A new stochastic and state space model of human colon cancer incorporating multiple pathways
 Wai Y Tan^{1}Email author and
 Xiao W Yan^{1}Email author
DOI: 10.1186/17456150526
© Tan and Yan; licensee BioMed Central Ltd. 2010
Received: 19 January 2010
Accepted: 20 April 2010
Published: 20 April 2010
Abstract
Background and Purpose
Studies by molecular biologists and geneticists have shown that tumors of human colon cancer are developed from colon stem cells through two mechanisms: The chromosomal instability and the microsatellite instability. The purpose of this paper is therefore to develop a new stochastic and state space model for carcinogenesis of human colon cancer incorporating these biological mechanisms.
Results
Based on recent biological studies, in this paper we have developed a state space model for human colon cancer. In this state space model, the stochastic system is represented by a stochastic model, involving 2 different pathwaysthe chromosomal instability pathway and the microsatellite instability pathway; the observation, cancer incidence data, is represented by a statistical model. Based on this model we have developed a generalized Bayesian approach to estimate the parameters through the posterior modes of the parameters via Gibbs sampling procedures. We have applied this model to fit and analyze the SEER data of human colon cancers from NCI/NIH.
Conclusions
Our results indicate that the model not only provides a logical avenue to incorporate biological information but also fits the data much better than other models including the 4stage single pathway model. This model not only would provide more insights into human colon cancer but also would provide useful guidance for its prevention and control and for prediction of future cancer cases.
Reviewers
This article was reviewed by M.P. Little and M. Kimmel
Background
In the past 15 years, molecular biologists and geneticists have revealed the basic molecular and genetic mechanisms for human colon cancer. These mechanisms have been linked to two avenues: The chromosomal instability (CIN) involving chromosomal aberrations and loss of heterozygosity (LOH), and the microsatellite instability (MSI) involving mismatch repair genes and the creation of mutator phenotype ([1–9]). The pathway of the CIN avenue (also referred to as LOH pathway) involves inactivation through genetic and/or epigenetic mechanisms, or loss, or mutation of the suppressor APC gene in chromosome 5q (about 85% of all human colon cancers) whereas the pathway of the MSI avenue involves mutation or epigenetic inactivation of the mismatch repair suppressor genes (about 15% of all colon cancers). This leads to multiple pathways for the generation of human colon cancer tumors with each pathway following a stochastic multistage model and with intermediate transformed cells subjecting to stochastic proliferation (birth) and differentiation (death). The goal of this paper is to develop a stochastic model for human colon cancer to incorporate these biological information and pathways. This paper is an extension of Tan et al. [10], Little and Wright [11] and Little et al. [12]. We note that besides the multiple pathways considered above, Little and Wright [11], Little et al. [12] and Little [13] have also included mixture type of multiple pathways; however, because the mutation rates are very small, the chance of mixture type of pathways will be extremely small in which case the Little model is equivalent to the model in Section 3.
For developing biologically supported stochastic model of carcinogenesis, in Section 2 we present the most recent cancer biology of human colon cancer. Using results from Section 2, we develop in Section 3 a stochastic model for carcinogenesis of human colon cancer involving multiple pathways. In Section 4 we derive a statistical model for cancer incidence data of human colon cancer. By combining models from Sections 3 and 4, in Section 5 we develop a state space model for human colon cancer. In Section 6, by using the state space model in Section 5, we develop a generalized Bayesian inference procedure to estimate unknown parameters and to predict state variables. To illustrate the applications of the model and methods, in Section 7 we apply the model and methods to the colon cancer incidence data from SEER. Finally in Section 8, we discuss the usefulness of the model and methods and provide some conclusions.
A Brief Summary of Colon Cancer Biology
As discussed in the introduction, genetic studies have indicated that there are two major avenues by means of which human colon cancer is derived: The Chromosomal Instability (CIN) and the MicroSatellite Instability (MSI). The first avenue is associated with the LOH pathway involving the APC gene in chromosome 5q and the latter associated with the microsatellite pathway involving mismatch repair genes. The most important oncogene is the βcatenin gene in chromosome 3p22.
The CIN (LOH) Pathway of Human Colon Cancer (The APCβcatenin  Tcf  mycpathway)
The CIN pathway involves loss or inactivation of the tumor suppressor genes  the APC gene in chromosome 5q, the Smad4 gene in chromosome 18q and the p53 gene in chromosome 17p; see Remark 1. This pathway accounts for about 85% of all colon cancers. It has been referred to as the LOH pathway because it is characterized by aneuploidy/or loss of chromosome segments (chromosomal instability); see Remark 2. This pathway has also been referred to as APCβ  catenin  Tcf  myc pathway because it involves the destruction complex GSK3β AxinAPC which phosphorylates the βcatenin protein leading to its degradation; when both copies of the APC gene are inactivated or mutated, the destruction complex is then inactive leading to accumulation of free βcatenin proteins in the cytoplasm which move to the nucleus to complex with Tcf/Lef transcription factor to activate and transcript oncogenes myc, cyclin D and CD44. (Free βcatenin protein in the cytoplasm also binds with Ecadherin and αcatenin to disrupt the gap junction between cells, leading to migration and metastasis of cancer tumors.)
