Carcinogenicity predictions for a group of 30 chemicals undergoing rodent cancer bioassays based on rules derived from subchronic organ toxicities.

Rodent carcinogenicities for a group of 30 chemicals which form the subject of the Second NIEHS Predictive-Toxicology Evaluation Experiment are predicted based on their subchronic organ toxicities. Predictions are made by rules learned by the rule learning (RL) induction program.


Introduction
We have been using a rule induction program to analyze the relationship between rodent carcinogencinity and many features of chemicals such as responses in shortterm assays and physical chemical properties. Recently we investigated the predictive strength of organ-specific toxicity for rodent carcinogenicity and noncarcinogenicity (1). The organ-specific toxicity was modeled by the presence or absence of 124 lesions observed at the end of subchronic studies by oral administration. Each lesion relates an organ with a morphological effect, and the 124 lesions were regrouped into a total of 32 organs and 43 morphological effects. We used the rule learning (RL) induction program to learn rules predicting rodent carcinogenicity and noncarcinogenicity from the organ-specific toxicities of 88 chemicals consisting of 60 carcinogens and 28 noncarcinogens. Four sets of rules were learned, each ofwhich was obtained using different combinations of assumptions and features in rule formation. We applied two of the four sets of rules to a group of 30 chemicals undergoing rodent cancer bioassays, which also forms the subject of the Second NIEHS Predictive-Toxicology Evaluation (PTE-2) experiment. In the present article, we report predictions of rodent carcinogenicity and noncarcinogenicity of these 30 chemicals.
Organ-specific Toxicity of 30 Chemicals Table 1 shows the organ-specific toxicity along with the response in the Salmonella mutagenicity assay (SAL) of each of the 30 chemicals. "+", "-", and "?" refer to positive, negative, and unknown responses in SAL, respectively. For each chemical, observed lesions (a pair of an organ and a morphological effect) are listed. Note that not all observed lesions were applicable because they were not among the 124 lesions on which rules were based.

