Multivariate distributions of clinical covariates at the time of cancer detection.

Many screening trials conducted in the past have generated a wealth of interesting data. These data represent an invaluable source of information for furthering our knowledge about the natural history of the disease. The traditional approach to modeling cancer screening tends to describe the process of tumor development in only one dimension, that is, the time natural history. A broader methodological idea is to construct a stochastic model of cancer development and detection that yields the multivariate distribution of observable variables at the time of diagnosis. By focusing on such multivariate observations, rather than just on the age of patients at diagnosis, this idea seeks to invoke an additional source of information (available only at the time of detection) in order to improve an estimation of unobservable quantitative parameters of cancer latency. In this article, we discuss modeling techniques that make the above-mentioned problems approachable. A special focus is placed on analytical tools for deriving joint distributions of clinical covariates at the time of cancer detection under an arbitrary screening protocol. In addition, some future research avenues and public health implications of the proposed approach are discussed.

Estimating Cure Rates From Survival Data: An Alternative to Two-Component Mixture Models.

This article considers the utility of the bounded cumulative hazard model in cure rate estimation, which is an appealing alternative to the widely used two-component mixture model. This approach has the following distinct advantages: (1) It allows for a natural way to extend the proportional hazards regression model, leading to a wide class of extended hazard regression models. (2) In some settings the model can be interpreted in terms of biologically meaningful parameters. (3) The model structure is particularly suitable for semiparametric and Bayesian methods of statistical inference. Notwithstanding the fact that the model has been around for less than a decade, a large body of theoretical results and applications has been reported to date. This review article is intended to give a big picture of these modeling techniques and associated statistical problems. These issues are discussed in the context of survival data in cancer.

Variable selection and pattern recognition with gene expression data generated by the microarray technology.

Lack of adequate statistical methods for the analysis of microarray data remains the most critical deterrent to uncovering the true potential of these promising techniques in basic and translational biological studies. The popular practice of drawing important biological conclusions from just one replicate (slide) should be discouraged. In this paper, we discuss some modern trends in statistical analysis of microarray data with a special focus on statistical classification (pattern recognition) and variable selection. In addressing these issues we consider the utility of some distances between random vectors and their nonparametric estimates obtained from gene expression data. Performance of the proposed distances is tested by computer simulations and analysis of gene expression data on two different types of human leukemia. In experimental settings, the error rate is estimated by cross-validation, while a control sample is generated in computer simulation experiments aimed at testing the proposed gene selection procedures and associated classification rules.

A survival model for fractionated radiotherapy with an application to prostate cancer.

This paper explores the applicability of a mechanistic survival model, based on the distribution of clonogens surviving a course of fractionated radiation therapy, to clinical data on patients with prostate cancer. The study was carried out using data on 1,100 patients with clinically localized prostate cancer who were treated with three-dimensional conformal radiation therapy. The patients were stratified by radiation dose (group 1: <67.5 Gy; group 2: 67.5-72.5 Gy; group 3: 72.5-77.5 Gy; group 4: 77.5-87.5 Gy) and prognosis category (favourable, intermediate and unfavourable as defined by pre-treatment PSA and Gleason score). A relapse was recorded when tumour recurrence was diagnosed or when three successive prostate specific antigen (PSA) elevations were observed from a post-treatment nadir PSA level. PSA relapse-free survival was used as the primary end point. The model, which is based on an iterated Yule process, is specified in terms of three parameters: the mean number of tumour clonogens that survive the treatment, the mean of the progression time of post-treatment tumour development and its standard deviation. The model parameters were estimated by the maximum likelihood method. The fact that the proposed model provides an excellent description both of the survivor function and of the hazard rate is prima facie evidence of the validity of the model because closeness of the two survivor functions (empirical and model-based) does not generally imply closeness of the corresponding hazard rates. The estimated cure probabilities for the favourable group are 0.80, 0.74 and 0.87 (for dose groups 1-3, respectively); for the intermediate group: 0.25, 0.51, 0.58 and 0.78 (for dose groups 1-4, respectively) and for the unfavourable group: 0.0, 0.27, 0.33 and 0.64 (for dose groups 1-4, respectively). The distribution of progression time to tumour relapse was found to be independent of prognosis group but dependent on dose. As the dose increases the mean progression time decreases (41, 28.5, 26.2 and 14.7 months for dose groups 1-4, respectively). This analysis confirms that, in terms of cure rate, dose escalation has a significant positive effect only in the intermediate and unfavourable groups. It was found that progression time is inversely proportional to dose, which means that patients recurring in higher dose groups have shorter recurrence times, yet these groups have better survival, particularly long-term. The explanation for this seemingly illogical observation lies in the fact that less aggressive tumours, potentially recurring after a long period of time, are cured by higher doses and do not contribute to the recurrence pattern. As a result, patients in higher dose groups are less likely to recur; however, if they do, they tend to recur earlier. The estimated hazard rates for prostate cancer pass through a clear-cut maximum, thus revealing a time period with especially high values of instantaneous cancer-specific risk; the estimates appear to be nonproportional across dose strata.