The model in Figure 1 is a 6stage model. However, because of the haploinsufficiency of the Smad4 gene (see Alberici et al.[15]) and the haploinsufficiency of the p53 gene ([16]), one may reduce this 6stage model into a 4stage model by combining the third stage and the fourth stage into one stage and by combining the fifth stage and the sixth stage into one stage. This may help explain why for single pathway models, the 4stage model fits the human colon cancer better than other single pathway multistage models ([17]). Recent biological studies by Green and Kaplan [4] and others have also shown that the inactivation or deletion or mutation of one copy of the APC gene in chromosome 5 can cause defects in microtubule plusend attachment during mitosis dominantly, leading to aneuploidy and chromosome instability. This would speed up the mutation or inactivation of the second copy of the APC gene and increase fitness of the APCcarrying cells in the microevolution process of cancer progression. This could also help explain why the APC LOH pathway is more frequent than other pathways.
Remark 1: As observed by Sparks et al. [8], instead of the APC gene, this pathway can also be initiated by mutation of the oncogene βcatenin gene; however, the proportion of human colon cancer due to mutation of βcatenin is very small (less than 1%) as compared to the APC gene, due presumably to the contribution of the APC on chromosome instability ([4]). Similarly, the destruction complex can become inactive either by the inhibition of GSK3β through the Wnt signalling pathway (see [18]) or the inactivation or mutation of the Axin protein, leading to accumulation of the β Catenin proteins in the cytoplasm; but the proportion of colon cancer caused by inhibition of GSK3β is also very small as compared to the colon cancer cases caused by the CIN and the MSI pathways.
Remark 2: The APC gene in chromosome 5q acts both as a tumor suppressor gene and an oncogene in initiating and promoting colon carcinogenesis. As an oncogene, the APC gene acts dominantly in regulating microtubule plusend attachment during mitosis ([4]). Thus, the inactivation or deletion or mutation of one copy of the APC gene in chromosome 5 can cause defects in microtubule plusend attachment during mitosis, leading to aneuploidy and chromosome instability. This would speed up the mutation or inactivation of the second copy of the APC gene and increase fitness of the APCcarrying cells in the microevolution process of cancer progression. This could also help explain why the APC LOH pathway is more frequent than other pathways.
The MSI (MicroSatellite Instability) Pathway of Human Colon Cancer
Morphologically, mutation or methylation silencing of the MMR gene hMLH1 or hMSH2 generates hyperplastic polyps which lead to the generation of serrated adenomas. These adenomas develop into carcinomas after the inactivation, or loss or mutations of the TGFβ RII gene and the Bax gene, thus abrogating differentiation and apoptosis. (Bax is an antiapoptosis gene.) In what follows, we let N denote the normal stem cells, J_{ i }the ith stage cells in the MSI pathways. Then for sporadic MSI, the model is N → J_{1} → J_{2} → J_{3} → J_{4} → J_{5} → cancer tumor.
The Major Signalling Pathways for Human Colon Cancer
Recent biological studies ([18, 19]) have shown that both the CIN and the MSI pathways involve the Wnt signalling pathway and the destruction complex (this complex is a downstream of the Wnt signalling pathway), the TGFβ inhibiting signalling pathway and the p53Bax apoptosis signalling pathway, but different genes in the CIN and MSI pathways are affected in these signalling processes. In the CIN pathway, the affected gene is the APC gene in the Wnt signalling, the Smad4 in the TGFβ signalling and the p53 gene in the p53Bax signalling; on the other hand, in the MSI pathway, the affected gene is the Axin 2 gene in the Wnt signalling, the TGFβ Receptor II in the TGFβ signalling and the Bax gene in the p53Bax signalling.
Because the probability of point mutation or genetic changes of genes are in general very small compared to epigenetic changes, one may speculate that colon cancer may actually be initiated by some epigenetic mechanisms ([18, 20, 21]). In fact, Breivik and Gaudernack [20] showed that in human colon cancer, either methylating carcinogens or hypermethylation at C_{ p }G islands would lead to G/T mismatch which in turn leads to Mismatch Repair (MMR) gene deficiency or epigenetic silencing of the MMR genes and hence MSI (Microsatellite Instability); alternatively, either hypomethylation, or bulkyadduct forming (BAF) carcinogens such as alkylating agents, UV radiation and oxygen species promote chromosomal rearrangement via activation of mitotic check points (MCP), thus promoting CIN (Chromosomal Instability). A recent review by Baylin and Ohm [18] have demonstrated that epigenetic events may lead to LOH and mutations of many genes which may further underline the importance of epigenetic mechanisms in cancer initiation and progression.
Based on the above biological studies, in this paper we thus postulate that the incidence data of human colon cancer are described and generated by a multistage model involving 2 pathways as defined above. In this paper, because of haploidinsufficiency of the tumor suppressor genes {Smad_{4}, p53, Axin, Bax, TGF  β  ReceptorII}, the number of stages for the CIN pathway and MSI are assumed as 4 and 5 respectively.
Methods
Stochastic MultiStage Model of Carcinogenesis for Human Colon Cancer Involving Multiple Pathways
Transition rates and transition probabilities for human colon carcinogenesis
1 N  →  1 N, 1 I_{1}  α_{0}(t)Δt 
1 N  →  1 N, 1 J_{1}  β_{0}(t)Δt 
1 I_{ l }  →  2 I_{ l } 

