Department of Molecular and Medical Pharmacology, University of California at Los Angeles, Los Angeles, CA 90095, USA

Department of Mechanical and Aerospace Engineering, University of California at Los Angeles, Los Angeles, CA 90095, USA

Department of Medicine, University of California at Los Angeles, Los Angeles, CA 90095, USA

Current Address: Johnson & Johnson Pharmaceutical Research & Development LLC, San Diego, CA 92121, USA

School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA

Abstract

Background

Cells constantly sense many internal and environmental signals and respond through their complex signaling network, leading to particular biological outcomes. However, a systematic characterization and optimization of multi-signal responses remains a pressing challenge to traditional experimental approaches due to the arising complexity associated with the increasing number of signals and their intensities.

Results

We established and validated a data-driven mathematical approach to systematically characterize signal-response relationships. Our results demonstrate how mathematical learning algorithms can enable systematic characterization of multi-signal induced biological activities. The proposed approach enables identification of input combinations that can result in desired biological responses. In retrospect, the results show that, unlike a single drug, a properly chosen combination of drugs can lead to a significant difference in the responses of different cell types, increasing the differential targeting of certain combinations. The successful validation of identified combinations demonstrates the power of this approach. Moreover, the approach enables examining the efficacy of all lower order mixtures of the tested signals. The approach also enables identification of system-level signaling interactions between the applied signals. Many of the signaling interactions identified were consistent with the literature, and other unknown interactions emerged.

Conclusions

This approach can facilitate development of systems biology and optimal drug combination therapies for cancer and other diseases and for understanding key interactions within the cellular network upon treatment with multiple signals.

Background

Understanding how multiple signals affect cellular functions is necessary in order to be able to understand and control these functions. Extensive studies have been done to address how the activation/inhibition of a particular cellular signaling pathway leads to a specific response. Several challenges limit the ability to study the simultaneous effects of multiple signaling. The complexity and lack of detailed knowledge of cellular systems prevent, in many cases, accounting for the effects of some unknown interactions among pathways or among non-primary signal targets. In addition, genetic or epigenetic alterations between otherwise similar cells can cause a significant difference in their responses. This places additional constraints on the experimental outcomes obtained by analyzing individual components. Furthermore, a critical challenge in the investigation of the effects of multiple signals is the arising complexity associated with the increasing number of signals and their various intensities. Without a systematic approach to replace a large number of time and resource consuming experimental tests, it is difficult to characterize the effects of these signals, to identify appropriate signal combinations.

There has been an increasing interest in examining how various biological activities are regulated by multiple interacting signals

Other recent work on identifying drug combination focuses on identifying mixtures of drugs where the the search space is reduced to use only a single concentration while the search space is increased by searching through a larger number of potential drugs

In this work, we achieve the desired goals through the integration of data-driven mathematical tools with biological measurements to generate quantitative models of cellular functions (Figure

Summary of approach

**Summary of approach**. (A) The proposed approach uses input-output data to generate mathematical models capable of predicting the cellular responses to input combinations. The models enable using other mathematical tools for analyzing the cellular responses and for selecting the appropriate combinations of the input signal to drive the system to respond favorably. (B) The desired drugs are combined in certain concentrations and a few combinations are chosen and evaluated experimentally. A predictive mathematical model is generated that can predict the response to all possible combinations. The model can be used to analyze and predict drug interactions and their effects on the observed cellular response and can also be used to determine effective combinations.

In comparison to the approaches previously mentioned, the approach introduced in this work overcomes their limitations by identification of the complete response function and carrying out the optimization

Results

We utilize mathematical tools to characterize the effects of three and four agents on the differential response of cancer and normal cells. The cellular ATP levels of a non-small cell lung cancer cells A549 and primary lung fibroblast culture of AG02603 cells in response to combinations of three chemical agents are measured. Mathematical tools are used to construct predictive models of the cellular ATP levels in response to the combinations. We examine the ability to utilize relatively small numbers of combinations for model generation. The results are extended to study systems of higher complexity with the addition of a fourth chemical agent. The resulting models are used to compare the responses of normal and cancer cell to the same set of combinations. We show that a properly selected combination can result in a significant difference in the respective performance of normal and cancer cells. We also experimentally validate the selected combinations. Furthermore, we show that a combination of chemical agents, if properly chosen, can be more effective than a single agent in inducing a differential response between normal and cancer cells. Using the models, we will examine all the possible lower order mixtures of the four drugs. Moreover, we extract key system-level signaling interactions and compare these interactions between different cell types. We also compare these interactions to known interactions between the drugs.

