A new view of radiation-induced cancer: integrating short- and long-term processes. Part I: Approach
Mathematical models of radiation carcinogenesis are important for understanding mechanisms and for interpreting or extrapolating risk. There are two classes of such models: (1) long-term formalisms that track pre-malignant cell numbers throughout an entire lifetime but treat initial radiation dose–response simplistically and (2) short-term formalisms that provide a detailed initial dose–response even for complicated radiation protocols, but address its modulation during the subsequent cancer latency period only indirectly. We argue that integrating short- and long-term models is needed. As an example of this novel approach, we integrate a stochastic short-term initiation/inactivation/repopulation model with a deterministic two-stage long-term model. Within this new formalism, the following assumptions are implemented: radiation initiates, promotes, or kills pre-malignant cells; a pre-malignant cell generates a clone, which, if it survives, quickly reaches a size limitation; the clone subsequently grows more slowly and can eventually generate a malignant cell; the carcinogenic potential of pre-malignant cells decreases with age.
Short- and long-term biologically based models
Biologically motivated mathematical modeling of background and ionizing radiation-induced carcinogenesis has a history spanning more than 50 years (Nordling 1953). Many of the models can be characterized as short-term, in that they focus on processes occurring during and shortly after irradiation (Hahnfeldt and Hlatky 1998; Radivoyevitch et al. 2001; Mebust et al. 2002; Schollnberger et al. 2002; Sachs and Brenner 2005; Hofmann et al. 2006; Shuryak et al. 2006; Little 2007; Sachs et al. 2007; Schneider and Walsh 2008). The main advantage of such models is that they provide a detailed initial dose–response relation for short-term endpoints, which are used as surrogates for carcinogenesis. The main disadvantage is that the possibly substantial modulations of the magnitude and shape of this initial dose–response during the lengthy period (multiple years-decades) between irradiation and manifestation of typical solid tumors are not considered mechanistically.
In contrast, another class of biologically motivated models can be characterized as long-term, in the sense that they track carcinogenesis rates throughout the entire life span, e.g., the Armitage-Doll model (Armitage and Doll 1954; Armitage 1985), the Moolgavkar-Venzon-Knudson two-stage clonal expansion (TSCE) model (Moolgavkar 1978, 1980; Moolgavkar and Knudson 1981), the two-stage logistic model (Sachs et al. 2005), and many others (Yakovlev and Polig 1996; Pierce and Mendelsohn 1999; Wheldon et al. 2000; Little and Wright 2003; Pierce and Vaeth 2003; Ritter et al. 2003; Little and Li 2007). The main advantage of long-term models is the more detailed treatment of slow carcinogenesis processes, including the modulation of the radiation dose–response during the long latency period. The main disadvantage is that the initial dose–response is typically treated in a non-mechanistic, phenomenological manner.
The need for a new approach: integration of long- and short-term models
Motivations for the specific unification used here
Here, we provide a specific example of integrating short- and long-term radiation carcinogenesis models. In that a variety of short- and long-term models have been proposed, our approach should be considered as illustrative, intended to investigate the practicality of integrating the two model types. The goal was to produce a novel formalism, which can describe typical patterns of background and radiogenic carcinogenesis with the smallest possible number of adjustable parameters. This goal required multiple simplifying assumptions about the complex multi-step carcinogenesis process. We suggest that for the purposes of generating a preliminary integrated model, the consequent reduction in biological realism is compensated by the reduction in model complexity and number of parameters. More complicated examples of unifying long- and short-term models are certainly possible, e.g., models analyzing multiple pre-malignant cell stages, analyzing genomic instability, and/or analyzing stochastic effects at both time scales, rather than just at short-time scales as we shall do.
The short-term part of our model (Fig. 1a) is the more mathematically intensive part of the integrated formalism. It is based on existing initiation, inactivation and repopulation (iir) models originally designed for analyzing second cancers induced by radiotherapy (Lindsay et al. 2001; Sachs and Brenner 2005; Shuryak et al. 2006; Little 2007; Sachs et al. 2007). This short-term part analyzes normal and pre-malignant stem cells during complex radiation exposure regimens and during the following weeks of tissue recovery. It tracks individual pre-malignant cell clones rather than just the total number of pre-malignant cells. The probability that a pre-malignant clone becomes extinct during radiotherapy can be substantial, so a fully stochastic formalism is used for the population dynamics of pre-malignant cells.
Normal and pre-malignant stem cells are located in specific tissue niches (Slack 2000; Borthwick et al. 2001; Potten and Booth 2002; Bennett et al. 2003; Fuchs et al. 2004; Ghazizadeh and Taichman 2005; Li and Xie 2005). We here have in mind a general concept of a niche, including, e.g., a clone whose further growth is prevented or slowed by any kind of micro-environmental constraints. The number of niches per organ, and the number of normal or pre-malignant stem cells per niche, is homeostatically regulated (Fuchs et al. 2004; Li and Xie 2005). We allow for the possibility that pre-malignant stem cells are regulated somewhat less stringently than their normal counterparts, so that a niche filled with pre-malignant cells can contain more such cells on average than a niche filled with normal cells.