1 I_{ l }  →  death 

1 I_{ l }  →  1 I_{ l },1 I_{l+1}  a_{ l }(t)Δt 
l = 1, ..., k_{1}  1  
1 J_{ r }  →  2 J_{ r } 

1 J_{ r }  →  death 

1 J_{ r }  →  1 J_{ r },1 J_{r+1}  β_{ r }(t)Δt 
r = 1, ..., k_{2}  1 
where λ_{ I }(t) = N(t)α_{0}(t), λ_{ J }(t) = N(t) β_{0}(t), , , and the initial condition is ψ ( , , z; t_{0}, t_{0}) = 1 given normal individuals at risk at time t_{0} .
The above pde is in general very difficult to solve; further, even if the solution of this equation can be derived, the results are very difficult to apply to estimate the unknown parameters and to predict future cancer cases. Most importantly, T(t) may not be Markov so that this theory is not applicable (Fakir et al.[22, 23]). In this paper, we will thus propose an alternative approach through stochastic equations. It can easily be shown through the method of pgf that if T(t) is Markov, then the stochastic equation method is equivalent to the method of Markov theory; as we shall see, however, the stochastic equation method is more powerful and does not need to assume Markov for T(t).
The Stochastic Equation for State Variables
To derive stochastic equations for the state variables, let be the number of births of the I_{ l }(J_{ r }) initiated cells during (t, t + Δt] {l = 1, ..., k_{1} 1 (r = 1, ..., k_{2}  1)}, the number of deaths of the I_{ l }(J_{ r }) initiated cells during (t, t + Δt] {l = 1, ..., k_{1} 1 (r = 1, ..., k_{2}  1)} and the number of mutation (I_{ l }→ I_{l+1}) (J_{ r }→ J_{r+1}) of I_{ l }(J_{ r }) cells during (t, t + Δt] {l = 1, ..., k_{1}  1 (r = 1, ..., k_{2}  1)}.
where λ_{ I }(t) = N(t)α_{0}(t), λ_{ J }(t) = N(t)β_{0}(t).
independently of and other transition variables.
independently of and other transition variables.
where .
where I_{0}(t) = J_{0}(t) = N(t).
The Expected Numbers
where .
The Probability Distribution of State Variables and Transition Variables
where I_{0}(t) = J_{0}(t) = N(t), a(l_{ u }, i_{ u }; t) = I_{ u }(t + Δt)  I_{ u }(t)  l_{ u }+ i_{ u }, u = 1, ..., k_{1}  1 and where b(m_{ v }, j_{ v }; t) = J_{ v }(t + Δt)  J_{ v }(t)  m_{ v }+ j_{ v }, v = 1, ..., k_{2}  1.
The above transition probability and hence the probability distribution of (t) is too complicated to be of much use. For implementing the Gibbs sampling procedure to estimate parameters and to predict state variables, we use data augmentation method to expand the model. Thus, we define the augmented variables . (In what follows we will refer these variables as the transition variables, unless otherwise stated.)
where for l = 1, ..., k_{1}  1 and for r = 1, ..., k_{2}  1.
The probability distribution given by equation (15) will be used to derive estimates and predicted numbers of state variables. This is discussed in Section 6.
A Statistical Model and The Probability Distribution of the Number of Detectable Tumors
Colon Cancer Data from SEER(overall population)
Age Group  Number of People at Risk  Observed Colon Cancer Cases  Total Prediced Colon Cancer 