Signal-Cellular Response Modeling with a Complete Data Set

Inhibition of cell survival and proliferation has been a widely-used approach in cancer treatment

Simplified pathway and drug interactions

**Simplified pathway and drug interactions**. Shown are simplified pathways targeted in the three and four-drug combination treatment of nonsmall cell lung cancer cells A549 and primary fibroblast AG02603 cells that are already known or reported. The dashed arrows indicate indirect connections.

One of the goals of this work is to identify differences in the responses between cancer and non-cancer cells upon the drug treatment. The interactions in Figure

Cellular ATP represents one of the most common and essential markers for all live cells. Measuring ATP is a generally accepted quantitative and sensitive assay for assessing the inhibition of cellular growth, proliferation, and induction of cell killing by drugs. The cellular ATP level, which is regulated by multiple cellular pathways, was experimentally quantitated

Single-drug dose response curves

**Single-drug dose response curves**. Shown are the experimental single-drug dose response curves for the four drugs used in the study. The data was used to identify the drug concentrations to be used in combination studies.

Concentrations of the drugs used in the three and four drug treatments.

**A**

**Drug Name**

**Drug Concentration**

AG490

0, 0.3, 1, 3, 10, 30, 100, 300 (uM)

U0126

0, 0.1, 0.3, 1, 3, 10, 30, 100 (uM)

Indirubin-3'-monoxime

0, 0.3, 1, 3, 10, 30, 100, 300 (uM)

**B**

**Drug Name**

**Drug Concentration**

AG490

0, 1, 3, 10, 30, 100, 300 (uM)

U0126

0, 0.1, 0.3, 1, 3, 10, 30 (uM)

Indirubin-3'-monoxime

0, 1, 3, 10, 30, 100, 300 (uM)

GF109203X

0, 0.3, 1, 3, 10, 30, 100 (uM)

(A) Shown are the eight concentrations (including zero) of each of the three drugs (AG490, U0126, Indirubin-3'-monoxime) used in the three-drug treatments of A549 and AG02603 cells. (B) Shown are the seven concentrations (including zero) of each of all four drugs (AG490, U0126, Indirubin-3'-monoxime, GF109203X) used in the four-drug treatments of A549 and AG02603 cells.

The ATP level in response to all of the 512 drug combinations was experimentally measured in lung cancer A549 cells and in primary lung fibroblast AG02603 cells. The fibroblast cells were derived from normal healthy tissue and are not cancer cells. There are several mathematical methods that can be used to generate models of input-output data. Here we provide a comparison of some of these methods in view of the function approximation problem considered. The methods include two neural network structures and two linear regression models. The neural network structures are a single layer multi-layer perceptron (MLP) and a cascaded neural network

The different models were trained against 80 out of 512 points with the goal of minimizing the mean square error of prediction. The outputs of the models are processed through a saturation function to limit outputs to the interval [0,1]. The trained models can predict the responses to all 512 combinations with high fidelity. The correlation coefficients between the predicted normalized ATP levels and their corresponding experimentally measured values are higher than R = 0.91 (Figure

Three-drug model fitting

**Three-drug model fitting**. The figure shows an evaluation of modeling the responses of A549 and AG02603 to combinations of three drugs. (A) The panels show a plot of the model predicted cellular ATP levels versus the experimentally measured values. Cellular ATP-level predictive models for A549 and AG02603 cells were developed using a number of different methods. a. A linear regression model that uses pairwise products of concentrations and quadratic terms (QRF). b. A linear regression model with n-wise products of concentrations (LR). c. A cascaded neural network with two single-neuron layers (Cascaded NNet). d. A four-neuron single layer multi-layer perceptron artificial neural network (MLP). The models are based on fitting 80 out of 512 combinations and the figure shows the predicted versus experimental values for all 512 combinations. The correlation between the experimentally tested cellular ATP-level (x-axis) and the predicted cellular ATP-level (y-axis) is shown. The circles in the graphs represent individual data points. The diagonal line represents a perfect fit between the experimental and predicted data. (B) Comparison between the predicted normalized ATP levels of different models. The predicted ATP levels for different models are plotted against each other. a. QRF versus LR. b. Cascaded NNet versus MLP. c. QRF versus MLP. d. LR versus MLP. The correlation coefficients between the different methods are shown.

