Algorithmic guided screening of drug combinations of arbitrary size for activity against cancer cells

The standard treatment for most advanced cancers is multidrug therapy. Unfortunately, combinations in the clinic often do not perform as predicted. Therefore, to complement identifying rational drug combinations based on biological assumptions, we hypothesized that a functional screen of drug combinations, without limits on combination sizes, will aid the identification of effective drug cocktails. Given the myriad possible cocktails and inspired by examples of search algorithms in diverse fields outside of medicine, we developed a novel, efficient search strategy called Medicinal Algorithmic Combinatorial Screen (MACS). Such algorithms work by enriching for the fitness of cocktails, as defined by specific attributes through successive generations. Because assessment of synergy was not feasible, we developed a novel alternative fitness function based on the level of inhibition and the number of drugs. Using a WST-1 assay on the A549 cell line, through MACS, we screened 72 combinations of arbitrary size formed from a 19-drug pool across four generations. Fenretinide, suberoylanilide hydroxamic acid, and bortezomib (FSB) was the fittest. FSB performed up to 4.18 SD above the mean of a random set of cocktails or “too well” to have been found by chance, supporting the utility of the MACS strategy. Validation studies showed FSB was inhibitory in all 7 other NSCLC cell lines tested. It was also synergistic in A549, the one cell line in which this was evaluated. These results suggest that when guided by MACS, screening larger drug combinations may be feasible as a first step in combination drug discovery in a relatively small number of experiments. [Mol Cancer Ther 2009;8(3):521–32]

Combining chemotherapeutic agents is an established way to improve on single-drug efficacy, and small (two- to three-drug) combinations are now the standard of treatment for most metastatic cancers (1). Recent successes with larger multiple-agent regimens in the clinic (1, 2) encourage a continued search for improved combination chemotherapies; however, only a few of the virtually infinite possible combinations have been evaluated to date (3, 4) due to concerns about the potential toxicity of the combinations and the daunting logistics of testing all feasible combinations of available drugs (5).

The improved tolerability of new, potentially effective molecular targeted agents has encouraged combination of these agents with standard chemotherapy regimens and, more recently, with each other in early-phase clinical trials (1, 2, 6). In a recent review, it was observed that there is “tremendous” potential for improved therapy of cancer from combinations of molecular targeted agents designed to modulate different aspects of the same or different targets (7). Traditionally, strategies of combination drug development have been based on insights into non–cross-resistant molecular mechanisms of action, evidence of synergy, preclinical or clinical insights into nonoverlapping toxicities, and, more recently, insights into their effects on signaling pathways and host-tumor interactions (8, 9).

As a complement to these methods, we investigated the utility of a functional, laboratory-based screening of combinations. Laboratory-based screens have long been used to identify active single agents, and one company explored doublets using synergy (10). However, to our knowledge, there are no ongoing efforts to screen large spaces of potential combinations of more than two drugs. Extending a screen to larger cocktails may be useful given the time required to compile clinical evidence. However, in vitro assessments of any given cocktail become more challenging as drug combinations use larger numbers of agents. Moreover, even with highly efficient assessments of individual combinations, only a small fraction of the total possible cocktails can be evaluated given that they are virtually countless. We made an effort to meet these challenges on two levels. First, we developed a rapid method, or fitness function alternative to synergy, to speed evaluation of individual cocktails. Second, we used a search algorithm intended to sequentially direct the screen toward increasingly more promising sets of drug combinations.

Fitness is a term used in search algorithms adapted from evolutionary theory. In biology, higher fitness describes an individual's greater capacity to reproduce. In search algorithms, it indicates a more optimal solution; what defines fitness is determined by the investigator and the problem that he or she is trying to solve. These algorithms guide a process through a finite list of operations intended to maximize fitness. Such algorithms have been extensively used outside medicine in applications such as data fitting, scheduling, trend spotting, and budgeting (11–13). They have been evaluated in medicine in breast cancer diagnosis (14) and in an antiviral in vitro model assessing a large space of dose permutations of a few cocktails (15). Here, a well-established strategy is applied to a novel setting: the discovery of drug cocktails. The algorithm works through a stepwise process of enriching successive pools of cocktails for increased fitness. It is this application we designate Medicinal Algorithmic Combinatorial Screen (MACS).