0  9934747  1  0 
04  38690768  2  0 
59  48506058  2  6 
1014  49881935  35  44 
1519  50447512  104  164 
2024  51612785  337  370 
2529  54071811  847  965 
3034  54194486  1829  2080 
3539  50363957  3420  3534 
4044  46029771  6174  6698 
4549  40674188  10950  11072 
5054  36070434  18716  18256 
5559  31084543  27438  25875 
6064  26507762  37155  34867 
6569  22772688  47202  45156 
7074  18785224  53190  52810 
7579  14592602  52887  53479 
8084  9751212  42589  41517 
The Probability Distribution of the Number of Detectable Tumors for Colon Cancer
To derive the probability distribution of time to tumors, one needs the probability distribution of T(t). For deriving this probability distribution, we observe that malignant cancer tumors arise by clonal expansion from primary cells and primary cells, where primary cells are cells derived from cells by mutation of cells and primary cells are cells derived from cells by mutation of cells.
where R_{ i }(t_{j1}, t_{ j }) = ω_{ i }(t_{j1})  ω_{ i }(t_{ j }).
Similarly, it can readily be shown that to the order of Min{o( (t)), o( (t)}, Q_{ T }(t) ~Q_{1}(t) + Q_{2}(t).
To further simplify the calculation of Q_{ T }(j), we observe that in studying human cancers, one time unit (i.e. Δt = 1) is usually assumed to be 3 months or 6 months or longer. In these cases, one may practically assume and if t  s ≥ 1.
A Statistical Model for Cancer Incidence Data
Because n_{ j }is very large and Q_{ T }(j) is very small, approximately Y_{ j }is Possion with mean τ_{ j }= n_{ j }Q_{ T }(j). Notice that to the order of Max{o( (t)), o( (t))}, τ_{ j }(and hence the probability distribution of Y_{ j }) depends on the stochastic model of colon carcinogenesis through the expected number {E [ (t)], E [ (t)]} of { (t), (t)} and the parameters { (t), (t)} over the time period (t_{j1}, t_{ j }].
The State Space Model of Human Colon Cancer
State space model is a stochastic model which consists of two submodels: The stochastic system model which is the stochastic model of the system and the observation model which is a statistical model based on available observed data from the system. Thus, the state space model of a system takes into account the basic mechanisms of the system and the random variation of the system through its stochastic system model and incorporates all these into the observed data from the system; furthermore, it validates and upgrades the stochastic model through its observation model and the observed data of the system. As illustrated in Tan ([25], Chapters 89), the state space model has many advantages over both the stochastic model and the statistical model when used alone since it combines information and advantages from both of these models.
For human colon cancer, the stochastic system model of the state space model is the stochastic model consisting of 2 pathways with each pathway following a multistage model as described in Section 3; the observation model of this state space model is a statistical model based on the observed number of colon cancer cases as described in Section 4.
The Stochastic System Model and the State Variables
where P{ (t  1) (t  1)} and P{ (t) (t  1), (t  1)} are given by equations (16) and (17) respectively and where Θ = {λ_{ I }, λ_{ J }, α_{ l }(t), β_{ r }(y), , d_{ l }(t)^{(I)}(t), , d_{ r }(t)^{(J)}(t), l = 1, ..., k_{1}  1, r = 1, ..., k_{2}  1}.
Notice that this probability distribution is basically a product of Poisson distributions and multinomial distributions.
The Observation Model Using SEER Data
where h(Y_{ j }; τ_{ j }) is the density at Y_{ j }of the Poisson distribution with mean τ_{ j }.
where and is the maximum likelihood estimate of τ_{ j }.
To apply the above distribution to estimate unknown parameters and to fit real data, we also make the following assumptions: (a) From biological observations ([1–9]), one may practically assume that {α_{ l }(t) = α_{ l }, l = 0, 1, 2, 3; β_{ r }(t) = β_{ r }, r = 0, 1, 2, 3, 4, }. (b) Because the colon polyps are generated by proliferation of I_{2} cells and J_{3} cells and because the polyps can only grow to a maximum size of about 10 mm^{3}, we assume that and for some small (δ_{ i }> 0, i = 1, 2). (c) Because colon cell divisions are mainly due to action of the βcatenin gene, one may also assume . In this case, one has approximately and , r = 1, 2. Under these assumptions, the unknown parameters of interest are Θ = {Θ_{1}, Θ_{2}}, where and Θ_{2} = {α_{3}, β_{4}).
The Generalized Bayesian Method and the Gibbs Sampling Procedure