Correlation coefficients between the 432 points not used for training and the corresponding experimentally measured values.

**Cell Type**

**QRF**

**LR**

**Cascaded NNet**

**MLP**

AG02603

R = 0.96

R = 0.95

R = 0.95

R = 0.96

A549

R = 0.96

R = 0.92

R = 0.97

R = 0.98

Experimental testing of all possible combinations can be a costly process. Whenever the response of the biological system is smooth enough, we can utilize a smaller number of combinations to map the entire response surface. In this regard, we have examined the effects of using varying numbers of points to fit the models on the accuracy of prediction. The four methods discussed above were considered and models using 10, 20, 40, 80, 160, and 320 points were fitted. To mimic an actual experimental setup, the points were randomly selected out of the 512 possible combinations using a uniform distribution. The mean square error of prediction for each of the methods and fitting data for both cell types indicate that increasing the number of points reduces the mean square error (Figure

Model accuracy as a function of modeling method and data size

**Model accuracy as a function of modeling method and data size**. The figure shows the effects of using different sizes of data sets for fitting the models and the effects on the model accuracy as measured by the mean square error of prediction. The mean square error of predicted data using four different methods and using 10, 20, 40, 80, 160, and 320 data points, for both A549 and AG02603 cells are displayed. The four methods are two linear regression methods and two neural network methods. The linear regression methods include one that uses pairwise products of concentrations and quadratic terms of the concentrations (QRF), and the other uses the n-wise products of concentrations of drugs as interaction terms (LR). The two neural network methods are a cascaded neural network with two single-neuron layers, and a four-neuron single-layer multi-layer perceptron (MLP). As the number of points used to generate the model increases, the mean square error decreases.

However, no significant improvement in the errors are observed for models with more than 80 points. Using a small number of points results in poor prediction with the linear regression models, the interaction model and the quadratic models. In the absence of post-processing of the data by passing through a saturation function, the mean square error of prediction of the linear regression models become significantly worse. Between the two linear regression models, the quadratic model performs better than the interaction model, potentially due to the added quadratic terms, suggesting the nonlinearity of the response of these cells to the drug combinations used. These models also have higher mean square errors than the neural network models. The differences between the mean square error of these models tend to diminish as the number of points used increases. The two neural network models were comparable overall.

The number of data points required to generate a valid model relies on factors including intrinsic signal-response relationships for individual cell cultures and the experimental measurement error. In addition, the smoothness of many signal-response relationships enables the modeling to rely on less dense mapping over small ranges of signal concentrations. Our results suggest that with a proper mathematical modeling method, the effect of signal combinations can be systematically described through randomly testing a relatively small percentage of signal combinations within specified concentration ranges.

Characterization of Signal-Cellular Response Relationships in Systems of Higher Complexity

The simulated models capable of systematically describing the signal-cellular response relationships for various cells enable the comparison of the cell-type specific differences in cellular responses to multi-drug treatments. An interesting question is whether our approach can simulate more complex systems with a relatively small set of experimental data, and whether efficient multi-drug combinations that lead to a high level of A549 cell inhibition while preserving AG02603 cells can be identified among more drugs. In this regard, we added GF 109203X, a PKC inhibitor, to the three drugs tested in the above section. The addition of a fourth drug increases the search space to 2401 combinations (four drugs with seven concentrations each). Models representing the relationship between the four drugs and the ATP-levels of A549 and AG02603 cells were generated based on 148 experimentally tested combinations. We fitted a single-layer neural network with four neurons using the experimental data. The correlation coefficients between the predicted data from this model and the experimental data were 0.98 for A549 and 0.97 for AG02603 (Figure

Four-drug model fitting

**Four-drug model fitting**. The figure shows the model-predicted ATP-levels versus the experimentally obtained values for four drugs combinations. The model used is a four neuron single layer neural network model that was fitting using a data set of size 148 out of 2401 combinations. The correlation between the experimentally tested cellular ATP-level (x-axis) and the predicted cellular ATP-level (y-axis) is shown. The circles in the graphs represent individual data points. The diagonal line represents a perfect fit between the experimental and predicted data.

