consensus data. Results were compared to the same analysis on true replicates, when available, to confirm the accuracy of this approach.

For each pair of assays in a screen, we compared four SI distributions: (i) the synergistic combinations with S_cut > S_cut, all combinations without synergy filtering, and most selective single agent from each combination (Fig. 1). For each pairwise comparison of assays, we report B with its 95% range (2 σ_B), and a confidence P-value from Poisson statistics³⁰ representing the probability that χ² differences as large as those between the SI distributions for the synergistic and unfiltered combinations are indistinguishable.

When analyzing the significance of selectivity bias measurements across all the screens (Fig. 3), we accounted for the number of pairwise comparisons using a sequential Bonferroni correction⁴⁷ to adjust the error bars. Each screen's B estimates were assigned an integer rank with increasing signal-to-noise ratio B/σ_B, and then each σ_B was increased by a factor corresponding to 95% confidence after adjusting for multiple tests equal to their rank. In doing this, the strongest bias values are given the largest correction to account for their being the best of all of the assay pairs, and the weakest B retains the 95% confidence error bars corresponding to its standard error from the SI distribution.

Hepatitis C assays.

Huh7 (human hepatoma) cells expressing a subgenomic RNA replicon of Con1 (genotype 1b) sequence origin and expressing the reporter enzyme luciferase were obtained from ReBLikon. Antiviral assays were performed by seeding 4,000 cells/well in a 384-well plate in a total volume of 30 μl/well and incubating at 37 °C, 5% CO₂ overnight. Prediluted compounds were added at a 10 × concentration to each well to achieve the desired final concentration. Assay plates were than incubated for 48 h or 37 °C, 5% CO₂. To equilibrate plates to 20 °C, assay plates were removed from the incubator for 30 min to 1 h before the addition of 25 μl/well of SteadyLite luciferase assay reagent from PerkinElmer. Cells were incubated with SteadyLite reagent for 10 min before collecting data with a luminometer (PerkinElmer Envision). Antiviral activity is quantified by the inhibition of luciferase activity. For the antiproliferation assay Huh7 parental cells, which do not express HCV replicon RNA are treated similarly to the above replicon cells; briefly, seed cells on a 384-well plate at 4,000 cells/well, as described above. Compounds are added the following day and, after subsequent 48-h incubation at 37 °C, 5% CO₂, 15 μl/well of ATPlite (PerkinElmer) is added after plates have been equilibrated to 20 °C. The ATPlite assay provides a quantitative measure of the levels of ATP in the cell cultures in each well, where higher levels of ATP correlate with greater cellular viability.

RSV assays.

Human lung-derived HEp-2 cells were obtained from Diagnostic Hybrids, and maintained in DMEM modified with 10% FBS, 1% P/S (penicillin/streptomycin) and 1% GlutaMAX-1. Antiviral assays were performed by seeding 3,000 cells/well in a 384-well plate in a total volume of 30 μl (in DMEM modified with 2% FBS and 1% GlutaMAX-1) and incubating at 37 °C, 5% CO₂ overnight. Prediluted compounds were added at a 10 × concentration to each well to achieve the desired final concentration. RSV was added to each well (save 12 no-virus added wells per plate) at 50 times the tissue culture infective dose. Assay plates were than incubated for 96 h at 37 °C, 5% CO₂. To equilibrate plates to room temperature, assay plates were removed from the incubator for 45 min before the addition of 25 μl/well of ATPlite assay reagent from PerkinElmer. Cells were incubated with ATPlite reagent for 10 min before collecting data with a luminometer. Antiviral activity was quantified by comparing the host cell viability in the presence of virus and drug treatment to uninfected cells. For the antiproliferation assay, HEp-2 cells were treated similarly to the above conditions, without the addition of virus. The antiproliferation assay provides a quantitative measure of the levels of ATP in the cell cultures in each well, where higher levels of ATP correlate with greater cellular viability.

Bacterial proliferation.

Methicillin-resistant S. aureus (American Type Culture Collection (ATCC)) bacteria were cultured in flasks with Basic Mueller Hinton broth overnight in a 37 °C shaker at 250 r.p.m. On the next day, the cultures were further diluted into Basic Mueller Hinton broth to have an absorbance equivalent to 50% of McFarland standard, corresponding to a density of ∼10⁸ colony-forming unit (CFU)/ml, and transferred with automated dispensers into 384-well assay plates (35 μl wells containing media and test compounds) to yield ∼10⁶ CFU/ml. The plates were incubated in the presence of drugs and media at 37 °C for 18 h (∼30 doublings). Upon completion, cell populations were measured at a single time point, with a turbidity (laser scattering) readout using a BMG Labtech Nephelometer. For the anthrax selectivity example, Bacillus thuringiensis (ATCC) was grown as above, but activity was determined by the addition of basic Mueller Hinton medium containing 10% Alamar Blue fluorescence dye. After a 4 h incubation at 37 °C, bacterial populations measurements were made at a single time point using a PerkinElmer Victor II plate reader (excitation at 535/590 nm emission).