The generalized Bayesian inference is based on the posterior distribution P{ΘX, U, } of Θ given {X, U, Y= }. This posterior distribution is derived by combining the prior distribution P{Θ} of Θ with the probability distribution P{X, U, YΘ} given by equation (20) with Ybeing replaced by . It follows that this inference procedure would combine information from three sources: (1) Previous information and experiences about the parameters in terms of the prior distribution P{Θ}, (2) Biological information represented by the stochastic system equations of the stochastic system (P{X, UΘ}) and (3) Information from observed data, represented by the statistical model through the conditional likelihood L(Θ , X, U).
Because of additional information from the stochastic system model, this inference procedure is advantageous over the standard Bayesian procedure in that it can avoid the identifiability problems associated with standard Bayesian method. For example, we have shown that to the order of Max{o(α_{3}(t)), o(β_{4}(t))} the probability distribution of the Y_{ j }'s depends on the stochastic model through the expected numbers of I_{3}(t) and J_{4}(t), which depend on the birth rates and death rates only through the difference of these rates. It follows that it is not possible to estimate the birth rates and death rates separately by the traditional Bayesian method. Most importantly, the number of parameters is very large and the number of data points is limited. Thus, without information from the stochastic system model, it is virtually impossible to estimate all unknown parameters; for more examples, see Tan ([25, 26]).
The Prior Distribution of the Parameters
where c is a positive constant if these parameters satisfy some biologically specified constraints; and equal to zero otherwise. These biological constraints are:
(i) For the mutation rates of the I _{ i }cells in the LOH pathway, 1 <λ _{ I }< 1000 (N → I _{1}), 10^{6} < α _{ i }< 10^{4}, i = 1, 2, 3. For the proliferation rates of I _{ i }cells in the LOH pathway, γ _{1}(t) = 0, 0 < < 0.5, i = 2, 3, γ _{2}(t) = , 10^{4} <γ _{2} < 2 * 10^{2}, 10^{5} <δ _{1} < 5 * 10^{3}, 10^{2} <γ _{3}< 0.5.
(ii) For the mutation rates in the MSI pathway, 1 <λ _{ J }< 1000 (N → I _{1}), 10^{8} <β _{1} < 10^{5}, 10^{6} <β _{ j }< 10^{2}, j = 2, 3, 4. For the proliferation rates in the MSI pathway, .
We will refer the above prior as a partially informative prior which may be considered as an extension of the traditional non informative prior given in Box and Tiao [27].
The Posterior Distribution of the Parameters Given {Y= , X, U}
Combining the prior distribution given in (6.1) with the density of P{X, U, YΘ} given in equation (20), one can readily derive the conditional posterior distribution of Θ given {X, U, Y= }. For (l = 2, 3), denote by: and ; similarly, for r = 3, 4, we define {B_{ rJ }, D_{ rJ }, N_{ rJ }} by replacing by respectively. Then, we have the following results for the conditional posterior distributions:
The Multilevel Gibbs Sampling Procedure For Estimating Parameters
Given the above probability distributions, the multilevel Gibbs sampling procedure for deriving estimates of the unknown parameters are given by:
(a) Step 1: Generating (X, U) Given (Y= , Θ) (The DataAugmentation Step):
Given Y= and given Θ, use the stochastic equations (3)(4) and the probability distributions given by equations (5)(10) in Section 3 to generate a large sample of (X, U). Then, by combining this sample with P{Y= X, U, Θ} to select (X, U) through the weighted bootstrap method due to Smith and Gelfant [28]. This selected (X, U) is then a sample from P{X, UY= , Θ} even though the latter is unknown. (For proof, see Tan [25], Chapter 3.) Call the generated sample ( , ).
(b) Step 2: Estimation of Θ = {Θ_{1}, Θ_{2}} Given {Y= , X, U}:
Given Y= and given (X, U) = ( , ) from Step 1, derive the posterior mode of the parameters by maximizing the conditional posterior distribution P{Θ , , }. Denote the generated mode as .
(c) Step 3: Recycling Step.
With {(X, U) = ( , ), Θ = } given above, go back to Step (a) and continue until convergence. The convergence of the above steps can be proved using procedure given in Tan ([25], Chapter 3). At convergence, the are the generated values from the posterior distribution of Θ given Y= independently of (X, U) (for proof, see Tan [25], Chapter 3). Repeat the above procedures one then generates a random sample of Θ from the posterior distribution of Θ given Y= ; then one uses the sample mean as the estimates of (Θ) and use the sample variances and covarainces as estimates of the variances and covariances of these estimates.
Results
Application to Fit the SEER Data
Estimates of Parameters for Each Pathway
LOH Pathway  