For MACS to be meaningful, fitness needs to predict a value of interest, and to be feasible, the fitness function needs to be efficient. We intended fitness to predict cocktail efficacy. Synergy is an established fitness function used for this purpose. Indeed, synergy was used by CombitoRx in evaluating doublets (10). However, synergy can become exponentially more labor-intensive with each drug added to a given cocktail. As an example, when six doses per drug are used, as done by CombitoRx, cocktails containing two, three, four, or five drugs generate 36, 216, 1,296, and 7,776 possible dose permutations, respectively. Although substantially smaller numbers of permutations could be managed by fixed dosing ratios or other ways of sampling, synergy may nonetheless remain daunting. Thus, we developed an alternative fitness function that required one assay per cocktail to improve the efficiency on the level of the individual cocktail. It was a composite of inhibition at a fixed dose of each drug and the number of drugs. Through the use of this novel fitness function, we lost information about a vast number of potentially informative dose combinations, but it enabled us to screen hundreds of unique cocktails per week. Other fitness functions could incorporate drug cost, anticipated toxicities, molecular insight, dose sequence, alternative dosing schemes, or any other characteristics of interest.

As with fitness functions that assess individual cocktails, there are countless possible search algorithms to guide screens of large sets of cocktails with different ones suited to varied combinatorial landscapes (Supplementary Fig. S1).⁸

We present a simple example of such an algorithm called hill climbing (Fig. 1A; Supplementary Fig. S1).⁸ This example does not represent actual data but rather illustrates a method. As can be seen in the figure, one can perform a MACS by sampling a relatively small subset of the possible cocktails (Supplementary Fig. S1).⁸ Although the rules can vary among MACSs, they all entail a cycling between experimental evaluation of drug cocktails and a rule-based creation of the next generation (Fig. 1B). These rules involve selection and alteration. The selection function directs the choice of a subset of the cocktails tested, and in the case of the hill climb, the fittest one. However, it could choose more than one cocktail and even incorporate a stochastic element analogous to natural evolution in which fitness improves the odds of survival but offers no guarantees (as presented previously).^9,10 The alteration function can create all of the nearest neighbors of the selected cocktail(s), analogous to mutations as in the hill climb, or could combine pieces of different selected cocktails analogous to recombination.

Because any search of a large enough potential space will, for practical reasons, be confined to a small fraction of the total possible cocktails, it is likely that many of the fittest combinations will never be tested even when a search algorithm guides the process. The MACS needs only to find combinations that have fitness levels that are “good enough” more efficiently than a random screen to be useful.

In this proof-of-principle study, we developed a novel fitness function to evaluate cocktails and tested a number of different kinds of MACSs operating on combinations all derived from the same pool of 19 available anticancer drugs in an in vitro non–small-cell lung cancer (NSCLC) model. Based on data from earlier MACSs showing that a local search method such as a hill climb would be efficient (as compared with the more complex algorithms we evaluated earlier),^9,10 here we tested a hill climb-MACS. We then ran validation assays on the fittest identified cocktail in line with the general strategy in most drug development screens.

Unlike the model (Fig. 1A), the actual laboratory experiments composing the hill climb entailed 19 real agents creating a combinatorial space of 2¹⁹ or 524,288 possible cocktails. Eighteen cocktails were randomly formed for its generation 0 (as previously described^9,10) and subsequent generations were created based on the assay results and the rules of the hill climb. The assay systems and conditions as well as fitness function were held constant throughout the hill climb-MACS.

Our fitness function permitted one assay per cocktail to be done in triplicate. It was composed of a measured assay value representing inhibition and a nonmeasured value, a factor of the number of drugs in the cocktail. Starting at 0, 0.1 was subtracted for each 10% of cells remaining compared with the no-treatment controls. Thus, 70% inhibition yields 30% of the no treatment control value, or 0.3 to be subtracted. A “penalty” of 0.1 was assigned for each drug composing the cocktail based on the observation that each drug dosed at IC₁₀ added to a cocktail increased inhibition by an average of 10% (data not shown). This penalty was designed to control for cocktail size to allow fitness to be determined by the cocktail composition.