Anthrax toxicity.

Raw 264.7 cells were plated on 384-well plates at a density of 15,000 cells/well and are incubated overnight at 37 °C, with 5% CO₂. Next, 4.5 μl of diluted compound was added to each well followed by the addition of 10 μl of anthrax lethal toxin (500 ng/ml PA-63 and 500 ng/ml Anthrax Lethal Factor from List Biological Laboratories). Assay plates were than incubated for 4.5 h at 37 °C, with 5% CO₂. Cell cytotoxicity was then determined by the release of lactose dehydrogenase (LDH) using a fluorescent CytoTox-One assay (Promega). After assay plates had equilibrated at 20 °C for 30 min, 30 μl/well of CytoTox reagent was added and assay plates were incubated for 10 min. Then 15 μl/well of Stop Solution was added to assay plates and the amount of LDH released was measured using PerkinElmer Envision.

Cancer proliferation.

A549 (ATCC #CLL-185), Colo-205 (ATCC #CLL-222), H929 (ATCC #CRL-9068), HCT116 (ATCC #CCL-247), NCI-H460 (ATCC #HTB-177), SK-MEL-28 (ATCC #HTB-72), SKOV-3 (ATCC #HTB-77), RPMI-8226 (ATCC #CCL-155), MM.1R and MM.1S (kindly provided by Steven Rosen) were cultured in RPMI-1640 media Supplemented with 10% FBS for between 2 and 20 passages, drawing test populations along the way. Cells were seeded into 384-well plates at a density of 1,500 cells per well in 35 μl media using automated dispensers. Plated cells were incubated at 37 °C, with 5% CO₂ overnight, after which the compounds were added and the plates were incubated again for 72 h (∼1–2 doublings). After incubation, cells were assayed for viability by measuring relative ATP levels using ATPLite 1Step (PerkinElmer) as per the manufacturer's instructions or 10.5% Alamar Blue fluorescence dye (in 40 μl media) was added and plates were incubated for 6 h. Cell viability measurements were taken at a single time using a PerkinElmer Envision plate reader.

Cardiovascular homogeneous time-resolved fluorescence (HTRF) assays.

Human coronary artery endothelial cells (HCAEC, Cell Applications) or smooth muscle cells (HCASMC, Cell Applications) were seeded into 384-well plates at a density of 1,500 cells/well in 35 μl supplemented media using automated dispensers. Cells were incubated at 37 °C, 5% CO₂ overnight to allow the cells to adhere, treated with compound, and stimulated with interleukin (IL)-1β (BD Biosciences) 1 h later at 3 ng/ml. The plates were incubated again for 24 h, after which 9 μl of cell supernatant was transferred into a corresponding low-profile, 384-well plate containing 9 μl of premixed HTRF reagents (functional grade purified anti-human MCP-1 antibody, Clone: MD3-F7 eBioscience, custom labeled with Eu-cryptate by CisBio, at a f.c. of 0.88 nM; biotin anti-human MCP-1 antibody, Clone: 2H5, eBioscience, at 3.35 nM f.c.; streptavidin labeled XL-665, CisBio, at 16.75 nM f.c.). HTRF signal was read from the low-profile plates after an overnight incubation at 20 °C using the LANCE-HTRF Eu/APC Dual protocol on PerkinElmer Envision plate reader. To assess viability based on ATP levels after removing the supernatant for the HTRF assay, the cell culture plate was treated with ATPLite 1Step reagent, and the luminescence for each well was read on the Envision plate reader.

Cardiovascular antiproliferative assays.

Adult normal human dermal fibroblast(Lonza), human aortic smooth muscle cells (ATCC), or human umbilical vein endothelial cells (ATCC) cells were cultured in fibroblast basal medium with provided supplements and growth factors(Lonza) for 2–10 passages. Test populations were seeded at 250 cells/well in 384-well plates and allowed to recover overnight at 37 °C with 5% CO₂. Diluted compounds were added to the cells and incubated at 37 °C with 5% CO₂ for 72 h. Cells were then assayed for viability by measuring relative ATP levels as described above.

IL-1β enzyme-linked immunosorbent assay (ELISA).