I _{0}  I _{1}  I _{2}  I _{3}  
Mutation Rate  1.4E06  2.2E04  3.2E03  1.2E06  
± 1.69E08  ± 1.32E05  ± 3.33E04  ± 2.06E07  
Proliferation Rate  0  0  3.6E03  1.6E02  
N/A  N/A  ± 1.12E03  ± 4.78E04  
Birth Rate Para.  0  0  7.4E03  1.9E02  
N/A  N/A  ± 1.03E03  ± 4.08E04  
Growth Limiting Para.  N/A  N/A  8.3E05  N/A  
N/A  N/A  ± 1.4E05  N/A  
MSI Pathway  
J _{ 0 }  J _{ 1 }  J _{ 2 }  J _{ 3 }  J _{ 4 }  
Mutation Rate  8.3E07  3.5E04  1.4E03  9.3E03  7.7E06 
± 1.38E08  ± 1.89E05  ± 8.57E05  ± 1.22E03  ± 1.7 9E06  
Proliferation Rate  0  0  0  2.8E03  2.0E02 
N/A  N/A  N/A  ± 7.01E04  ± 3.31E04  
Birth Rate  0  0  0  9.6E03  2.6E02 
N/A  N/A  N/A  ± 6.08E04  ± 2.88E04  
Growth Limiting Para.  N/A  N/A  N/A  1.6E03  N/A 
N/A  N/A  N/A  ± 3.7E04  N/A 
From these results, we have made the following observations:
(a) As shown by results in Table 2, the predicted number of cancer cases are very close to the observed cases in all age groups. This indicates that the model fits the data well and that one can safely assume that the human colon cancer can be described by a model of 2 pathways. The AIC (Akaike Information Criterion) and the BIC (Bayesian Information Criterion) from the model are 55.96 and 81.30 which are smaller than the AIC of 816.0667 and the BIC value of 827.1513 from a single pathway 4stage model respectively (Luebeck and Moolgavkar [17]). This shows that the multiple pathway model fits better than the single pathway 4stage model as proposed by Luebeck and Moolgavkar [17].
(b) From Table 2, it is observed that the largest number of cancer cases is in the age group between 70 and 75 years old. Comparing the values of Q _{ i }(j) between the CIN pathway (i = 1) and the MSI pathway (i = 2), it appears that the largest cancer cases is between the age group 65 and 70 years old for the CIN pathway and is between 85 and 90 years old for the MSI pathways. Presumably this might be due to the fact that the MSI pathway has one more stage than the CIN pathway.
(c) Reflecting the contribution of the APC gene on chromosomal instability, results in Table 3 showed that the mutation rates of the I _{ r }cells from I _{1} → I _{2} and from I _{2} → I _{3} had increased about 100 times and 1000 times respectively than the mutation rate from N → I _{1} cells. Similarly, due to the contribution to genomic instability by the mismatch repair genes, the mutation rates from J _{1} → J _{2}, from J _{2} → J _{3} and J _{3} → J _{4} had increased about 5 * 10^{2}, 0.5 * 10^{4} and 10^{4} times respectively than the mutation rate from N → J _{1}. Notice also from Table 3 that the mutation rates from J _{1} → J _{2} → J _{3} → J _{4} are about 2 to 3 times of those from I _{1} → I _{2} → I _{3}. As shown in probability plots (not shown here), these increases have speeded up the time to cancer in the MSI pathway by about 510 years.
(d) Results in Table 3 showed that the mutation rates from I _{3} → I _{4} and from J _{4} → J _{5} are of the order 10^{6} which were about 10^{2} → 10^{3} times smaller than the mutation rates from I _{1} → I _{2} → I _{3} and from J _{1} → J _{2} → J _{3} → J _{4}. These results might be the consequence that we had ignored the stages of vascular carcinogenesis (i.e. angiogenesis and metastasis; see Hanahan and Weinberg [29] and Weinberg [30]) by merging these stages into the last stage. From Weinberg ([30], Chapters 1314), notice that the angiogenesis and metastasis are also multistage processes.
(e) Results in Table 3 showed that the proliferation rates (birth rate  death rate) of the I _{3} cells and the J _{4} cells are of order 10^{2} which are much larger than the proliferation rates of the I _{2} cells and the J _{3} cells, due presumably to the effects of the silencing or inactivation of the cell cycle inhibition genes (Smad4 and TGFβRII) and the apoptosis inhibition genes (p53 and Bax). Notice from Table 3 that the estimates of the proliferation rates of the I _{2} and I _{3} cells are approximately equal to those of the J _{3} and J _{4} cells respectively. These results seemed to suggest that the genomic instabilities had little effects on cell proliferations.
Conclusions and Discussion
Recent studies of cancer molecular biology have indicated very clearly that human colon cancer is developed through multiple pathways ([1–9]). This indicates that single pathway models are not realistic and hence may lead to incorrect prediction and confusing results. For developing efficient prevention and controlling procedures for human colon cancer and for prediction of future human colon cancer, in this paper we have developed a stochastic model and a state space model for carcinogenesis of human colon cancer involving multiple pathways with each pathway being a multistage model. Using this model, we have derived for the first time the probability distribution of the numbers of initiated cells and the probability distribution of time to cancer tumors. Such derivation by the traditional approach is extremely difficult and had not been attempted previously for multiple pathway models. Based on the state space model of colon cancer, we have developed a generalized Bayesian procedure to estimate the unknown parameters and to predict future cancer cases. This approach combines information from three sources: The stochastic system model via P{X, UΘ}, the prior information via P{Θ} and information from data via L{Θ , X, U}. Because of additional information from the stochastic system model, our procedure is advantageous over the standard Bayesian procedure and the sampling theory procedure. Notice that there are a large number of unknown parameters in the model and only a limited amount of data are available. Without this additional information, it is then not possible to estimate all unknown parameters. Notice also that through the stochastic system model, one can incorporate biological mechanism into the model. Because the number of stages and the mutation rates of intermediate cells in different pathways are different and different drugs may affect different pathways, we believe that this is important and necessary.
We have applied these models and procedure to the NCI SEER data (upto November, 2007). Our results showed that the proposed multiple pathways model fitted better than the single pathway 4stage model as proposed by Luebeck and Moolgavkar [17]. (The respective AIC and BIC for the multiple pathways model are 55.96 and 81.30 which are ten times smaller than those of the AIC (816.0667) and BIC (827.1513) respectively of the single pathway 4stage model.)
In this preliminary study, we have not yet compared the multiple pathways model with the single pathway model regarding prediction of future cancer cases and evaluation of treatment protocols for human colon cancer. This will be our future research, we will not go any further here.
Reviewers' comments
Reviewer 1 (M.P. Little)
General comments
This is a generally wellwritten paper, describing a model very similar to that recently developed by Little et al. [12], generalising the model of Little and Wright ([11]). Arguably thislatest model should be referenced, with discussion of mathematical differences between the model outlined here and that one. There could also be discussion of the somewhat different conclusions reached in fits to more or less the same SEER colon cancer data.
Response: The paper by Little et al. has been added to the reference; see Little et al. [12].
Specific comments (page/line)
(1) The 6stage model is suddenly drawn from the hat here, but is then almost immediately reduced to a 4stage model! Why does the haploinsufficiency of Smad4 and p53 justify combining the stages in this way? Is there evidence that Smad4 and p53 are without function at half gene dose? Also, what is the evidence for mutations being in the order given in Figure 1?
Response: References documenting haploinsufficiency of p53 and smad4 have been given in the paper. The reason why we can combine the twostages involving P53 into onestage is based on these papers. To illustrate, let B denote the P53 gene and b the mutant of p53. Then, under haploinsufficiency of P53, the level and effects of the P53 protein has been reduced significantly (at least 4fold or more) so that the phenotype of genotype B/b (or B/) is closely approximated by that of genotype b/b.