Methodology: Rule Induction
The RI Program RL (2) is a knowledge-based inductive rule learning program that induces one or more IF-condition-THEN-class rules from specific examples of classes. For example, to make predictions of rodent carcinogenicity, the RL program is given a set of carcinogens and noncarcinogens and induces one or more rules that classify them. RL's heuristic search can examine a much larger number of identification (or classification) criteria than can be examined by manual analysis. Also, prior domain knowledge such as facts, heuristics, or assumptions used by scientists can be included during the search to learn rules that are plausible biologically as well as statistically.
The main strength of RL is its flexibility. Given a learning problem, many prob- Given a learning problem, i.e., the names of one or more target classes, a set of their examples, and a partial domain model of the problem, RL searches for rules by examining a large but limited number of combinations of features. An example is represented as a vector of attribute-value pairs, each of which describes a feature of the example. For example, the representation of methyleugenol is shown below, which means that methyleugenol is associated with degeneration of testes; necrosis, hyperplasia, and inflammation of liver; but did not cause inflammation of kidney, .... and its rodent carcinogenicity is unknown.
((Name methyleugenol) (testes degeneration +) (liver necrosis +) (liver hyperplasia +) (liver inflammation +) (kidney inflammation -) ... (rodent ?)) The RL induction program searches for rules by generating and evaluating many No data available Degeneration, metaplasia, and inflammation in nasal cavity; necrosis and hemosiderin pigment in liver; hyperplasia in forestomach Cytoplasmic vacuolization in liver; regeneration in kidney No data available Hyperplasia and metaplasia in nasal cavity Inflammation and metaplasia in larynx; inflammation in lung; metaplasia in trachea; inflammation,degeneration, and metaplasia in nasal cavity; degeneration in testes No lesions Pigmentation and degeneration in liver; pigmentation in kidney Ulceration, hyperkeratoses, and inflammation on skin; nephropathy, necrosis, and mineralization in kidney; degeneration in brain; necrosis in liver; degeneration in heart Inflammation and hyperkeratoses on skin; cytoplasmic vacuolization in liver Nephropathy, hyperplasia, and pigmentation in kidney No lesions Regeneration, hemosiderin pigment, necrosis, and degeneration in liver; hematopoiesis and hemosiderin pigment in spleen; degeneration and hemosiderin pigment in kidney; hyperplasia in bone marrow; necrosis, ulceration, inflammation, and hyperplasia in forestomach; degeneration in testes Metaplasia, hyperplasia, degeneration, exudate, and inflammation in nasal cavity Hemosiderin pigment in spleen; inflammation and hyperplasia in lung; hemosiderin pigment and regeneration in liver; metaplasia in larynx; hyperplasia in bone marrow; hyperplasia in lymph node No applicable lesions Necrosis, inflammation, degeneration, and metaplasia in nasal cavity; inflammation and metaplasia in larynx; inflammation and metaplasia in trachea; necrosis in spleen; degeneration in testes Inflammation, degeneration, and necrosis in stomach glandular; pigmentation, necrosis, hyperplasia, and inflammation in liver; atrophy in uterus; degeneration in testes 8+, -, and ? indicate positive, negative, or unknown response, respectively, in the Salmonella mutagenicity assay. blnapplicable lesions are lesions that were observed among the 30 chemicals but were not used in learning rules. combinations of features. It starts with single features and successively specializes rules by adding new features or specializing values associated with features. The choice and order of feature combination are in general dependent on their concordance with training data. For example, a combination of features with higher positive predictive values may be evaluated prior to other combinations of features. On the other hand, it is also possible to provide RL with a set of assumptions that guide it such that specific types of rules are excluded or included. For example, in predicting rodent carcinogenicity of nongenotoxic chemicals, rules including responses in short-term assays were preferred (3). Thus, the plausibility of a rule is determined by its performance (how accurately it classifies examples) and its concordance with assumptions, constraints, and domain knowledge.
The result of rule search is a set of IFcondition-THEN-class rules, where condition is a conjunction ("AND") of features. For example, the following rule uses two features to predict rodent carcinogenicity: IF (liver +) AND (kidney +) THEN (rodent C), which is interpreted as if a chemical causes any morphological effect on liver and kidney, then it is classified as a carcinogen. Such IF-THEN rules are very easy to understand, unlike numerical weights and nodes in a neural network; and the comprehensibility of rules permits the facile verification of rules by experts. Unless a learning problem is simple enough to classify all training examples with a single rule, RL finds a disjunctive set of rules, each of which classifies a subset of training examples. Such rules are then used collectively to make predictions on new cases.
RL is a descendant of the Meta-DEN-DRAL system (4), which specialized in finding rules of mass spectrometry in chemistry. However, unlike the Meta-DENDRAL, RL is a general purpose learning program that can be applied to many problems in different domains. RL has been applied successfully in several realworld problems, including predicting rodent carcinogenicity with short-term assays (3), predicting human developmental toxicity based on the results of animal toxicity assays (5), triggering design in high energy physics (6), and analyzing massive quantities of data on infant mortality (7). RL is written in C programming language. An export version of RL is available on request. However, the export version does not include new features that are under further development and some of which were used in the present study.