Analysis of drug combination differential responses

**Analysis of drug combination differential responses**. The figure presents the analysis results to identify combinations of drugs that can maximize a performance function reflecting the desired objective of minimizing the killing of AG02603 cells and maximizing the killing of A549 cels (see methods section for the details on the performance functions. (A) Results of k-means clustering analysis of the performance of four-drug combinations. Points are grouped together to minimize the distance to the mean of that group. Each group is given a identification number (ID). The x-axis shows cluster IDs; the y-axis shows the performance score of each drug combination. The heat map is a function of the performance for each point. (B) Cell ATP-level (x-axis: AG02603; y-axis: A549) upon four-drug combination treatments. The red stars represent the experimentally tested 148 drug combinations. The squares represent the whole set of 2401 possible drug combinations with the heat map colors reflecting the performance of each combination.

We have also investigated the ability of these models to predict the experimental data obtained in the three drug combinations experiment. These combinations correspond to four-drug combinations where the concentration of the fourth drug is set to zero.

Correlation coefficients between the 512 data points and their predicted values were 0.86 for A549 cells and 0.88 for AG02603. These correlation coefficients are quite reasonable given that the two experiments were conducted several months apart and the model was trained on a separate and independent data set. This provides evidence of the ability of our models to predict cellular responses despite the relatively small number of data points used to generate these models.

Effects of Single Signal Versus Multiple Signals

Although all four drugs inhibit cellular functions common in both cell types, combinations of these drugs may result in a significant difference in cellular ATP-levels between A549 and AG02603. Inhibition of A549 cells and preservation of AG02603 cells using the same drug combination constitute two conflicting objectives. Our goal of identifying the drug combinations that effectively satisfy the above objectives can be realized by utilizing a multi-objective search or optimization technique. A performance function combining the relative importance of each of the two conflicting objectives is introduced (Materials and Methods). The drug combinations resulting in the highest performance are considered the best drug combinations of a given set of drugs with corresponding concentrations.

Identification of the best performing subset is achievable through various methods. Enumeration and sorting of all possible combinations and their corresponding performances is one method. Alternatively, we can use a clustering algorithm such as a k-means clustering algorithm to group combinations with similar performances

Clustering the points into 20 different groups (Figure

Single-drug response curves

**Single-drug response curves**. The figure shows an evaluation of the model-predicted ATP-levels of A549 and AG02603 when the concentrations of three drugs are held constant while the concentration of the fourth drug is varied. The values of the concentration at which the three drugs are held constant are shown in Table 3. Shown are the single-drug effects of increasing concentrations of (A) AG490, (B) U0126, (C) indirubin-3'-monoxime and (D) GF 109203X on cellular ATP-level with respect to different levels of the other three drugs in the treatment. Here the second row fourth column subfigure corresponds to the cellular ATP-levels of A549 and AG02603 to varying concentrations of U0126 while the remaining three other drugs are held at the selected concentration in Table 3.

Shown are the concentrations used for for generating the drug response curves.

**Drug Name**

**Low**

**Medium**

**Maximum**

**Selected**

AG490

0

10

300

30

U0126

0

1

30

0.3

Indirubin-3'-monoxime

0

10

300

10

GF109203X

0

3

100

0.3

Additionally, we investigated the effects of pairwise combinations of drugs on the responses of both A549 and AG02603 cells. The remaining two drugs were fixed at one of two cases, zero or a selected optimal concentration (Table

Two-drug interactions

**Two-drug interactions**. The figure shows an evaluation of the model-predicted ATP-levels of A549 and AG02603 when the concentrations of two drugs are held constant while the concentration of the other two drugs are varied. The first and third columns (A) and (C) correspond to ATP levels of AG02603 in response to variations in different pairs of drug. The concentrations of the other two drugs are fixed at zero for column (A) and a selected combination (Table 3) for column (C). The second and fourth columns (B) and (D) correspond to ATP levels of A549 in response to variations in different pairs of drug (rows). The concentrations of the other two drugs are fixed at zero for column (B) and a selected combination (Table 3) for column (D). The figure shows that at the selected combination, the difference between the response surface of AG02603 and A549 is significantly higher than when the difference at zero concentrations.