Modeling cancer detection: tumor size as a source of information on unobservable stages of carcinogenesis.

This paper is concerned with modern approaches to mechanistic modeling of the process of cancer detection. Measurements of tumor size at diagnosis represent a valuable source of information to enrich statistical inference on the processes underlying tumor latency. One possible way of utilizing this information is to model cancer detection as a quantal response variable. In doing so, one relates the chance of detecting a tumor to its current size. We present various theoretical results emerging from this approach and illustrate their usefulness with numerical examples and analyses of epidemiological data. An alternative approach based on a threshold type mechanism of tumor detection is briefly described.

Distribution of the number of clonogens surviving fractionated radiotherapy: a long-standing problem revisited.

PURPOSE: A long-standing problem is addressed: what form of the probability distribution for the number of clonogenic tumor cells remaining after fractionated radiotherapy should be used in the analysis aimed at evaluating the efficacy of cancer treatment? Over a period of years, a lack of theoretical results leading to a closed-form analytic expression for this distribution, even under very simplistic models of cell kinetics in the course of fractionated radiotherapy, was the most critical deterrent to the development of relevant methods of data analysis. MATERIALS AND METHODS: Rigorous mathematical results associated with a model of fractionated irradiation of tumors based on the iterated birth and death stochastic process are discussed. RESULTS: A formula is presented for the exact distribution of the number of clonogenic tumor cells at the end of treatment. It is shown that, under certain conditions, this distribution can be approximated by a Poisson distribution. An explicit formula for the parameter of the limiting Poisson distribution is given and sample computations aimed at evaluation of the convergence rate are reported. Another useful limit that retains a dose-response relationship in the distribution of the number of clonogens has been found. Practical implications of the key theoretical findings are discussed in the context of survival data analysis. CONCLUSIONS: This study answers some challenging theoretical questions that have been under discussion over a number of years. The results presented in this work provide mechanistic motivation for parametric regression models designed to analyze data on the efficacy of radiation therapy.

An alternative stochastic model of generation of oligodendrocytes in cell culture.

According to our previous model, oligodendrocyte--type 2 (O-2A) astrocyte progenitor cells become competent for differentiation in vitro after they complete a certain number of critical mitotic cycles. After attaining the competency to differentiate, progenitor cells divide with fixed probability p in subsequent cycles. The number of critical cycles is random; analysis of data suggests that it varies from zero to two. The present paper presents an alternative model in which there are no critical cycles, and the probability that a progenitor cell will divide again decreases gradually to a plateau value as the number of completed mitotic cycles increases. In particular all progenitor cells have the ability to differentiate from the time of plating. The Kiefer-Wolfowitz procedure is used to fit the new model to experimental data on the clonal growth of purified O-2A progenitor cells obtained from the optic nerves of 7 day old rats. The new model is shown to fit the experimental data well, indicating that it is not possible to determine whether critical cycles exist on the basis of these experimental data. In contrast to the fit of the previous model, which suggested that the addition of thyroid hormone increased the limiting probability of differentiation as the number of mitotic cycles increases, the fit of the new model suggests that the addition of thyroid hormone has almost no effect on the limiting probability of differentiation.

Estimation problems associated with stochastic modeling of proliferation and differentiation of O-2A progenitor cells in vitro.

Our previous research effort has resulted in a stochastic model that provides an excellent fit to our experimental data on proliferation and differentiation of oligodendrocyte type-2 astrocyte progenitor cells at the clonal level. However, methods for estimation of model parameters and their statistical properties still remain far away from complete exploration. The main technical difficulty is that no explicit analytic expression for the joint distribution of the number of progenitor cells and oligodendrocytes, and consequently for the corresponding likelihood function, is available. In the present paper, we overcome this difficulty by using computer-intensive simulation techniques for estimation of the likelihood function. Since the output of our simulation model is essentially random, stochastic optimization methods are necessary to maximize the estimated likelihood function. We use the Kiefer-Wolfowitz procedure for this purpose. Given sufficient computing resources, the proposed estimation techniques significantly extend the spectrum of problems that become approachable. In particular, these techniques can be applied to more complex branching models of multi-type cell systems with dependent evolutions of different types of cells.