An initiated stem cell either dies out, or grows into a pre-malignant clone, which quickly (e.g., within a year, Campbell et al. 1996) fills the entire stem cell niche in which it originated, i.e., “initiates” the whole niche. To expand beyond the first niche, the pre-malignant clone needs to invade an adjacent niche; alternatively, the niche containing the clone can divide into two daughter niches, e.g., colon crypt fission (Greaves et al. 2006; Johnston et al. 2007; Edwards and Chapman 2007). These processes require years or decades, and may involve acquisition of new mutations (Spencer et al. 2006).
Pre-malignant cells in all niches are assumed to gradually lose their carcinogenic potential with age, so that at old age they have a progressively smaller probability of being transformed to malignant cells. This assumption is suggested by a decrease in incidence of most cancers at very old age, e.g., >80 years in humans [Surveillance Epidemiology and End Results (SEER) database, http://seer.cancer.gov] and >800 days in mice (Pompei et al. 2001). The likely mechanism is senescence of stem cells and/or deterioration of stem cell function, or niche function, with age (Brunet and Rando 2007; Carlson and Conboy 2007; Sharpless and DePinho 2007).
To reduce the number of model parameters, clonal expansion of pre-malignant stem cell niches is treated deterministically, using a net proliferation rate. In other words, it is estimated by a deterministic “exponential model”, instead of the stochastic TSCE model. Given realistic parameter values, the deterministic and stochastic versions of the two-stage model are numerically very similar from birth until old age (Heidenreich and Hoogenveen 2001). In our analyses, by the time irradiation is over and cell repopulation in the exposed organ has occurred, surviving pre-malignant clones have grown to substantial sizes, making extinction very unlikely. This reduces the need for stochastic treatment of clone dynamics during the decades following radiation exposure. At old age, the deterministic approximation predicts an increasing hazard, whereas the stochastic TSCE model predicts a plateau. Considering the evidence cited above about a decrease in cancer incidence at old age, neither behavior is fully realistic. We here instead model the downturn in incidence at old age explicitly, according to the assumption above.
Unification of short- and long-term models
The integration of long- and short-term models will have two features typical of multi-timescale modeling (Engquist and Runborg 2005): (1) information is passed in both directions between the two components; and (2) a formally infinite time interval in the short-term model represents a short time interval, here typically several months, in the long-term model (Fig. 1).
The deterministic long-term equations provide the mean number of niches filled with pre-malignant stem cells and the mean number of pre-malignant cells per niche just before irradiation. The stochastic short-term equations then provide the number of these niches that are eradicated by radiation as well as the number of pre-malignant clones that are induced by, and survive, the radiation exposure. Each of these clones is assumed to be independent (i.e., located in a different part of the organ), and to be capable of quickly filling a niche with pre-malignant cells. Therefore, the total number of pre-malignant niches soon after irradiation can be calculated by the stochastic short-term model. The mean of this number is the initial condition for the deterministic long-term equations, which are applied from this point onwards until old age (Fig. 1). Cancer incidence is assumed to be proportional to the total number of pre-malignant cells in all niches, shifted by a lag time. A mathematical description of the model is provided below.
Materials and methods
Summary of model parameters
Spontaneous stem cell initiation and transformation
Pre-malignant niche replication
Age-dependent pre-malignant cell senescence
Homeostatic regulation of pre-malignant cell number per niche
Carrying capacity for pre-malignant cells per niche
Stem cell inactivation by radiation
Maximum net stem cell proliferation (repopulation) rate
Lag period between the first malignant cell and cancer
Long-term model for background cancers
The long-term model was intended mainly to place the results of the stochastic short-term calculations in an appropriate context, enabling estimation of the effects of age at exposure and time since exposure on predicted cancer risk. Simplicity and parsimony were emphasized. The long-term model in the absence of radiation consists just of a few rather simple deterministic equations, as follows.
The functions A and H are numerically very similar for realistic parameter values; they diverge substantially only if cancer incidence is high. We use the exact hazard H for data fitting in the companion paper, but use the simpler expression for A in the equations below, keeping in mind its interpretation and limitations.
Effects of radiation
The exposure scenario analyzed here is radiotherapy for an existing malignancy, where there are K daily dose-fractions, all of equal size d in some nearby organ, with treatment gaps during the weekends. Straightforward generalizations to variable doses per fraction, to other temporal patterns, and/or to cases where one must consider dose-volume histograms are omitted for brevity. A single acute dose exposure is a simple special case, where the number of fractions is K = 1.
The short-term part of the model considers initiation, inactivation, and repopulation (iir) of normal and pre-malignant stem cells during the radiation regimen and a recovery period of about a month (Fig. 1). Niche takeover by pre-malignant cells that survive the irradiation and modulation of niche size by radiogenic promotion (discussed below) are here also considered short-term processes. They are assumed to be essentially completed by a few months following exposure, before the long-term processes of spontaneous initiation and niche replication, which operate on the time scale of multiple years, start to have an appreciable effect (Fig. 1).