Examining the Efficacy of Mixtures of Two, Three, and Four drugs

The availability of the model enables examining all lower order mixtures of drugs and their potential performances. Testing four or more drugs at time can enable efficient identification of effective lower order mixtures of the drugs. Based on the model, we evaluated the performances of all possible mixtures of drugs at varying concentrations. We identified the best achievable performance for each mixture (Figure

Evaluation of the best combinations of all lower order mixtures of the drugs

**Evaluation of the best combinations of all lower order mixtures of the drugs**. (A) The performance of the best combination for single drugs as well as combinations of two, three, four drugs are shown. Also shown are the performances of the worst possible combinations. (B) The normalized ATP level for the best combination for all single, two, three, and four-drug mixtures are shown.

However, this improvement becomes less as four drugs are used instead of three. The results also show that the improvement is more significant for certain mixtures of drugs, e.g., the performance of the mixture of AG490 and I-3-M is better than the performance of AG490 and U0126. A similar observation can be made about three-drug mixtures where the mixture of AG490, I-3-M and GF109203X performs better than the three other three-drug mixtures. This information is not known a priori and without this approach, determining the best mixtures requires testing of each mixture independently. This can be a lengthy and costly process. Moreover, if one elects to randomly select combinations for the mixtures then there is no guarantee that the combination or mixture can have a good performance. In fact the random combination and mixture can perform quite poorly (Figure

The potency of our approach can be illustrated by examining the performance of this approach versus more simple approaches. An example of such approaches is combining the drugs at the best single-drug concentrations. The best concentrations for each single drug were 30uM, 30uM, 30uM, and 100uM for AG490, U0126, I-3-M, and GF109203X respectively. Combining the drugs at the best single-drug concentrations results in a poor performance of 44.48 corresponding to zero normalized ATP levels of A549 and AG02603. One can argue that this combination can be very toxic because of the combined high concentration of each of the drugs. However, there is no simpler way to reduce the concentration of some or all drugs to achieve a better performance. Additionally, an argument can be made for the selection of the mixture of two drugs by picking the two best single drugs. In our case, this would correspond to mixing U0126 and I-3-M. This mixture does not perform as well as the mixture of AG490 and I-3-M (Figure

Validation of selected top performing combinations (30 uM of AG490, 0.3 uM of U0126, 10 uM of I-3-M, and 0.3 uM of GF109203X), (30 uM of AG490, 0.3 uM of U0126, 30 uM of I-3-M, and 1 uM of GF109203X), and (30 uM of AG490, 1 uM of U0126, 10 uM of I-3-M, and 0.3 uM of GF109203X) showed that the experimental ATP levels of A549 and AG02603 were 0.2 and 0.65, 0.13 and 0.51, and 0.13 and 0.5 respectively. This is in agreement with the predicted values for A549 and AG02603 of 0.21 and 0.59, 0.08 and 0.51, and 0.2 and 0.58 respectively.

Drug-Drug Interactions Affect the Observed System Responses

Modeling of multi-signal induced cellular responses can also be used to study key system-level signaling interactions between the applied signals and their effects on the system outputs. To that end, consider the linear regression model with interaction terms. This model uses a total of 15 regressors that are the concentrations of the drugs, six pairwise interaction terms, four three-drug interaction terms, and one four-drug interaction term. This is a large number of regressors and, potentially, only a portion of these regressors are necessary to describe the variation in the data. Reduction of the regressors matrix ^{15 }- 1 possible models.

The residual sum of squares of the best models shows that there is no significant reduction in the residual sum of squares for models with more than 10 regressors for the A549 data, and for models with more than 7 regressors for the AG02603 data (Figure

Analysis of models with varying numbers of regressors

**Analysis of models with varying numbers of regressors**. Shown are the results of best subset regression algorithm which evaluates the best models of 1, 2, . . .,

System-level signaling interactions in A549 and AG02603 cells

**System-level signaling interactions in A549 and AG02603 cells**. Shown are some model-predicted system-level signaling interactions upon treatment with multiple drugs. The figure shows the interactions that are specific to A549, the interactions that are specific to AG02603, and the interactions common to both A549 and AG02603 cells. Direct arrows between the drugs and the ATP levels exist and are omitted to simplify the diagram.