A stochastic model of temporally regulated generation of oligodendrocytes in cell culture.

The results of our previous analyses suggest that O-2A progenitor cells become competent for differentiation in vitro after they complete a certain number of critical mitotic cycles. The number of critical cycles varies from clone to clone and should be thought of as a random variable. We propose an approach to the analysis of oligodendrocyte generation in vitro based on a stochastic model allowing for an arbitrary distribution of this random variable with a finite support. When applied to experimental data on clonal growth and differentiation of purified O-2A progenitor cells obtained from optic nerves of 1 and 7 day-old rats, the model provides a good quantitative description not only of the first two moments (mean and variance) of the number of O-2A progenitor cells and oligodendrocytes at different times after the start of experiment, but of the corresponding distributions as well. As our estimates show, there are scarcely any O-2A progenitor cells that divide in vitro more than twice before they acquire the competence for differentiation. Those O-2A cells that have undergone the critical divisions differentiate into an oligodendrocyte in each of the subsequent mitotic cycles with a certain probability. We give estimates of this probability for O-2A cells under different growth conditions. Our analysis suggests that the effect of thyroid hormone is twofold: it reduces the mean duration of the mitotic cycle for progenitor cells, and it increases the probability of their transformation into oligodendrocytes.

The shape of the hazard function in breast carcinoma: curability of the disease revisited.

BACKGROUND: The question of curability of breast carcinoma remains controversial. Because the probability of cure essentially is an asymptotic notion, the corresponding estimation problems call for special statistical methods. Such methods should account for an intimate connection between the probability of cure and the shape of the hazard function. METHODS: The study was performed on survival data for 13,166 women with breast carcinoma identified through the Utah Cancer Registry and stratified by clinical stage and age at diagnosis. For these patients, the follow-up period was 30 years. Three estimation procedures were used for estimating the hazard function from the data: the life table estimator, a kernel counterpart of the Nelson-Aalen estimator, and a parametric estimator specifically designed for two-component hazards. The parametric estimate of the hazard function was used to provide estimates of cure rates for each category of patients. RESULTS: For all categories of patients under study, the estimated hazard functions passed through a clear-cut maximum, showing a tendency to decrease as time approached the end of a follow-up period. The hazards appeared to be nonproportional across the strata. The estimated values of the cure rate and the corresponding confidence intervals were determined for each stratum of patients with breast carcinoma. CONCLUSIONS: The results of the current study strongly suggest that cure is a possible outcome of breast carcinoma treatment. The condition of proportionality of risks is not met in breast carcinoma survival data.

Quantitative insight into proliferation and differentiation of oligodendrocyte type 2 astrocyte progenitor cells in vitro.

As part of our attempts at understanding fundamental principles that underlie the generation of nondividing terminally differentiated progeny from dividing precursor cells, we have developed approaches to a quantitative analysis of proliferation and differentiation of oligodendrocyte type 2 astrocyte (O-2A) progenitor cells at the clonal level. Owing to extensive previous studies of clonal differentiation in this lineage, O-2A progenitor cells represent an excellent system for such an analysis. Previous studies have resulted in two competing hypotheses; one of them suggests that progenitor cell differentiation is symmetric, the other hypothesis introduces an asymmetric process of differentiation. We propose a general model that incorporates both such extreme hypotheses as special cases. Our analysis of experimental data has shown, however, that neither of these extreme cases completely explains the observed kinetics of O-2A progenitor cell proliferation and oligodendrocyte generation in vitro. Instead, our results indicate that O-2A progenitor cells become competent for differentiation after they complete a certain number of critical mitotic cycles that represent a period of symmetric development. This number varies from clone to clone and may be thought of as a random variable; its probability distribution was estimated from experimental data. Those O-2A cells that have undergone the critical divisions then may differentiate into an oligodendrocyte in each of the subsequent mitotic cycles with a certain probability, thereby exhibiting the asymmetric type of differentiation.

A stochastic model of brain cell differentiation in tissue culture.