Formally, fitness was −K(s) − 0.1|s|, where K(s) is the proportion of cells surviving relative to the no-treatment control and |s| was the number of drugs in the cocktail. Fitness values thus can range from −0.1 (one drug kills all cells) to −2.9 (all 19 drugs with no inhibition). Even lower fitness values would be possible with growth stimulation relative to the untreated controls.

Nineteen agents, 16 molecularly targeted and 3 chemotherapy, were used based on their availability, affordability, and ease of preparation and storage and the authors' experience in their use (Table 1; Supplementary Methods). Selected agents were required to inhibit A549 growth by at least 10% (IC₁₀) under the conditions used in the MACS (data not shown). A549 was chosen because it grows rapidly under uncomplicated laboratory conditions. Drugs requiring excessive concentrations of DMSO were excluded given the independent ability of DMSO to inhibit A549 at doses of 1.7% by volume (data not shown). Drugs dissolved in ethanol were excluded because ethanol potentiated inhibition of A549 by DMSO.

Eight human NSCLC cell lines were used to test the activity of the lead drug combination. A549, H226, Calu-3, H441, H1975, and HCC827 were purchased from American Type Culture Collection. HCC2279 was a gift from John Minna (Southwestern Medical School, Dallas, TX), and PC-14 was purchased from Riken BioResource Center. A549, H226, and Calu-3 cells were incubated in DMEM/F-12 (Invitrogen Corp.) containing 10% fetal bovine serum and 1% penicillin/streptomycin. H441, H1975, HCC827, HCC2279, and PC-14 cells were incubated in RPMI 1640 plus 10% fetal bovine serum.

Cells were isolated from BD Falcon 75-cm² tissue culture–treated flasks after a rinse with Cellgro PBS (Mediatech, Inc.) and the addition of Gibco trypsin-EDTA (0.05% trypsin, with EDTAx4Na) from Invitrogen. The cells were then plated at a volume of 50 μL/well in 96-well MicroWell Nunclon-Δ plates (Nalge Nunc International). Plated cells were then stored at 37°C in a humidified atmosphere containing 5% CO₂ until the drugs were added. The numbers of cells per well were as follows: A549, 1,500; H226, 3,000; PC-14, 2,500; HCC2279, 4,000; HCC827, 7,000; H441, 2,700; H1975, 6,000; and Calu-3, 15,000. All cell lines were incubated on Nunc 167008 96-well plates (Fisher Scientific International, Inc.).

Drugs for a given combination were mixed in a 96-well 2-mL sterile polypropylene block from Denville Scientific. All combinations were mixed at four times the volume needed for one well to allow for triplicate experiments and pipetting error. The final volume of DMSO (0.66%) in all wells, including the no treatment controls, was based on the maximum possible DMSO dose derived from the one cocktail containing all 19 drugs. Because each cocktail had its own individual DMSO volume, we added the unique additional DMSO individually to each cocktail, yielding 0.66% of a final volume of 250 μL. Medium was added to obtain a volume of 200 μL, which was in turn added to a well already containing the cells and media plated immediately before occupying 50 μL, yielding 250 μL.

We used the single-agent IC₁₀ dose (1× dose) regardless of the size of the combination because we anticipated that a higher IC value would result in maximal inhibition from many of the small cocktails, thereby reducing our ability to compare cocktails. All combinations were tested in triplicate in adjacent wells on the same plate, and each plate contained its own six untreated controls. All cocktails and controls were within the 60 non-edge wells to avoid an edge effect.^9,10 With each generation, we included 1×, 2×, and 4× single-agent controls in triplicate for all 19 drugs on a separate plate to check for consistency.

Cells were isolated from BD Falcon 75-cm² tissue culture–treated flasks after a rinse with Cellgro PBS (Mediatech) and the addition of Gibco trypsin-EDTA (Invitrogen). The cells were plated at a volume of 50 μL/well in 96-well MicroWell Nunclon-Δ plates (Nalge Nunc International). The numbers of cells per well were as follows: A549, 1,500; H226, 3,000; PC-14, 2,500; HCC2279, 4,000; HCC827, 7,000; H441, 2,700; H1975, 6,000; and Calu-3, 15,000. Cell lines were incubated on Nunc 167008 96-well plates (Fisher Scientific International). After the drug mixture was added, the plates were incubated for 44 h at 37°C in a humidified atmosphere containing 5% CO₂. This permitted reproducible measurements and a repeat assay during the workweek.