Discussion and Conclusions

We demonstrated the integration of an efficient mathematical approach for a systematic quantitative characterization of the effect of multi-signal combinations in two different cell types. Our method enables the establishment of accurate models to directly connect multi-signal combinations and their effects through a learning process. Only a small percentage of total data points are required to be experimentally tested to establish a predictive model that is capable of simulating the effect of all possible signal combinations.

The resulting predictive model is also able to systematically reveal the inter-drug interactions which are often non-linear relationships. Such a model is necessary for multi-drug combination optimization as the optimal combination, upon the addition of a new drug, cannot be achieved simply by testing various amounts of the new drug added to the previously optimized combination. Moreover, the approach allows for examining all lower order mixtures of the drugs and for evaluating the effectiveness without added experimental overhead. This enables efficient selection of drug combinations of different drugs particularly as the interactions between the different drugs vary. The approach enables systematic selection of input signals (drug combinations) that can achieve desired therapeutic goals. Experimental validation of selected top performing combinations presents additional evidence of the validity and utility of the approach introduced.

Our approach requires a relatively small number of experimental measurements. With the development of large scale-high throughput measurement systems, our approach will be necessary and more efficient, e.g., when the number of inputs increases. Additionally, it will allow for integration into automated machines for testing and analysis of various biological systems. The addition of other cellular outputs can be another factor in favor of using this approach.

Mathematical Modeling Methods

We demonstrated the utility of various mathematical modeling approaches for the purposes of modeling and predicting multi-signal induced cellular responses. Linear regression with models including interaction and quadratic terms were capable of producing powerful predictive models. The usefulness of linear regression methods and subset selection algorithms was also demonstrated as they enabled determination of key system-level signal interactions that result in the observed cellular responses. Alternately, neural network models performed generally better for fitting the data particularly when fewer data points were available to fit the models. Additionally, different structures of neural networks were almost equally powerful in producing the models. This present ample opportunity to exploit different structure of networks to mimic more realistic interactions within the cell given such a priori data. The choice of the best method for fitting the model is dependent on many factors. While in some cases simple linear regression can be used, in other cases nonlinear regression methods are necessary to get satisfactory models. Neural networks represent powerful computational models with great flexibility. Different types and configuration renders them useful for a wide range of problems including dynamic prediction using, e.g., recurrent neural networks. This presents an additional use for neural network to build on a wide range of applications in biological and medical research

It is important to note here that the novelty in this work is not this particular use of artificial neural networks. Rather, it is the use of a system-level model-based approach to study the effects of a large number of signals on various cellular processes and to use that as a basis for selection of the retrospective optimal input signals. Our method resembles a general systems biology approach that can be utilized to address a broad range of biological questions. The method enables comparison of multiple system performances through modeling of their responses. The advantage of our method becomes critical when input signal combinations are characterized for the development of effective

For the drugs and cell types chosen, our results showed that four drugs did not provide a significant improvement over three drugs. However, this is not a general result and the causes of this observation are not well known and pose an interesting problem to be examined in a separate study. A potential cause is that the signaling pathways affected by some of the drugs are saturated potentially due to sharing of a common target between two of the drugs. In other studies we are working on with different drugs and cells, four drugs result in an improvement over three drugs and five drugs also improve on four drugs.

Mechanistic Reasoning

Our modeling approach enabled identification of key system-level signaling interactions that contribute to the observed responses. These interactions highlighted potential differences in the signaling networks of different cell types as demonstrated. While some signaling interactions were common to the two cell types studied, there were other interactions that were specific to each cell type. Many of these interactions were known a priori. For example, the interactions between U0126 and GF109203X, I-3-M and GF109203X, AG490 and I-3-M, as well as AG490 and U0126 and GF109203X. There are other interactions that were not known a priori such as the interaction between U0126 and I-3-M. Thus the prediction of such interactions can guide experimental identification of the molecules mediating the interactions.