(2) How would a model incorporating epigenetic effects differ from the (DNA mutational) models already outlined? I suspect that mathematically the formalism would be exactly the same, although the implied "mutation rates" would be very much higher for epigenetic events.
Response: From modeling viewpoint, it is difficult to tell the difference between epigenetic changes and mutation except that the former is much more frequent and very often reversible. However, epigenetic changes can help the modeler to incorporate biological information into the biological process. Many biological papers (just during 2008, there are hundreds of biological papers published) support the epigenetic changes and the view that epigenetic changes are the driving force for cancer initiation, progression and metastasis, more important than gene mutations in cancer initiation and progression. Epigenetic changes include methylation (hypo and hypermethylation), microRNA (noncoding RNA), loss of imprinting, Histone acetylation, HDAC, tissue disorganization and gap junction disruption, etc.; epigenetic changes can also lead to gene mutations; for a brief review, see Tan and Hanin [31] (Chapter 3).
(3) This is slightly confusing. I assume that the two pathways referred to are CIN and MSI, but coming after discussion of epigenetic effects, perhaps this was meant as well or instead.
Response: The genetic sequence of the CIN and MSI pathways were determined by molecular biology of colon cancer and have been published in cancer journals. (There are a large number of biological papers documenting this. I have just listed a few of them in our paper. I can provide many more published papers from cancer journals if one wishes.) This sequence appears to be logical from biological mechanism. Notice that the APCβ Catenin activates myc and cyclin D to push the cell into cell cycle; for the cell cycle to progress, the inhibition effects of p15, p16, p18, p19, p27 have to be abrogated through the inhibition of the TGFβ signaling pathway (epigenetic silencing or inactivation or mutation or deletion of smad2/smad4 (CIN pathway) or TGFβ Receptor II (MSI pathway).). When the number of cells has increased to certain level, then the apoptosis p53Bax pathway is activated. For the progression of carcinogenesis, the p53 (CIN pathway) or the Bax gene (MSI pathway) have to be epigenetically silenced or inactivated or mutated or deleted. This is the reason why only the late stage involves silencing or inactivation of the gene p53. This is illustrated in our paper; see also Tan and Hanin [31] (Chapter 3, Chapter 11 and Chapter 12). There are no biological supports singlestage or two stage models. Hence it is not logical to accept single stage or two stage models for colon cancer for people who are born normal. Also, it is extremely difficult for me to accept that there are no proliferation for I_{2} cells as assumed in some of the cancer model papers simply because of the observation of polyps in colon which are derived from proliferation of secondstage cells and is the basis for colon cancer screening procedure "Colonoscopy" practiced by medical doctors.
(4) I assume that asymmetric mutations are assumed, in which a cell produces one normal daughter cell and one mutant daughter cell, as assumed by Little et al. [12] and Little and Wright [11] and many others, but this could perhaps be clarified. It might be useful to derive this Kolmogorov forward equation (2) in an Appendix. Clearly this forward equation (2) is in general intractable. I suspect that as in the papers of Little et al. [12] and Little and Wright [11] it is much easier to attempt to solve the Kolmogorov backward equations.
Response: As in Little, I assume that under genetic changes, a normal (or one I_{ j }cell) cell will give rise to one normal cell (or one I_{ j }cell) and one mutant cell (or one I_{j+1}cell)(asymmetric change or mutation). This is logical because mutation or genetic changes occur during cell division.
(5) These conclusions are somewhat at odds with those of Little et al. [12] and Little and Li [32] who fitted models to very similar SEER colon cancer data and demonstrated that there was little evidence of better fit of models that allowed for genomic instability compared with those that did not. In particular the 4stage model of Luebeck and Moolgavkar [16] provided as good fit as models that allowed for genomic instability [32], or as here multiple types of instability (CIN and MSI) [12]. The authors may care to discuss this.
Response: On November 13, 2009, Dr. Little had sent me his comments about our paper and a copy of his 2008 paper which I was not aware of before I wrote our paper. Dr. Little claimed that their fitting of the data could not differentiate between many different models. When I examined Dr. Little's paper, I found out that our estimation approach is very different from Dr. Little's. While Dr. Little's approach used the classical sampling theory through maximum likelihood estimates; our approach is the state space modeling and the generalized Bayesian inference incorporating information from three different sources: (1) The biological information from cancer molecular biology, (2) prior information from epigenetic and genetic mechanism, and (3) information from the likelihood function of observed data involving cancer incidence. Furthermore, because cancer incidence are derived only from people who are less than 100 years old (his/her life time), we first change the SEER data from (n_{ j }, y_{ j }) to (m_{ j }, y_{ j }), where m_{ j }is the number of people who can develop colon cancer during his life time (i.e., less that 100 yrars old.). Notice that n_{ j }is of the order 10^{7} while m_{ j }is of the order 10^{5}. This is described in detail in our paper in Tan and Hanin [31] (Chapter 11). Because of this and because of additional information from biological mechanism and prior distribution, we do not have the problem of identification of parameters which statisticians usually encounter in analyzing data without using information from biological mechanism. Thus, using classical approach, the parameters are not identifiable so that one has to make some assumptions such as that growth rates (birth rates) of different stage cancer initiated cells are equal which can hardly be realistic in carcinogenesis. (Biological studies by cancer biologists clearly demonstrated that the proliferation rates of cancer initiated cells with different genetic changes are very different in most of the cases.)
Finally, I want to emphazise that many models can fit the data but some fit better than others. Finally I like to emphasize what Dr. Van Ryzin had concluded 20 years ago that for cancer risk assessment, many models can fit the data but only models which are biologically supported can give correct results for cancer prediction and cancer risk assessment. It is important to list the predicted numbers along with the observed numbers in cancer modeling research. EPA has revised the guidelines to require that cancer risk assessment models should be biologically supported.
Reviewer 2 (M. Kimmel)
The paper by Tan and Yan, proposes a new stochastic model of colon cancer progression, involving the chromosomal instability pathway and the microsatellite instability pathway. This model not only might provide more insights into human colon cancer but also might provide useful guidance for its prevention and control and for prediction of future cancer cases. One interesting question is as to whether Tan and Yang model might help in deciding about the value of early detection of colon cancer by screening examination. Although colonoscopy is an accepted procedure, there are questions that linger concerning the impact of early detection on mortality reduction.
Response to Reviewer 2
Thank you very much for your comments and suggestion. We are currently collecting data on screening by colonoscopy. We will apply the model to estimate cancer incidence under screening and will examine if colonoscopy will help reduce cancer incidence. We will do computer simulation to find this out. This is our next research on colon cancer.
Abbreviations
 CIN:

Chromsomal Instability
 LOH:

Loss of Heterozygocity
 MSI:

MicroSatellite Instability
 APC:

Adenomatous Polyposis Coli
 MMR:

MisMatch Repair
 ML:

Multinomial
 Dev:

Deviance
 AIC:

Akaike Information Criterion
 BIC:

Bayesian Information Criterion
 SEER:

Surveillance Epidemiology and End Results.
Declarations
Acknowledgements
The research of this paper is supported by an NCI/NIH grant, R15 CA11334701 and an NSF grant, DMS0805829. As a onetime exception to the publishing policy of Biology Direct the articles in this series are being published with two reviewers.
Authors’ Affiliations
References
 de la Chapelle A: Genetic predisposition to colorectal cancer. Nature Review (Cancer). 2004, 4: 769780. 10.1038/nrc1453.View ArticleGoogle Scholar
 Fodde R, Smit R, Clevers H: APC, signal transduction and genetic instability in colorectal cancer. Nature Review (Cancer). 2001, 1: 5567. 10.1038/35094067.View ArticleGoogle Scholar
 Fodde R, Kuipers J, Rosenberg C, et al: Mutations in the APC tumor suppressor gene cause chromosomal instability. Nature Cell Biology. 2001, 3: 433438. 10.1038/35070129.PubMedView ArticleGoogle Scholar
 Green RA, Kaplan KB: Chromosomal instability in colorectal tumor cells is associated with defects in microtubule plusend attachments caused by a dominant mutation in APC. The Journal of Cell Biology. 2003, 163: 949961. 10.1083/jcb.200307070.PubMedPubMed CentralView ArticleGoogle Scholar
 Hawkins NJ, Ward RL: Sporadic colorectal cancers with microsatellite instability and their possible origin in hyperplastic polyps and serrated adenomas. J Natl Cancer Institute. 2001, 93: 13071313. 10.1093/jnci/93.17.1307.View ArticleGoogle Scholar
 Hisamuddin IM, Yang VW: Genetics of colorectal cancer. Medscape General Medicine. 2004, 6 (3): 2004Google Scholar
 Peltomaki P: Deficient DNA mismatch repair: A common etiologic factor for colon cancer. Human Molecular Genetics. 2001, 10: 735740. 10.1093/hmg/10.7.735.PubMedView ArticleGoogle Scholar
 Sparks AB, Morin PJ, Vogelstein B, Kinzler KW: Mutational analysis of the APC/betacatenin/Tcf pathway in colorectal cancer. Cancer Res. 1998, 58: 11301134.PubMedGoogle Scholar
 Ward R, Meagher A, Tomlinson I, O'Connor T, et al: Microsatellite instability and the clinicopathological features of sporadic colorectal cancer. Gut. 2001, 48: 821829. 10.1136/gut.48.6.821.PubMedPubMed CentralView ArticleGoogle Scholar
 Tan WY, Zhang LJ, Chen W, Zhu JM: A stochastic model of human colon cancer involving multiple pathways. "Handbook of Cancer Models with Applications". Edited by: Tan WY, Hanin L. 2008, World Scientific, Singapore and River Edge, NJ, Chapter 3:Google Scholar
 Little MP, Wright EG: A stochastic carcinogenesis model incorporating genomic instability fitted to colon cancer data. Math Biosci. 2003, 183: 111134. 10.1016/S00255564(03)000403.PubMedView ArticleGoogle Scholar
 Little MP, Vineis P, Li G: A stochastic carcinogenesis model incorporating multiple types of genomic instability fitted to colon cancer data. J Theor Biolpgy. 2008, 254: 229238. 10.1016/j.jtbi.2008.05.027.View ArticleGoogle Scholar
 Little MP: Cancer models, ionization and genomic instability: A review. "Handbook of Cancer Models with Applications". Edited by: Tan WY, Hanin L. 2008, World Scientific, Singapore and River Edge, NJ, Chapter 5:Google Scholar
 Jessup JM, Gallic GG, Liu B: The molecular biology of colorectal carcinoma: Importance of the Wg/Wnt signal transduction pathway. "The Molecular Basis of Human Cancer". Edited by: Coleman WB, Tsongalis GJ. 2002, Humana Press, Totowa, Chapter 13: 251268.View ArticleGoogle Scholar
 Alberici P, JagmohanChangur S, De Pater E, et al: Smad4 haploinsufficiency in mouse models for intestinal cancer. Oncogene. 2006, 25: 18411851. 10.1038/sj.onc.1209226.PubMedView ArticleGoogle Scholar
 Lynch CJ, Milner J: Loss of one p53v allele results in fourfold reduction in p53 mRNA and protein: A basis for p53 haploinsufficiency. Oncogene. 2006, 25: 34633470. 10.1038/sj.onc.1209387.PubMedView ArticleGoogle Scholar
 Luebeck EG, Moolgavkar SH: Multistage carcinogenesis and colorectal cancer incidence in SEER. Proc Natl Acad Sci USA. 2002, 99: 1509515100. 10.1073/pnas.222118199.PubMedPubMed CentralView ArticleGoogle Scholar
 Baylin SB, Ohm JE: Epigenetic silencing in cancera mechanism for early oncogenic pathway addiction. Nature Rev Cancer. 2006, 6: 107116. 10.1038/nrc1799.View ArticleGoogle Scholar
 Koinuma K, Yamashita Y, Liu W, Hatanaka H, Kurashina K, Wada T, et al: Epigenetic silencing of AXIN2 in colorectal carcinoma with microsatellite instability. Oncogene. 2006, 25: 139146.PubMedGoogle Scholar
 Breivik J, Gaudernack G: Genomic instability, DNA methylation, and natural selection in colorectal carcinogenesis. Semi Cancer Biol. 1999, 9: 245254. 10.1006/scbi.1999.0123.View ArticleGoogle Scholar
 Jones PA, Baylin SB: The fundamental role of epigenetic events in cancer. Nat Rev Genet. 2002, 3: 415428. 10.1038/nrg962.PubMedView ArticleGoogle Scholar
 Fakir H, Tan WY, et al: Stochastic population dynamic effects for lung cancer progression. Radiation Research. 2009, 172: 383393. 10.1667/RR1621.1.PubMedView ArticleGoogle Scholar
 Yakovlev AY, Tsodikov AD: Stochastic Models of Tumor Latency and Their Biostatistical Applications. 1996, World Scientific, Singapore and River Edge, New JerseyGoogle Scholar
 Tan WY: Stochastic Models of Carcinogenesis. 1991, Marcel Dekker, New YorkGoogle Scholar
 Tan WY: Stochastic Models With Applications to Genetics, Cancers, AIDS and Other Biomedical Systems. 2002, World Scientific, Singapore and River Edge, New JerseyGoogle Scholar
 Tan WY: Stochastic Modeling of AIDS Epidemiology and HIV Pathogenesis. 2000, World Scientific, Singapore and River Edge, New JerseyGoogle Scholar
 Box GEP, Tiao GC: Bayesian Inference in Statistical Analysis. 1973, AddisonWesley, Reading, MAGoogle Scholar
 Smith AFM, Gelfand AE: Bayesian statistics without tears: A samplingresampling perspective. American Statistician. 1992, 46: 8488. 10.2307/2684170.Google Scholar
 Hanahan D, Weinberg RA: The hallmarks of cancer. Cell. 2000, 100: 5770. 10.1016/S00928674(00)816839.PubMedView ArticleGoogle Scholar
 Weinberg RA: The Biology of Cancer. 2007, Garland Sciences, Taylor and Frances Group, New YorkGoogle Scholar
 Tan WY, Hanin L: Handbook of Cancer Models with Applications. 2008, World Scientific Publishing Co. Pte. Ltd, 9:Google Scholar
 Little MP, Li G: Stochastic modelling of colon cancer: is there a role for genomic instability?. Carcinogenesis. 2007, 28: 479487. 10.1093/carcin/bgl173.PubMedView ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.