Rules
In our previous study, which investigated the relationship between organ-specific toxicity and rodent carcinogenicity, we reported four sets of rules, two of which were learned using the responses in SAL in addition to organ-specific toxicity (1). Of the four sets learned in the previous study, we used two sets to make predictions for the 30 chemicals presented in the second PTE experiment. Both rule sets were learned using only the organ-specific toxicity and did not include responses in SAL. Since these two rule sets correspond to the first and third rule sets reported previously (1), we will refer to them as RI and R3, respectively. Both rule sets were learned from 88 chemicals, of which 60 were carcinogens and 28 were noncarcinogens. None of the 30 chemicals in the second PTE experiment were included in the training data. The main differences between RI and R3 are the assumptions the RL induction program used in evaluating and learning rules. For RI, RL gave greater weight to liver and kidney toxicities than to other organs, while RL gave equal weights to all organs for rules in R3. In other words, when a chemical caused an effect in liver or kidney, it was taken more seriously (for carcinogenicity) than an effect in another organ. The choice of kidney and liver was made because it was found that they were the two organs most indicative of carcinogenicity, and the battery of these two organs was even more accurate in classifying the 88 chemicals. The details of assumptions as well as the analysis have been described (1). Tables 2 and 3 show rules in RI and R3, respectively. Both rule sets contain eight rules, three of which predict rodent carcinogenicity. For each rule, its condition and the class it predicts are shown along with three statistics. #C and #NC refer to the number of carcinogens and noncarcinogens in the training data that were covered by a rule. CF refers to a certainty factor. For a rule predicting rodent carcinogenicity, CF is calculated by (#C -0. 5) / (#C + #NC); similarly, for a rule predicting noncarcinogenicity, CF is obtained by (#NC-0.5)/(#C + #NC). Table 4 shows the predictions for the 30 chemicals in the second PTE experiment      made by the two rule sets, Ri and R3. The table also contains the responses in the Salmonella mutagenicity assay. The model of organ toxicity from which RL learned rules is not exactly equal to the model under which the organ toxicities of the 30 chemicals in the PTE were observed. In fact, in Table 1, we already indicated that some lesions caused by some chemicals were not applicable because there were no matching lesions in the 124 lesions on which the rules were based.

Predictions
Neither RI nor R3 makes predictions for all 30 chemicals; RI and R3 made predictions for 22 and 23 chemicals, respectively. Of the 22 chemicals predicted by RI, 13 were predicted to be carcinogens and 9 were to be noncarcinogens. R3 predicted 15 of 23 chemicals to be carcinogens and 8 to be noncarcinogens. RI and R3 agreed on predictions for 21 chemicals but disagreed on two chemicals, nitromethane and sodium nitrite. While RI was not able to make any  Tables 2 and 3. predictions for nitromethane, R3 predicted it to be a carcinogen. Also, while RI predicted sodium nitrate to be a noncarcinogen, R3 predicted it to be a carcinogen.
The rules in each set that match each chemical are shown in Table 5. Rules are referred to by the numbers assigned to them in Tables 2 and 3. For example, in RI, chloroprene matches rule-2 (which predicts carcinogenicity) and no rules predicting noncarcinogenicity. In other words, rule-2 in RI provides the evidence that chloroprene is a carcinogen, and there is no evidence that it is a noncarcinogen. On the other hand, in R3, while evidence for carcinogenicity is provided by rule-1, rule-8 also provides the evidence for noncarcinogenicity. However, since the certainty (CF) of rule-I is greater than that for rule-8, chloroprene was predicted to be a carcinogen.
Let us look at the two chemicals, nitromethane and sodium nitrite, for which the predictions of Ri and R3 did not agree. While there are no rules in RI matching nitromethane, in R3, rule-3 matches the chemical. Thus, RI did not make a prediction for nitromethane, and R3 predicted it to be a carcinogen because of the evidence provided by the training data, i.e., there were 12 (of 60) carcinogens that caused spleen toxicity but only 2 (of 28) noncarcinogens that caused spleen toxicity. For sodium nitrite, RI predicted it to be a noncarcinogen because of the evidence for noncarcinogenicity provided by rule-5, i.e., a chemical is more likely to be a noncarcinogen if it does not affect liver and kidney but causes hyperplasia on other organs. On the other hand, R3 predicted sodium nitrate to be a carcinogen because the evidence for carcinogenicity provided by rule-3 (i.e., the presence of spleen toxicity) is greater than the evidence for noncarcinogenicity provided by rule-S (i.e., no effects on liver and kidney and the presence of hyperplasia on other organs). In other words, despite the fact that sodium nitrite did not cause liver or kidney toxicity, it is predicted to be a carcinogen by R3 due to the lesion in the spleen.