Cell proliferation was measured using the WST-1 Cell Proliferation Reagent (Roche Diagnostics). After 44 h of incubation, 25 μL of WST-1 reagent were added constituting 9% of the well volume and the plates were incubated for another 4 h. The absorbance was measured on a scanning multiwell spectrophotometer at 440 nm with a 600-nm reference.

All eight possible subcombinations of FSB—fenretinide, suberoylanilide hydroxamic acid, and bortezomib—(F, S, B, FS, FB, SB, FSB, no treatment control) and ASB in which fenretinide is replaced by all-trans retinoic acid (A, AS, AB, ASB) were tested in triplicate in eight NSCLC cell lines including A549. The doses were held stable in all cell lines at the IC₁₀ for A549, and no edge wells were used. Each plate contained six untreated controls. An additional plate of A549 cells containing 1×, 2×, and 4× doses in triplicate of each drug and six no treatment controls was done with each experiment.

We assessed the synergy of FSB in A549 and H2226 using all 64 possible dose combinations derivable from four dose levels: zero, 1× (high), 0.66× (medium), and 0.44× (low), done in triplicate. We used the same doses in each cell line to mirror the clinical case in which the drug titer in the serum affects the tumor as a whole unvaried from one clone to another. No edge wells were used and each plate contained six untreated controls. We also did single-agent controls on a separate plate at the 1× (high), 2×, and 4× doses for both cell lines. These values were also used in the synergy calculations. The combination index was calculated by the Chou-Talalay equation. In cases in which single agents stimulated growth, the value 0.000,001 was used to allow a calculation.

No formal statistics were used to estimate the fitness. Statistical analyses of the MACS data and proliferation assays were carried out in Excel (Microsoft, Inc.) and SPlus (Insightful, Inc.) using ANOVA. The Kolmogorov-Smirnov test was run on the fitness values on sets of randomly formed cocktails. Analyses of synergy were done using CalcuSyn software (Biosoft Ltd.; refs. 16, 17).

An assessment of two sets of randomly formed cocktails showed that each additional drug increased inhibition by ∼10%, supporting the 10% penalty for the fitness function (Supplementary Fig. S2).⁸

The hill climb-MACS arrived at the combination, FSB, in only three generations (generations 0–2). It was confirmed in the 4th generation (generation 3) as fittest in its immediate neighborhood after assaying a total of 72 unique combinations over 2 weeks (Figs. 2A and 3D). This is the same cocktail identified from MACSs described elsewhere.^9,10 We compared FSB and its nearest neighbors to two separate sets of randomly formed cocktails: 18 cocktails forming generation 0 of the hill climb and 30 cocktails derived from earlier experiments.^9,10 To compare to the 18 cocktails, we used the lower of two fitness values for FSB from generation 3 in the hill climb [K(s) = 0.26; fitness, −0.56]. FSB was 2.57 SD above the mean of its generation 0 (mean, −1.09; SD, 0.207) or in the upper 0.51% of the distribution. We had previously compared this same set (FSB and its nearest neighbors) by testing it the same day as the 30 randomly formed cocktails.^9,10 FSB [K(s) = 0.24; fitness, −0.54] performed 4.18 SD above the mean or in the upper 0.0014% of the distribution of these 30 cocktails assuming a normal distribution of fitness levels (Fig. 2B).^9,10 A Kolmogorov-Smirnov test on multiple sets of randomly formed cocktails derived from these earlier experiments (including the 30 mentioned above) produced a P value of 0.5, indicating a close approximation of normality.

In the hill climb, there was stability of the single-agent controls across the generations (Fig. 2C; Supplementary Fig. S3)⁸ but a statistical improvement in the mean performance of the latter generations compared with the mean of the baseline generation (Figs. 2D and 3). Likewise, FSB and its nearest neighbors had a mean fitness value of −0.64 (SD, 0.07) compared with a mean fitness value of −1.18 (SD, 0.15) for the 30 randomly formed cocktails (Fig. 2B). Thus, the hypothesis that the means of the two distributions were the same was easily rejected (t = 17.1).