Normal stem cells are treated deterministically because their number per organ is assumed to be large. In contrast, the number of pre-malignant stem cells per clone is much smaller, and extinction of some pre-malignant clones is a real possibility. Consequently, a stochastic approach described below is used to estimate the average number of live pre-malignant niches when post-irradiation long-term processes begin.
Equations (7–9) allow the calculation of the normal stem cell number n(t) in the entire organ at all relevant times t throughout the radiation exposure and recovery periods. Explicit results for n(t) are provided in the Appendix.
We assume that the long-term growth advantage of pre-malignant cells manifests itself only on the scale of years and decades, and is negligible on the much shorter time scale of a few weeks of radiotherapy. Consequently, the maximum net proliferation rates for normal and pre-malignant stem cells are assumed to be equal, described by the parameter λ.
Unifying short- and long-term processes
Calculation of absolute and relative cancer risks
Note that the senescence parameter (c) cancels out of this ERR expression; due to the way we have defined our radiation initiation parameter (X), the spontaneous initiation parameter (a) also cancels out. The term Q1 can be interpreted as the normalized size of old pre-malignant stem cell niches. Q2 is proportional to the number of such niches. Q3 is proportional to the number of new pre-malignant niches, and Q4 is proportional to the total number of pre-malignant niches under background conditions.
Figure 4 shows an example of a model-generated age-dependent background incidence curve for an adult-onset solid cancer, using the four relevant parameters a, b, c, and L. The shape of this curve agrees well with the data for many cancers (e.g., SEER database, http://seer.cancer.gov), including the downturn in incidence at old age, which is not described as well by standard models.
Promotion-driven ERR can also be modulated by time after exposure. This occurs due to the assumption that the number of pre-malignant cells per niche is homeostatically regulated (by parameter δ), so that radiation-induced hyper-proliferation of cells within their niches can be reversed, as pre-irradiation cell birth/death rates, and hence niche sizes, are restored. If δ > 0, ERR due to promotion will decrease over time following exposure. This effect is seen in Figs. 6, 7: If the risk is measured at some constant age (e.g., 70 years), which is a sum of age at exposure and time since exposure, a decrease in ERR with time since exposure due to a δ > 0 will appear as an increase in the ERR with age at exposure (Fig. 6). A decrease in promotion-driven ERR with time since exposure also can have a conceptually important effect on the dose–response—at longer times after exposure, not only the magnitude, but also the shape, of the dose–response can change (Fig. 7).
The formalism presented here is the first comprehensive attempt to unify short- and long-term modeling approaches. The short-term part of the model belongs to the previously discussed iir category. It tracks the numbers of pre-malignant cells throughout irradiation stochastically. The long-term part of the model builds on the concepts developed in previous two-stage formalisms by adding an analysis of some aspects of tissue architecture (i.e., stem cell niches/compartments) and aging of pre-malignant stem cells. The particular short and long-term models that we have chosen to use are not crucial—the real issue is the integration. Certainly other long-term models could perfectly well be integrated into this short-/long-term framework—and we hope they will be.
The unified formalism has a number of advantages. The short-term part can generate reasonable predictions even at high doses, such as those in cancer radiotherapy. The long-term part analyzes the entire lifetime of the individual, putting the short-term predictions in an appropriate context by estimating the effects of age at exposure and time since exposure. The combined approach therefore allows the dose–response for the number of pre-malignant cells to be examined at any time point, from the start of irradiation until development of cancer years to decades later, which is not possible using either short- or long-term models alone.
Our model can be used for estimating risks of second malignancies induced by radiotherapy. This issue is growing in importance (Brenner et al. 2000; Ron 2006) as patients are treated earlier in life and the number of cancer survivors increases (Editorial 2004); the lifetime risk of radiation-induced second cancers is not negligible (Brenner et al. 2007). Direct measurement of second cancer incidence requires decades of follow-up because the latency period for radiogenic solid tumors is long (Tokunaga et al. 1979; Brenner et al. 2000; Ivanov et al. 2004, 2009). Meanwhile, radiotherapy protocols are rapidly changing. Our model, calibrated using data from older protocols, can produce risk predictions for any modern or prospective radiotherapeutic protocol. This application of the model is discussed in the accompanying paper (Shuryak et al. 2009).
Research supported by National Cancer Institute grant 5T32-CA009529 (IS), Department Of Energy DE-FG02-03ER63668 and National Institutes of Health CA78496 (PH), National Aeronautics and Space Administration 03-OBPR-07-0059-0065(LH), National Aeronautics and Space Administration NSCOR NNJ04HJ12G/NNJ06HA28G (RKS), National Institutes of Health grants P41 EB002033-09 and P01 CA-49062 (DJB).
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.