The timing of cell differentiation can be controlled both by cellintrinsic mechanisms and by cell-extrinsic signals. Oligodendrocyte type-2 astrocyte progenitor cells are known to be the precursor cells that give rise to oligodendrocytes. When stimulated to divide by purified cortical astrocytes or by platelet-derived growth factor, these progenitor cells generate oligodendrocytes in vitro with a timing like that observed in vivo. The most widely accepted model of this process assumes a cell-intrinsic biological clock that resides in the progenitor cell. The intrinsic clock model originally proposed in 1986 remains as the dominant theoretical concept for the analysis of timed differentiation in this cell lineage. However, the results of a recent experimental study (Ibarrola et al., Developmental Biology, vol. 180, 1-21, 1996) are most consistent with the hypothesis that the propensity of a clone of dividing O-2A progenitor cells initially to generate at least one oligodendrocyte may be regulated by cell-intrinsic mechanisms, but that environmental signals regulate the extent of further oligodendrocyte generation. We propose a stochastic model of cell differentiation in culture to accommodate the most recent experimental findings. Our model is an age-dependent branching stochastic process with two types of cells. The model makes it possible to derive analytical expressions for the expected number of progenitor cells and of oligodendrocytes as functions of time. The model parameters were estimated by fitting these functions through data on the average (sample mean) number of both types of cells per colony at different time intervals from start of experiment. Using this method we provide a biologically meaningful interpretation of the observed pattern of oligodendrocyte generation in vitro and its modification in the presence of thyroid hormone.

A model of multiple tumorigenesis allowing for cell death: quantitative insight into biological effects of urethane.

This paper considers the utility of a stochastic model of carcinogenesis proposed by Yakovlev and Polig [Math. Biosci. 132 (1996) 1-33] in the analysis of experimental data on multiple tumors induced by chemical carcinogens. The model provides a good description of published data on multiple tumors developing in the lungs of mice in response to different schedules of urethane. The distribution of pulmonary tumor counts appears to be negative binomial for each period of time after exposure to urethane. Our results suggest that the rate of administration of urethane has little effect both on the mean number of initiated cells per unit dose and on the rate of formation of lesions responsible for cell death. As our estimates show, more than 80% of initiated cells die in the course of tumor promotion. The model is robust to variations in the rate of urethane excretion given a fixed total dose of the carcinogen. Some prospects for further development of the model to allow for expansion of promoted cell clones are discussed.

Estimating the probability of initiated cell death before tumor induction.

The effects of cell toxicity are known to be inherent in carcinogenesis induced by radiation or chemical carcinogens. The event of cell death precludes tumor induction from occurring. A long standing problem is to estimate the proportion of initiated cells that die before tumor induction. No experimental techniques are currently available for directly gauging the rate of cell death over extended periods of time. The obstacle can be surmounted by newly developed theoretical methods of carcinogenesis modeling. In this paper, we apply such methods to published data on multiple lung tumors in mice receiving different schedules of urethane. Bioassays of this type play an important role in testing environmental chemicals for carcinogenic activity. Our estimates for urethane-induced carcinogenesis show that, unexpectedly, many initiated cells die early in the course of tumor promotion. We present numerical estimates for the probability of initiated cell death for different schedules (and doses) of urethane administration.

Mechanistic modeling of multiple tumorigenesis: an application to data on lung tumors in mice exposed to urethane.

This paper discusses an application of the stochastic model of carcinogenesis proposed by Yakovlev and Polig to the analysis of experimental data on multiple pulmonary tumors induced by urethane. The results of this analysis suggest that the rate of administration of urethane has no effect on the mean number of initiated cells per unit dose. Likewise, the formation of lesions responsible for cell death appears to be unaffected by the dose rate.

A distribution of tumor size at detection: an application to breast cancer data.

This paper discusses a method of estimating numerical characteristics of unobservable stages of carcinogenesis from data on tumor size at detection. To this end, a stochastic model of spontaneous carcinogenesis has been developed to allow for a simple pattern of tumor growth kinetics. It is assumed that a tumor becomes detectable when its size attains some threshold level, which is treated as a random variable. The model yields a parametric family of joint distributions for tumor size and age at detection. Some estimation problems associated with the proposed model appear to be tractable. This is illustrated with an application to the statistical analysis of data on primary breast cancer.

Queueing models of potentially lethal damage repair in irradiated cells.

Some of the ideas arising in queueing theory are applied to describe the repair mechanisms responsible for recovery of cells from potentially lethal radiation damage. Two alternative versions are presented of a queueing model of damage repair after a single dose of irradiation. The first version represents a linear misrepair model, and the second invokes the idea of spontaneous lesion fixation. They are pieced together in the third model, allowing for both mechanisms. The consistency of the proposed models with published experimental data is tested.

A distribution of tumor size at detection and its limiting form.

A distribution of tumor size at detection is derived within the framework of a mechanistic model of carcinogenesis with the object of estimating biologically meaningful parameters of tumor latency. Its limiting form appears to be a generalization of the distribution that arises in the length-biased sampling from stationary point processes. The model renders the associated estimation problems tractable. The usefulness of the proposed approach is illustrated with an application to clinical data on premenopausal breast cancer.