We have constructed an operational ePD model to illustrate its salient features. This model tracks the effect of an epidermal growth factor receptor (EGFR) inhibitor (such as gefitinib) on tumor growth and accounts for multiple genomic variations within the cellular regulatory network that controls tumor response (). The simplified regulatory network in this model is a linear pathway with a single additional coherent feed-forward motif from EGFR to RAF. In this ePD model, the link to a PK model is simplified by using the extent of fractional inhibition of an activated receptor as the input.
Operational ePD model for EGFR inhibitor therapeutics
A number of simplified patient scenarios demonstrate how genomic and epigenomic variations in individual cancer patients can be modeled within the dynamics of anti-EGFR therapy to demonstrate the potential power of an ePD model. In this model, all hypothetical patients have driver mutations that activate the EGFR or copy number variations that increase amounts of EGFR, either of which results in increased proliferation and tumor formation. All patients are being treated with an EGFR tyrosine kinase inhibitor such that there is 80% inhibition of receptor activity. Details of the model and simulations can be found in the Supplementary Materials
, and the results of the simulations at the tissue/organ level (that is, tumor size) are shown in .
Enhancing prediction of drug efficacy
These simulations show how the response to a drug in each patient can vary on the basis of the number and type of genomic or epigenomic alterations. Consider a hypothetical standard patient (SP) who has the driver EGFR alterations but no additional changes in the genome and epigenome related to this tumorigenetic network. This patient would show decreased tumor growth upon drug treatment [, black bar (SP)]. The profiles of activated RAS, RAF, and MEK1/2 amounts and levels of the cell cycle activator cyclin D are shown in fig. S1
Now consider patient B (), who displays the following features: (i) hypermethylation of RASAL1
, which results in lowered amounts RasGAP (table S3
) and (ii) a single-nucleotide polymorphism (SNP) (such as rs55716409) in the RKIP
). The RKIP/PEBP
gene encodes the protein RKIP, an inhibitor of RAF1 (). We assume that this SNP results in an RKIP protein with greatly lowered affinity for protein kinase C (PKC) and that RKIP is not phosphorylated by PKC. Thus the increased signal from KRAS that results from the reduced amount of RasGAP is effectively suppressed at the level of RAF by the increased amounts of active RKIP, resulting in levels of active MEK1/2 that are similar to what is seen in the standard patient. Therefore, in patient B, the SNP in RKIP/PEBP
effectively cancels out the epigenetic (DNA methylation) change in the RASAL1
gene and results in the same partial-remission response to the drug as the SP [, black bar (B)]. The profiles of activated RAS, RAF, and MEK1/2 amounts and cyclin D levels for patient B are shown in fig. S2
In contrast, patient A has (i) a hypermethylated RASAL1
gene, but (ii) no other changes in RKIP/PEBP
or alterations in the amount of microRNA-221 (miR-221), which is known to modulate the cell cycle regulatory protein p27kip
). In this patient, the decrease in RasGAP protein levels leads to an uncompensated increase in RAS activity, which eventually results in over-expression of the cell cycle activator cyclin D in response to EGFR stimulation, even in the presence of the EGFR inhibitor. As a result, patient A has a tumor variant that is resistant to drug therapy and proliferates even when treated with a drug dose that blocks 80% of EGFR activity [, purple bar (A)]. The profiles of activated RAS, RAF, and MEK1/2 amounts and cyclin D levels of patient A are shown in fig. S3
. If patient A accumulates additional deleterious changes, such as increased amounts of miR-221, which lead to decreased amounts of p27kip
, then such a patient (A′) would show even greater increases in cyclin-dependent kinase 4/6 (CDK 4/6) activity (); this change leads to increased tumor growth even when the patient is treated with the drug [, left-most bar, purple (A′)].
Now consider patient C, who is represented in as a green bar. This patient has (i) a normal RASAL1
gene, (ii) a SNP in the RKIP/PEBP
gene that produces a hyperactive RKIP because of its lack of responsiveness to PKC regulation, and (iii) decreased amounts of miR-221. The increased amounts of active RKIP inhibit RAF activity, and the decreased miR-221 amounts lead to an increase in p27kip
; together, these changes yield greatly decreased amounts of functional cyclin D and CDK4/6, which result in cancer cell proliferation and tumor growth (). The profiles of activated RAS, RAF, and MEK1/2 amounts and cyclin D levels for patient C are shown in fig. S4
. Patient C′, who has (i) hypomethylated RAS-AL1
(which leads to increased amounts of RasGAP), (ii) a normal RKIP/PEBP1
gene, and (iii) increased amounts of miR-221 (which leads to decreased amounts of p27 kip
), is also highly responsive to drug therapy [, green bar (C′)]. For patient C′, the epigenetic changes in the RASAL1
gene compensate for increased miR-221 amounts and the subsequent decrease in the cell cycle inhibitor p27 kip
These hypothetical cases clearly indicate how multiple genomic and epigenomic changes can produce a wide range of responses to drug therapy. Even when patients have the same initial oncogenic driver mutation, distinct genomic and epigenomic changes can profoundly affect drug response. As shown in , operational ePD models, such as the one described here, provide a mechanistic understanding of why such variability occurs and can handle a range of genomic and epigenomic variations and predict drug response.
It is likely that the complexity of patient responses will be far more diverse than the simple example depicted in , because many drugs have multiple targets and are used in combination therapies. ePD models can make the problem of deciphering and predicting patient responses in these complex cases explicitly tractable when such computational models are developed in a data-driven manner and computational analyses become an integral part of clinical decision-making. Although we are not there yet, current progess is slow but steady.
At a conceptual level, the myriad genomic, epigenomic, translational, and posttranslational changes that are possible appear to be very complicated. However, all alterations at the various levels of biological regulation fall into only two types of changes in an ePD model. Mutations, SNPs, and posttranslational modifications such as protein phosphorylation can change biochemical reaction rates. Missense mutations and changes in DNA methylation, histone modifications, microRNA expression, ubiquitination, and protein turnover all alter the concentrations of the reactants. With the appropriate data, a range of genomic, epigenomic, and posttranslational changes can be readily represented with precision in ePD models. Hence, the complexity of ePD models is operational rather than conceptual.