In the work presented, we used mathematical tools to construct a predictive model of cellular outcomes. The method was developed in experimental systems that only involved single output measurements. However, the method is a general method that can be used to construct models of the effects of multiple signals on various cellular outcomes, including signaling intermediates. Molecules from various cellular pathways, "intermediates", can serve as candidates for measurement and quantification

Implications on the Design of Combination Therapies

Combinational therapies have attracted significant research efforts. Combination therapy for cancer is one example. The challenges associated with cancer treatment range from a lack of understanding of fundamental pathogenic mechanisms to practical experimentation limitations. Cancer is usually caused by multiple mutations and alterations of multiple signaling pathways which pose an extra challenge when defining the mechanisms underlining cancer development

However, the design of combination therapies based on studies of the response to individual drugs might not lead to the desired outcomes as interactions between the various drugs can lead to unknown outcomes. The emerging toxicities can be a major hurdle to the development of combination therapies. Therefore, it is important to consider system level signals or outputs that are representative of the state of the cell in addition to cellular intermediates targeted by the drugs. As such, it is important to investigate whether and how a combination of drugs can lead to a better system response compared to any of the individual single-drug responses.

Development of drug combination therapies for cancer can lead to more effective therapies to overcome drug resistance and achieve maximal drug efficacy. The demanding task is to find a combination that is maximally effective on cancer cells with minimal effect on non-cancer cells. The method introduced in this work provides the ability to study the system-level effects of combinations of drugs and use the data to select combinations that can lead to desired outcomes by comparing the mathematical models of multiple cellular types. Furthermore, the method is capable of studying and characterizing experimental problems of higher complexity such as the order and timing of administration of multiple inputs into the system, which may help further reduce cytotoxicity and enhance efficacy with the same drugs utilized in clinical treatment of diseases. Moreover, employing our method with emerging technologies such as micro-fluidic devices or "lab-on-chip" will enable high-throughput investigation of multi-drug combinations, and become a promising platform for developing personalized medicine.

Methods

Cell Culture and Experimental Measurements

A549 cells were cultured in RPMI1640 medium (CellGro) supplemented with 10% heat-inactivated FBS, 100U/ml penicillin and 100g/ml streptomycin (CellGro). AG02603 cells were in MEM (Gibco) supplemented with non-essential amino acid (Gibco), 15%heat-inactivated FBS, 100U/ml penicillin, and 100g/ml streptomycin. Cells were seeded one day before drug treatment at a density of 1000cells/well for A549, and 8000cells/well for AG02603, in 96 well plates. ATP assays were conducted with the ATPlite 1step assay system (PerkinElmer) following the manufacturer's instructions and the luminescence signal was measured by Lmax microplate luminometer (Molecular Devices).

Neural Networks Models

We used two types of neural networks to fit the data. The first is a single layer four-neuron multilayer perceptron and the other is a two-layer cascaded neural network with a single neuron per layer. The networks were constructed and trained using the neural network toolbox of MATLAB

Linear Regression Models

We used two linear regression models to fit the data. The two models are

Performance Function

In our work, we used a single performance function reflecting the relative importance of each criteria to our setup. The criteria are: maximize AG02603 cellular ATP-level and minimize AG02603 cellular ATP-level. The performance function used is defined as follows. Let _{nc }be the ATP-level of AG02603 and _{cc }be the ATP-level of A549 cells. Then, the performance function used is given by Perf(_{nc}, _{cc})

where

Authors' contributions

IA, FY, and RS wrote the paper. IA, FY, and RS analyzed the data. IA, CMH, and JSS devised the mathematical solutions. FY, JF, and KY conducted the wet experiments. IA carried out the computational experiments. IA, FY, SD, CMH, JSS, and RS conceived the idea. All authors read and approved the final manuscript.

Acknowledgements

We thank Shiva Portonovo and Drs. Christopher Denny, Desmond Smith, Helen Brown, Elliot Landaw, and Henry Huang for critical reading of the manuscript; all Sun lab members for helpful discussions. This study is supported by NIH (CA091791 and CA127042), the Stop Cancer Foundation, Burroughs Wellcome Fund, NSF grant (ECS-0501394) and NIH Roadmap for Medical Research (PN2 EY018228).