A closer look shows a steady enrichment for FSB. Among all 18 randomly formed cocktails from generation 0, the 6 most inhibitory cocktails contained a doublet derived from FSB (Fig. 3). In generation 1, if the parent, Bor,Ly,Sa, had inhibited slightly more, it would have been fittest and the hill climb would have stopped. By generation 2, all the nearest neighbors of BS were tested and FSB was fittest. In generation 3, FSB as parent was the fittest in its immediate neighborhood and the hill climb stopped. Note that subtraction of any of the drugs from FSB increased the K(s) sufficiently to render the doublets less fit. Moreover, addition of any drugs did not yield additional improvements in K(s) because FSB was maximally inhibitory under the laboratory conditions we used.

We compared FSB to 22 other triplets tested in an earlier MACSs.^9,10 The K(s) values were 0.43 to 0.99 (fitness, −0.73 to −1.29), compared with 0.18 − 0.26 (−0.48 to −0.56), among the six values for FSB in all MACSs.^9,10 FSB was also superior to an additional 20 triplets examined through the hill climb with K(s) 0.28 to 0.88 (−0.58 to −118; Fig. 3) although these triplets were highly enriched for FSB.

FSB was found to be synergistic in A549 (Fig. 4). In H226, synergy could not be calculated for most dose combinations given the single-agent stimulatory activity of suberoylanilide hydroxamic acid despite assigning a value of 0.000,001 to allow calculations (Fig. 4). This stimulatory behavior confirms earlier dose response results (Fig. 5A) and experiments discussed below (Fig. 5B). Although suberoylanilide hydroxamic acid increasingly stimulated H226, as its doses increased as a single agent and in combination with fenretinide, it had minimal effect when paired with bortezomib and magnified inhibition when combined with SB (Fig. 4).

Across all eight NSCLC cell lines, FSB was uniformly inhibitory, whereas the doublets were not and the single agents even less so (Fig. 5; Supplementary Table S1).⁸ In all cases, FSB was more inhibitory than its constituent doublets. The substitution of all-trans retinoic acid, a congener of fenretinide, when averaging across all eight cell lines, yielded less inhibitory doublets and triplet despite the fact that all-trans retinoic acid, on average, was given at a more inhibitory single-agent dose than fenretinide (Supplementary Table S1).⁸

To the best of our knowledge, this is the first study that shows the feasibility for screening drug combinations of arbitrary size. It may also be the first using the strategy of a search algorithm–guided screen of cocktails (i.e., a MACS as described earlier^9,10). The hill climb-MACS quickly and automatically identified a highly fit combination, FSB, the same cocktail identified in an earlier MACS.^9,10 It did so without the benefit of prior molecular insight into the interactions of the combined drugs. In addition, use of a novel fitness function substantially increased efficiencies at the level of the individual cocktail as compared with synergy thereby facilitating MACS in our lab.

Enrichment through the generations of sets of cocktails and the superior fitness of the FSB itself indicate that the hill climb-MACS was operating effectively (Fig. 2). This enrichment is also represented by the progressively increased presence of cocktails containing drugs derived from FSB, which were more inhibitory, a behavior that also indicates the upward directedness of the screen (Fig. 3). Similar behavior was also observed in the earlier MACSs.^9,10 Indeed, a MACS, whatever its rules, depends on the presence of a relationship between cocktail composition and fitness because the search has to “learn” from prior experience.

In the case of the hill climb-MACS, a triplet in the initial generation was highly enriched for FSB, by chance, placing it relatively high up a peak on the fitness landscape (Fig. 3A). This may in part explain the rapidity with which the hill climb identified FSB as well as its relatively low value of 2.57 SD above the mean fitness of its generation 0. Indeed, of 969 possible triplets, only 4 (0.41%), including FSB, contain at least two of the drugs. FSB at 4.18 SD above the mean needs to be interpreted with caution because the random set of 30 cocktails and its controls included edge wells whereas FSB and its nearest neighbors and controls did not (despite both sets being tested on the same day).^9,10 A number of post hoc analyses controlling for the edge effect yielded SD values ranging from 3.4 to >5 SD above the mean (data not shown). Even at 3.4 SD, FSB is still higher than would have been expected if cocktails were randomly tested (368 cocktails were evaluated through the earlier MACSs). The interpretation of the significance of the number of SD also depends on the assumption of a normal distribution of fitness values. Although an analysis of a larger set of 60 randomly formed cocktails showed a close statistical approximation to normality, we expect the tail of the distribution to “fatten” as cocktails fully inhibit by being quite large. Nonetheless, it does seem that FSB performed better than expected based on the above and other considerations.