The inhibitory effect of certain compounds or combinations on the secretion of IL-1β was assayed using the standard sandwich ELISA. Assay plates were prepared by adding lipopolysaccharide-stimulated complete RPMI media into 384-well plates. Master plates were appropriately diluted into assay plates and purified PBMC (20,000 cells/well) were added. Plates were incubated at 37 °C, 5% CO₂ overnight and spun to clear the supernatant, which was transferred to ELISA plates coated with IL-1β capture antibody. After a 2h incubation at RT, plates were washed with 1 × PBS, 0.1% Tween 20, and a second IL-1β antibody and the detection agent (Europium-labeled streptavidin) were added. Plates were incubated again for 1 h at RT, and after several washes, enhancement solution was added. After an overnight incubation at 4 °C, plates were read on a PerkinElmer Envision plate reader.

Inflammation assays.

The ability of compounds or combinations to suppress the secretion of pro-inflammatory cytokines was assayed as follows. Compound stocks dissolved in DMSO (or H₂O as appropriate) were serially diluted on master plates using liquid-handling automation. Master plates were diluted into plates with aqueous media, stimulants (phorbol-12-myristate 13-acetate and ionomycin) and serum. Human buffy coat preparations from peripheral blood were obtained fresh daily from the blood bank and diluted in supplemented media before addition to the assay plates. Plates were incubated overnight at 37 °C and 5% CO₂ and spun to pellet the cells. After transferring the supernatant to an ELISA plate coated with a capture antibody specific for the target cytokine, plates were washed and probed with a second antibody and detection reagent. Data were read from the ELISA plate using PerkinElmer Envision readers.

Htt protein translocation assay.

Immortalized striatal cells derived from mutant huntingtin (Htt) knockout STHdhQ111 mice were seeded at 8,000 cells/well in a collagen I coated 384-well plate (BD Biosciences). Compounds were added and cells were grown overnight in 33 °C incubator with 5% CO₂. Cells were fixed with formaldehyde, permeabilized with 0.5% Triton X-100, blocked with 0.5% BSA and stained with 1F8 antibody from Main Biotechnology Services(1:2,000). Nuclei were defined by Hoechst staining. Htt localization used a Cy3-labeled anti-mouse secondary antibody followed by PBS wash. Images were acquired using Cellomics AssayScan VTi 5.0 under a 10 × or 20 × objectives, using automatic focus every other well. Perinuclear staining was quantified using the Compartmental Analysis BioApplication and “RingSpotTotalInten” as endpoint. Cell viability and cell numbers were quantified using the Compartmental Analysis BioApplication and “ObjectPerField” as endpoint.

Norepinephrine induction assay.

For the NE experiment (Supplementary Note 4), purified human T-cells were cultured as triplicate samples for each condition. First, an inflammatory response was stimulated with αCD3 and αCD28 antibodies for 30 min, after which cells were treated with drugs as indicated. After incubating for 18 h at 37 °C with 5% CO₂, supernatants were collected and TNFα was quantified using cytometric bead arrays.

Rat asthma model.

Brown Norway rats (n = 10/cohort) were sensitized with intraperitoneal ovalbumin (OVA) given on days 0, 7 and 14. Animals were treated with test agents by oral gavage 1 h before intranasal ovalbumin (OVA) challenge for 1 h on day 21. Lungs were lavaged 72 h after exposure of OVA to quantify total leukocytes and specific cell populations in bronchoaveolar lavage fluid.

Rat pain assay.

Male Sprague Dawley rats (200–225 g) were dosed with drugs three times (2 d, 1 d and 30 min) before the pain induction. Cohorts of eight rats were treated for each treatment level, along with vehicle and diclofenac (positive control, 25 mg/kg) cohorts. After administering carrageenan, the average retraction force in response to a pain stimulus for the injected and uninjected hind paws was measured with Von Frey filaments at 20, 40, 60, 80 and 120 min post injection. The data for both hind paws were averaged after subtracting a baseline for each animal from measurements on day 2, and the area under the curve was integrated over the full time range.

Rat toxicity model.

Cohorts of five male Sprague Dawley rats(Charles River Laboratories), with average starting body weight 178–186 g, were given intraperitoneal doses for 10 d, the final dose given 2 h before authorization. Livers were removed and stored in an RNAlater (Ambion) at 4 °C. Samples were homogenized using TissueRuptor (Qiagen) and total RNA was isolated using the RNeasy-Plus Mini kit (Qiagen). Equal amounts of total RNA were used for one step RT-PCR (QuantiTect Probe, Qiagen). Taqman gene expression assays, Applied Biosystems reagents were used for detection of TAT and β-actin (endogenous control) mRNA using the Applied Biosystems 7300 Real-Time PCR System. The TAT expression values were calculated using the formula relative expression = 2^ΔTAT/2^ΔArbp, where Δgene is the difference between the drug- and vehicle-treated expression levels, as measured by the cycle threshold levels for that gene.