In Silico Prediction of Carcinogenicity

Cancer is a disorder in which the mechanisms that control proliferation of cells no longer function adequately. Cancer is considered to be an all-or-none effect (i.e., a tumor is or is not present). Because most cancers are considered irreversible life-threatening changes, chemical carcinogenesis (chemically induced cancer development) has hitherto been perceived by many to be different from other forms of chemical toxicity.

Currently there are no good in vitro assays that can be used to identify carcinogens. Most testing needs to be carried out in vivo, with bioassays for carcinogenicity consisting of exposing adequate numbers of animals to the test substance for a significant proportion of their normal life span. Typically, this is performed using both male and female rats and male and female mice over a 2 year test period. This combination helps to distinguish species and/or gender specificity for the chemical being tested. This test is expensive and time-consuming to conduct, and therefore alternatives or surrogate tests are often employed in advance of carcinogenicity testing, which thereby reduces the number of chemicals for which carcinogenicity data are generated.

One of these surrogate paradigms is genetic toxicity testing, as discussed in the previous section. However, some chemicals can cause cancer by mechanisms other than through direct interaction with DNA,48,49 and are therefore not detected through the use of genotoxicity tests. These classes of compounds are commonly referred to as nongenotoxic or epigenetic carcinogens. This diversity of mechanisms and high cost of experimental testing has been a significant driving force for the development of in silico models for carcinogenicity.

Despite the high cost to test chemicals, there are in fact many sources of data for carcinogenicity. For example, the International Agency for Cancer Research (IARC) has been compiling and reviewing data on chemicals suspected of causing carcinogenicity in humans for over three decades.50 Surprisingly there are only 68 unique substances that have been classified as IARC Group 1 Carcinogens (known human carcinogens) and only a further 56 substances categorized as IARC Group 2A Carcinogens (probable human carcinogens). The US NTP has been conducting and reviewing animal toxicity data for chemicals, and to date has compiled carcinogenicity data on over 530 chemicals and mixtures. Other sources of data include the Carcinogenic Potency Database (CPDB), compiled by Gold etal.,51,52 which contains over 1300 chemicals and mixtures, both naturally occurring and synthetic, compiled from public literature sources. It is worth noting that there is significant overlap between the NTP and Gold databases.

A review of the CPDB shows that around 50% of the compounds analyzed were in fact considered to be carcinogenic in rodents.53 The explanation for this probably lies in the fact that the rodents are given the chemical the maximum tolerated dose (i.e., at chronic, near-toxic doses). Evidence is accumulating to suggest that the high dose itself, rather than the chemical per se, can contribute toward the expression of cancer.53 The relevance therefore of rodent carcinogenicity testing to the situation where humans are exposed to relatively low levels of these chemicals has to be questioned. The high proportion of positive chemicals has also been attributed to the selective testing of chemicals suspected to cause cancer. One advantage that this bias has yielded is that it makes the data set much more amenable to the development of in silico models.

Many methods have been applied to create predictive models of carcinogenicity with varying degrees of success. While the carcinogenicity endpoint is usually expressed as a binary classification (i.e., positive or negative), some methods have used estimates of carcinogenic potency to differentiate between those that cause cancer at relatively low exposures versus those that require high doses and prolonged exposures. One common method used for determining potency is the use of the TD50 approach. The TD50 approach was described by Peto et al. in 1984. Put simply, the TD50 is the calculated daily dose per unit body weight causing a 50% incidence above control of tumors at the most sensitive site.

As with the prediction of genotoxicity, some developers of in silico models have focused their attention on specific classes of chemicals such as aromatic amines, nitroarenes, or nitrosamines. For a comprehensive review of structure-based approaches for predicting mutagenicity and carcinogenicity, refer to Benigni.55

5.39.3.2.1 Prospective challenges for carcinogenicity prediction

While most developers use a test set to validate the performance of a model, the true test of its predictivity comes from a prospective analysis, that is, where predictions are made in advance of the experimental tests being conducted. The NTP has issued two such prospective challenges to date for chemicals it planned on testing for rodent carcinogenicity. The first of these exercises took place in 1990, when Tennant et al.56 published predictions for 44 chemicals that were due to be tested by the NTP, and invited others to do the same.57 Seven developers in addition to Tenant et al. published their respective predictions prior to completion to testing.58-64

In 1993, once testing had been completed, the US National Institute of Environmental Health Sciences conducted an international workshop to evaluate what had been learned from the first predictive exercise. The workshop reached two main conclusions.65 First, SAR-based models do not perform as accurately as models that utilize biological attributes and, second, models that used multiple attributes to represent the chemical carcinogenicity endpoint performed better than models that were based on one or two attributes.

The comparison between predicted and actual results has been the subject for many discussions and publications.66'67 Benigni concludes that all systems currently suffer from an inability to distinguish those chemicals that contain a structural alert and are carcinogenic from those that also contain a structural alert but are inactive with the predictions made by the human experts, Ashby and Tennant being the most accurate, getting 75% of the predictions correct (Table 1).

In 1996, the NTP issued a second challenge to system developers to predict the outcome for a further 30 chemicals due to be tested for rodent carcinogenicity.68 As before, most predictions were published prior to the completion of testing.69-82 In this exercise there was a greater variety in the systems used to predict the outcome of the rodent carcinogenicity studies. They included several human expert opinions, the use of existing or new in vitro assays, as well as the use of computational approaches. This second set of chemicals presented developers with a greater challenge by avoiding chemicals with obvious modes of action or structural alerts. Many chemicals in this second exercise, in contrast to the first, were Ames-negative, implying that carcinogenicity would not be induced through direct interactions with DNA. To compensate for this additional challenge, developers were allowed access to the in vitro and short-term in vivo test results to aid them in making their predictions.

Benigni in his analysis of the results from this second prospective exercise concluded that at present the best systems for prediction still rely to some extent on human judgment and that the upper limit for accuracy from computational approaches is 65%. Overall accuracy results for each system used for prediction are summarized in Table 2.

In general, the models for noncongeneric chemicals are inferior in their predictive power when compared with the classical QSARs for congeneric series. This is not entirely surprising, since a QSAR for congeners is aimed at modeling only one mechanistic phenomenon, whereas the general models for the noncongeneric sets try to model several mechanisms of action at the same time, each relative to a whole class of carcinogens. Obviously, this broad-sweeping

Table 1 Concordance between predictions and rodent results

System Accuracy

Tennant et al.56 0.75

Jones and Easterly; RASH62 0.68

Bakale and McCreary64 0.65

Sanderson and Earnshaw; DEREK63 0.59

Benigni61 0.58

Sanderson and Earnshaw; DEREK hybrid63 0.57

Enslein et al.; TOPKAT58 0.57

Lewis et al.; COMPACT60 0.54

Rosenkranz and Klopman; MCASE59 0.49

Table 2 Accuracy of the QSAR predictions compared with the rodent bioassay results

System

Accuracy

OncoLogic82

10 Ways To Fight Off Cancer

10 Ways To Fight Off Cancer

Learning About 10 Ways Fight Off Cancer Can Have Amazing Benefits For Your Life The Best Tips On How To Keep This Killer At Bay Discovering that you or a loved one has cancer can be utterly terrifying. All the same, once you comprehend the causes of cancer and learn how to reverse those causes, you or your loved one may have more than a fighting chance of beating out cancer.

Get My Free Ebook


Post a comment