In addition, FSB was substantially fitter than any of the other unrelated 22 triplets tested; thus, its fitness is not just a simple consequence of the fact that it is a triplet. Some caution is warranted because some of these triplets and their respective controls were tested using edge wells whereas FSB and its controls were not. However, the edge effect is substantially less than the observed differences in inhibition between FSB and the other triplets (data not shown).

The fact that FSB, which had been identified in an earlier MACS,^9,10 was identified a second time in the hill climb after screening only 72 cocktails was highly unlikely by chance given the space of a total of >500,000 potential cocktails and the many 10,000s of smaller potential cocktails. This also suggests that this particular fitness landscape is composed of a few higher peaks given the improbability of ascending the same peak twice by chance otherwise (Supplementary Fig. S1).⁸ Interestingly, the hill climb nearly stopped at the triplet, Bor,LY,Sa, as it ascended the peak. Indeed, if by chance it had been a slightly more inhibitory value, the hill climb would have stopped. This indicates the potential benefit in performing a number of hill climb-MACS runs even with simpler landscapes.

With complex multipeaked landscapes, a hill climb-MACSs may arrive at different highly fit cocktails on different runs, thereby supplying a number of varied candidate cocktails (Supplementary Fig. S1).⁸ However, such fitness landscapes, especially when composed of many lesser peaks, are often more successfully canvassed by more complex search algorithm–guided screens such as the hybrid algorithms we initially tested.^9,10 Indeed, future research can entail a determination of how factors such as assay type, constituent drugs, and varied fitness functions can influence the topography of fitness landscapes.

To speed the evaluation of individual cocktails, we used a fitness function alternative to synergy during the MACS. However, synergy remains important to consider given that it is a mainstay of current cocktail assessments and arguably more reliably predicts clinical efficacy because it integrates multiple dosing results per cocktail into a single synergy value. However, synergy experiments are increasingly labor-intensive as cocktails get larger with exponential increases in dose permutations. Moreover, such cocktails will likely have complex dose-response landscapes as hinted at by our results (Fig. 4). These likely cannot be neatly summarized by a synergy value. In addition, stimulatory values are not accommodated by standard synergy software, an issue that can arise even with single agents when using more than one cell line (Figs. 4 and 5A and B). One alternative to synergy to evaluate such dose-response combinatorial spaces is to use a search algorithm to guide a screen. This was recently reported by Wong et al. (15) on a space of 1,000,000 dose permutations derived from six drugs at 10 different doses per drug. Their strategy was similar to that used by us in this article and as previously described^9,10 but in their case instead directed at dose permutations. They used viral inhibition of a single-dose combination as a fitness value rather than synergy and applied a stochastic search algorithm to arrive at a highly fit dose combination. However, one could examine dose-response fitness data to identify regions of the space simple enough to perform standard synergy analyses. Such spaces may be yet more amenable to assessment using novel analytic synergy methods (18). Molecular insights may further direct analyses of this space.

Although the method reported by Wong et al. was highly efficient and reproducible, it required weeks to complete a single run to arrive at the highly fit dose combination for a given cocktail. Their method applied to a large space of possible cocktails would thus be less speedy per cocktail, although more informative, than the fitness function we had used to evaluate individual cocktails. Future study can explore how to prioritize and optimize search efforts between the large space of different potential cocktails, as we have done, and the space of dose permutations per cocktail as done by Wong et al. (15). Yet other large combinatorial spaces formed from drug sequencing and altered dose durations, among others, may also merit exploration.

For MACS to be meaningful, the fitness result needs to be predictive. Future study could evaluate this including our fitness function, synergy, or any other function. Validation studies showing activity of FSB in multiple cell lines suggest some predictive potential of our simple fitness function. Moreover, it supports the hypothesis that combination therapy is more effective against a diversity of clones across patients and within heterogeneous tumors, a plausible observation given the experience in the clinic. There are also data from other labs supporting the positive interactions of some of the constituent doublets of FSB, although FSB seems to be a novel cocktail not previously studied (19–21). These reports indicate the power of existing and improving molecular insights in developing drug cocktails such as the doublets mentioned above. MACS is intended to complement current discovery methods for combination therapy based on molecular insights thereby speeding the existing process. It may also identify otherwise unanticipated candidate cocktails. Indeed, the potential to find cocktails unexpected by available insights is also indicated by results from a massive unguided screen of doublets by CombitoRx through which they found highly promising anticancer doublets formed from noncancer drugs (10).

The observation that there is a wide spectrum of NSCLC lines inhibited by FSB suggests a potential to inhibit host cells thereby drawing attention to the limitations of in vitro assays as predictors of toxicity. In the case of FSB, there are some clinical data indicating that FSB may be tolerable in the clinic (22–24). However, concerns for high risk of toxicity of drug cocktails argue the interest in development of in vitro toxicity assays, which in turn could be integrated into fitness functions and the MACSs. Indeed, the penalty we affixed to larger cocktails in our fitness function was a limited attempt to represent toxicity by favoring smaller cocktails. Whether FSB should be further developed or not, although there are substantial predictive limitations of existing in vitro assays, they are a foundation of drug development because they potentially offer some information of clinical value.

Separate from the question of feasibility and predictiveness of a fitness function are barriers to exploring the entire space of potential cocktails posed by limits in an assay. For example, we were unable to explore larger cocktails embedded with FSB, which, like FSB, maximally inhibited A549 but by definition were less fit. Alterations in assays or its conditions may expand the space of cocktails that can be meaningfully compared.

Although F, S, and B are not dosed at high single-agent inhibitory values relative to the other agents, they do have a disproportionately steep dose-response curve, with 4× dose inhibitory levels ranked 1st, 3rd, and 4th (all-trans retinoic acid was 16th; Table 1). MACS was not embedded with rules entailing prior knowledge of this attribute and yet efficiently found FSB. Nonetheless, such drug activity information or other functional data obtained through interrogating cocktail data sets generated during a MACS may usefully be embedded in a new algorithm or fitness function to improve MACS efficiencies.

Likewise, future MACSs could integrate biological insights by assigning a fitness “reward” for combinations that contain pairs of agents that are biologically tenable, or alternatively, one could choose pools of drugs that fit a molecular theme (or one could even use molecular rather than functional assays in the MACS). Contrariwise, MACSs founded in functional assays can be used to identify not only candidate therapeutics but also highly relevant candidate molecular probes. Insights derived from these probes could in turn be integrated into a new MACS, thereby forming a positive feedback loop of alternating MACSs and molecular evaluations. Limitations to molecular insights are exemplified by the fact that we would not likely have studied a fenretinide-containing regimen without a MACS given the recent diminished clinical interest in retinoids in lung cancer. They are also illustrated by the knowledge that all-trans retinoic acid is a congener of fenretinide and enabled only a limited capacity to predict efficacy (Supplementary Table S1).⁸ Thus, successful MACSs would likely not limit exploration to molecularly tenable cocktails.

Further studies of combinatorial landscapes and search algorithms as well as the development of efficient fitness functions and improvements in laboratory gold standards and clinical models to evaluate promising candidates need to be developed. Even if MACS is shown elsewhere to be a useful strategy, the need to conduct a lengthy series of laboratory tests will remain, and if still promising, a series of evaluations through clinical trials of candidate cocktails. In addition, we note that there may be some challenges for widespread adoption of MACS, particularly the contractual or legal difficulties that may present when testing cocktails with agents owned by different entities. Principles derived from the development of MACS in this oncology setting may also be applied on a wider basis, such as modeling and assessing therapeutic candidates for infectious, rheumatologic, or other diseases, providing additional impetus to further development of this potentially advantageous screening strategy.

No potential conflicts of interest were disclosed.

We thank Jean P. Issa, M.D., for kindly providing critical reagents; Dafna Lotan, M.S., for her superb assistance in drug preparation; Frank Fossella, M.D., Mandri Obeyesekere, Ph.D., and Gabor Balazsi, Ph.D., for carefully reading the text and their highly insightful editorial suggestions; and Kathryn B. Carnes and Suzanne Davis, M.B.A., for excellent editorial assistance.